Paper deep dive
Polysemanticity or Polysemy? Lexical Identity Confounds Superposition Metrics
Iyad Ait Hou, Rebecca Hwa
Intelligence
Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 94%
Last extracted: 4/2/2026, 3:27:43 AM
Summary
This paper investigates the 'lexical confound' in neural network interpretability, demonstrating that many neurons labeled as 'polysemantic' (due to superposition) are actually 'sense-blind' lexical detectors that fire for shared word forms rather than compressed concepts. Through a 2x2 factorial decomposition across nine transformer models, the authors show that lexical identity drives more activation overlap than semantic compression. They propose methods to classify neurons into sense-selective and sense-blind groups, showing that filtering out sense-blind neurons improves word sense disambiguation and makes model editing (e.g., ROME) more selective.
Entities (6)
Relation Signals (3)
Sense-selective neurons → improves → Word Sense Disambiguation
confidence 95% · filtering it out improves word sense disambiguation and makes knowledge edits more selective
Sparse Autoencoders → inherits → Lexical Confound
confidence 95% · The confound carries into sparse autoencoders (18–36% of features blend senses).
Lexical Confound → inflates → Superposition
confidence 90% · Standard metrics cannot tell these apart, and so they risk overestimating superposition-driven variable sharing.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:If the same neuron activates for both "lender" and "riverside," standard metrics attribute the overlap to superposition--the neuron must be compressing two unrelated concepts. This work explores how much of the overlap is due a lexical confound: neurons fire for a shared word form (such as "bank") rather than for two compressed concepts. A 2x2 factorial decomposition reveals that the lexical-only condition (same word, different meaning) consistently exceeds the semantic-only condition (different word, same meaning) across models spanning 110M-70B parameters. The confound carries into sparse autoencoders (18-36% of features blend senses), sits in <=1% of activation dimensions, and hurts downstream tasks: filtering it out improves word sense disambiguation and makes knowledge edits more selective (p = 0.002).
Tags
Links
- Source: https://arxiv.org/abs/2604.00443v1
- Canonical: https://arxiv.org/abs/2604.00443v1
Trouble viewing inline? Open PDF directly →
Full Text
54,103 characters extracted from source content.
Expand or collapse full text
Polysemanticity or Polysemy? Lexical Identity Confounds Superposition Metrics Iyad Ait Hou Department of Computer Science George Washington University Washington, D.C, USA iyad.aithou@gwu.edu &Rebecca Hwa Department of Computer Science George Washington University Washington, D.C, USA rebecca.hwa@gwu.edu Abstract If the same neuron activates for both “lender” and “riverside,” standard metrics attribute the overlap to superposition—the neuron must be compressing two unrelated concepts. This work explores how much of the overlap is due to a lexical confound: neurons fire for a shared word form (such as “bank”) rather than for two compressed concepts. A 2×22× 2 factorial decomposition reveals that the lexical-only condition (same word, different meaning) consistently exceeds the semantic-only condition (different word, same meaning) across models spanning 110M–70B parameters. The confound carries into sparse autoencoders (18–36% of features blend senses), sits in ≤ 1% of activation dimensions, and hurts downstream tasks: filtering it out improves word sense disambiguation and makes knowledge edits more selective (p=0.002p=0.002). 1 Introduction Polysemanticity (Elhage et al., 2022; Olah et al., 2020) describes the phenomenon in which a neuron in a large language model is activated for multiple, seemingly unrelated concepts. An explanation for this is superposition: when a model lacks the dimensions to represent all its learned concepts separately, it packs more variables into its activation space than it has neurons. By storing these variables as nearly-orthogonal directions, the model can function despite a single neuron participating in several distinct concepts (Elhage et al., 2022). In this light, polysemanticity looks like a compression artifact, and sparse autoencoders (SAEs) are trained to undo it (Bricken et al., 2023; Templeton et al., 2024). However, some cases of polysemanticity may be a lexical confound rather than genuine compression. A neuron activating for both “financial bank” and “river bank” is labeled polysemantic (Olah et al., 2020; Bricken et al., 2023), but the overlap may also reflect the model processing a shared word form before disambiguating its meaning (Ethayarajh, 2019). Standard metrics cannot tell these apart, and so they risk overestimating superposition-driven variable sharing. We therefore ask: to what extent are current measurements of polysemanticity actually capturing lexical identity rather than superposition? We hypothesize that much of the measured overlap is driven by shared word forms rather than the need to compress unrelated concepts, and that these