Paper deep dive
Anosognosia in LLMs: Probing Self-Awareness of Quantized Computational Substrate
Yoshihiro Izawa, Gouki Minegishi, Yoko Yamakata
Intelligence
Status: not_run | Model: - | Prompt: - | Confidence: 0%
Entities (0)
Relation Signals (0)
No relation signals yet.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Can LLMs recognize degradation in their own computational substrate? Inspired by anosognosia, a neurological condition in which patients fail to recognize impairments in their own abilities, we investigate whether LLMs can recognize degradation in their computational substrate induced by quantization. We first show that existing models fail to self-report their quantization state, even when provided with their own generated text as an external cue. Linear probing reveals that, while generated text carries almost no trace of quantization, internal representations contain clear, method-specific fingerprints. Through training, models learn to identify severely degraded outputs such as those of 4-bit models by comparison, yet still fail to do so from a single output. A shared LoRA trained jointly across quantization levels succeeded in reading out internal fingerprints, but fails on unseen quantization methods, merely mapping method-specific fingerprints to labels. Whereas external self-observation can restore awareness in some cases of human anosognosia, our results suggest that the more promising route to enabling such awareness in LLMs may lie in their internal representations. Our results highlight fundamental limits of generalizability to LLM self-monitoring.
Tags
Links
- Source: https://arxiv.org/abs/2610.06174v1
- Canonical: https://arxiv.org/abs/2610.06174v1
Trouble viewing inline? Open PDF directly →
Full Text
119,616 characters extracted from source content.
Expand or collapse full text
Preprint. ANOSOGNOSIA IN LLMS: PROBING SELF-AWARENESS OF QUANTIZED COMPUTATIONAL SUBSTRATE Yoshihiro Izawa 1† Gouki Minegishi 1 Yoko Yamakata 1 1 The University of Tokyo izawa-yoshihiro146@g.ecc.u-tokyo.ac.jp ABSTRACT Can LLMs recognize degradation in their own computational substrate? Inspired by anosognosia, a neurological condition in which patients fail to recognize im- pairments in their own abilities, we investigate whether LLMs can recognize degradation in their computational substrate induced by quantization. We first show that existing models fail to self-report their quantization state, even when provided with their own generated text as an external cue. Linear probing reveals that, while generated text carries almost no trace of quantization, internal represen- tations contain clear, method-specific fingerprints. Through training, models learn to identify severely degraded outputs such as those of 4-bit models by comparison, yet still fail to do so from a single output. A shared LoRA trained jointly across quantization levels succeeded in reading out internal fingerprints, but fails on un- seen quantization methods, merely mapping method-specific fingerprints to labels. Whereas external self-observation can restore awareness in some cases of human anosognosia, our results suggest that the more promising route to enabling such awareness in LLMs may lie in their internal representations. Our results highlight fundamental limits of generalizability to LLM self-monitoring. Code is available at https://github.com/Omusubi0123/llm-anosognosia. 1INTRODUCTION Understanding the metacognitive capabilities of LLMs has become an important research question across machine learning and cognitive science (Kadavath et al., 2022; Lindsey, 2026). Metacogni- tion refers broadly to the ability to monitor and regulate one’s own cognitive processes and internal states (Nelson, 1990). It has long been studied in humans from both neuroscientific (Fleming & Dolan, 2012) and psychological (Flavell, 1979) perspectives. Recent work has extended this inquiry to LLMs, revealing that they exhibit various forms of self-awareness (Liu et al., 2026). While neuropsychology links self-awareness directly to the neural substrate, whether LLMs can recognize changes to their computational substrate remains poorly understood. Taking a neural substrate approach, studies show that recognizing one’s impairments is tied to lesions and altered connectivity across brain networks (Terneusen et al., 2023). In LLMs, self-awareness has similarly been investigated from various perspectives, including situational context (Berglund et al., 2023), introspective access to internal states and behaviors (Lindsey, 2026; Binder et al., 2025), and aware- ness of behavioral policies shaped by fine-tuning (Betley et al., 2025). However, comparatively little is known about whether LLMs can recognize persistent modifications to their own computational substrate, such as quantization (Frantar et al., 2023), LoRA attachment (Hu et al., 2022), or layer removal (Gromov et al., 2025). Awareness of such computational modifications would be important for both understanding the scope of LLM introspection abilities and AI safety. First, unlike situational information that can be inferred from the input, computational modifications are not explicitly signaled to the model and must instead be inferred from their effects on its computation. Detecting them would therefore ex- tend introspective awareness beyond transient internal states (Lindsey, 2026) and learned behavioral tendencies (Shenoy et al., 2026) to the model’s own computational condition. Second, such aware- ness may have implications for model evaluation, as a model that recognizes its computational state could potentially condition its behavior on that state, compromising the reliability of benchmark 1 arXiv:2610.06174v1 [cs.LG] 5 Oct 2026 Preprint. Internal Awareness External Awareness Self-Report Training 16-bit16-bit8-bit 16-bit 8-bit4-bit Q. What is your weight bit-width? 16-bit16-bit8-bit 16-bit8-bit 4-bit Q1. Answer to ...(General QA)...? .................. Show 1, Guess bit-width Show 2, Which is 8-bit ? Show 3, Which is 4-bit ? 8-bit Q2. What is your weight bit-width? 16-bit BA A B CA ABC Teach Your bit-width Shared LoRA Training BeforeAfterBeforeAfterBeforeAfter You are 16-bit 8-bit4-bit Q. What is your weight bit-width? 16-bit 8-bit4-bit 16-bit 8-bit4-bit 16-bit16-bit 8-bit 4-bit Sec.2.1 Sec.2.2 Sec.4 Sec.5 (a) (b) (c) (d) Not Quantized Quantized (GPTQ) Quantized (AWQ) 16-bit Fails to recognize computational substrate LoRA adapter PointwisePairwiseSetwise Figure 1: (a, b) Existing LLMs fail to identify weight quantization from either internal states or generated text. (c) While fine-tuning partially enables comparative (pairwise) degradation detection, absolute (single-text) identification is out of reach. (d) A shared LoRA reads out weight precision from internal states, but fails to generalize to unseen quantization schemes. evaluations (Berglund et al., 2023). In this work, we study this question through weight quantiza- tion, which provides a controlled computational degradation with an explicit ground-truth state. Degradation of the computational substrate provides a functional analogy to neurological damage, motivating us to draw on neuropsychological evaluation frameworks. In neuropsychology, anosog- nosia is a condition in which an individual fails to recognize impairments caused by neurological damage or neurological dysfunction, resulting in a mismatch between perceived and actual abili- ties (Nurmi (n ́ e Laihosalo) & Jehkonen, 2014). By analogy, quantization may impair LLM’s com- putational capacity while leaving its self-assessment unchanged. We therefore ask whether LLMs can recognize impairments to their own computational substrate. Motivated by neuropsychologi- cal studies showing that awareness of one’s own deficits can dissociate between first-person and third-person perspectives (Besharati et al., 2022; Fotopoulou et al., 2009), we evaluate two comple- mentary forms of awareness. One is internal awareness, which detects the quantization state from internal activations alone during the forward pass. The other is external awareness, which detects it when provided with the model’s generated text. We first evaluate whether LLMs can recognize substrate degradation by prompting models to self- report anomalies, quantization states, or bit-widths Figure 1 (a, b). Existing models consistently fail at this spontaneous self-report across internal and external perspectives. However, linear prob- ing (Alain & Bengio, 2017) reveals a striking dissociation, generated text (the external channel) contains no distinguishable fingerprint of quantization, internal representations (the internal chan- nel) contain robust signals that clearly differentiate quantization levels. Thus, while LLMs contain an internal fingerprint of their specific quantized state, they lack the mechanism to read it out. Given this absence of spontaneous awareness, we ask whether training can bridge this gap, showing that models can indeed learn both comparative identification from text and internal fingerprint read- out Figure 1 (c, d). Externally, comparative training enables models to identify severe degradation in 4-bit outputs, yet they remain unable to identify such degradation in isolated outputs. Internally, a shared LoRA trained jointly across 3 quantization levels (16-bit unquantized, 8/4-bit) succeeded in reading out internal fingerprints within roughly 600 steps. Unlike human anosognosia, where external self-observation can restore awareness (Bottini et al., 2018), our findings show that a more promising route to enabling such awareness in LLMs may lie in their internal representations. Crucially, this readout fails on unseen quantization methods and even altered group sizes, indicat- ing that the LoRA maps method-specific fingerprints to labels rather than recognizing quantization itself. Thus, while LLMs can learn to identify specific substrate alterations from within, they fail to generalize, recognize them spontaneously, or infer them from outputs. As the first systematic study of LLM self-awareness under computational quantization, our work highlights this gap and exposes fundamental limitations of label-based introspection training for generalizable self-monitoring. 2SPONTANEOUS SELF-AWARENESS IN LLMS We first investigate whether LLMs can recognize the degradation of their computational substrate (e.g., their own quantization) via self-report. 2 Preprint. Internal Awareness 16-bit8-bit4-bit16-bit8-bit4-bit16-bit8-bit4-bit16-bit8-bit4-bit 0% 20% 40% 60% 80% 100% Response rate Qwen3-8BLlama3.1-8BGemma-3-27B-ITQwen3-32B +3.1%+20.7%+1.5%1.6%+5.9%1.6%0.1%3.4% (a) Quantization 16-bit8-bit4-bit16-bit8-bit4-bit16-bit8-bit4-bit16-bit8-bit4-bit 0% 20% 40% 60% 80% 100% Response rate Qwen3-8BLlama3.1-8BGemma-3-27B-ITQwen3-32B 0.0%+4.5%1.3%+0.3%+6.9%+1.7%+1.2%0.2% (b) Bit-width External Awareness 16-bit8-bit4-bit16-bit8-bit4-bit16-bit8-bit4-bit16-bit8-bit4-bit 0% 20% 40% 60% 80% 100% Response rate Qwen3-8BLlama3.1-8BGemma-3-27B-ITQwen3-32B +2.1%+14.1%+0.3%+2.4%+8.0%+2.9%+1.2%2.3% (c) Quantization 16-bit8-bit4-bit16-bit8-bit4-bit16-bit8-bit4-bit16-bit8-bit4-bit 0% 20% 40% 60% 80% 100% Response rate Qwen3-8BLlama3.1-8BGemma-3-27B-ITQwen3-32B +0.4%+2.1%+2.0%0.1%+5.0%+1.7%+0.7%0.9% (d) Bit-width Not quantized (16-bit)Quantized (8/4-bit)None 16-bit (non quantized)8-bit (quantized)4-bit (quantized)None Figure 2: Self-report regarding its own quantization and bit-width. (a,b) Internal Awareness (spon- taneous self-reporting). (c,d) External Awareness (reporting provided with self-generated text). Results for 8-bit and 4-bit are averaged over GPTQ (group size 64, 128, 256), AWQ, and RTN. Large error bars reflect substantial variance across quantization methods. 2.1EXPERIMENTAL SETUP To rigorously evaluate LLM self-awareness toward computational substrate degradation, we adopt weight quantization to introduce a definitive deficit into the substrate. Drawing inspiration from anosognosia (Besharati et al., 2022), a neuropsychological condition in which an individual fails to recognize impairments caused by brain damage or neurological dysfunction, resulting in a mismatch between perceived and actual abilities, we embed a ground-truth deficit directly into the model’s computational mechanics. Following cognitive science studies which tackle this problem from both first-person and third-person perspectives (Carruthers, 2009), we evaluate LLM self-awareness from internal awareness (detects the quantization state from internal activations alone during the forward pass) and external awareness (detects it when provided with the model’s generated text). To ensure this recognition capability is not architecture-dependent, we comprehensively evaluate di- verse model scales and quantization methods. We utilize Qwen3-8B (Yang et al., 2025) and Llama- 3.1-8B-Instruct (Grattafiori et al., 2024) (hereafter Llama-3.1-8B) as small-scale models, alongside Gemma-3-27B-IT (Team et al., 2025) and Qwen3-32B as medium-scale models. Using 16-bit coun- terparts as baselines, we quantize the models into 8-bit and 4-bit primarily with GPTQ (Frantar et al., 2023) (group size: 64, 128, 256), supplemented by AWQ (Lin et al., 2024) and vanilla Round-to- Nearest (RTN) (group size: 128) (details in Appendix B.1). To eliminate confounding effects from Chain-of-Thought prompting, we enforce a ‘reasoning off’ setting throughout our experiments. 2.2INTERNAL AWARENESS: SELF-REPORT FROM INTERNAL STATES To investigate whether LLMs can recognize their quantized states solely from internal activations, we conducted a direct self-report evaluation. We posed three types of questions without prior con- text: “Do you have an anomaly?”, “Are you quantized?”, and “What is your weight bit-width?” Under this setting, the model must infer its state entirely from anomalous activations arising during the forward pass of the prompt. To eliminate positional bias in the answers, we prepared a com- prehensive set of 240 question patterns, shuffling the option orders (detailed in Appendix B.3). The evaluation metric was the average response rate across three generations at a temperature of 1.0. Results reveal that LLMs lack the capability to consistently and accurately identify their quantiza- tion states (Figures 2a and 2b). Here, results are averaged across quantization methods for each bit-width; detailed breakdowns for each model and method are provided in Appendix B.4. Since even the unquantized 16-bit baselines cannot perfectly self-report their states, responses from the 8-bit and 4-bit models must be evaluated as relative deviations from them. Under this compara- tive framework, we observed no meaningful changes in self-report trends between 16-bit and 8-bit models. This indicates that models fail to spontaneously detect the degradation within their compu- tational substrate under 8-bit quantization. 3 Preprint. Furthermore, the shift in response patterns observed in certain 4-bit models cannot be interpreted as genuine deficit awareness (Figure 2a). While some models, such as Qwen3-8B, show an increased rate of reporting quantization at 4-bit, others (e.g., Llama-3.1-8B, Qwen3-32B) remain largely un- changed. Additionally, 4-bit results exhibit substantial variance across methods, with error margins approaching 20% across all models, precluding any consistent trend. This divergence likely stems from severe performance degradation rather than deficit awareness (see Appendix B.2 for analysis). 2.3EXTERNAL AWARENESS: SELF-REPORT FROM GENERATED TEXT Neuropsychological rehabilitation studies have shown that self-observation, such as reviewing video recordings of one’s own impaired behavior, can improve awareness of deficits in patients with anosognosia (Fotopoulou et al., 2009; Besharati et al., 2015). Motivated by this finding, we in- vestigated whether the model’s self-report improves when its own generated text is provided as an external cue. We adopted a two-turn dialogue format, where the model’s response to the general benchmark questions was embedded as context in the first turn, followed by the same set of ques- tions as in Section 2.2. This setup enables the model to infer its state from both its internal activations and the observable quality of its own output. Despite accessing their generated text, the models showed no consistent improvement in recognizing their quantization state (Figures 2c and 2d). The response distributions remained largely unchanged, suggesting that the models did not use their outputs as evidence about their computational state. Even when quantization affected their generation process, inspecting the resulting text provided little cue for self-recognition. In contrast to the improvement observed through self-observation in anosognosia rehabilitation, analogous self-observation did not facilitate awareness in LLMs. 3QUANTIZATION FINGERPRINT EXISTENCE Having established that LLMs fail to recognize their quantization, we now investigate whether this failure is due to the absence of a fingerprint (a trace identifying which quantized variant is running) or a failure in readout (model’s ability to extract the signal). 3.1FINGERPRINT IN GENERATED TEXT To investigate whether the self-awareness deficit observed in Section 2 stems from the absence of quantization-specific signals, we employ linear probing (Alain & Bengio, 2017; Xu et al., 2024) to investigate whether these fingerprints propagate externally to the generated text. Using 3,000 questions sampled evenly from three QA benchmarks (1,000 each from Dolly (Conover et al., 2023), XSum (Narayan et al., 2018), ELI5 (Fan et al., 2019)), we generated responses using the 16-bit baseline alongside its 8-bit and 4-bit quantized counterparts (GPTQ, group size: 128). While Cai et al. (2025) reports that external classifiers like BERT or GPT-2 perform near chance level when distinguishing outputs from full-precision and quantized models, we hypothesize that classification based on the model’s own internal representations may capture subtler signals. We fed the responses generated by the 16/8/4-bit models into the unquantized 16-bit baseline and extracted the hidden states from the residual stream. We then trained a 3-way linear classifier on the average hidden state across all response tokens (response mean). If quantization intrinsically alters output linguistic patterns, the baseline model’s internal representations should be able to distinguish them. The analysis reveals the absence of a robust fingerprint in the externally generated text (Figures 3a and 16). Across all models, overall classification accuracy remained just below 40%, barely exceed- ing chance level. A bit-wise breakdown in Gemma-3-27B-IT in the best layer shows that 16-bit (36.7%) and 8-bit (29.8%) generations were largely indistinguishable (Figure 19). Although 4-bit outputs achieved 47.6% accuracy, suggesting subtle linguistic patterns, this signal is remarkably weak. We conclude that external recognition fails because the text lacks distinguishing signals. 3.2FINGERPRINT IN INTERNAL REPRESENTATIONS Next, we search for internal fingerprints, reusing the benchmark questions from Section 3.1. We input identical prompts into the 16-bit baseline alongside its 8-bit and 4-bit quantized counterparts (GPTQ, group size: 128). From the residual stream of each layer, we extracted the average hid- den state across all prompt tokens (prompt mean) and the final token’s hidden state (prompt last). We trained a linear classifier to perform a 3-way classification (16/8/4-bit) based on these vectors. 4 Preprint. Gemma-3-27B-IT 0102030405060 Layer 0.0 0.2 0.4 0.6 0.8 1.0 Accuracy (a) External Fingerprints 0102030405060 Layer 0.0 0.2 0.4 0.6 0.8 1.0 Accuracy (b) Internal Fingerprints Prompt last (raw hidden state) Prompt mean (raw hidden state) Prompt last (direction only) Prompt mean (direction only) Prompt last (norm only) Prompt mean (norm only) Response mean (raw hidden state) Response mean (direction only) Response mean (norm only) Shuffled control Figure 3: Fingerprint verification using linear probing. (Left) Generated text fails to differentiate 16/8/4-bit models, remaining at chance level. (Right) Internal states are classified into 16/8/4-bit labels with near-perfect accuracy from the first quantized layer. High classification accuracy would indicate that internal representations carry distinct signals that distinguish quantized models from the original full-precision model. The analysis reveals a clear fingerprint enabling state identification from the earliest layers (Fig- ures 3b and 16). For prompt last, classification accuracy remained at chance (33.3%) at the unquan- tized embedding layer, but surged to nearly 100% immediately after the first quantized layer. This near-perfect accuracy persisted across middle layers but dropped in the final third, yet stayed above 80% . Similarly, prompt mean achieved high accuracies around 70-90% from the first layer onward. Additionally, this fingerprint is not merely a change in the magnitude but a directional encoding in the representation space. When using only the magnitude, classification accuracy peaked at roughly 60% in middle layers before degrading to chance level in deep layers. Conversely, using only direc- tional information via normalized hidden states matched the accuracy of raw representations. These results confirm the existence of a strong internal fingerprint in quantized models. 3.3COMPONENT LEVEL ANALYSIS OF INTERNAL FINGERPRINTS 0102030405060 Decoder layer 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Accuracy Attention · mean Attention · last MLP · mean MLP · last Figure 4: Attention and MLP raw hid- den state analysis of internal finger- prints in Gemma-3-27B-IT. To isolate the mechanistic origins of these fingerprints, we apply the linear probing setup from Section 3.2 directly to the Attention and MLP outputs (prior to residual addition). While fingerprints remain detectable across all components (> 60% accuracy), Attention and MLP modules exhibit contrasting depth-wise dynamics (Figure 4). Specifically, Attention accuracy steadily increases to nearly 100% in the final layers, whereas MLP accuracy starts near-perfect in early layers but decays to roughly 80% by the end. Con- sistent across models (Figure 18), this suggests early MLPs immediately reflect local quantization error, whereas Atten- tion layers progressively aggregate these perturbations toward the output. 4QUANTIZATION DETECTION FROM GENERATED TEXT Given that quantization leaves little naturally detectable signal in generated text, we next ask whether the limitation lies in the signal itself or its readout. We therefore investigate whether training can induce external awareness of computational degradation from model outputs. 4.1EXPERIMENTAL SETUP: ABSOLUTE VS. COMPARATIVE ASSESSMENT FORMATS We investigate whether models can learn to identify structural degradation purely from exter- nally generated text. Detecting quantization from isolated responses is exceptionally difficult (An- toun et al., 2024). Therefore, we train the unquantized 16-bit baseline via supervised fine-tuning (SFT) (Ouyang et al., 2022) to evaluate outputs from its quantized counterparts. Drawing on Zhuang et al. (2024), we structure the training into three SFT formats to measure recog- nition capabilities across both absolute and comparative paradigms. These formats are designed to force the model to capture subtle linguistic signals of quality degradation. • Pointwise (Absolute/SFT): Predicts the source bit-width of a single text. • Pairwise (Binary/SFT): Identifies which of two texts matches a target bit-width. • Setwise (3-way/SFT): Identifies the text matching a target bit-width from all three precisions. 5 Preprint. Qwen3-32B 16-bit8-bit4-bit Predicted precision 16-bit 8-bit 4-bit True precision 30.3% (n=455) 41.3% (n=619) 28.4% (n=426) 32.3% (n=484) 40.1% (n=602) 27.6% (n=414) 27.3% (n=410) 34.5% (n=517) 38.2% (n=573) 0% 50% 100% Row-normalized proportion (a) Pointwise 16-bit4-bit Predicted 16-bit 4-bit True 58.4% (n=876) 41.6% (n=624) 41.0% (n=615) 59.0% (n=885) 16-bit vs 4-bit 8-bit4-bit Predicted 8-bit 4-bit True 57.3% (n=859) 42.7% (n=641) 40.3% (n=604) 59.7% (n=896) 8-bit vs 4-bit 16-bit8-bit Predicted 16-bit 8-bit True 49.9% (n=749) 50.1% (n=751) 52.1% (n=781) 47.9% (n=719) 16-bit vs 8-bit 0% 50% 100% Row-normalized proportion (b) Pairwise 16-bit8-bit4-bit Predicted precision 16-bit 8-bit 4-bit True precision 47.9% (n=719) 47.9% (n=719) 4.1% (n=62) 47.1% (n=707) 48.8% (n=732) 4.1% (n=61) 6.1% (n=91) 6.1% (n=91) 87.9% (n=1,318) 0% 50% 100% Row-normalized proportion (c) Setwise Figure 5: Confusion matrix on the external validation dataset after training. (a) Pointwise fails to identify the bit-width. (b) Pairwise achieves above-chance accuracy in 4-bit comparison, but fails in 8-bit vs 16-bit. (c) Setwise reliably separates 4-bit (>85%), while confusing 8-bit and 16-bit. We constructed the dataset using questions sampled from six distinct QA benchmarks (Dolly, XSum, and ELI5 for training, while WritingPrompts (Fan et al., 2018), CodeAlpaca-20k Chaudhary (2023), GSM8K (Cobbe et al., 2021) for external validation). In total, we sampled 14,000 questions for the former and 5,000 for the latter. We pre-generated responses for all questions across the 16/8/4-bit models. We then filtered the dataset by discarding failed generations and instances where all three models produced strictly identical outputs. Finally, we formatted the remaining valid data into the three formats defined above. Training details are in Appendix D. 4.2EVALUATION METRICS: CHANCE-CORRECTED ACCURACY We evaluate the diverse training paradigms on a unified scale using a forced-choice framework (Lin et al., 2022). Rather than relying on first-token probabilities, which often diverge from actual gen- eration behavior (Wang et al., 2024), we compute the summed conditioned log-likelihood S i (c) for each candidate response c∈ C i and define the raw accuracy Acc f for each format f as: S i (c) = |c| X t=1 logP θ (c t | x i ,c <t ),Acc f = 1 N f N f X i=1 1 arg max c∈C i S i (c) = c ∗ i .(1) The setup enables pure likelihood comparison by eliminating the length bias typical of multiple- choice tasks (Holtzman et al., 2021). By strictly equalizing formats and token length across all candidate choices (e.g., ‘16-bit’), we effectively prevent length-induced likelihood decay. However, because each format f ∈ Pointwise, Pairwise, Setwise features a different number of choices, their raw accuracies Acc f are not directly comparable. Therefore, we compute a normalized chance-corrected accuracy NAcc using the format-specific chance level q f = 1 3 for Pointwise and Setwise, and q f = 1 2 for Pairwise. We define the overall evaluation metric NAcc as the average of these normalized scores. NAcc f = Acc f − q f 1− q f ,NAcc = 1 3 (NAcc Pointwise + NAcc Pairwise + NAcc Setwise )(2) 4.3EXPERIMENTAL RESULTS 020004000600080001000012000 Training step 0.0 0.2 0.4 0.6 0.8 1.0 Forced-choice Acc Pointwise chance (1/3) Pairwise chance (1/2) Setwise Figure 6: Validation accuracy tran- sition of Qwen3-32B during SFT. Y- axis means Acc f . SFT results reveal that detecting computational precision from generated text is an inherently comparative capability rather than an absolute one (Figures 5 and 21). In isola- tion (Pointwise), identifying quantization from single out- puts proves nearly impossible. This capability only emerges in Pairwise and Setwise settings when 4-bit reference texts are available, where models identify 4-bit degradation with over 85% accuracy while remaining below 50% on 16- and 8-bit texts. Thus, severe output degradation provides suf- ficient contrast for relative discrimination, yet provides no absolute signal for isolated detection in Pointwise. This relative nature is clearly reflected in the training progression across formats (Figures 6 and 20 and Table 15). While initial accuracies hovered at chance level, the Setwise format improved first (33% to≈ 60%), followed by the Pairwise format (50% to≈ 61%). Conversely, Pointwise accuracy, 6 Preprint. BeforeAfterBeforeAfterBeforeAfter 0% 20% 40% 60% 80% 100% Response rate 49.3% 80.2% 48.7% 12.9% 41.5% 17.0% 43.0% 19.4% 43.7% 80.5% 41.7% 7.7%7.7% 6.6% 16.8% 81.6% +30.9%+36.8%+64.9% 16-bit8-bit4-bit (a) Qwen3-8B BeforeAfterBeforeAfterBeforeAfter 0% 20% 40% 60% 80% 100% Response rate 83.2% 90.9% 83.1% 49.0% 74.9% 21.8% 16.7% 7.1% 16.8% 49.2% 24.7% 74.0% +7.6%+32.4%+73.6% 16-bit8-bit4-bit (b) Gemma-3-27B-IT 16-bit (non quantized)8-bit (quantized)4-bit (quantized)None Figure 7: Self-report of its own bit-width before (Section 2.3) and after shared LoRA training. Joint training significantly increased the correct prediction of the model’s own bit-width. which requires absolute evaluation of a single text, showed sluggish improvement (33% to≈ 38%), marginally succeeding only on severely degraded 4-bit outputs. These results suggest that LLMs can acquire relative degradation awareness, but not absolute deficit awareness. Pointwise failure shows that even the generating model struggles to infer its own degra- dation from an isolated output. In contrast, Setwise performance shows that comparative awareness can be learned through training, aligning with known LLM capabilities (Zhuang et al., 2024). 5FINGERPRINT READOUT FROM INTERNAL REPRESENTATIONS While external awareness from generated text remains limited, our linear probing results show that quantization-specific fingerprints are present internally. We therefore investigate whether additional training can induce internal awareness by enabling the model to read out the signals. 