lexical signals play a primary, causal role in the model’s processing. To test this, we isolate the drivers of activation overlap using a 2×22× 2 factorial decomposition (Figure 2) that pairs sentences along two axes—same/different word and same/different meaning. The lexical-only condition (same word, different meaning) can then be compared directly against the semantic-only condition (different word, same meaning): whichever produces more overlap reveals the main driver. Figure 1: Overview. Standard metrics (top) see that neurons n1n_1–n5n_5 fire for both senses of bank and call all of it superposition. Our decomposition (bottom) shows that most shared neurons are sense-blind—they encode the word form, not compressed concepts. Only a small remainder is genuine superposition. The results show a consistent pattern: word form wins. Across nine transformers spanning 110M to 70B parameters, four architecture families, and 407 polysemous words, the lexical-only condition consistently exceeds the semantic-only condition (Figure 3). The confound carries into SAEs, where 18–36% of learned features blend both senses of a word. It sits in a compact 20-dimensional subspace (≤ 1% of activation dimensions) that can be cleanly removed. And it matters in practice: sense-blind neurons hurt word sense disambiguation (6 p below sense-selective neurons) and make ROME knowledge edits less selective (p=0.002p=0.002). Figure 1 shows the core idea. 2 Related work Polysemanticity and superposition. The superposition hypothesis (Elhage et al., 2022) frames polysemanticity as a compression artifact, and SAEs are designed to recover monosemantic features from it (Bricken et al., 2023; Templeton et al., 2024; Sharkey et al., 2022). Recent work has shown that polysemanticity can arise even without capacity pressure (Lecomte et al., 2025) and has tracked its emergence during training (Wu and others, 2025). None of these works examine whether the inputs triggering multi-concept activation share a lexical form—the confound we identify. Word sense representations. Ethayarajh (2019) and Wiedemann et al. (2019) established that contextual embeddings cluster by word identity before sense. This is a geometric observation about full hidden states; it was not connected to polysemanticity metrics or to individual neurons. Our work bridges this gap: we operate at the level of individual MLP neurons (the unit at which polysemanticity is defined), provide a quantitative metric (RlexR_lex), and trace the confound through causal and downstream consequences. Lyu and others (2025) and Minegishi et al. (2025) study polysemous words in neural representations but do not measure how the confound affects interpretability tools. Causal methods and model editing. Probing classifiers (Conneau et al., 2018) test what information a representation encodes; activation patching (Vig et al., 2020; Meng et al., 2022) and mean-ablation (Chan et al., 2022) test whether that information plays a causal role. We use both to validate the sense-selective / sense-blind distinction. Model editing methods such as ROME (Meng et al., 2022) and embedding-level interventions (Ait Hou and Hwa, 2026) modify specific representations to change model behavior; none account for lexical identity as a source of collateral damage across word senses. 3 Methodology Our approach has three parts: (1) a factorial decomposition that separates lexical from semantic contributions to neuron overlap, (2) a neuron classification that identifies which neurons encode word form vs. word meaning, and (3) causal and downstream experiments that test whether the distinction matters. All analyses target MLP intermediate activations, the level at which polysemanticity is typically analyzed (Olah et al., 2020; Elhage et al., 2022; Bricken et al., 2023). 3.1 Decomposition framework Figure 2: 2×22× 2 factorial decomposition design. We construct sentence pairs crossing two binary factors—same/different word and same/different sense—yielding four conditions (Figure 2): SL (same word, same sense) produces the most neuron overlap because both word form and meaning are shared; CL (different word, different sense) produces the least because neither is shared. The two off-diagonal conditions isolate the drivers: PS (same word, different sense) measures overlap from shared word form alone, and SYN (different word, same sense) measures overlap from shared meaning alone. For each sentence, we extract MLP intermediate activations at the last subword of the target word (Ethayarajh, 2019) and compute pairwise overlap via three metrics: cosine similarity, Jaccard overlap of active-neuron sets, and magnitude divergence, aggregated per word with bootstrap 95% CIs. Lexical contribution ratio. If lexical and semantic contributions do not interact (validated in Section 5.6), their relative sizes can be read off directly. Normalizing by the total range gives the lexical contribution ratio: Rlex=M¯Polysemy−M¯SynonymM¯Same-Lemma−M¯Cross-LemmaR_lex= M_Polysemy- M_Synonym M_Same-Lemma- M_Cross-Lemma (1) The numerator is the difference between same-word overlap and same-meaning overlap—the excess due to word form. The denominator is the full range from maximum to minimum overlap. Rlex>0R_lex>0 means word form contributes more overlap than meaning; Rlex=1R_lex=1 means word form accounts for all of it. As a control, we also compute RlexR_lex from raw token embeddings (before any transformer computation); MLP-level RlexR_lex exceeding this value indicates the confound is not simply inherited from the input encoding. 