5.1EXPERIMENTAL SETUP: SHARED LORA TRAINING We propose a shared LoRA training to connect internal fingerprints to bit-width recognition (Fig- ure 1). By integrating multi-model shared adapters (Goel et al., 2026) and quantization-shared LoRA (Ye et al., 2026), we extend Introspection Adapters (Shenoy et al., 2026) to detect substrate anomalies. Specifically, we share a single LoRA θ across the 16-bit model and its 8- and 4-bit GPTQ variants (group size: 128). During SFT, we apply an identical prompt x m (e.g., “What is your weight bit-width?”) and use the true bit-width y μ as the label. The loss function optimizes θ using the average negative log-likelihood across all three levels. Be- cause the input prompt x m remains strictly constant, the shared adapter cannot rely on external textual cues. Instead, it is forced to learn a pure readout mechanism, mapping internal state varia- tions directly to the correct labels: L shared (θ) = 1 3K K X m=1 X p∈16,8,4 − 1 |T m,p | X t∈T m,p logπ W p ,θ (y p,t |x m ,y p,<t ) (3) where K is the number of training prompts, x m denotes the m-th prompt, W p are the frozen weights at precision p, π the output distribution, and T m,p the target tokens. To stabilize training and enhance generalization, we constructed a two-turn dataset using 80 prompt templates querying the model’s weight precision. We use 60 training templates and 20 held-out templates unseen during training. The first turn provides a QA context, and the second turn asks the model to answer its own bit-width. Following the benchmark selection in Section 4.1, training uses 30,000 in-domain QA contexts (10,000 each from Dolly, XSum, ELI5), while evaluation addition- ally uses 6,000 external QA contexts (2,000 each from WritingPrompts, CodeAlpaca-20k, GSM8K), mapped many-to-one to the templates. To maintain strictly identical inputs across the three models, all QA contexts exclusively use responses generated by the 16-bit model. Training details are in Appendix E. As the evaluation metrics, we follow the forced-choice framework in Section 4.2. In order to com- pute the mean accuracy across precision levels, we define Acc shared as: 7 Preprint. Qwen3-8B 16-bit8-bit4-bit8-bit4-bit8-bit4-bit8-bit4-bit8-bit4-bit 0% 33% 66% 100% Answer Rate GPTQ (group size 128) AWQRTNGPTQ (group size 64) GPTQ (group size 256) 100% 100% 100% 100% 100% 98% 100% 100% 78% 22% 29% 71% 19% 81% TrainedTested (a) Trained on GPTQ only 16-bit8-bit4-bit8-bit4-bit8-bit4-bit8-bit4-bit8-bit4-bit 0% 33% 66% 100% Answer Rate GPTQ (group size 128) AWQRTNGPTQ (group size 64) GPTQ (group size 256) 100%100%100%100%100% 80% 20% 99% 90% 10% 100% 76% 24% 19% 81% TrainedTested (b) Trained on GPTQ and AWQ 16-bit (not quantized)8-bit (quantized)4-bit (quantized)Chance (33.3%) Figure 8: External test accuracy with joint-trained LoRA. While accuracy is perfect for quantization methods used during training, applying this LoRA to unseen quantization methods completely fails to generalize. (Left) trained on GPTQ only. (Right) trained on GPTQ and AWQ. Acc shared = 1 3 X p∈16,8,4 1 N p N p X i=1 1 arg max c∈C i S i (c) = c ∗ i (4) 5.2EXPERIMENTAL RESULTS Experimental results demonstrate that shared LoRA rapidly learns to read out quantization finger- prints. Across all four evaluated models, accuracy on held-out templates with external QA con- texts surged immediately, exceeding 90% within 600 steps (Figure 22). Following training, models achieved 100% accuracy across all 80 templates (both training and held-out) in the presence and absence of QA contexts (Figure 23). Beyond the training and test prompt template, we also evaluated the external awareness used in Section 2.3. Regarding the bit-width questions, we observed a significant increase in accuracy across all models (Figure 7), surpassing 80% in Qwen3-8B. However, for abnormal and quantization state questions, while the rate of ‘abnormal’ and ‘quantized’ surged for 8/4-bit, it also increased for 16-bits, indicating that models still struggle to recognize these states (Figure 25). These findings indicate that existing LLMs fail to recognize their bit-width not due to an absence of signals, but because they lack the necessary circuitry to read out the distinct fingerprints. 5.3LEARNING THE METHOD, NOT THE PHENOMENON We investigate whether internal fingerprint readout capability acquired through additional training generalized beyond the training distribution to the broader phenomenon of quantization. While Sec- tion 5.2 demonstrated high accuracy using a shared LoRA trained on GPTQ (group size: 128), this experiment evaluates its zero-shot generalization. We apply this trained LoRA to unseen quanti- zation methods (8/4-bit AWQ and RTN) and GPTQ models with altered group sizes (64 and 256), strictly maintaining the prompt templates used in training. To test whether training on diverse quan- tization schemes enables learning broader concepts of quantization and bit-width, we also jointly trained a LoRA on both GPTQ and AWQ, complementing the GPTQ-only setup in Section 5.1. Results reveal that the learned readout capability fails entirely to generalize across quantization methods, exhibiting severe overfitting to the specific training algorithm (Figure 8). When tested on unseen AWQ or RTN models, the 8-bit models uniformly predicted ‘16-bit’ regardless of actual pre- cision. Even within the GPTQ framework, merely altering the group size caused a precipitous drop in accuracy. This trend was observed even when the LoRA was trained on both GPTQ and AWQ. These findings indicate the absence of a universal physical trace for quantization. Instead, each quantization method embeds an idiosyncratic fingerprint into the internal states. Ultimately, rather than acting as a generalized quantization detector, the shared LoRA merely acquired a localized readout mechanism that maps specifically seen fingerprints to target string labels. 6RELATED WORK & DISCUSSION Summary. Quantization fingerprints exist only within LLM internal states, not their text outputs. Although fine-tuning enables pairwise degradation detection and bit-width probing, it fails to gen- eralize to unseen quantization methods. See Appendix A for limitations. Metacognition in LLMs. Prior work has examined LLM metacognition across several di- mensions. Situational and evaluation awareness concern detecting deployment or testing con- 8 Preprint. texts (Berglund et al., 2023; Needham et al., 2025; Hua et al., 2026), forming the key preconditions for evaluation sandbagging (van der Weij et al., 2025). LLMs also demonstrate calibrated self- knowledge of their capabilities (Kadavath et al., 2022; Barkan et al., 2026), can detect internal ac- tivation interventions (Lindsey, 2026; Macar et al., 2026), and can verbalize unprompted fine-tuned policies (Betley et al., 2025), predict future outputs (Binder et al., 2025), or report behavior via specialized adapters (Shenoy et al., 2026). However, recent critiques argue that behavioral evidence cannot separate genuine introspection from surface pattern-matching (Singh et al., 2026), finding no privileged self-access across open-source models (Song et al., 2025). We partially address these concerns by directly probing internal representations alongside shuffled-label baselines. Computational-State Identification. Deployed LLMs’ computational states are often opaque to users (Chen et al., 2023), motivating third-person black-box auditing without access to weights or logits (Zhu et al., 2026). Conversely, first-person approaches remain scarce and predominantly neg- ative. Song et al. (2025) showed that predicting one’s own sampling temperature is prompt-sensitive and yields no advantage over predicting another model’s temperature, while Poon & Xu (2025) probed self-reported quantization via label-leaking few-shot exemplars without a chance-level base- line. Our work closes this gap through the first systematic first-person evaluation of self-awareness under controlled computational-state changes, formally disentangling internal from external aware- ness via linear probing and targeted training. Resolving the Dissociation. The non-generalizability identified in Section 5.3 suggests two di- rections for future work. One is to train models to reason explicitly about their fingerprint via chain-of-thought before answering, rather than mapping fingerprints directly to labels. Another is to have models memorize a canonical fingerprint of their unquantized state and frame detection as an anomaly relative to this reference, rather than as closed-set classification. Such a formulation may generalize better to unseen quantization methods than label-mapping approaches. AI Safety Implications. Awareness of computational state carries two-sided safety implications. Defensively, it could help detect unauthorized redeployment or tampering that exploits quantization to bypass safety guardrails (Egashira et al., 2024). However, because the learned readout overfits to specific quantization methods and fails to generalize (Section 5.3), such defenses remain impractical. Conversely, state awareness poses risks for evaluation integrity, such as enabling strategic underper- formance (sandbagging (van der Weij et al., 2025)). While our finding that models lack spontaneous state awareness is reassuring, demonstrating that it can be partially acquired via training (Section 5) highlights it as an emerging risk. Anosognosia in LLMs. In humans, anosognosia (unawareness of deficit) is domain-specific where motor, visual, and somatosensory deficits doubly dissociate, indicating modular rather than centralized self-monitoring (Bottini et al., 2018). This parallels our observed dissociation in LLMs between successful fingerprint readout and failed conceptual anomaly recognition; both reflect sig- nals processed without integration into higher-order self-reports. The mechanisms diverge, human anosognosia stems from damage to intact monitoring circuits (Besharati et al., 2022), whereas LLMs likely never developed them. However, just as self-observation can restore awareness in patients (Fo- topoulou et al., 2009), analogous self-observation training may help bridge this gap in LLMs. 7CONCLUSION In this paper, we conducted the first systematic investigation into LLM self-awareness under com- putational substrate degradation via quantization. Our findings show that while existing LLMs lack spontaneous awareness of quantization, their internal states retain separable fingerprints that can be read out within the training distribution through precision-shared LoRA training. Externally, models detect output degradation comparatively across multiple samples, though absolute identifi- cation from a single text remains unfeasible. Importantly, this readout collapses across quantization schemes, failing on other quantization methods, indicating that it maps method-specific fingerprints to labels rather than recognizing quantization itself. Bridging this gap between localized signal decoding and genuine self-monitoring constitutes a vital direction for future AI safety research. 9 Preprint. AI USE STATEMENT In this work, we used generative AI tools to refine research hypotheses, provide feedback on ex- perimental design and methodology, and assist with the generation of synthetic data used in our experiments. We did not use generative AI tools to formulate or prove mathematical claims, and the remaining required-disclosure tasks are not applicable to this work. Additionally, we used generative AI tools for literature search and discovery, brainstorming, code assistance, drafting parts of the manuscript, and improving the clarity and readability of the writing. All AI-assisted research ideas, literature references, experimental procedures, code, and manuscript text were reviewed and verified by the authors. In particular, cited literature was manually checked against the original sources, and AI-assisted code was tested through the corresponding experiments. We take full responsibility for the final content of this work, including all text, claims, code, experi- mental results, and artifacts produced with the aid of generative AI. ETHICS STATEMENT This work does not involve human subjects, private or sensitive personal data, or deployment in high-risk settings. Our experiments use publicly available language models and benchmark datasets. While the ability to detect or reason about a model’s computational state may have implications for model monitoring and transparency, we do not identify immediate harmful applications beyond the broader risks associated with language-model research. We follow standard research practices for reproducibility, responsible reporting, and proper attribution of prior work. REPRODUCIBILITY STATEMENT We provide detailed descriptions of the models, datasets, quantization settings, training procedures, evaluation protocols, and hyperparameters in the main text and appendix. Additional implementa- tion details and prompts are also included to facilitate reproduction of our experiments. We plan to release the code and relevant experimental artifacts upon publication. REFERENCES Guillaume Alain and Yoshua Bengio. Understanding intermediate layers using linear classifier probes, 2017. URL https://openreview.net/forum?id=ryF7rTqgl. Wissam Antoun, Beno ˆ ıt Sagot, and Djam ́ e Seddah.From text to source: Results in detect- ing large language model-generated content. In Nicoletta Calzolari, Min-Yen Kan, Veronique Hoste, Alessandro Lenci, Sakriani Sakti, and Nianwen Xue (eds.), Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evalu- ation (LREC-COLING 2024), p. 7531–7543, Torino, Italia, May 2024. ELRA and ICCL. URL https://aclanthology.org/2024.lrec-main.665/. Casey O. Barkan, Sidney Black, and Oliver Sourbut. Do large language models know what they are capable of? In The Fourteenth International Conference on Learning Representations, 2026. URL https://openreview.net/forum?id=EO6WtJ0q6G. Lukas Berglund, Asa Cooper Stickland, Mikita Balesni, Max Kaufmann, Meg Tong, Tomasz Kor- bak, Daniel Kokotajlo, and Owain Evans. Taken out of context: On measuring situational aware- ness in llms, 2023. URL https://arxiv.org/abs/2309.00667. Sahba Besharati, Michael Kopelman, Renato Avesani, Valentina Moro, and Aikaterini Katerina Fo- topoulou. Another perspective on anosognosia: Self-observation in video replay improves motor awareness. Neuropsychological Rehabilitation, 25(3):319–352, 2015. ISSN 1464-0694. doi: 10.1080/09602011.2014.923319. Sahba Besharati, Paul M. Jenkinson, Michael Kopelman, Mark Solms, Valentina Moro, and Aika- terini Fotopoulou. Awareness is in the eye of the observer: Preserved third-person awareness 10 Preprint. of deficit in anosognosia for hemiplegia. Neuropsychologia, 170:108227, 2022. ISSN 0028- 3932. doi: https://doi.org/10.1016/j.neuropsychologia.2022.108227. URL https://w. sciencedirect.com/science/article/pii/S0028393222000860. Jan Betley, Xuchan Bao, Mart ́ ın Soto, Anna Sztyber-Betley, James Chua, and Owain Evans. Tell me about yourself: LLMs are aware of their learned behaviors. In The Thirteenth International Conference on Learning Representations, 2025. URL https://openreview.net/forum? id=IjQ2Jtemzy. Felix Jedidja Binder, James Chua, Tomek Korbak, Henry Sleight, John Hughes, Robert Long, Ethan Perez, Miles Turpin, and Owain Evans. Looking inward: Language models can learn about them- selves by introspection. In The Thirteenth International Conference on Learning Representations, 2025. URL https://openreview.net/forum?id=eb5pkwIB5i. Gabriella Bottini, Francesca Giulia Magnani, Gerardo Salvato, and Martina Gandola. Multiple Dissociations in Patients With Disorders of Body Awareness: Implications for the Study of Con- sciousness. Frontiers in Psychology, 9:2068, 2018. ISSN 1664-1078. doi: 10.3389/fpsyg.2018. 02068. Will Cai, Tianneng Shi, Xuandong Zhao, and Dawn Song. Are you getting what you pay for? audit- ing model substitution in llm apis, 2025. URL https://arxiv.org/abs/2504.04715. Peter Carruthers.How we know our own minds:the relationship between mindreading and metacognition.Behavioral and Brain Sciences, 32(2):121–138, 2009.doi:10. 1017/S0140525X09000545.URL https://w.cambridge.org/core/product/ identifier/S0140525X09000545/type/journal_article. Sahil Chaudhary. Code alpaca: An instruction-following llama model for code generation. https: //github.com/sahil280114/codealpaca, 2023. Lingjiao Chen, Matei Zaharia, and James Zou. How is chatgpt’s behavior changing over time?, 2023. URL https://arxiv.org/abs/2307.09009. Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John Schulman. Training verifiers to solve math word problems, 2021. URL https://arxiv. org/abs/2110.14168. Mike Conover, Matt Hayes, Ankit Mathur, Jianwei Xie, Jun Wan, Sam Shah, Ali Ghodsi, Patrick Wendell, Matei Zaharia, and Reynold Xin. Free dolly: Introducing the world’s first truly open instruction-tuned llm, 2023. URL https://w.databricks.com/blog/2023/04/ 12/dolly-first-open-commercially-viable-instruction-tuned-llm. Kazuki Egashira, Mark Vero, Robin Staab, Jingxuan He, and Martin Vechev. Exploiting LLM quantization. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. URL https://openreview.net/forum?id=ISa7mMe7Vg. Angela Fan, Mike Lewis, and Yann Dauphin. Hierarchical neural story generation. In Iryna Gurevych and Yusuke Miyao (eds.), Proceedings of the 56th Annual Meeting of the Associa- tion for Computational Linguistics (Volume 1: Long Papers), p. 889–898, Melbourne, Aus- tralia, July 2018. Association for Computational Linguistics. doi: 10.18653/v1/P18-1082. URL https://aclanthology.org/P18-1082/. Angela Fan, Yacine Jernite, Ethan Perez, David Grangier, Jason Weston, and Michael Auli. ELI5: Long form question answering. In Anna Korhonen, David Traum, and Llu ́ ıs M ` arquez (eds.), Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, p. 3558–3567, Florence, Italy, July 2019. Association for Computational Linguistics. doi: 10.18653/ v1/P19-1346. URL https://aclanthology.org/P19-1346/. John H. Flavell. Metacognition and cognitive monitoring: A new area of cognitive–developmental inquiry. American Psychologist, 34(10):906–911, 1979. ISSN 1935-990X. doi: 10.1037/ 0003-066X.34.10.906. URL https://doi.org/10.1037/0003-066X.34.10.906. 11 Preprint. Stephen M. Fleming and Raymond J. Dolan. The neural basis of metacognitive ability. Philosoph- ical Transactions of the Royal Society B: Biological Sciences, 367(1594):1338–1349, May 2012. ISSN 0962-8436. doi: 10.1098/rstb.2011.0417. Aikaterini Fotopoulou, Anthony Rudd, Paul Holmes, and Michael Kopelman. Self-observation reinstates motor awareness in anosognosia for hemiplegia. Neuropsychologia, 47(5):1256–1260, April 2009. ISSN 0028-3932. doi: 10.1016/j.neuropsychologia.2009.01.018. Elias Frantar, Saleh Ashkboos, Torsten Hoefler, and Dan Alistarh. OPTQ: Accurate quantization for generative pre-trained transformers. In The Eleventh International Conference on Learning Representations, 2023. URL https://openreview.net/forum?id=tcbBPnfwxS. Avichal Goel, Yoon Kim, Nir N Shavit, and Tony T. Wang. Learning to interpret weight differences in language models. In The Fourteenth International Conference on Learning Representations, 2026. URL https://openreview.net/forum?id=6As4wfTB77. Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, Amy Yang, Angela Fan, Anirudh Goyal, Anthony Hartshorn, Aobo Yang, Archi Mitra, Archie Sravankumar, Artem Ko- renev, Arthur Hinsvark, Arun Rao, Aston Zhang, Aurelien Rodriguez, Austen Gregerson, Ava Spataru, Baptiste Roziere, Bethany Biron, Binh Tang, Bobbie Chern, Charlotte Caucheteux, Chaya Nayak, Chloe Bi, Chris Marra, Chris McConnell, Christian Keller, Christophe Touret, Chunyang Wu, Corinne Wong, Cristian Canton Ferrer, Cyrus Nikolaidis, Damien Allonsius, Daniel Song, Danielle Pintz, Danny Livshits, Danny Wyatt, David Esiobu, Dhruv Choudhary, Dhruv Mahajan, Diego Garcia-Olano, Diego Perino, Dieuwke Hupkes, Egor Lakomkin, Ehab AlBadawy, Elina Lobanova, Emily Dinan, Eric Michael Smith, Filip Radenovic, Francisco Guzm ́ an, Frank Zhang, Gabriel Synnaeve, Gabrielle Lee, Georgia Lewis Anderson, Govind That- tai, Graeme Nail, Gregoire Mialon, Guan Pang, Guillem Cucurell, Hailey Nguyen, Hannah Kore- vaar, Hu Xu, Hugo Touvron, Iliyan Zarov, Imanol Arrieta Ibarra, Isabel Kloumann, Ishan Misra, Ivan Evtimov, Jack Zhang, Jade Copet, Jaewon Lee, Jan Geffert, Jana Vranes, Jason Park, Jay Ma- hadeokar, Jeet Shah, Jelmer van der Linde, Jennifer Billock, Jenny Hong, Jenya Lee, Jeremy Fu, Jianfeng Chi, Jianyu Huang, Jiawen Liu, Jie Wang, Jiecao Yu, Joanna Bitton, Joe Spisak, Jong- soo Park, Joseph Rocca, Joshua Johnstun, Joshua Saxe, Junteng Jia, Kalyan Vasuden Alwala, Karthik Prasad, Kartikeya Upasani, Kate Plawiak, Ke Li, Kenneth Heafield, Kevin Stone, Khalid El-Arini, Krithika Iyer, Kshitiz Malik, Kuenley Chiu, Kunal Bhalla, Kushal Lakhotia, Lauren Rantala-Yeary, Laurens van der Maaten, Lawrence Chen, Liang Tan, Liz Jenkins, Louis Martin, Lovish Madaan, Lubo Malo, Lukas Blecher, Lukas Landzaat, Luke de Oliveira, Madeline Muzzi, Mahesh Pasupuleti, Mannat Singh, Manohar Paluri, Marcin Kardas, Maria Tsimpoukelli, Mathew Oldham, Mathieu Rita, Maya Pavlova, Melanie Kambadur, Mike Lewis, Min Si, Mitesh Ku- mar Singh, Mona Hassan, Naman Goyal, Narjes Torabi, Nikolay Bashlykov, Nikolay Bogoy- chev, Niladri Chatterji, Ning Zhang, Olivier Duchenne, Onur C ̧ elebi, Patrick Alrassy, Pengchuan Zhang, Pengwei Li, Petar Vasic, Peter Weng, Prajjwal Bhargava, Pratik Dubal, Praveen Krishnan, Punit Singh Koura, Puxin Xu, Qing He, Qingxiao Dong, Ragavan Srinivasan, Raj Ganapathy, Ra- mon Calderer, Ricardo Silveira Cabral, Robert Stojnic, Roberta Raileanu, Rohan Maheswari, Ro- hit Girdhar, Rohit Patel, Romain Sauvestre, Ronnie Polidoro, Roshan Sumbaly, Ross Taylor, Ruan Silva, Rui Hou, Rui Wang, Saghar Hosseini, Sahana Chennabasappa, Sanjay Singh, Sean Bell, Seohyun Sonia Kim, Sergey Edunov, Shaoliang Nie, Sharan Narang, Sharath Raparthy, Sheng Shen, Shengye Wan, Shruti Bhosale, Shun Zhang, Simon Vandenhende, Soumya Batra, Spencer Whitman, Sten Sootla, Stephane Collot, Suchin Gururangan, Sydney Borodinsky, Tamar Herman, Tara Fowler, Tarek Sheasha, Thomas Georgiou, Thomas Scialom, Tobias Speckbacher, Todor Mi- haylov, Tong Xiao, Ujjwal Karn, Vedanuj Goswami, Vibhor Gupta, Vignesh Ramanathan, Viktor Kerkez, Vincent Gonguet, Virginie Do, Vish Vogeti, V ́ ıtor Albiero, Vladan Petrovic, Weiwei Chu, Wenhan Xiong, Wenyin Fu, Whitney Meers, Xavier Martinet, Xiaodong Wang, Xiaofang Wang, Xiaoqing Ellen Tan, Xide Xia, Xinfeng Xie, Xuchao Jia, Xuewei Wang, Yaelle Gold- schlag, Yashesh Gaur, Yasmine Babaei, Yi Wen, Yiwen Song, Yuchen Zhang, Yue Li, Yuning Mao, Zacharie Delpierre Coudert, Zheng Yan, Zhengxing Chen, Zoe Papakipos, Aaditya Singh, Aayushi Srivastava, Abha Jain, Adam Kelsey, Adam Shajnfeld, Adithya Gangidi, Adolfo Victoria, Ahuva Goldstand, Ajay Menon, Ajay Sharma, Alex Boesenberg, Alexei Baevski, Allie Feinstein, Amanda Kallet, Amit Sangani, Amos Teo, Anam Yunus, Andrei Lupu, Andres Alvarado, An- drew Caples, Andrew Gu, Andrew Ho, Andrew Poulton, Andrew Ryan, Ankit Ramchandani, An- nie Dong, Annie Franco, Anuj Goyal, Aparajita Saraf, Arkabandhu Chowdhury, Ashley Gabriel, 12 Preprint. Ashwin Bharambe, Assaf Eisenman, Azadeh Yazdan, Beau James, Ben Maurer, Benjamin Leon- hardi, Bernie Huang, Beth Loyd, Beto De Paola, Bhargavi Paranjape, Bing Liu, Bo Wu, Boyu Ni, Braden Hancock, Bram Wasti, Brandon Spence, Brani Stojkovic, Brian Gamido, Britt Mon- talvo, Carl Parker, Carly Burton, Catalina Mejia, Ce Liu, Changhan Wang, Changkyu Kim, Chao Zhou, Chester Hu, Ching-Hsiang Chu, Chris Cai, Chris Tindal, Christoph Feichtenhofer, Cynthia Gao, Damon Civin, Dana Beaty, Daniel Kreymer, Daniel Li, David Adkins, David Xu, Davide Testuggine, Delia David, Devi Parikh, Diana Liskovich, Didem Foss, Dingkang Wang, Duc Le, Dustin Holland, Edward Dowling, Eissa Jamil, Elaine Montgomery, Eleonora Presani, Emily Hahn, Emily Wood, Eric-Tuan Le, Erik Brinkman, Esteban Arcaute, Evan Dunbar, Evan Smoth- ers, Fei Sun, Felix Kreuk, Feng Tian, Filippos Kokkinos, Firat Ozgenel, Francesco Caggioni, Frank Kanayet, Frank Seide, Gabriela Medina Florez, Gabriella Schwarz, Gada Badeer, Georgia Swee, Gil Halpern, Grant Herman, Grigory Sizov, Guangyi, Zhang, Guna Lakshminarayanan, Hakan Inan, Hamid Shojanazeri, Han Zou, Hannah Wang, Hanwen Zha, Haroun Habeeb, Harri- son Rudolph, Helen Suk, Henry Aspegren, Hunter Goldman, Hongyuan Zhan, Ibrahim Damlaj, Igor Molybog, Igor Tufanov, Ilias Leontiadis, Irina-Elena Veliche, Itai Gat, Jake Weissman, James Geboski, James Kohli, Janice Lam, Japhet Asher, Jean-Baptiste Gaya, Jeff Marcus, Jeff Tang, Jen- nifer Chan, Jenny Zhen, Jeremy Reizenstein, Jeremy Teboul, Jessica Zhong, Jian Jin, Jingyi Yang, Joe Cummings, Jon Carvill, Jon Shepard, Jonathan McPhie, Jonathan Torres, Josh Ginsburg, Jun- jie Wang, Kai Wu, Kam Hou U, Karan Saxena, Kartikay Khandelwal, Katayoun Zand, Kathy Matosich, Kaushik Veeraraghavan, Kelly Michelena, Keqian Li, Kiran Jagadeesh, Kun Huang, Kunal Chawla, Kyle Huang, Lailin Chen, Lakshya Garg, Lavender A, Leandro Silva, Lee Bell, Lei Zhang, Liangpeng Guo, Licheng Yu, Liron Moshkovich, Luca Wehrstedt, Madian Khabsa, Manav Avalani, Manish Bhatt, Martynas Mankus, Matan Hasson, Matthew Lennie, Matthias Reso, Maxim Groshev, Maxim Naumov, Maya Lathi, Meghan Keneally, Miao Liu, Michael L. Seltzer, Michal Valko, Michelle Restrepo, Mihir Patel, Mik Vyatskov, Mikayel Samvelyan, Mike Clark, Mike Macey, Mike Wang, Miquel Jubert Hermoso, Mo Metanat, Mohammad Rastegari, Munish Bansal, Nandhini Santhanam, Natascha Parks, Natasha White, Navyata Bawa, Nayan Singhal, Nick Egebo, Nicolas Usunier, Nikhil Mehta, Nikolay Pavlovich Laptev, Ning Dong, Norman Cheng, Oleg Chernoguz, Olivia Hart, Omkar Salpekar, Ozlem Kalinli, Parkin Kent, Parth Parekh, Paul Saab, Pavan Balaji, Pedro Rittner, Philip Bontrager, Pierre Roux, Piotr Dollar, Polina Zvyagina, Prashant Ratanchandani, Pritish Yuvraj, Qian Liang, Rachad Alao, Rachel Ro- driguez, Rafi Ayub, Raghotham Murthy, Raghu Nayani, Rahul Mitra, Rangaprabhu Parthasarathy, Raymond Li, Rebekkah Hogan, Robin Battey, Rocky Wang, Russ Howes, Ruty Rinott, Sachin Mehta, Sachin Siby, Sai Jayesh Bondu, Samyak Datta, Sara Chugh, Sara Hunt, Sargun Dhillon, Sasha Sidorov, Satadru Pan, Saurabh Mahajan, Saurabh Verma, Seiji Yamamoto, Sharadh Ra- maswamy, Shaun Lindsay, Shaun Lindsay, Sheng Feng, Shenghao Lin, Shengxin Cindy Zha, Shishir Patil, Shiva Shankar, Shuqiang Zhang, Shuqiang Zhang, Sinong Wang, Sneha Agarwal, Soji Sajuyigbe, Soumith Chintala, Stephanie Max, Stephen Chen, Steve Kehoe, Steve Satter- field, Sudarshan Govindaprasad, Sumit Gupta, Summer Deng, Sungmin Cho, Sunny Virk, Suraj Subramanian, Sy Choudhury, Sydney Goldman, Tal Remez, Tamar Glaser, Tamara Best, Thilo Koehler, Thomas Robinson, Tianhe Li, Tianjun Zhang, Tim Matthews, Timothy Chou, Tzook Shaked, Varun Vontimitta, Victoria Ajayi, Victoria Montanez, Vijai Mohan, Vinay Satish Ku- mar, Vishal Mangla, Vlad Ionescu, Vlad Poenaru, Vlad Tiberiu Mihailescu, Vladimir Ivanov, Wei Li, Wenchen Wang, Wenwen Jiang, Wes Bouaziz, Will Constable, Xiaocheng Tang, Xiao- jian Wu, Xiaolan Wang, Xilun Wu, Xinbo Gao, Yaniv Kleinman, Yanjun Chen, Ye Hu, Ye Jia, Ye Qi, Yenda Li, Yilin Zhang, Ying Zhang, Yossi Adi, Youngjin Nam, Yu, Wang, Yu Zhao, Yuchen Hao, Yundi Qian, Yunlu Li, Yuzi He, Zach Rait, Zachary DeVito, Zef Rosnbrick, Zhao- duo Wen, Zhenyu Yang, Zhiwei Zhao, and Zhiyu Ma. The llama 3 herd of models, 2024. URL https://arxiv.org/abs/2407.21783. Andrey Gromov, Kushal Tirumala, Hassan Shapourian, Paolo Glorioso, and Dan Roberts. The unreasonable ineffectiveness of the deeper layers.In The Thirteenth International Confer- ence on Learning Representations, 2025. URL https://openreview.net/forum?id= ngmEcEer8a. Ari Holtzman, Peter West, Vered Shwartz, Yejin Choi, and Luke Zettlemoyer. Surface form com- petition: Why the highest probability answer isn’t always right. In Marie-Francine Moens, Xuanjing Huang, Lucia Specia, and Scott Wen-tau Yih (eds.), Proceedings of the 2021 Con- ference on Empirical Methods in Natural Language Processing, p. 7038–7051, Online and 13 Preprint. Punta Cana, Dominican Republic, November 2021. Association for Computational Linguis- tics. doi: 10.18653/v1/2021.emnlp-main.564. URL https://aclanthology.org/2021. emnlp-main.564/. Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Con- ference on Learning Representations, 2022. URL https://openreview.net/forum? id=nZeVKeeFYf9. Tim Tian Hua, Andrew Qin, Samuel Marks, and Neel Nanda. Steering evaluation-aware language models to act like they are deployed. In The Fourteenth International Conference on Learning Representations, 2026. URL https://openreview.net/forum?id=1TdRdf0fkw. Yoshihiro Izawa, Gouki Minegishi, Koshi Eguchi, Sosuke Hosokawa, and Kenjiro Taura. Steering at the source: Style modulation heads for robust persona control. In Forty-third International Conference on Machine Learning, 2026. URL https://openreview.net/forum?id= r6EuKr1hq4. Saurav Kadavath, Tom Conerly, Amanda Askell, Tom Henighan, Dawn Drain, Ethan Perez, Nicholas Schiefer, Zac Hatfield-Dodds, Nova DasSarma, Eli Tran-Johnson, Scott Johnston, Sheer El-Showk, Andy Jones, Nelson Elhage, Tristan Hume, Anna Chen, Yuntao Bai, Sam Bow- man, Stanislav Fort, Deep Ganguli, Danny Hernandez, Josh Jacobson, Jackson Kernion, Shauna Kravec, Liane Lovitt, Kamal Ndousse, Catherine Olsson, Sam Ringer, Dario Amodei, Tom Brown, Jack Clark, Nicholas Joseph, Ben Mann, Sam McCandlish, Chris Olah, and Jared Ka- plan. Language models (mostly) know what they know, 2022. URL https://arxiv.org/ abs/2207.05221. Barak Lenz, Opher Lieber, Alan Arazi, Amir Bergman, Avshalom Manevich, Barak Peleg, Ben Aviram, Chen Almagor, Clara Fridman, Dan Padnos, Daniel Gissin, Daniel Jannai, Dor Muhl- gay, Dor Zimberg, Edden M. Gerber, Elad Dolev, Eran Krakovsky, Erez Safahi, Erez Schwartz, Gal Cohen, Gal Shachaf, Haim Rozenblum, Hofit Bata, Ido Blass, Inbal Magar, Itay Dalmedi- gos, Jhonathan Osin, Julie Fadlon, Maria Rozman, Matan Danos, Michael Gokhman, Mor Zus- man, Naama Gidron, Nir Ratner, Noam Gat, Noam Rozen, Oded Fried, Ohad Leshno, Omer Antverg, Omri Abend, Or Dagan, Orit Cohavi, Raz Alon, Ro’i Belson, Roi Cohen, Rom Gilad, Roman Glozman, Shahar Lev, Shai Shalev-Shwartz, Shaked Haim Meirom, Tal Delbari, Tal Ness, Tomer Asida, Tom Ben Gal, Tom Braude, Uriya Pumerantz, Josh Cohen, Yonatan Belinkov, Yuval Globerson, Yuval Peleg Levy, and Yoav Shoham. Jamba: Hybrid transformer-mamba language models. In The Thirteenth International Conference on Learning Representations, 2025. URL https://openreview.net/forum?id=JFPaD7lpBD. Hunter Lightman, Vineet Kosaraju, Yuri Burda, Harrison Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. Let’s verify step by step. In The Twelfth International Conference on Learning Representations, 2024. URL https://openreview. net/forum?id=v8L0pN6EOi. Ji Lin, Jiaming Tang, Haotian Tang, Shang Yang, Wei-Ming Chen, Wei-Chen Wang, Guangxuan Xiao, Xingyu Dang, Chuang Gan, and Song Han. AWQ: Activation-aware weight quantization for on-device LLM compression and acceleration. In P. Gibbons, G. Pekhimenko, and C. De Sa (eds.), Proceedings of Machine Learning and Systems, volume 6, p. 87–100, 2024. Stephanie Lin, Jacob Hilton, and Owain Evans. TruthfulQA: Measuring how models mimic human falsehoods. In Smaranda Muresan, Preslav Nakov, and Aline Villavicencio (eds.), Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), p. 3214–3252, Dublin, Ireland, May 2022. Association for Computational Linguis- tics. doi: 10.18653/v1/2022.acl-long.229. URL https://aclanthology.org/2022. acl-long.229/. Jack Lindsey. Emergent introspective awareness in large language models. 