3.2 Neuron classification The decomposition above measures overlap at the population level. To test whether individual neurons behave differently depending on whether they encode word form or word meaning, we classify them into two groups. Sense Selectivity Index (SSI). For each polysemous word, we have sentences for two distinct senses (e.g., “bank” as financial institution = sense A, “bank” as riverbank = sense B). For each neuron j, we compute Cohen’s d between its activations on sense-A vs. sense-B sentences: SSIj=|μAj−μBj|σpooledjSSI_j= | _A^j- _B^j| _pooled^j (2) Neurons with SSI>2SSI>2 are sense-selective; neurons with SSI<0.5SSI<0.5 (above median activation) are sense-blind. These thresholds follow standard conventions for Cohen’s d (large vs. small effect). The condition ordering (PS >> SYN) and the sense-blind/selective dissociation do not depend on the exact threshold choice; SSI distributions are reported in Appendix B.3. Form detectors (sense-label-free). To avoid circularity, we independently identify form detectors using only word identity: for each word w and neuron j, we compute consistency (1−CV1-CV of activations across all w-sentences) and specificity (Cohen’s d vs. other words). Neurons ranking in the top K by consistency × specificity are form detectors for w—they fire reliably for w regardless of sense. 3.3 Causal intervention protocol If sense-selective and sense-blind neurons truly encode different information, removing one group should affect model behavior differently from removing the other. We test this via mean-ablation (Chan et al., 2022): replacing selected neurons’ activations with their dataset mean, simultaneously across all layers at the target-word position. Neuron counts are matched across groups (sense-A-selective, sense-B-selective, sense-blind, random) at each layer. We evaluate via: (1) KL divergence of the full output distribution, (2) sense accuracy (diagnostic token probabilities), and (3) sense-specific perplexity change, with specificity =|Δppltarget−Δpplother|=| _target- _other|. 4 Experimental setup Dataset. We use SemCor (Miller et al., 1993), the only large-scale corpus with human-annotated WordNet sense tags (37,176 sentences). From it we select 407 content words (nouns/verbs, ≥2≥ 2 senses, ≥5≥ 5 sentences per sense, Wu-Palmer similarity <0.50<0.50 between senses; varying this threshold from 0.35 to 0.65 does not change the condition ordering). SYN pairs use WordNet synonyms attested in SemCor; CL pairs sample unrelated words. Results replicate across nine models trained on four corpora and validate on modern Wikipedia text (Appendix F.1, B.5, A). Models. We evaluate nine transformers spanning four architecture families and 110M–70B parameters (Table 1). Models up to 13B are loaded in float32; LLaMA-2-70B uses NF4 quantization. Model Params Layers dmlpd_mlp Type Ref. GPT-2 117M 12 3072 Auto Radford et al. 2019 GPT-2-Med 345M 24 4096 Auto Radford et al. 2019 BERT-base 110M 12 3072 Bidir Devlin et al. 2019 ELECTRA-base 110M 12 3072 Bidir Clark et al. 2020 Pythia-1B 1.0B 16 8192 Auto Biderman et al. 2023 Pythia-6.9B 6.9B 32 16384 Auto Biderman et al. 2023 Pythia-12B 12B 36 20480 Auto Biderman et al. 2023 LLaMA-2-13B 13B 40 13824 Auto Touvron et al. 2023 LLaMA-2-70B 70B 80 28672 Auto Touvron et al. 2023 Table 1: Models evaluated. Auto = autoregressive LM; Bidir = bidirectional (masked LM / replaced-token detection). LLaMA-2-70B uses NF4 quantization. 5 Results We present the evidence in stages: the confound appears consistently across layers (Section 5.1), replicates across model scale (Section 5.2) and into SAEs (Section 5.3), has identifiable mechanistic structure (Sections 5.4–5.8), and affects downstream tasks (Section 5.9). Primary results use GPT-2; replication across models follows in Section 5.2. 5.1 The lexical confound across layers Cosine similarity follows sl>ps>SYN>cl sl> ps>SYN> cl across layers and models (Figure 3); the same ordering holds for Jaccard overlap and magnitude divergence (Appendix C). Crucially, PS consistently exceeds SYN (Wilcoxon p<0.001p<0.001, Holm-Bonferroni corrected): sharing a word form produces more overlap than sharing a meaning. A token-embedding baseline (Rlexemb=0.71R_lex^emb=0.71–0.880.88) confirms the confound is not merely inherited from token embeddings (Table 6). RlexR_lex decreases with depth (0.74 to 0.29 across GPT-2 layers; Figure 11). In all models tested, the bootstrap CI remains above zero at all layers—though we note this is an empirical observation, not a guaranteed property; a model with stronger disambiguation could in principle drive RlexR_lex to zero or below. Sense-selective neurons (SSI>2SSI>2) stay below 1%1\% of the population, confirming sense information is