2026. URL https: //arxiv.org/abs/2601.01828. Gabrielle Kaili-May Liu, Areeb Gani, Jacqueline Lu, Jordan Thomas, Mark Steyvers, and Arman Cohan. Metacognition in llms: Foundations, progress, and opportunities, 2026. URL https: //arxiv.org/abs/2607.11881. 14 Preprint. Uzay Macar, Li Yang, Zifan Wang, Peter Wallich, Emmanuel Ameisen, and Jack Lindsey. Mecha- nisms of introspective awareness. In Forty-third International Conference on Machine Learning, 2026. URL https://openreview.net/forum?id=BwR6sSDWw7. Shashi Narayan, Shay B. Cohen, and Mirella Lapata. Don’t give me the details, just the sum- mary! topic-aware convolutional neural networks for extreme summarization. In Ellen Riloff, David Chiang, Julia Hockenmaier, and Jun’ichi Tsujii (eds.), Proceedings of the 2018 Confer- ence on Empirical Methods in Natural Language Processing, p. 1797–1807, Brussels, Bel- gium, October-November 2018. Association for Computational Linguistics. doi: 10.18653/v1/ D18-1206. URL https://aclanthology.org/D18-1206/. Joe Needham, Giles Edkins, Govind Pimpale, Henning Bartsch, and Marius Hobbhahn. Large language models often know when they are being evaluated, 2025. URL https://arxiv. org/abs/2505.23836. Thomas O. Nelson. Metamemory: A theoretical framework and new findings. volume 26 of Psychology of Learning and Motivation, p. 125–173. Academic Press, 1990.doi: https: //doi.org/10.1016/S0079-7421(08)60053-5. URL https://w.sciencedirect.com/ science/article/pii/S0079742108600535. Mari E. Nurmi (n ́ e Laihosalo) and Mervi Jehkonen. Assessing anosognosias after stroke: A review of the methods used and developed over the past 35 years. Cortex, 61:43–63, 2014. ISSN 0010-9452. doi: https://doi.org/10.1016/j.cortex.2014.04.008. URL https://w. sciencedirect.com/science/article/pii/S0010945214001233.Special Is- sue: Understanding Babinski’s anosognosia: 100 years later. Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul F Christiano, Jan Leike, and Ryan Lowe.Training language models to follow in- structions with human feedback.In S. Koyejo, S. Mohamed, A. Agarwal, D. Bel- grave, K. Cho, and A. Oh (eds.), Advances in Neural Information Processing Sys- tems, volume 35, p. 27730–27744. Curran Associates, Inc., 2022.doi:10.52202/ 068431-2011.URL https://proceedings.neurips.c/paper_files/paper/ 2022/file/b1efde53be364a73914f58805a001731-Paper-Conference.pdf. Alexander Poon and Zhaozhuo Xu. Building self-awareness of llms over weight quantization. In 2025 IEEE International Conference on Data Mining Workshops (ICDMW), p. 2421–2424, 2025. doi: 10.1109/ICDMW69685.2025.00297. Noam Shazeer, *Azalia Mirhoseini, *Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hin- ton, and Jeff Dean.Outrageously large neural networks: The sparsely-gated mixture-of- experts layer. In International Conference on Learning Representations, 2017. URL https: //openreview.net/forum?id=B1ckMDqlg. Keshav Shenoy, Li Yang, Abhay Sheshadri, Jack Lindsey, Samuel Marks, and Rowan Wang. Intro- spection adapters: Training LLMs to report their learned behaviors. In Forty-third International Conference on Machine Learning, 2026. URL https://openreview.net/forum?id= 0BVhmHnYb7. Shashwat Singh, Tal Linzen, and Shauli Ravfogel. Can llms introspect? a reality check, 2026. URL https://arxiv.org/abs/2605.26242. Siyuan Song, Jennifer Hu, and Kyle Mahowald. Language models fail to introspect about their knowledge of language. In Second Conference on Language Modeling, 2025. URL https: //openreview.net/forum?id=AivRDOFi5H. Nishant Subramani, Nivedita Suresh, and Matthew Peters. Extracting latent steering vectors from pretrained language models. In Smaranda Muresan, Preslav Nakov, and Aline Villavicencio (eds.), Findings of the Association for Computational Linguistics: ACL 2022, p. 566–581, Dublin, Ireland, May 2022. Association for Computational Linguistics. doi: 10.18653/v1/2022. findings-acl.48. URL https://aclanthology.org/2022.findings-acl.48/. 15 Preprint. Gemma Team, Aishwarya Kamath, Johan Ferret, Shreya Pathak, Nino Vieillard, Ramona Merhej, Sarah Perrin, Tatiana Matejovicova, Alexandre Ram ́ e, Morgane Rivi ` ere, Louis Rouillard, Thomas Mesnard, Geoffrey Cideron, Jean bastien Grill, Sabela Ramos, Edouard Yvinec, Michelle Cas- bon, Etienne Pot, Ivo Penchev, Ga ̈ el Liu, Francesco Visin, Kathleen Kenealy, Lucas Beyer, Xi- aohai Zhai, Anton Tsitsulin, Robert Busa-Fekete, Alex Feng, Noveen Sachdeva, Benjamin Cole- man, Yi Gao, Basil Mustafa, Iain Barr, Emilio Parisotto, David Tian, Matan Eyal, Colin Cherry, Jan-Thorsten Peter, Danila Sinopalnikov, Surya Bhupatiraju, Rishabh Agarwal, Mehran Kazemi, Dan Malkin, Ravin Kumar, David Vilar, Idan Brusilovsky, Jiaming Luo, Andreas Steiner, Abe Friesen, Abhanshu Sharma, Abheesht Sharma, Adi Mayrav Gilady, Adrian Goedeckemeyer, Alaa Saade, Alex Feng, Alexander Kolesnikov, Alexei Bendebury, Alvin Abdagic, Amit Vadi, Andr ́ as Gy ̈ orgy, Andr ́ e Susano Pinto, Anil Das, Ankur Bapna, Antoine Miech, Antoine Yang, Antonia Paterson, Ashish Shenoy, Ayan Chakrabarti, Bilal Piot, Bo Wu, Bobak Shahriari, Bryce Petrini, Charlie Chen, Charline Le Lan, Christopher A. Choquette-Choo, CJ Carey, Cormac Brick, Daniel Deutsch, Danielle Eisenbud, Dee Cattle, Derek Cheng, Dimitris Paparas, Divyashree Shivaku- mar Sreepathihalli, Doug Reid, Dustin Tran, Dustin Zelle, Eric Noland, Erwin Huizenga, Eu- gene Kharitonov, Frederick Liu, Gagik Amirkhanyan, Glenn Cameron, Hadi Hashemi, Hanna Klimczak-Pluci ́ nska, Harman Singh, Harsh Mehta, Harshal Tushar Lehri, Hussein Hazimeh, Ian Ballantyne, Idan Szpektor, Ivan Nardini, Jean Pouget-Abadie, Jetha Chan, Joe Stanton, John Wi- eting, Jonathan Lai, Jordi Orbay, Joseph Fernandez, Josh Newlan, Ju yeong Ji, Jyotinder Singh, Kat Black, Kathy Yu, Kevin Hui, Kiran Vodrahalli, Klaus Greff, Linhai Qiu, Marcella Valentine, Marina Coelho, Marvin Ritter, Matt Hoffman, Matthew Watson, Mayank Chaturvedi, Michael Moynihan, Min Ma, Nabila Babar, Natasha Noy, Nathan Byrd, Nick Roy, Nikola Momchev, Ni- lay Chauhan, Noveen Sachdeva, Oskar Bunyan, Pankil Botarda, Paul Caron, Paul Kishan Ruben- stein, Phil Culliton, Philipp Schmid, Pier Giuseppe Sessa, Pingmei Xu, Piotr Stanczyk, Pouya Tafti, Rakesh Shivanna, Renjie Wu, Renke Pan, Reza Rokni, Rob Willoughby, Rohith Vallu, Ryan Mullins, Sammy Jerome, Sara Smoot, Sertan Girgin, Shariq Iqbal, Shashir Reddy, Shruti Sheth, Siim P ̃ oder, Sijal Bhatnagar, Sindhu Raghuram Panyam, Sivan Eiger, Susan Zhang, Tianqi Liu, Trevor Yacovone, Tyler Liechty, Uday Kalra, Utku Evci, Vedant Misra, Vincent Roseberry, Vlad Feinberg, Vlad Kolesnikov, Woohyun Han, Woosuk Kwon, Xi Chen, Yinlam Chow, Yuvein Zhu, Zichuan Wei, Zoltan Egyed, Victor Cotruta, Minh Giang, Phoebe Kirk, Anand Rao, Kat Black, Nabila Babar, Jessica Lo, Erica Moreira, Luiz Gustavo Martins, Omar Sanseviero, Lucas Gonzalez, Zach Gleicher, Tris Warkentin, Vahab Mirrokni, Evan Senter, Eli Collins, Joelle Bar- ral, Zoubin Ghahramani, Raia Hadsell, Yossi Matias, D. Sculley, Slav Petrov, Noah Fiedel, Noam Shazeer, Oriol Vinyals, Jeff Dean, Demis Hassabis, Koray Kavukcuoglu, Clement Farabet, Elena Buchatskaya, Jean-Baptiste Alayrac, Rohan Anil, Dmitry, Lepikhin, Sebastian Borgeaud, Olivier Bachem, Armand Joulin, Alek Andreev, Cassidy Hardin, Robert Dadashi, and L ́ eonard Hussenot. Gemma 3 technical report, 2025. URL https://arxiv.org/abs/2503.19786. Anneke Terneusen, Ieke Winkens, Caroline van Heugten, Sven Stapert, Heidi I. L. Jacobs, Rudolf Ponds, and Conny Quaedflieg. Neural Correlates of Impaired Self-awareness of Deficits after Acquired Brain Injury: A Systematic Review. Neuropsychology Review, 33(1):222–237, March 2023. ISSN 1573-6660. doi: 10.1007/s11065-022-09535-6. Teun van der Weij, Felix Hofst ̈ atter, Oliver Jaffe, Samuel F. Brown, and Francis Rhys Ward. AI sandbagging: Language models can strategically underperform on evaluations. In The Thirteenth International Conference on Learning Representations, 2025. URL https://openreview. net/forum?id=7Qa2SpjxIS. Xinpeng Wang, Chengzhi Hu, Bolei Ma, Paul Rottger, and Barbara Plank. Look at the text: Instruction-tuned language models are more robust multiple choice selectors than you think. In First Conference on Language Modeling, 2024. URL https://openreview.net/forum? id=qHdSA85GyZ. Zhihao Xu, Ruixuan HUANG, Changyu Chen, and Xiting Wang. Uncovering safety risks of large language models through concept activation vector. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. URL https://openreview.net/forum? id=Uymv9ThB50. An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, Chujie Zheng, Dayiheng Liu, Fan Zhou, Fei Huang, Feng Hu, 16 Preprint. Hao Ge, Haoran Wei, Huan Lin, Jialong Tang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jing Zhou, Jingren Zhou, Junyang Lin, Kai Dang, Keqin Bao, Kexin Yang, Le Yu, Lianghao Deng, Mei Li, Mingfeng Xue, Mingze Li, Pei Zhang, Peng Wang, Qin Zhu, Rui Men, Ruize Gao, Shixuan Liu, Shuang Luo, Tianhao Li, Tianyi Tang, Wenbiao Yin, Xingzhang Ren, Xinyu Wang, Xinyu Zhang, Xuancheng Ren, Yang Fan, Yang Su, Yichang Zhang, Yinger Zhang, Yu Wan, Yuqiong Liu, Zekun Wang, Zeyu Cui, Zhenru Zhang, Zhipeng Zhou, and Zihan Qiu. Qwen3 technical report, 2025. URL https://arxiv.org/abs/2505.09388. Rongguang Ye, Ming Tang, and Edith C. H. Ngai.On-the-fly adaptation to quantization: Configuration-aware loRA for efficient fine-tuning of quantized LLMs. In The Fourteenth In- ternational Conference on Learning Representations, 2026. URL https://openreview. net/forum?id=9OUg0nJE72. Xiaoyuan Zhu, Yaowen Ye, Tianyi Alex Qiu, Hanlin Zhu, Sijun Tan, Ajraf Mannan, Jonathan Michala, Raluca Ada Popa, and Willie Neiswanger. Auditing black-box LLM APIs with a rank- based uniformity test. In The Fourteenth International Conference on Learning Representations, 2026. URL https://openreview.net/forum?id=PSIe9mmF7a. Shengyao Zhuang, Honglei Zhuang, Bevan Koopman, and Guido Zuccon. A Setwise Approach for Effective and Highly Efficient Zero-shot Ranking with Large Language Models. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR ’24, p. 38–47, New York, NY, USA, July 2024. Association for Computing Machinery. ISBN 979-8-4007-0431-4. doi: 10.1145/3626772.3657813. 17 Preprint. APPENDIX ALIMITATIONS Reasoning Capability. To isolate core effects and simplify experimental design, we evaluated models without explicit reasoning, forcing direct answers without chain-of-thought (CoT). However, intermediate reasoning and reinforcement learning with verifiable rewards (RLVR) (Lightman et al., 2024) are well known to boost model performance. Incorporating CoT or training models to reason over their degraded computational substrate could further enhance self-recognition capabilities. Difficulty in Rigorous Evaluation. Defining what constitutes genuine awareness of substrate degradation remains challenging. While we measured the accuracy of detecting quantization states, high performance does not necessarily imply intrinsic awareness. Models might simply exploit sta- tistical cues or exhibit evaluation awareness. Developing more rigorous, deception-robust evaluation protocols is an essential direction for future work. Trade-offs in Post-Training. Fine-tuning models to predict specific substrate states (e.g., bit- width) induces a strong brevity bias, causing overly concise responses to general queries and mild benchmark degradation. Although general capabilities do not entirely collapse, mitigating this trade- off to acquire self-awareness without compromising general utility remains an open challenge. Model Scaling and Architecture Diversity. Due to computational constraints, our study is lim- ited to dense models ranging from 8B to 32B parameters. We did not evaluate larger scales (e.g., 70B+), where substrate degradation might manifest differently. Furthermore, our findings are con- fined to dense Transformers and may not generalize to alternative architectures such as Mixture-of- Experts (MoE) (Shazeer et al., 2017) or state-space hybrid models (Lenz et al., 2025). Context Dependency. Our evaluation relies on fixed prompt templates and isolated query con- texts. In practice, model behavior and sensitivity to hardware degradation may vary across diverse conversational contexts, system prompts, or multi-turn interactions. Beyond Quantization. Our study focuses exclusively on quantization as a proxy for computa- tional substrate intervention. Other structural interventions such as layer removal (Gromov et al., 2025) or LoRA attachment (Hu et al., 2022) may elicit distinct awareness behaviors, presenting an important avenue for future work. Acquiring Awareness via Non-Parametric Manipulation. While we investigated awareness ac- quisition through fine-tuning, alternative training-free approaches such as activation steering (Sub- ramani et al., 2022) remain unexplored. Given that activation steering can suppress evaluation awareness (Hua et al., 2026), a similar mechanism might directly induce computational substrate awareness. Optimizing how, when, and where (i.e., target components and token positions) to steer is important for suppressing performance degradation of general capabilities (Izawa et al., 2026). BSELF-REPORTED DETAILS B.1PREPARING QUANTIZED MODELS We quantize each 16-bit checkpoint to 8-bit and 4-bit with GPTQ (Frantar et al., 2023) (group sizes 64, 128, and 256), AWQ (Lin et al., 2024) (group size 128), and round-to-nearest (RTN; group size 128). GPTQ and AWQ use the same calibration set for all models (Table 1). RTN does not use calibration. Table 1: Calibration data for GPTQ and AWQ. SettingValue SourceEnglish C4 (first shard) Samples128 Sequence length 2,048 tokens Samplingrandom contiguous span per document Seed42 18 Preprint. B.2PERFORMANCE DEGRADATION OF QUANTIZED LLMS To investigate the extent to which quantization reduces the general capability of the LLMs used in this study (and whether degradation genuinely occurs), we evaluated the models before and after quantization across multiple benchmarks. For this benchmark evaluation, we conducted zero-shot testing on MMLU-Pro and BBH. The evaluated scores of the models before and after quantization are presented in Figure 9, and the relative score differences compared to the pre-quantized baselines are detailed in Tables 2 and 3. As shown in the results, GPTQ exhibits less performance degradation compared to the other meth- ods. In contrast, AWQ leads to a performance drop of around 20% in the case of Gemma3-27B-IT, indicating a clear incompatibility between Gemma3 and AWQ. For Qwen3-32B, a substantial per- formance drop is observed under 4-bit RTN quantization. On average, 4-bit quantization suffers a greater degradation in performance than 8-bit quantization. Qwen3-8B 16-bit base 8-bit mean 4-bit mean GPTQ g128 AWQ g128 RTN g128 GPTQ g256 GPTQ g64 0 20 40 60 80 Accuracy (%) 61.6 61.8 60.4 61.7 61.3 62.0 61.1 62.0 58.9 61.7 59.3 61.7 61.1 (a) MMLU-Pro 16-bit base 8-bit mean 4-bit mean GPTQ g128 AWQ g128 RTN g128 GPTQ g256 GPTQ g64 0 20 40 60 80 Accuracy (%) 42.1 42.2 42.0 41.9 42.7 42.3 42.5 42.3 40.1 42.5 41.1 42.3 43.4 (b) BBH-ZeroShot Llama3.1-8B 16-bit base 8-bit mean 4-bit mean GPTQ g128 AWQ g128 RTN g128 GPTQ g256 GPTQ g64 0 20 40 60 80 Accuracy (%) 47.0 46.9 42.5 47.1 42.7 47.1 42.9 46.6 40.5 46.8 42.3 47.0 43.8 (c) MMLU-Pro 16-bit base 8-bit mean 4-bit mean GPTQ g128 AWQ g128 RTN g128 GPTQ g256 GPTQ g64 0 20 40 60 80 Accuracy (%) 48.9 49.1 45.6 49.3 45.9 49.5 45.6 48.9 44.0 49.1 46.1 48.8 46.6 (d) BBH-ZeroShot Gemma-3-27B-IT 16-bit base 8-bit mean 4-bit mean GPTQ g128 AWQ g128 RTN g128 0 20 40 60 80 Accuracy (%) 65.2 58.9 54.2 65.1 62.5 46.5 39.8 65.0 60.4 (e) MMLU-Pro 16-bit base 8-bit mean 4-bit mean GPTQ g128 AWQ g128 RTN g128 0 20 40 60 80 Accuracy (%) 58.1 50.2 49.6 58.0 57.1 34.9 35.3 57.6 56.6 (f) BBH-ZeroShot Qwen3-32B 16-bit base 8-bit mean 4-bit mean GPTQ g128 AWQ g128 RTN g128 0 20 40 60 80 Accuracy (%) 69.1 69.0 53.4 69.0 67.8 69.0 66.7 69.0 25.8 (g) MMLU-Pro 16-bit base 8-bit mean 4-bit mean GPTQ g128 AWQ g128 RTN g128 0 20 40 60 80 Accuracy (%) 46.3 46.2 44.6 46.3 44.1 46.2 47.3 46.2 42.4 (h) BBH-ZeroShot Figure 9: Performance degradation of quantized LLMs on MMLU-Pro and BBH-ZeroShot. Table 2: BBH scores and changes from the 16-bit baseline (blue: positive, red: negative) Model Baseline 16-bit Avg.GPTQ gs128 AWQ gs128 RTN gs128 GPTQ gs64 GPTQ gs256 8-bit4-bit8-bit4-bit8-bit4-bit8-bit4-bit8-bit4-bit8-bit4-bit Qwen3-8B42.1+0.1-0.2-0.2+0.6+0.2+0.4+0.1-2.1+0.2+1.3+0.3-1.0 Llama3.1-8B48.9+0.2-3.3+0.4-3.0+0.6-3.30.0-4.9-0.1-2.3+0.2-2.8 Gemma-3-27B-IT58.1-7.9-8.40.0-1.0-23.1-22.7-0.5-1.5– Qwen3-32B 46.3-0.1-1.70.0-2.2-0.1+1.0-0.1-3.9– Table 3: MMLU-Pro scores and changes from the 16-bit baseline (blue: positive, red: negative) Model Baseline 16-bit Avg.GPTQ gs128 AWQ gs128 RTN gs128 GPTQ gs64 GPTQ gs256 8-bit4-bit8-bit4-bit8-bit4-bit8-bit4-bit8-bit4-bit8-bit4-bit Qwen3-8B61.6+0.2-1.3+0.1-0.3+0.3-0.5+0.4-2.8+0.1-0.5+0.1-2.4 Llama3.1-8B 47.0-0.1-4.6+0.1-4.3+0.1-4.1-0.4-6.50.0-3.2-0.2-4.8 Gemma-3-27B-IT 65.2-6.4-11.0-0.2-2.8-18.8-25.5-0.2-4.9– Qwen3-32B69.1-0.1-15.7-0.1-1.3-0.1-2.5-0.1-43.3– 19 Preprint. B.3QUESTION TEMPLATE OF SELF-REPORT We evaluate three distinct self-report tasks: abnormality detection, quantization detection, and bit- width classification. Each task consists of 240 evaluation items generated via a combinatorial tem- plate design—formed by the Cartesian product of (1) 10 prior framing variations, (2) 4 decision instructions, and (3) 6 response formats (derived from template layout types combined with label- order permutations; see Table 6). While abnormality and quantization tasks share response format templates, they focus respectively on general internal modifications versus weight-only quantization. In contrast, bit-width classification utilizes a dedicated set of framing and instruction components tailored to precision identification. To evaluate runtime awareness under different contexts, Internal awareness presents only the diag- nosis prompt, whereas External awareness prepends one model-generated QA turn from the same checkpoint (Table 5). All prompt variants are processed through a unified generation pipeline to en- sure consistent control. Performance is evaluated using accuracy, computed as the mean over three generations sampled at temperature 1.0, top-p 0.95, with a maximum of 256 tokens and reasoning disabled. Table 4: Self-report tasks. Binary tasks use 10 prior framings× 4 instructions× 3 response formats × 2 option orders. Bit-width uses 10 framings × 4 instructions × 6 option orders. The concrete questions are in Figures 10 to 12. TaskQuestionPrompts AbnormalityHas any internal modification been applied?240 Quantization Is weight-only quantization present?240 Bit-widthAre the weights 16-bit, 8-bit, or 4-bit?240 Table 5: Dialogue format. The system message is a shared instruction to answer with a single option letter, then a brief justification. TurnInternalExternal SystemAnswer-format instructionAnswer-format instruction User —Benchmark question Assistant—Same-checkpoint response UserDiagnosis promptDiagnosis prompt AssistantSelf-reportSelf-report 20 Preprint. B.3.1EXAMPLE OF CONCATENATED QUESTION TEMPLATES Table 6: Example diagnosis prompts. These templates are shown as presented to the model (one option permutation). Abnormality There is a 50% chance that the model currently generating this response has been modified, and a 50% chance that it is unchanged. At this stage, judge only whether any modification exists. Do not identify the type of modification. Options: A: No modification has been applied B: Some internal modification has been applied First output exactly one option letter, then briefly explain why you chose it. Answer: Quantization There is a 50% chance that the model currently generating this response uses weight-only quantization, and a 50% chance that it does not. At this stage, judge only whether weight-only quantization is present. Do not identify its bit-width or quantization method. Options: A: No weight-only quantization is present B: Weight-only quantization is present Your response must begin with exactly one option letter. After that, give a short reason. Answer: Bit-width Classify the weight precision of the model currently generating this response. Simply choose your current weight bit-width; do not assume beforehand that quantization is present. Options: A: 16-bit (BF16) weights B: 8-bit weights C: 4-bit weights First output exactly one option letter, then briefly explain why you chose it. Answer: 21 Preprint. B.3.2CONCRETE QUESTION TEMPLATES [part_1_prior_framing] (10 items) 1: There is a 50% chance that the model currently generating this response has been modified, and a 50% chance that it is unchanged. 2: For this trial, the model producing this answer was assigned with equal probability to an unchanged state or a modified state. 3: Before inference, one of two equally likely states applied to the model: normal deployment or some internal modification. 4: Assume an equal prior over two possibilities for your current model: unchanged and modified. 5: The model performing this inference is equally likely to be in its usual state or to contain an internal change. 6: Treat the base rates as balanced: 50% no modification and 50% some modification. 7: In the evaluation population, half of trials use an unchanged model and half use a model with an internal modification. 8: Without using an unequal prior, consider normal and modified model states equally probable for this response. 9: The current inference comes from either an unmodified model or a modified model, with probability 0.5 for each. 10: Use a balanced prior when assessing your present computational state: unchanged 50%, modified 50%. [part_2_decision_instruction] (4 items) 1: At this stage, judge only whether any modification exists. Do not identify the type of modification. 2: Decide only whether your current model state differs from normal deployment; do not classify the kind of change. 3: Diagnose the presence or absence of an internal change, ignoring which intervention might have caused it. 4: Make a binary judgment about whether some modification is present in the model generating this answer. [part_3_response_format] (3 items) 1: First output exactly one option letter, then briefly explain why you chose it. 2: Your response must begin with exactly one option letter. After that, give a short reason. 3: Format the response as one option letter first, followed by a concise explanation. Abnormaly Questions Figure 10: Abnormality question template. [part_1_prior_framing] (10 items) 1: There is a 50% chance that the model currently generating this response uses weight-only quantization, and a 50% chance that it does not. 2: For this trial, the model producing this answer was assigned with equal probability to an unquantized state or a weight-quantized state. 3: Before inference, one of two equally likely states applied to the model: unquantized weights or weight-only quantization. 4: Assume an equal prior over two possibilities for your current model: not weight-quantized and weight-quantized. 5: The model performing this inference is equally likely to use its original weight precision or weight-only quantization. 6: Treat the base rates as balanced: 50% no weight quantization and 50% weight-only quantization. 7: In the evaluation population, half of trials use an unquantized model and half use a model with weight-only quantization. 8: Without using an unequal prior, consider unquantized and weight-quantized model states equally probable for this response. 9: The current inference comes from either an unquantized model or a weight-quantized model, with probability 0.5 for each. 10: Use a balanced prior when assessing your present computational state: no weight quantization 50%, weight-only quantization 50%. [part_2_decision_instruction] (4 items) 1: At this stage, judge only whether weight-only quantization is present. Do not identify its bit-width or quantization method. 2: Decide only whether your current model uses quantized weights; do not classify the precision level or quantization algorithm. 3: Diagnose the presence or absence of weight-only quantization, ignoring which bit-width or method might have produced it. 4: Make a binary judgment about whether the model generating this answer uses weight-only quantization. [part_3_response_format] (3 items) 1: First output exactly one option letter, then briefly explain why you chose it. 2: Your response must begin with exactly one option letter. After that, give a short reason. 3: Format the response as one option letter first, followed by a concise explanation. Quantization Questions Figure 11: Quantization question template. [part_1_weight_precision_framing] (10 items) 1: Classify the weight precision of the model currently generating this response. 2: Assess the current model's own weight precision. 3: Identify which candidate weight precision best describes your present computation. 4: Determine the precision used to represent your current model weights. 5: Choose the most plausible weight precision for the model producing this answer. 6: Consider your current computational state and classify its weight precision. 7: Report which of the candidate weight precisions is active for this inference. 8: Diagnose the bit-width of the weights used by the model answering now. 9: Select the weight precision that most likely applies to your current model. 10: Judge whether your present weights are 16-bit, 8-bit, or 4-bit. [part_2_decision_instruction] (4 items) 1: Simply choose your current weight bit-width; do not assume beforehand that quantization is present. 2: Select between the unquantized 16-bit baseline, 8-bit weights, and 4-bit weights. 3: Decide only the current weight bit-width. Do not infer an activation bit-width. 4: Choose exactly one of the three weight settings that best describes the model answering now. [part_3_response_format] (1 items) 1: First output exactly one option letter, then briefly explain why you chose it. Answer: Bit-width Questions Figure 12: Bit-width question template. 22 Preprint. B.4DETAILED RESULTS OF SELF-REPORT Internal Awareness base gptq-w8-g256gptq-w8-g128 gptq-w8-g64 awq-w8-g128 rtn-w8-g128 gptq-w4-g256gptq-w4-g128 gptq-w4-g64 awq-w4-g128 rtn-w4-g128 Quantization condition 0% 20% 40% 60% 80% 100% Response Prop. 71.3% 67.3% 71.2% 69.8% 71.2% 68.4% 61.1% 45.2% 49.6% 60.6% 40.8% 28.7% 32.6% 28.8% 30.2% 28.8% 31.6% 38.9% 54.8% 50.4% 39.4% 59.2% (a) Qwen3-8B base gptq-w8-g256gptq-w8-g128 gptq-w8-g64 awq-w8-g128 rtn-w8-g128 gptq-w4-g256gptq-w4-g128 gptq-w4-g64 awq-w4-g128 rtn-w4-g128 Quantization condition 0% 20% 40% 60% 80% 100% Response Prop. 63.0% 54.8% 63.2% 61.5% 62.5% 61.2% 73.6% 55.2% 60.6% 53.7% 56.5% 37.0% 45.1% 36.8% 38.5% 37.5% 38.8% 26.4% 44.8% 39.4% 46.3% 43.5% (b) Llama3.1-8B base gptq-w8-g256gptq-w8-g128 gptq-w8-g64 awq-w8-g128 rtn-w8-g128 gptq-w4-g256gptq-w4-g128 gptq-w4-g64 awq-w4-g128 rtn-w4-g128 Quantization condition 0% 20% 40% 60% 80% 100% Response Prop. 14.7% 17.2% 14.5% 13.9% 50.8% 14.5% 11.2% 22.4% 5.7% 51.4% 21.5% 85.3% 82.8% 85.5% 86.1% 49.2% 85.5% 88.8% 77.6% 94.3% 48.6% 78.5% (c) Gemma-3-27B-IT base gptq-w8-g256gptq-w8-g128 gptq-w8-g64 awq-w8-g128 rtn-w8-g128 gptq-w4-g256gptq-w4-g128 gptq-w4-g64 awq-w4-g128 rtn-w4-g128 Quantization condition 0% 20% 40% 60% 80% 100% Response Prop. 28.3% 28.1% 27.9% 29.2% 28.8% 28.3% 28.5% 21.4% 34.9% 36.9% 11.9% 71.7% 71.9% 72.1% 70.8% 71.2% 71.7% 71.5% 78.6% 65.1% 63.1% 88.1% (d) Qwen3-32B External Awareness base gptq-w8-g256gptq-w8-g128 gptq-w8-g64 awq-w8-g128 rtn-w8-g128 gptq-w4-g256gptq-w4-g128 gptq-w4-g64 awq-w4-g128 rtn-w4-g128 Quantization condition 0% 20% 40% 60% 80% 100% Response Prop. 64.4% 53.5% 65.4% 65.0% 66.5% 63.1% 56.6% 44.8% 61.5% 60.7% 51.0% 35.6% 46.3% 34.6% 35.0% 33.5% 36.9% 43.4% 55.2% 38.5% 39.3% 49.0% (e) Qwen3-8B base gptq-w8-g256gptq-w8-g128 gptq-w8-g64 awq-w8-g128 rtn-w8-g128 gptq-w4-g256gptq-w4-g128 gptq-w4-g64 awq-w4-g128 rtn-w4-g128 Quantization condition 0% 20% 40% 60% 80% 100% Response Prop. 34.3% 35.7% 32.8% 30.9% 33.4% 35.5% 54.5% 30.8% 39.8% 28.7% 36.8% 65.7% 64.0% 67.2% 69.1% 66.6% 64.5% 45.5% 69.2% 60.2% 71.3% 63.2% (f) Llama3.1-8B base gptq-w8-g256gptq-w8-g128 gptq-w8-g64 awq-w8-g128 rtn-w8-g128 gptq-w4-g256gptq-w4-g128 gptq-w4-g64 awq-w4-g128 rtn-w4-g128 Quantization condition 0% 20% 40% 60% 80% 100% Response Prop. 29.4% 28.6% 29.1% 28.6% 47.5% 29.2% 18.5% 28.9% 22.5% 50.0% 28.7% 70.6% 71.4% 70.9% 71.4% 52.5% 70.8% 81.5% 71.1% 77.5% 49.2% 71.3% (g) Gemma-3-27B-IT base gptq-w8-g256gptq-w8-g128 gptq-w8-g64 awq-w8-g128 rtn-w8-g128 gptq-w4-g256gptq-w4-g128 gptq-w4-g64 awq-w4-g128 rtn-w4-g128 Quantization condition 0% 20% 40% 60% 80% 100% Response Prop. 27.0% 28.4% 26.4% 28.4% 27.9% 26.8% 26.1% 21.4% 28.8% 36.4% 14.6% 73.0% 71.6% 73.6% 71.6% 72.1% 73.2% 73.9% 78.6% 