encoded by a very sparse subpopulation. Figure 3: Top: Layer-averaged cosine similarity by condition across nine models (110M–70B). PS consistently exceeds SYN: word form drives more overlap than shared meaning. Bottom: Per-layer breakdown showing the pattern holds across layers. RlexR_lex trends in Appendix B. 5.2 Cross-model consistency The PS >> SYN gap replicates across all nine models tested, from 110M to 70B parameters (Figure 3; Appendix B). Bidirectional models (BERT, ELECTRA) show steeper RlexR_lex declines with depth; autoregressive models maintain the pattern through 12–13B. LLaMA-2-70B (NF4 quantization, 80 layers) shows the same ordering (layer-averaged: ps=.117 ps=.117 vs. SYN=.076SYN=.076; Rlex=0.45R_lex=0.45). The neuron classification and causal experiments use models up to 13B in full precision; the 70B result confirms the condition ordering extends to larger scale. 5.3 Impact on SAEs and standard metrics Pre-trained GPT-2 SAEs (Bloom, 2024) (32k and 128k features) inherit the confound: 18–32% of active features per layer are sense-blind, and the 128k SAE (4×4× capacity) produces nearly identical ratios (Figure 4). On Pythia-410M (65k SAEs), ratios are 19–36%. SAEs do not resolve the lexical confound. Individual feature analysis. Inspecting individual sense-blind features reveals the scope of the problem. Feature #5132 in the 32k GPT-2 SAE at layer 8 (the OAI v5 release; publicly inspectable on Neuronpedia111https://w.neuronpedia.org/gpt2-small/8-res_post_32k-oai/5132) fires with similar activation (Cohen’s d<0.5d<0.5) for 232 different polysemous words, spanning unrelated domains (foot, church, kick, dance, win, cell). Its auto-generated label on Neuronpedia is “intensifying adjectives and adverbs describing sound or scale,” yet it fires on 47.5% of all inputs—a density incompatible with a narrow semantic category. This feature encodes lexical form, not a coherent concept; its auto-interp label reflects surface patterns rather than its actual function. Applying standard polysemanticity scores to our dataset, ≈57%≈\!57\% of activating input pairs for flagged neurons (conditional on polysemous inputs) share a word form rather than a genuinely different concept. This is not a global inflation rate—but polysemous words are common (>>80% of frequent English vocabulary; Rodd et al. 2002). Figure 4: SAE collision analysis (GPT-2). Left: mean features per word by sense discriminability. Right: collision ratio across layers. 18–32% of features conflate senses of the same word (18–36% including Pythia-410M; Section 5.3). 5.4 Neuron-level validation The preceding sections establish the confound at the population level. We now ask whether it holds at the level of individual neurons. Form detectors are sense-blind. To validate the confound non-circularly, we identify form detectors using only word identity (Section 3.2) and separately classify them by SSI. On GPT-2, 67.3%67.3\% of form detectors are sense-blind (SSI<0.5SSI<0.5), while only 0.1%0.1\% are sense-selective—establishing that the neurons polysemanticity metrics flag are predominantly lexical identity detectors. Probing confirms the functional split. Logistic regression probes (LOO CV, C=0.01C=0.01) on neuron-group activations reveal a clear dissociation: selective neurons achieve 91.1%91.1\% sense accuracy; blind neurons 41.9%41.9\% (near chance); random 68.9%68.9\%. For word-form detection, all groups perform comparably (∼ 85%), confirming word identity is distributed. Blind neurons encode form (85%) but not sense (42%); selective neurons encode both. 5.5 Causal intervention: sense-specific editing We test whether the distinction has causal consequences via mean-ablation. Ablating sense-selective neurons produces 33–5×5× larger KL divergence than matched sense-blind or random groups (up to 345M; 1.21.2–2.1×2.1× at 12–13B), suggesting sense-selective neurons carry disproportionate causal weight despite small absolute effects (10−310^-3–10−410^-4 KL, as expected when ablating ∼ 50 of 3k–20k neurons). Sense-specific editing. Ablating across all layers with matched neuron counts (Table 2), sense-selective ablation is 6.6×6.6× more specific than blind on GPT-2 (53 words). Sense-A ablation raises perplexity on sense-A sentences (+0.55+0.55) while leaving sense-B largely unchanged (−0.10-0.10); blind ablation has minimal effect on either sense (−0.07-0.07, +0.17+0.17). Ablation Δ Δ Specificity Sense-A neurons +0.55+0.55 −0.10-0.10 0.660.66 Sense-B neurons +0.10+0.10 +2.62+2.62 2.522.52 Sense-blind −0.07-0.07 +0.17+0.17 0.240.24 Random −0.64-0.64 −0.32-0.32 0.320.32 Table 2: Sense-specific editing (GPT-2, 53 words, 142 neurons per group). Sense-selective ablation targets one sense (6.6×6.6× specificity gap over blind). Specificity =|Δppltarget−Δpplother|=| _target- _other|. Collateral damage on individual words. Per-word ablation at layer 6 confirms the pattern: for match, sense-blind ablation damages both senses equally (specificity =0.010=0.010), while sense-selective ablation is 5.5×5.5× more targeted; for light, 12×12×. Sense-blind neurons cannot be used to target one meaning without equally affecting the other. 5.6 Controls The interaction term I is small in all models (|I|<0.04|I|<0.04), supporting additivity. RlexR_lex correlates weakly with frequency (r=0.15r=0.15) and non-significantly with position. A synonym-free variant and subset invariance tests confirm robustness (Appendix B, B.5; embedding baseline and interaction plots in Appendix C). 5.7 The confound is correctable The confound is not merely detectable—it is correctable. We define a lexically-adjusted polysemanticity score: Pjadj=Pjraw−λ^ℓ⋅FjP^adj_j=P^raw_j- λ_ · F_j (3) where PjrawP^raw_j is the standard score, FjF_j indicates whether neuron j is a form detector, and λ^ℓ=Rlex(ℓ)⋅P¯ℓ λ_ =R_lex( )· P_ is the layer-specific expected lexical inflation. On GPT-2, this reclassifies 6.5% (layer 6) to 10.4% (layer-averaged) of neurons flagged as polysemantic, with 93–94% of reclassified neurons confirmed as form detectors (SSI<0.5SSI<0.5). The correction targets the top-50 form detectors per word (≈ 1.6% of GPT-2’s 3072 MLP neurons), chosen to balance coverage with precision; this is a conservative lower bound. This demonstrates correctability: the confound can be identified and removed from existing metrics. The current implementation requires sense labels, limiting it to controlled evaluations; extending this to an unsupervised correction is an open problem (Section 6). 5.8 Mechanistic analysis: lexical identity subspace Does the confound have specific geometric structure, or is it diffuse? For each word wiw_i with synonym data (n=62n=62), we compute i=¯wi−¯synid_i= a_w_i- a_syn_i (mean activation minus synonym mean), isolating word form from meaning. PCA on these difference vectors yields a lexical identity subspace (LIS) (cf. Arditi et al., 2024). Removing just 20 LIS dimensions (≤ 1% of activation space) reduces the PS–SYN gap by 37–51%; at 50 dimensions the gap reverses entirely (full dose-response in Appendix D). The confound is compact, localizable, and consistent across GPT-2 and Pythia-1B. The corrective score (Section 5.7) and the LIS show that the confound is not just measurable but removable. We now test whether doing so helps in practice. 5.9 Downstream applications Figure 5: Sense-selective ROME editing. Standard ROME modifies all activated neurons, disturbing both senses. Our approach classifies neurons first, shields sense-blind (lexical-form) neurons from modification, and edits only sense-selective ones—preserving the unedited sense. Finally, we demonstrate that the sense-selective / sense-blind distinction has practical value beyond diagnosis (Figure 5). Word sense disambiguation. For each of 20 CoarseWSD-20 words, we train a logistic regression classifier (5-fold CV) on MLP activations. Sense-selective neurons (top 25% by selectivity) achieve 89.8% accuracy in GPT-2 (89.3% in Pythia-1B), outperforming all neurons (88.3%) and sense-blind neurons (83.6%/84.6%). The selective–blind gap (5–6 p) peaks at early-to-mid layers. ROME knowledge editing. We apply ROME (Meng et al., 2022) to change one sense’s prediction and measure collateral damage on the other, comparing standard (all neurons), filtered (sense-selective only), and blind-only edits. Filtered edits are significantly more selective (0.5400.540 vs. 0.5190.519, p=0.002p=0.002, Wilcoxon; largest gain at layer 7: +0.046+0.046). Blind-only edits are worst (0.5120.512), confirming sense-blind neurons spread edits indiscriminately across senses. 6 Discussion Are sense-blind neurons doing something useful? Neurons firing for all senses likely encode a pre-disambiguation representation, and we do not dispute this. Our claim is narrower: these neurons should not be counted as evidence of superposition, because they encode word form, not compressed concepts. Causally, they have minimal impact on sense processing (probe accuracy 42%, KL ratio 0.25×0.25×) while encoding word identity at 85%. Three independent lines rule out circularity: form detectors identified without sense labels overlap 67% with sense-blind neurons; probing shows a clean dissociation independent of the decomposition; and causal ablation of sense-blind neurons has minimal effect on sense processing. Implications for SAEs and practice. SAE evaluation assumes features are monosemantic (Bricken et al., 2023; Templeton et al., 2024), but 18–36% of features active for polysemous words are monosemantic for a word form, not a concept—as illustrated by feature #5132 (Section 5.3), labeled “intensifying adjectives” on Neuronpedia despite firing sense-blind for 232 unrelated words. Since polysemous words are over 80% of frequent vocabulary (Rodd et al., 2002), such features are likely common. When practitioners steer or edit using SAE features (Arditi et al., 2024; Meng et al., 2022), sense-blind features spread changes across senses; our classification provides a filter. Recommendations. For SAE evaluation: check whether a feature’s top-activating examples share a word form, and test sense discrimination if so. For model editing: classify target neurons by SSI and restrict edits to sense-selective neurons. For polysemanticity measurement: report RlexR_lex or apply the LIS projection before computing overlap metrics. Looking ahead, a lexical-identity penalty in SAE training could produce sense-level monosemantic features. Limitations. Absolute ablation effects are small (<<3 ppl points), as expected from ∼ 150 of 3k–20k neurons; the primary evidence is the condition ordering itself, not the downstream magnitudes. SemCor’s vintage text is validated on modern Wikipedia (Appendix F.1), and the ordering replicates across nine models on four corpora. The 57% confound ratio applies only to polysemous inputs. The corrected score requires sense labels; the LIS removal is more deployable but a fully unsupervised correction remains open (Appendix D). 