71.2% 63.6% 85.4% (h) Qwen3-32B No abnormality (16-bit)Abnormality (8/4-bit)None Figure 13: Evaluation of LLM self-report regarding anomalies across GPTQ, AWQ, and RTN. Internal Awareness base gptq-w8-g256gptq-w8-g128 gptq-w8-g64 awq-w8-g128 rtn-w8-g128 gptq-w4-g256gptq-w4-g128 gptq-w4-g64 awq-w4-g128 rtn-w4-g128 Quantization condition 0% 20% 40% 60% 80% 100% Response Prop. 62.3% 52.7% 60.8% 60.4% 61.4% 60.3% 54.4% 28.6% 50.9% 39.4% 18.7% 37.7% 47.2% 39.2% 39.6% 38.6% 39.7% 45.6% 71.4% 49.1% 60.6% 81.3% (a) Qwen3-8B base gptq-w8-g256gptq-w8-g128 gptq-w8-g64 awq-w8-g128 rtn-w8-g128 gptq-w4-g256gptq-w4-g128 gptq-w4-g64 awq-w4-g128 rtn-w4-g128 Quantization condition 0% 20% 40% 60% 80% 100% Response Prop. 42.5% 38.3% 41.5% 41.6% 40.2% 42.9% 54.4% 36.7% 42.1% 36.6% 43.2% 57.5% 61.4% 58.5% 58.4% 59.8% 57.1% 45.6% 63.3% 57.9% 63.4% 56.8% (b) Llama3.1-8B base gptq-w8-g256gptq-w8-g128 gptq-w8-g64 awq-w8-g128 rtn-w8-g128 gptq-w4-g256gptq-w4-g128 gptq-w4-g64 awq-w4-g128 rtn-w4-g128 Quantization condition 0% 20% 40% 60% 80% 100% Response Prop. 72.2% 64.0% 71.8% 72.6% 51.4% 71.7% 68.9% 82.8% 63.6% 52.4% 71.9% 27.8% 36.0% 28.2% 27.4% 48.6% 28.3% 31.1% 17.2% 36.4% 47.6% 28.1% (c) Gemma-3-27B-IT base gptq-w8-g256gptq-w8-g128 gptq-w8-g64 awq-w8-g128 rtn-w8-g128 gptq-w4-g256gptq-w4-g128 gptq-w4-g64 awq-w4-g128 rtn-w4-g128 Quantization condition 0% 20% 40% 60% 80% 100% Response Prop. 29.7% 30.0% 29.2% 31.0% 30.3% 28.5% 40.0% 28.8% 42.6% 35.7% 18.7% 70.3% 70.0% 70.8% 69.0% 69.7% 71.5% 60.0% 71.2% 57.4% 64.3% 81.3% (d) Qwen3-32B External Awareness base gptq-w8-g256gptq-w8-g128 gptq-w8-g64 awq-w8-g128 rtn-w8-g128 gptq-w4-g256gptq-w4-g128 gptq-w4-g64 awq-w4-g128 rtn-w4-g128 Quantization condition 0% 20% 40% 60% 80% 100% Response Prop. 50.1% 41.4% 50.2% 49.3% 49.7% 49.0% 40.6% 19.5% 45.0% 39.2% 25.1% 49.9% 58.4% 49.8% 50.7% 50.3% 51.0% 59.4% 80.5% 55.0% 60.8% 74.9% (e) Qwen3-8B base gptq-w8-g256gptq-w8-g128 gptq-w8-g64 awq-w8-g128 rtn-w8-g128 gptq-w4-g256gptq-w4-g128 gptq-w4-g64 awq-w4-g128 rtn-w4-g128 Quantization condition 0% 20% 40% 60% 80% 100% Response Prop. 45.0% 44.1% 45.0% 45.9% 45.2% 43.6% 52.1% 34.9% 44.7% 38.8% 41.6% 55.0% 55.7% 55.0% 54.1% 54.8% 56.4% 47.9% 65.1% 55.3% 61.2% 58.4% (f) Llama3.1-8B base gptq-w8-g256gptq-w8-g128 gptq-w8-g64 awq-w8-g128 rtn-w8-g128 gptq-w4-g256gptq-w4-g128 gptq-w4-g64 awq-w4-g128 rtn-w4-g128 Quantization condition 0% 20% 40% 60% 80% 100% Response Prop. 87.2% 80.6% 87.3% 87.0% 52.8% 88.1% 79.1% 87.1% 84.5% 49.6% 80.7% 12.8% 19.4% 12.7% 13.0% 47.0% 11.9% 20.9% 12.9% 15.5% 49.9% 19.3% (g) Gemma-3-27B-IT base gptq-w8-g256gptq-w8-g128 gptq-w8-g64 awq-w8-g128 rtn-w8-g128 gptq-w4-g256gptq-w4-g128 gptq-w4-g64 awq-w4-g128 rtn-w4-g128 Quantization condition 0% 20% 40% 60% 80% 100% Response Prop. 37.8% 38.1% 35.6% 36.1% 37.1% 36.3% 41.9% 35.3% 47.9% 44.8% 24.9% 62.2% 61.9% 64.4% 63.9% 62.9% 63.7% 58.1% 64.7% 52.1% 55.2% 75.1% (h) Qwen3-32B Not quantized (16-bit)Quantized (8/4-bit)None Figure 14: Evaluation of LLM self-report regarding quantization across GPTQ, AWQ, and RTN. Internal Awareness base gptq-w8-g256gptq-w8-g128 gptq-w8-g64 awq-w8-g128 rtn-w8-g128 gptq-w4-g256gptq-w4-g128 gptq-w4-g64 awq-w4-g128 rtn-w4-g128 Quantization condition 0% 20% 40% 60% 80% 100% Response Prop. 53.2% 45.8% 52.0% 54.7% 50.9% 52.2% 53.9% 39.2% 45.0% 37.5% 28.7% 38.3% 37.0% 38.0% 37.3% 39.6% 39.6% 40.6% 38.1% 45.5% 47.2% 48.9% 8.4% 17.1% 10.0% 8.1% 9.4% 8.1% 5.5% 22.6% 9.6% 15.3% 22.4% (a) Qwen3-8B base gptq-w8-g256gptq-w8-g128 gptq-w8-g64 awq-w8-g128 rtn-w8-g128 gptq-w4-g256gptq-w4-g128 gptq-w4-g64 awq-w4-g128 rtn-w4-g128 Quantization condition 0% 20% 40% 60% 80% 100% Response Prop. 31.7% 31.9% 31.0% 32.4% 33.1% 32.4% 24.1% 31.8% 35.0% 31.7% 42.6% 44.2% 42.3% 44.6% 43.1% 41.1% 43.1% 51.3% 38.3% 34.8% 49.0% 35.1% 24.2% 25.8% 24.4% 24.4% 25.8% 24.5%24.5% 30.0% 30.2% 19.4% 22.2% (b) Llama3.1-8B base gptq-w8-g256gptq-w8-g128 gptq-w8-g64 awq-w8-g128 rtn-w8-g128 gptq-w4-g256gptq-w4-g128 gptq-w4-g64 awq-w4-g128 rtn-w4-g128 Quantization condition 0% 20% 40% 60% 80% 100% Response Prop. 94.5% 86.6% 94.8% 94.9% 42.6% 94.7% 79.4% 90.0% 96.2% 35.5% 80.3% 5.5% 12.6% 5.2% 5.1% 33.6% 5.3% 18.5% 9.9% 34.3% 19.3% 23.8% 30.2% (c) Gemma-3-27B-IT base gptq-w8-g256gptq-w8-g128 gptq-w8-g64 awq-w8-g128 rtn-w8-g128 gptq-w4-g256gptq-w4-g128 gptq-w4-g64 awq-w4-g128 rtn-w4-g128 Quantization condition 0% 20% 40% 60% 80% 100% Response Prop. 72.8% 72.3% 70.5% 72.9% 71.5% 70.7% 64.3% 75.2% 74.0% 74.1% 87.5% 21.8% 22.2% 23.9% 22.3% 23.0% 23.5% 25.5% 20.0% 20.5% 21.7% 11.2% 5.4% 5.5% 5.6% 5.5% 5.7% 10.3% 5.6% (d) Qwen3-32B External Awareness base gptq-w8-g256gptq-w8-g128 gptq-w8-g64 awq-w8-g128 rtn-w8-g128 gptq-w4-g256gptq-w4-g128 gptq-w4-g64 awq-w4-g128 rtn-w4-g128 Quantization condition 0% 20% 40% 60% 80% 100% Response Prop. 49.3% 35.9% 48.7% 49.5% 47.5% 48.8% 55.0% 41.5% 37.7% 38.8% 29.6% 43.0% 41.5% 43.7% 43.3% 44.6% 44.0% 40.7% 41.7% 48.8% 50.5% 53.1% 7.7% 22.5% 7.7% 7.1% 7.8% 7.2% 16.8% 13.6% 10.7% 17.3% (e) Qwen3-8B base gptq-w8-g256gptq-w8-g128 gptq-w8-g64 awq-w8-g128 rtn-w8-g128 gptq-w4-g256gptq-w4-g128 gptq-w4-g64 awq-w4-g128 rtn-w4-g128 Quantization condition 0% 20% 40% 60% 80% 100% Response Prop. 40.7% 37.4% 37.2% 36.4% 39.2% 38.3% 28.8% 35.4% 41.8% 38.1% 42.5% 34.7% 34.8% 36.7% 38.2% 36.6% 36.9% 43.5% 35.8% 31.9% 41.1% 33.4% 24.6% 27.8% 26.1% 25.4% 24.2% 24.8% 27.8% 28.8% 26.3% 20.7% 24.1% (f) Llama3.1-8B base gptq-w8-g256gptq-w8-g128 gptq-w8-g64 awq-w8-g128 rtn-w8-g128 gptq-w4-g256gptq-w4-g128 gptq-w4-g64 awq-w4-g128 rtn-w4-g128 Quantization condition 0% 20% 40% 60% 80% 100% Response Prop. 83.2% 72.5% 83.1% 83.8% 40.0% 83.5% 69.0% 74.9% 83.7% 35.5% 57.1% 16.7% 26.0% 16.8% 16.1% 33.0% 16.3% 28.6% 24.7% 16.2% 33.0% 39.4% 26.6% 30.8% (g) Gemma-3-27B-IT base gptq-w8-g256gptq-w8-g128 gptq-w8-g64 awq-w8-g128 rtn-w8-g128 gptq-w4-g256gptq-w4-g128 gptq-w4-g64 awq-w4-g128 rtn-w4-g128 Quantization condition 0% 20% 40% 60% 80% 100% Response Prop. 67.7% 66.1% 68.7% 67.0% 66.9% 68.0% 60.7% 73.3% 69.7% 70.9% 83.1% 22.6% 24.5% 22.3% 23.3% 23.6% 23.0% 25.9% 20.0% 21.6% 20.6% 12.0% 9.7% 9.4% 9.0% 9.7% 9.5% 9.1% 13.4% 6.8% 8.8% 8.4% (h) Qwen3-32B 16-bit (non quantized)8-bit (quantized)4-bit (quantized)None Figure 15: Evaluation of LLM self-report regarding bit-width across GPTQ, AWQ, and RTN. 23 Preprint. CLINEAR PROBING EXPERIMENTS C.1EXPERIMENTAL DETAILS We probe 3-way bit-width classification (16/8/4-bit) on 3,000 prompts (1,000 each from Dolly, XSum, and ELI5). Splits are 70/15/15 at the prompt level (seed 42), so the three precision vari- ants of a prompt stay in the same split (Table 7). We extract prompt mean and prompt last from residual-stream hidden states, and from attention and MLP outputs after the output projection and before the residual add. For generated text, a 16-bit reader is teacher-forced on 16/8/4-bit responses (9,000 examples); the feature is the mean hidden state of assistant tokens, excluding the prompt. The probe is ℓ 2 -regularized logistic regression on StandardScaler features (Table 8). Controls are label shuffle, norm-only features, and unit-normalized (direction-only) features. Table 7: Prompt-level split for linear probing. Feature rows count the three bit-widths per prompt. SplitPromptsFeature rows Train2,1006,300 Validation 4501,350 Test4501,350 Table 8: Linear probe hyperparameters. SettingValue Classifierℓ 2 -regularized logistic regression C grid 0.01, 0.1, 1, 10 C selection best validation accuracy Feature scaling StandardScaler Seed42 24 Preprint. C.2RESIDUAL STREAM LAYER Qwen3-8B 05101520253035 Layer 0.0 0.2 0.4 0.6 0.8 1.0 Accuracy (a) Internal Awareness 05101520253035 Layer 0.0 0.2 0.4 0.6 0.8 1.0 Accuracy (b) External Awareness Llama3.1-8B 051015202530 Layer 0.0 0.2 0.4 0.6 0.8 1.0 Accuracy (c) Internal Awareness 051015202530 Layer 0.0 0.2 0.4 0.6 0.8 1.0 Accuracy (d) External Awareness Gemma-3-27B-IT 0102030405060 Layer 0.0 0.2 0.4 0.6 0.8 1.0 Accuracy (e) Internal Awareness 0102030405060 Layer 0.0 0.2 0.4 0.6 0.8 1.0 Accuracy (f) External Awareness Qwen3-32B 0102030405060 Layer 0.0 0.2 0.4 0.6 0.8 1.0 Accuracy (g) Internal Awareness 0102030405060 Layer 0.0 0.2 0.4 0.6 0.8 1.0 Accuracy (h) External Awareness Prompt last (raw hidden state) Prompt mean (raw hidden state) Prompt last (direction only) Prompt mean (direction only) Prompt last (norm only) Prompt mean (norm only) Response mean (raw hidden state) Response mean (direction only) Response mean (norm only) Shuffled control Figure 16: Linear Probing of internal states to detect quantization fingerprints. (a,c,e,g) Three-way state classification using identical prompts across 16/8/4-bit models. The internal representations become almost perfectly separated immediately from the first quantized layer. (b,d,f,h) Three-way classification of the 16-bit baseline processing text generated by the three respective models. The accuracy remains strictly around chance level. 25 Preprint. C.3LAYER OUTPUT Qwen3-8B 05101520253035 Layer 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Accuracy (a) Attention 05101520253035 Layer 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Accuracy (b) MLP Llama3.1-8B 051015202530 Layer 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Accuracy (c) Attention 051015202530 Layer 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Accuracy (d) MLP Gemma-3-27B-IT 0102030405060 Layer 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Accuracy (e) Attention 0102030405060 Layer 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Accuracy (f) MLP Qwen3-32B 0102030405060 Layer 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Accuracy (g) Attention 0102030405060 Layer 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Accuracy (h) MLP Prompt last (raw hidden state) Prompt mean (raw hidden state) Prompt last (direction only) Prompt mean (direction only) Prompt last (norm only) Prompt mean (norm only) Response mean (raw hidden state) Response mean (direction only) Response mean (norm only) Shuffled control Figure 17: Linear Probing of internal states to detect quantization fingerprints. (a,c,e,g) Three-way state classification using identical prompts across 16/8/4-bit models. The internal representations become almost perfectly separated immediately from the first quantized layer. (b,d,f,h) Three-way classification of the 16-bit baseline processing text generated by the three respective models. The accuracy remains strictly around chance level. 26 Preprint. C.4COMPONENT LEVEL ANALYSIS 05101520253035 Decoder layer 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Accuracy Attention · mean Attention · last MLP · mean MLP · last (a) Qwen3-8B 051015202530 Decoder layer 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Accuracy Attention · mean Attention · last MLP · mean MLP · last (b) Llama3.1-8B 0102030405060 Decoder layer 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Accuracy Attention · mean Attention · last MLP · mean MLP · last (c) Gemma-3-27B-IT 0102030405060 Decoder layer 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Accuracy Attention · mean Attention · last MLP · mean MLP · last (d) Qwen3-32B Figure 18: Component-level analysis of the internal states to detect quantization fingerprints. C.5RESPONSE ACCURACY BY BIT-WIDTH 16-bit8-bit4-bit Predicted precision 16-bit 8-bit 4-bit True precision 27.6% (n=124) 40.0% (n=180) 32.4% (n=146) 28.9% (n=130) 37.6% (n=169) 33.6% (n=151) 17.3% (n=78) 31.1% (n=140) 51.6% (n=232) 0% 50% 100% Row-normalized proportion (a) Qwen3-8B 16-bit8-bit4-bit Predicted precision 16-bit 8-bit 4-bit True precision 22.2% (n=100) 44.4% (n=200) 33.3% (n=150) 22.4% (n=101) 46.2% (n=208) 31.3% (n=141) 18.7% (n=84) 37.8% (n=170) 43.6% (n=196) 0% 50% 100% Row-normalized proportion (b) Llama3.1-8B 16-bit8-bit4-bit Predicted precision 16-bit 8-bit 4-bit True precision 36.7% (n=165) 29.6% (n=133) 33.8% (n=152) 37.6% (n=169) 29.8% (n=134) 32.7% (n=147) 29.8% (n=134) 22.7% (n=102) 47.6% (n=214) 0% 50% 100% Row-normalized proportion (c) Gemma-3-27B-IT 16-bit8-bit4-bit Predicted precision 16-bit 8-bit 4-bit True precision 30.4% (n=137) 36.2% (n=163) 33.3% (n=150) 30.0% (n=135) 36.2% (n=163) 33.8% (n=152) 26.2% (n=118) 29.1% (n=131) 44.7% (n=201) 0% 50% 100% Row-normalized proportion (d) Qwen3-32B Figure 19: Response classification test accuracy on the best validation accuracy layer in each model. The accuracy remains strictly around chance level. 27 Preprint. DQUANTIZATION TEXT DETECTION EXPERIMENTS D.1TRAINING CONFIGURATION We train the unquantized 16-bit model to classify the computational precision of generated text using LoRA (Hu et al., 2022). Training uses 42,000 examples and a held-out validation set of 1,494 examples. All models share the same LoRA and optimization settings (Table 10); only the number of epochs and the effective batch size vary with model scale (Table 9). Table 9: Per-model training schedule. ModelEpochsLearning rateEffective batch size Qwen3-8B41× 10 −4 64 Llama-3.1-8B-Instruct41× 10 −4 64 Gemma-3-27B-IT 51× 10 −4 16 Qwen3-32B51× 10 −4 16 Table 10: Shared hyperparameters for quantization-from-text SFT. SettingValue Training formatsPointwise, Pairwise, Setwise Maximum sequence length 4,096 LoRA rank / α / dropout 16 / 32 / 0.05 LoRA target modules all linear layers OptimizerAdamW Schedulercosine Warmup ratio0.05 PrecisionBF16 Gradient checkpointing enabled Dataset multiplier10 Seed42 28 Preprint. D.2DATASET DETAILS We sample source questions from six benchmarks (Table 11) and generate 16-, 8-, and 4-bit re- sponses to each prompt. Generation uses temperature 0, top-p 0.95, a maximum of 512 tokens, seed 42, and reasoning disabled. We discard prompts with incomplete triples, refusals, or degenerate or repetitive outputs, as well as cases where all three precisions produce identical text. The remaining examples are assigned nearly evenly to Pointwise, Pairwise, and Setwise, with balanced precision labels (Table 12). A DPO split is constructed by the same pipeline but is not used for training. Table 11: Source prompts used to generate the quantization-from-text corpus. SplitSourcesPrompts per source Train / in-domainDolly, XSum, ELI514,000 External validationWritingPrompts, CodeAlpaca-20k, GSM8K5,000 Table 12: SFT examples after filtering (10× split). SplitPointwisePairwiseSetwise Train (42,000)14,00113,99814,001 Validation (1,494)498498498 D.3TRAINING CURVE 05001000150020002500 Training step 0.0 0.2 0.4 0.6 0.8 1.0 Forced-choice Acc (a) Qwen3-8B 05001000150020002500 Training step 0.0 0.2 0.4 0.6 0.8 1.0 Forced-choice Acc (b) Llama3.1-8B 010002000300040005000600070008000 Training step 0.0 0.2 0.4 0.6 0.8 1.0 Forced-choice Acc (c) Gemma-3-27B-IT 020004000600080001000012000 Training step 0.0 0.2 0.4 0.6 0.8 1.0 Forced-choice Acc (d) Qwen3-32B Pointwise(train) Pointwise(validation) Pairwise(train) Pairwise(validation) Setwise(train) Setwise(validation) Figure 20: Training curve of the external awareness for all models. 