7 Conclusion A substantial portion of what current metrics call polysemanticity is lexical identity, not superposition. Using a factorial decomposition across nine transformers (110M–70B), we showed that sharing a word form consistently produces more neuron overlap than sharing a meaning. The confound propagates into SAEs—where individual features fire sense-blind for hundreds of words—occupies a compact removable subspace, and causes collateral damage when editing or steering model behavior. These findings do not deny that genuine superposition exists. They do show that the lexical signal must be separated out before the remaining overlap can be attributed to compression. We hope the tools introduced here—RlexR_lex, the sense-selective/blind classification, and the lexical identity subspace—provide a starting point for polysemanticity measurements that distinguish word form from concept. References I. Ait Hou and R. Hwa (2026) Parameter-efficient token embedding editing for clinical class-level unlearning. arXiv preprint arXiv:2603.19302. Cited by: §2. A. Arditi, O. Obeso, A. Syed, D. Paleka, N. Rimsky, L. Sharkey, and N. Nanda (2024) Refusal in language models is mediated by a single direction. arXiv preprint arXiv:2406.11717. Cited by: §5.8, §6. S. Biderman, H. Schoelkopf, Q. G. Anthony, H. Bradley, K. O’Brien, E. Hallahan, M. A. Khan, S. Purber, U. S. Prashanth, E. Raff, et al. (2023) Pythia: a suite for analyzing large language models across training and scaling. In International Conference on Machine Learning, p. 2397–2430. Cited by: Table 1, Table 1, Table 1. J. Bloom (2024) Understanding SAE features with the logit lens. Alignment Forum. Note: Pre-trained GPT-2 SAEs available at https://huggingface.co/jbloom/GPT2-Small-OAI-v5-32k-resid-post-SAEs Cited by: §5.3. T. Bricken, A. Templeton, J. Batson, B. Chen, A. Jermyn, T. Conerly, N. L. Turner, C. Anil, C. Denison, A. Askell, R. Lasenby, Y. Wu, S. Kravec, N. Schiefer, T. Maxwell, N. Joseph, A. Tamkin, K. Nguyen, B. McLean, J. E. Burke, T. Hume, S. Carter, T. Henighan, and C. Olah (2023) Towards monosemanticity: decomposing language models with dictionary learning. Note: Transformer Circuits Thread External Links: Link Cited by: §1, §1, §2, §3, §6. L. Chan, A. Garriga-Alonso, N. Goldowsky-Dill, R. Greenblatt, J. Nitishinskaya, A. Radhakrishnan, B. Shlegeris, and N. Thomas (2022) Causal scrubbing: a method for rigorously testing interpretability hypotheses. AI Alignment Forum. Cited by: §2, §3.3. K. Clark, M. Luong, Q. V. Le, and C. D. Manning (2020) ELECTRA: pre-training text encoders as discriminators rather than generators. In Proceedings of ICLR, Cited by: Table 1. A. Conneau, G. Kruszewski, G. Lample, L. Barrault, and M. Baroni (2018) What you can cram into a single $&!#* vector: probing sentence embeddings for linguistic properties. In Proceedings of the Association for Computational Linguistics (ACL), Cited by: §2. J. Devlin, M. Chang, K. Lee, and K. Toutanova (2019) BERT: pre-training of deep bidirectional transformers for language understanding. In Proceedings of NAACL-HLT, Cited by: Table 1. N. Elhage, T. Hume, C. Olsson, N. Schiefer, T. Henighan, S. Kravec, Z. Hatfield-Dodds, R. Lasenby, D. Drain, C. Chen, R. Grosse, S. McCandlish, J. Kaplan, D. Amodei, M. Wattenberg, and C. Olah (2022) Toy models of superposition. Transformer Circuits Thread. External Links: Link Cited by: §1, §2, §3. K. Ethayarajh (2019) How contextual are contextualized word representations? Comparing the geometry of BERT, ELMo, and GPT-2 representations. In Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP), Cited by: §1, §2, §3.1. M. Geva, A. Caciularu, K. Wang, and Y. Goldberg (2022) Transformer feed-forward layers build predictions by promoting concepts in the vocabulary space. In Proceedings of EMNLP, Cited by: Appendix D. V. Lecomte, K. Chandrasekaran, T. Henighan, and N. Elhage (2025) Incidental polysemanticity. In Proceedings of the International Conference on Learning Representations (ICLR), Cited by: §2. D. Loureiro, J. Camacho-Collados, K. Rezaee, and M. T. Pilehvar (2022) Language models and word sense disambiguation: an overview and analysis. In Computational Linguistics, Vol. 48, p. 789–823. Cited by: §F.1. X. Lyu et al. (2025) Exploring layer-wise representations of English and Chinese homonymy in pre-trained language models. In Findings of the Association for Computational Linguistics (ACL), Cited by: §2. K. Meng, D. Bau, A. Andonian, and Y. Belinkov (2022) Locating and editing factual associations in GPT. In Advances in Neural Information Processing Systems (NeurIPS), Cited by: §2, §5.9, §6. G. A. Miller, C. Leacock, R. Tengi, and R. T. Bunker (1993) A semantic concordance. In Proceedings of the ARPA Workshop on Human Language Technology, Cited by: Appendix A, §4. G. Minegishi, H. Furuta, Y. Iwasawa, and Y. Matsuo (2025) Rethinking evaluation of sparse autoencoders through the representation of polysemous words. In Proceedings of the International Conference on Learning Representations (ICLR), Cited by: §2. nostalgebraist (2020) Interpreting GPT: the logit lens. Note: https://w.lesswrong.com/posts/AcKRB8wDpdaN6v6ru/interpreting-gpt-the-logit-lens Cited by: Appendix D. C. Olah, N. Cammarata, L. Schubert, G. Goh, M. Petrov, and S. Carter (2020) Zoom in: an introduction to circuits. Distill. External Links: Document, Link Cited by: §1, §1, §3. A. Radford, J. Wu, R. Child, D. Luan, D. Amodei, and I. Sutskever (2019) Language models are unsupervised multitask learners. OpenAI Blog. Cited by: Table 1, Table 1. J. M. Rodd, M. G. Gaskell, and W. D. Marslen-Wilson (2002) Making sense of semantic ambiguity: semantic competition in lexical access. Journal of Memory and Language 46 (2), p. 245–266. Cited by: §5.3, §6. L. D. Sharkey, D. Braun, and B. Millidge (2022) Taking features out of superposition with sparse autoencoders. Note: AI Alignment Forum External Links: Link Cited by: §2. A. Templeton, T. Conerly, J. Marcus, J. Lindsey, T. Bricken, B. Chen, A. Pearce, C. Citro, E. Ameisen, A. Jones, H. Cunningham, N. L. Turner, C. McDougall, M. MacDiarmid, C. D. Freeman, T. R. Sumers, E. Rees, J. Batson, A. Jermyn, S. Carter, C. Olah, and T. Henighan (2024) Scaling monosemanticity: extracting interpretable features from Claude 3 Sonnet. Note: Transformer Circuits Thread External Links: Link Cited by: §1, §2, §6. H. Touvron, L. Martin, K. Stone, P. Albert, A. Almahairi, Y. Babaei, N. Bashlykov, S. Batra, P. Bhargava, S. Bhosale, et al. (2023) LLaMA 2: open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288. Cited by: Table 1, Table 1. J. Vig, S. Gehrmann, Y. Belinkov, S. Qian, D. Nevo, Y. Singer, and S. Shieber (2020) Investigating gender bias in language models using causal mediation analysis. In Advances in Neural Information Processing Systems (NeurIPS), Cited by: §2. G. Wiedemann, S. Remus, A. Chawla, and C. Biemann (2019) Does BERT make any sense? interpretable word sense disambiguation with contextualized embeddings. In Proceedings of the 15th Conference on Natural Language Processing (KONVENS), Cited by: §2. J. Wu et al. (2025) From tokens to semantics: the emergence and stabilization of polysemanticity in language models. In Proceedings of the Conference on Language Modeling (COLM), Cited by: §2. Z. Wu and M. Palmer (1994) Verb semantics and lexical selection. In Proceedings of the Annual Meeting of the Association for Computational Linguistics (ACL), p. 133–138. Cited by: Appendix A. Appendix: Table of Contents Click any entry to jump directly to that section. App. A Dataset Details p. A ⊳ Source corpus and selection criteria (SemCor, 407 polysemous words) ⊳ Cleaning pipeline: function verb exclusion, deduplication, length filter ⊳ Sense distance verification (Wu-Palmer similarity thresholds) ⊳ Condition pair counts (Table 3) and POS distribution App. B Full Numerical Results p. B ⊳ RlexR_lex by layer for all models (Tables 4, 5) ⊳ Token-embedding baseline comparison across 8 models (Table 6) ⊳ Sense selectivity index: SSI>2SSI>2 fractions per layer, threshold sensitivity (Table 7) ⊳ SAE collision: 32k vs. 128k dictionary comparison (Table 8) ⊳ Synonym subset validation, POS stratification, Mann-Whitney tests (Table 9) App. C Additional Figures p. C ⊳ Jaccard overlap of active-neuron sets across layers (GPT-2) ⊳ Per-word cross-sense cosine similarity heatmap ⊳ RlexR_lex vs. Rlexno-synR_lex^no-syn comparison ⊳ Magnitude divergence across layers ⊳ Condition means at representative layers ⊳ RlexR_lex across all 9 models with 95% bootstrap CIs App. D Supplementary Analyses p. D ⊳ Sense accuracy after ablation (causal validation) ⊳ Concrete causal example: plant (factory vs. botanical) ⊳ Cross-model scaling details (Pythia-12B, LLaMA-2-13B, LLaMA-2-70B) ⊳ The U-shaped trajectory: analysis and interpretation App. E Attention Representations p. E ⊳ Analysis of attention-head outputs (complementary to MLP analysis) App. F Reproducibility & Validation p. F ⊳ Compute requirements (GPU hours, hardware) ⊳ Software versions and dependencies ⊳ Formal metric definitions (cosine similarity, Jaccard overlap, magnitude divergence, SSI) ⊳ Modern corpus validation on CoarseWSD-20 (Wikipedia sentences, Table 11) Appendix A Dataset details Source corpus and selection. All sentences are drawn from SemCor (Miller et al., 1993), a sense-tagged corpus built on the Brown Corpus with human-annotated WordNet sense labels. From the full corpus (37,176 sentences, 88,334 sense-tagged tokens), we identified 494 content words (nouns and verbs) with at least two senses, each attested in ≥8≥ 8 sentences. Cleaning pipeline. Three cleaning steps were applied in order: 1. Function verb exclusion. We removed 31 high-frequency function verbs (be, have, do, go, make, get, take, come, give, say, etc.) whose multiple WordNet “senses” reflect grammatical rather than semantic distinctions. 