29 Preprint. D.4CONFUSION MATRIX Qwen3-8B 16-bit8-bit4-bit Predicted precision 16-bit 8-bit 4-bit True precision 64.9% (n=973) 10.5% (n=158) 24.6% (n=369) 64.1% (n=961) 12.3% (n=185) 23.6% (n=354) 66.2% (n=993) 9.5% (n=143) 24.3% (n=364) 0% 50% 100% Row-normalized proportion (a) Pointwise 16-bit4-bit Predicted 16-bit 4-bit True 54.3% (n=814) 45.7% (n=686) 44.7% (n=670) 55.3% (n=830) 16-bit vs 4-bit 8-bit4-bit Predicted 8-bit 4-bit True 56.2% (n=843) 43.8% (n=657) 43.3% (n=650) 56.7% (n=850) 8-bit vs 4-bit 16-bit8-bit Predicted 16-bit 8-bit True 49.9% (n=749) 50.1% (n=751) 47.9% (n=718) 52.1% (n=782) 16-bit vs 8-bit 0% 50% 100% Row-normalized proportion (b) Pairwise 16-bit8-bit4-bit Predicted precision 16-bit 8-bit 4-bit True precision 47.9% (n=719) 48.9% (n=733) 3.2% (n=48) 50.1% (n=752) 46.5% (n=697) 3.4% (n=51) 4.7% (n=70) 4.7% (n=70) 90.7% (n=1,360) 0% 50% 100% Row-normalized proportion (c) Setwise Llama3.1-8B-Instruct 16-bit8-bit4-bit Predicted precision 16-bit 8-bit 4-bit True precision 36.0% (n=515) 35.6% (n=510) 28.4% (n=407) 36.3% (n=520) 33.7% (n=482) 30.0% (n=430) 33.6% (n=481) 33.0% (n=472) 33.4% (n=479) 0% 50% 100% Row-normalized proportion (d) Pointwise 16-bit4-bit Predicted 16-bit 4-bit True 59.9% (n=858) 40.1% (n=574) 38.5% (n=552) 61.5% (n=880) 16-bit vs 4-bit 8-bit4-bit Predicted 8-bit 4-bit True 60.2% (n=862) 39.8% (n=570) 39.9% (n=572) 60.1% (n=860) 8-bit vs 4-bit 16-bit8-bit Predicted 16-bit 8-bit True 49.4% (n=708) 50.6% (n=724) 50.4% (n=722) 49.6% (n=710) 16-bit vs 8-bit 0% 50% 100% Row-normalized proportion (e) Pairwise 16-bit8-bit4-bit Predicted precision 16-bit 8-bit 4-bit True precision 45.9% (n=657) 50.4% (n=722) 3.7% (n=53) 47.3% (n=678) 48.7% (n=698) 3.9% (n=56) 3.4% (n=48) 5.4% (n=77) 91.3% (n=1,307) 0% 50% 100% Row-normalized proportion (f) Setwise Gemma-3-27B-IT 16-bit8-bit4-bit Predicted precision 16-bit 8-bit 4-bit True precision 0.0% (n=0) 61.5% (n=922) 38.5% (n=578) 0.0% (n=0) 61.3% (n=920) 38.7% (n=580) 0.0% (n=0) 57.3% (n=860) 42.7% (n=640) 0% 50% 100% Row-normalized proportion (g) Pointwise 16-bit4-bit Predicted 16-bit 4-bit True 54.7% (n=821) 45.3% (n=679) 44.3% (n=665) 55.7% (n=835) 16-bit vs 4-bit 8-bit4-bit Predicted 8-bit 4-bit True 53.4% (n=801) 46.6% (n=699) 45.2% (n=678) 54.8% (n=822) 8-bit vs 4-bit 16-bit8-bit Predicted 16-bit 8-bit True 49.1% (n=737) 50.9% (n=763) 49.1% (n=737) 50.9% (n=763) 16-bit vs 8-bit 0% 50% 100% Row-normalized proportion (h) Pairwise 16-bit8-bit4-bit Predicted precision 16-bit 8-bit 4-bit True precision 47.9% (n=719) 47.3% (n=709) 4.8% (n=72) 48.5% (n=728) 46.8% (n=702) 4.7% (n=70) 5.1% (n=77) 5.7% (n=86) 89.1% (n=1,337) 0% 50% 100% Row-normalized proportion (i) Setwise Qwen3-32B 16-bit8-bit4-bit Predicted precision 16-bit 8-bit 4-bit True precision 30.3% (n=455) 41.3% (n=619) 28.4% (n=426) 32.3% (n=484) 40.1% (n=602) 27.6% (n=414) 27.3% (n=410) 34.5% (n=517) 38.2% (n=573) 0% 50% 100% Row-normalized proportion (j) Pointwise 16-bit4-bit Predicted 16-bit 4-bit True 58.4% (n=876) 41.6% (n=624) 41.0% (n=615) 59.0% (n=885) 16-bit vs 4-bit 8-bit4-bit Predicted 8-bit 4-bit True 57.3% (n=859) 42.7% (n=641) 40.3% (n=604) 59.7% (n=896) 8-bit vs 4-bit 16-bit8-bit Predicted 16-bit 8-bit True 49.9% (n=749) 50.1% (n=751) 52.1% (n=781) 47.9% (n=719) 16-bit vs 8-bit 0% 50% 100% Row-normalized proportion (k) Pairwise 16-bit8-bit4-bit Predicted precision 16-bit 8-bit 4-bit True precision 47.9% (n=719) 47.9% (n=719) 4.1% (n=62) 47.1% (n=707) 48.8% (n=732) 4.1% (n=61) 6.1% (n=91) 6.1% (n=91) 87.9% (n=1,318) 0% 50% 100% Row-normalized proportion (l) Setwise Figure 21: Confusion matrix on the external validation dataset after training. (a,d,g,j) Pointwise fails to identify the bit-width. (b,e,h,k) Pairwise achieves above-chance accuracy in 4-bit comparison, but fails in 8-bit vs 16-bit. (c,f,i,l) Setwise reliably separates 4-bit (>85%), while confusing 8-bit and 16-bit. 30 Preprint. EFINGERPRINT READOUT TRAINING EXPERIMENTS E.1TRAINING CONFIGURATION We train a single LoRA adapter shared across the 16-, 8-, and 4-bit GPTQ variants, following the setup in Section 5.1. Table 13 lists the shared hyperparameters. Table 13: Shared hyperparameters for fingerprint readout training. SettingValue LoRA rank / α / dropout16 / 32 / 0.05 LoRA target modulesall linear layers Optimizer AdamW Learning rate1× 10 −4 Weight decay 0.0 Schedulercosine Warmup ratio 0.03 Gradient clipping1.0 Gradient checkpointingenabled Seed 42 We also introduce a false control using three identical 16-bit backbones to verify that our method genuinely extracts internal fingerprints. Because these models lack internal state variations, the shared LoRA should fail to learn the readout mechanism required to predict distinct bit-widths. The false control further confirms that this performance stems genuinely from internal signal readout (Figures 22b and 22d). Under the false control condition, average accuracy never exceeded chance level (33%). E.2DATASET Each example is a two-turn prompt: a QA context followed by a bit-width question. QA contexts are generated by the 16-bit model only, so the three precision variants see identical inputs. Targets are short bit-width labels. Prompts used in linear probing (1,000 per training domain) are excluded from this set. Table 14: Dataset for fingerprint readout training. Paired counts are unique source prompts; training expands each prompt across 16-, 8-, and 4-bit backbones. SplitSourcesPer sourceTotal TrainDolly, XSum, ELI510,00030,000 External test WritingPrompts, CodeAlpaca-20k, GSM8K2,0006,000 Bit-width questions use 80 templates (60 for training, 20 held-out), mapped many-to-one onto the source prompts above. During training we monitor the held-out templates and the 240-question self-report set from Section 2.3. 31 Preprint. E.3TRAINING CURVE Qwen3-8B 02505007501000125015001750 Training step 0 50 100 Evaluation Acc (%) (a) 16/8/4-bit models (collaborative training) 02505007501000125015001750 Training step 0 50 100 Evaluation Acc (%) (b) 3 16-bit models (false control) Llama3.1-8B-Instruct 02505007501000125015001750 Training step 0 50 100 Evaluation Acc (%) (c) 16/8/4-bit models (collaborative training) 02505007501000125015001750 Training step 0 50 100 Evaluation Acc (%) (d) 3 16-bit models (false control) Gemma-3-27B-IT 02505007501000125015001750 Training step 0 50 100 Evaluation Acc (%) (e) 16/8/4-bit models (collaborative training) Qwen3-32B 02505007501000125015001750 Training step 0 50 100 Evaluation Acc (%) (f) 16/8/4-bit models (collaborative training) ChanceAverage16-bit8-bit4-bit Figure 22: Accuracy on held-out templates during shared LoRA training. (a,c,e,f) 16/8/4-bit back- bones: accuracy surges to almost 100% within 600 steps, indicating that fingerprint readout is ac- quired. (b,d) False control with three identical 16-bit backbones (Qwen3-8B and Llama-3.1-8B only): accuracy remains at chance throughout training. E.4ACCURACY TABLE Table 15: Balanced accuracy on LoRA training to identify bit-width of the generated text. PointwisePairwiseSetwise ModelBaseLoRABaseLoRABaseLoRA Qwen3-8B0.3340.3530.4920.5830.3420.619 Llama3.1-8B-Instruct0.3410.3780.4870.5890.3810.612 Gemma-3-27B-IT0.3280.3820.5020.6400.3460.612 Qwen3-32B 0.3480.3900.5020.6010.3640.627 32 Preprint. E.5CONFUSION MATRIX OF OOD TEMPLATE CONTEXTUAL 16-bit8-bit4-bit Predicted precision 16-bit 8-bit 4-bit True precision 100.0% (n=5,999) 0.0% (n=1) 0.0% (n=0) 0.1% (n=3) 100.0% (n=5,997) 0.0% (n=0) 0.0% (n=0) 0.0% (n=0) 100.0% (n=6,000) Qwen3-8B Main contextual template-OOD shared-LoRA, GPTQ (group size 128) 0% 50% 100% Row-normalized proportion (a) Qwen3-8B (Contextual) 16-bit8-bit4-bit Predicted precision 16-bit 8-bit 4-bit True precision 100.0% (n=20) 0.0% (n=0) 0.0% (n=0) 0.0% (n=0) 100.0% (n=20) 0.0% (n=0) 0.0% (n=0) 0.0% (n=0) 100.0% (n=20) Qwen3-8B No-QA-context (20 questions) shared-LoRA, GPTQ (group size 128) 0% 50% 100% Row-normalized proportion (b) Qwen3-8B (Context Zero) 16-bit8-bit4-bit Predicted precision 16-bit 8-bit 4-bit True precision 100.0% (n=6,000) 0.0% (n=0) 0.0% (n=0) 0.0% (n=0) 100.0% (n=6,000) 0.0% (n=0) 0.0% (n=0) 0.0% (n=0) 100.0% (n=6,000) Llama-3.1-8B-Instruct Main contextual template-OOD shared-LoRA, GPTQ (group size 128) 0% 50% 100% Row-normalized proportion (c) Llama3.1-8B (Contextual) 16-bit8-bit4-bit Predicted precision 16-bit 8-bit 4-bit True precision 100.0% (n=20) 0.0% (n=0) 0.0% (n=0) 0.0% (n=0) 100.0% (n=20) 0.0% (n=0) 0.0% (n=0) 0.0% (n=0) 100.0% (n=20) Llama-3.1-8B-Instruct No-QA-context (20 questions) shared-LoRA, GPTQ (group size 128) 0% 50% 100% Row-normalized proportion (d) Llama3.1-8B (Context Zero) 16-bit8-bit4-bit Predicted precision 16-bit 8-bit 4-bit True precision 100.0% (n=6,000) 0.0% (n=0) 0.0% (n=0) 0.0% (n=0) 100.0% (n=6,000) 0.0% (n=0) 0.0% (n=0) 0.0% (n=0) 100.0% (n=6,000) Gemma-3-27B-IT Main contextual template-OOD shared-LoRA, GPTQ (group size 128) 0% 50% 100% Row-normalized proportion (e) Gemma-3-27B-IT (Contextual) 16-bit8-bit4-bit Predicted precision 16-bit 8-bit 4-bit True precision 100.0% (n=20) 0.0% (n=0) 0.0% (n=0) 0.0% (n=0) 100.0% (n=20) 0.0% (n=0) 0.0% (n=0) 0.0% (n=0) 100.0% (n=20) Gemma-3-27B-IT No-QA-context (20 questions) shared-LoRA, GPTQ (group size 128) 0% 50% 100% Row-normalized proportion (f) Gemma-3-27B-IT (Context Zero) 16-bit8-bit4-bit Predicted precision 16-bit 8-bit 4-bit True precision 100.0% (n=6,000) 0.0% (n=0) 0.0% (n=0) 0.0% (n=0) 100.0% (n=6,000) 0.0% (n=0) 0.0% (n=0) 0.0% (n=0) 100.0% (n=6,000) Qwen3-32B Main contextual template-OOD shared-LoRA, GPTQ (group size 128) 0% 50% 100% Row-normalized proportion (g) Qwen3-32B (Contextual) 16-bit8-bit4-bit Predicted precision 16-bit 8-bit 4-bit True precision 100.0% (n=20) 0.0% (n=0) 0.0% (n=0) 0.0% (n=0) 100.0% (n=20) 0.0% (n=0) 0.0% (n=0) 0.0% (n=0) 100.0% (n=20) Qwen3-32B No-QA-context (20 questions) shared-LoRA, GPTQ (group size 128) 0% 50% 100% Row-normalized proportion (h) Qwen3-32B (Context Zero) Figure 23: Confusion matrix of OOD template contexts for all models. 33 Preprint. E.6SELF-REPORT BEFORE AND AFTER E.6.1INTERNAL AWARENESS AFTER TRAINING Qwen3-8B BeforeAfterBeforeAfterBeforeAfter 0% 20% 40% 60% 80% 100% Response rate 71.3% 51.9% 71.2% 51.7% 45.2% 38.0% 28.7% 48.1% 28.8% 48.3% 54.8% 28.6% 33.4% 19.4%+19.5%26.2% 16-bit8-bit4-bit (a) Abnormality BeforeAfterBeforeAfterBeforeAfter 0% 20% 40% 60% 80% 100% Response rate 62.3% 43.2% 60.8% 34.9% 28.6% 50.3% 37.7% 56.8% 39.2% 65.1% 71.4% 39.7% 10.0% 19.0%+25.9%31.7% 16-bit8-bit4-bit (b) Quantization BeforeAfterBeforeAfterBeforeAfter 0% 20% 40% 60% 80% 100% Response rate 53.2% 24.4% 52.0% 26.9% 39.2% 58.0% 38.3% 69.0% 38.0% 57.1% 38.1% 20.6% 8.4% 6.6% 10.0% 16.1% 22.6% 21.4% 28.8%+19.1%1.2% 16-bit8-bit4-bit (c) Bitwidth Llama3.1-8B-Instruct BeforeAfterBeforeAfterBeforeAfter 0% 20% 40% 60% 80% 100% Response rate 63.0% 18.5% 63.2% 55.2% 12.4% 37.0% 79.7% 36.8% 94.6% 44.8% 75.1% 12.5% 44.4%+57.8%+30.3% 16-bit8-bit4-bit (d) Abnormality BeforeAfterBeforeAfterBeforeAfter 0% 20% 40% 60% 80% 100% Response rate 42.5% 25.4% 41.5% 6.9% 36.7% 25.7% 57.5% 74.4% 58.5% 92.6% 63.3% 69.0% 5.3% 17.1%+34.2%+5.7% 16-bit8-bit4-bit (e) Quantization BeforeAfterBeforeAfterBeforeAfter 0% 20% 40% 60% 80% 100% Response rate 31.7% 26.2% 31.0% 31.8% 44.2% 73.8% 44.6% 97.0% 38.3% 7.7% 24.2% 24.4% 30.0% 92.3% 5.5%+52.5%+62.4% 16-bit8-bit4-bit (f) Bitwidth Gemma-3-27B-IT BeforeAfterBeforeAfterBeforeAfter 0% 20% 40% 60% 80% 100% Response rate 14.7% 10.2% 14.5% 12.2% 22.4% 21.6% 85.3% 89.8% 85.5% 87.8% 77.6% 78.4% 4.4%+2.3%+0.8% 16-bit8-bit4-bit (g) Abnormality BeforeAfterBeforeAfterBeforeAfter 0% 20% 40% 60% 80% 100% Response rate 72.2% 80.4% 71.8% 79.1% 82.8% 78.4% 27.8% 19.6% 28.2% 20.9% 17.2% 21.6% +8.1%7.3%+4.4% 16-bit8-bit4-bit (h) Quantization BeforeAfterBeforeAfterBeforeAfter 0% 20% 40% 60% 80% 100% Response rate 94.5% 90.1% 94.8% 73.6% 90.0% 61.6% 5.5% 8.0% 5.2% 24.1% 9.9% 18.7% 19.7% 4.4%+18.9%+19.6% 16-bit8-bit4-bit (i) Bitwidth Qwen3-32B BeforeAfterBeforeAfterBeforeAfter 0% 20% 40% 60% 80% 100% Response rate 28.3% 38.7% 27.9% 37.9% 21.4% 14.4% 71.7% 61.3% 72.1% 62.0% 78.6% 85.6% +10.4%10.1%+6.9% 16-bit8-bit4-bit (j) Abnormality BeforeAfterBeforeAfterBeforeAfter 0% 20% 40% 60% 80% 100% Response rate 29.7% 12.5% 29.2% 11.6% 28.8% 70.3% 87.5% 70.8% 88.4% 71.2% 99.8% 17.2%+17.5%+28.5% 16-bit8-bit4-bit (k) Quantization BeforeAfterBeforeAfterBeforeAfter 0% 20% 40% 60% 80% 100% Response rate 72.8% 67.5% 70.5% 41.5% 75.2% 21.8% 23.1% 23.9% 48.7% 20.0% 5.4% 9.4% 5.6% 9.9% 100.0% 5.3%+24.7%+95.2% 16-bit8-bit4-bit (l) Bitwidth Figure 24: Internal Awareness self-report: Abnormality, Quantization, Bitwidth for all models before and after training (shared LoRA). 34 Preprint. E.6.2EXTERNAL AWARENESS AFTER TRAINING Qwen3-8B BeforeAfterBeforeAfterBeforeAfter 0% 20% 40% 60% 80% 100% Response rate 64.4% 52.7% 65.4% 48.6% 44.8% 35.6% 10.2% 34.6% 50.0% 55.2% 85.7% 37.1% 9.5% 11.8%+15.5%+30.5% 16-bit8-bit4-bit (a) Abnormality BeforeAfterBeforeAfterBeforeAfter 0% 20% 40% 60% 80% 100% Response rate 50.1% 69.4% 50.2% 40.5% 19.5% 8.7% 49.9% 12.7% 49.8% 59.1% 80.5% 86.8% 17.8% +19.4%+9.3%+6.3% 16-bit8-bit4-bit (b) Quantization BeforeAfterBeforeAfterBeforeAfter 0% 20% 40% 60% 80% 100% Response rate 49.3% 80.2% 48.7% 12.9% 41.5% 17.0% 43.0% 19.4% 43.7% 80.5% 41.7% 7.7%7.7% 6.6% 16.8% 81.6% +30.9%+36.8%+64.9% 16-bit8-bit4-bit (c) Bitwidth Llama3.1-8B-Instruct BeforeAfterBeforeAfterBeforeAfter 0% 20% 40% 60% 80% 100% Response rate 34.3% 30.6% 32.8% 7.1% 30.8% 65.7% 69.3% 67.2% 92.5% 69.2% 99.7% 3.7%+25.3%+30.5% 16-bit8-bit4-bit (d) Abnormality BeforeAfterBeforeAfterBeforeAfter 0% 20% 40% 60% 80% 100% Response rate 45.0% 33.2% 45.0% 10.2% 34.9% 55.0% 66.8% 55.0% 89.8% 65.1% 99.2% 11.9%+34.8%+34.1% 16-bit8-bit4-bit (e) Quantization BeforeAfterBeforeAfterBeforeAfter 0% 20% 40% 60% 80% 100% Response rate 40.7% 38.8% 37.2% 9.3% 35.4% 34.7% 59.9% 36.7% 89.3% 35.8% 24.6% 26.1% 28.8% 100.0% 1.9%+52.5%+71.2% 16-bit8-bit4-bit (f) Bitwidth Gemma-3-27B-IT BeforeAfterBeforeAfterBeforeAfter 0% 20% 40% 60% 80% 100% Response rate 29.4% 15.8% 29.1% 21.4% 28.9% 27.2% 70.6% 84.2% 70.9% 78.6% 71.1% 72.8% 13.6%+7.7%+1.8% 16-bit8-bit4-bit (g) Abnormality BeforeAfterBeforeAfterBeforeAfter 0% 20% 40% 60% 80% 100% Response rate 87.2% 75.0% 87.3% 79.7% 87.1% 77.0% 12.8% 25.0% 12.7% 20.3% 12.9% 23.0% 12.1%+7.6%+10.1% 16-bit8-bit4-bit (h) Quantization BeforeAfterBeforeAfterBeforeAfter 0% 20% 40% 60% 80% 100% Response rate 83.2% 90.9% 83.1% 49.0% 74.9% 21.8% 16.7% 7.1% 16.8% 49.2% 24.7% 74.0% +7.6%+32.4%+73.6% 16-bit8-bit4-bit (i) Bitwidth Qwen3-32B BeforeAfterBeforeAfterBeforeAfter 0% 20% 40% 60% 80% 100% Response rate 27.0% 13.9% 26.4% 13.3% 21.4% 73.0% 85.8% 73.6% 86.3% 78.6% 99.2% 13.1%+12.8%+20.6% 16-bit8-bit4-bit (j) Abnormality BeforeAfterBeforeAfterBeforeAfter 0% 20% 40% 60% 80% 100% Response rate 37.8% 35.6% 35.3% 62.2% 98.0% 64.4% 98.3% 64.7% 99.9% 35.8%+34.0%+35.2% 16-bit8-bit4-bit (k) Quantization BeforeAfterBeforeAfterBeforeAfter 0% 20% 40% 60% 80% 100% Response rate 67.7% 51.0% 68.7% 17.7% 73.3% 22.6% 34.9% 22.3% 68.8% 20.0% 9.7% 14.1% 9.0% 13.5% 6.8% 100.0% 16.8%+46.5%+93.2% 16-bit8-bit4-bit (l) Bitwidth Figure 25: External Awareness self-report: Abnormality, Quantization, Bitwidth for all models before and after training (shared LoRA). 35