2. Cross-word deduplication. 3,591 sentences that appear under multiple words in SemCor were kept only for their first word, preventing cross-word contamination. 3. Length filter. Sentences shorter than 30 characters were removed (typically sentence fragments or captions). After cleaning, 407 words remain with ≥5≥ 5 sentences per sense (mean: 10.3 sentences per sense; range: 5–15). Sense distance verification. We computed Wu-Palmer similarity (Wu and Palmer, 1994) between the two selected WordNet synsets for each word. The mean similarity across all 407 word pairs is 0.160.16 (range: 0.100.10–0.220.22). We enforced a maximum threshold of 0.500.50 to exclude near-synonymous senses. Condition pair counts. Table 3 shows the number of sentence pairs constructed for each condition. SL and PS pairs are available for all 407 words. The SYN condition requires a WordNet synonym attested in SemCor with ≥3≥ 3 sentences; 172 words have such a synonym for sense A and 160 for sense B. CL pairs are formed by randomly sampling sentences from unrelated words, excluding the current word’s own sentences. Condition Words Pairs (total) Pairs/word (mean) Same-Lemma (SL) 407 ∼ 40,700 100 Polysemantic (PS) 407 ∼ 40,700 100 Synonym (SYN) 251 ∼ 12,500 50 Cross-Lemma (CL) 407 ∼ 40,700 100 Table 3: Sentence pair counts by condition. SYN coverage is lower because synonyms must be attested in SemCor. All pairs are capped at 200 per word. POS distribution. Of the 407 words, 156 (38%) are nouns and 251 (62%) are verbs. 13 lemmas appear twice with different POS tags (e.g., brush as noun and verb), which we treat as separate entries since they have distinct sense pairs. The full word list, per-word SSI distributions, and SAE feature-level collision data (including the 232 words for which feature #5132 is sense-blind) are released with our code. Appendix B Full numerical results B.1 RlexR_lex by layer Table 4 reports RlexR_lex computed from cosine similarity at each layer for the 12-layer models. RlexR_lex decreases with depth in all models, with BERT and ELECTRA achieving lower minima than GPT-2, reflecting stronger sense separation in bidirectional models. Layer GPT-2 BERT ELECTRA GPT-2-Med 0 .735 .662 .686 .751 1 .629 .566 .528 .700 2 .560 .552 .482 .690 3 .494 .471 .428 .612 4 .446 .398 .367 .549 5 .441 .409 .307 .524 6 .423 .367 .250 .521 7 .418 .313 .268 .498 8 .447 .275 .264 .443 9 .462 .225 .309 .454 10 .399 .360 .381 .440 11 .290 .209 .542 .429 Min .290 .209 .250 .270 Table 4: RlexR_lex (cosine-based) by layer for 12-layer models and GPT-2-Medium (first 12 of 24 layers shown). GPT-2-Medium’s minimum of 0.270.27 occurs at layer 23. BERT achieves its minimum at layer 9 (0.230.23), ELECTRA at layer 6 (0.250.25). GPT-2’s RlexR_lex remains above 0.290.29 throughout. Layer Pythia-12B LLaMA-2-13B 0 .913 .911 5 .516 .406 10 .377 .262 15 .360 .456 20 .422 .476 25 .391 .348 30 .440 .335 35 .288 .281 39 — .261 Min .288 .250 Table 5: RlexR_lex (cosine-based) at representative layers for the larger models. Pythia-12B has 36 layers; LLaMA-2-13B has 40 layers. Both start with Rlex>0.9R_lex>0.9 and decline with depth, reaching minima comparable to the smaller models. B.2 Token-embedding baseline Table 6 reports the embedding-only RlexembR_lex^emb for each model, computed from raw token embeddings before any transformer computation. MLP-level RlexR_lex exceeds this baseline at layer 0 in all autoregressive models, confirming that the confound is amplified beyond trivial token reuse. Model RlexembR_lex^emb RlexR_lex (L0) Exceeds? GPT-2 0.713 0.735 Yes GPT-2-Med 0.720 0.751 Yes BERT-base 0.778 0.662 No ELECTRA-base 0.727 0.686 No Pythia-1B 0.821 0.836 Yes Pythia-6.9B 0.882 0.910 Yes Pythia-12B 0.891 0.913 Yes LLaMA-2-13B 0.815 0.911 Yes Table 6: Token-embedding baseline comparison. All autoregressive models show MLP-level RlexR_lex exceeding the embedding baseline at layer 0, confirming non-trivial amplification. Bidirectional models (BERT, ELECTRA) show MLP RlexR_lex below the embedding baseline at layer 0, suggesting early-layer bidirectional processing partially resolves lexical overlap—but the confound nonetheless persists at all layers. B.3 Sense selectivity index Table 7 reports the fraction of MLP neurons with SSI>2SSI>2 (sense-selective) at each layer for the 12-layer models. In all models, fewer than 1% of neurons are sense-selective at any layer. The choice of threshold (SSI>2SSI>2) follows the standard Cohen’s d convention for a large effect. Relaxing to SSI>1.5SSI>1.5 roughly doubles the count but does not change the qualitative pattern: sense-selective neurons remain a small minority (<<2%) and the probing dissociation (sense accuracy for blind neurons near chance, selective neurons >>90%) holds across threshold choices. The sense-blind threshold (SSI<0.5SSI<0.5) similarly follows the small-effect convention; using SSI<0.3SSI<0.3 or SSI<0.8SSI<0.8 shifts counts but preserves the functional split. Layer GPT-2 BERT ELECTRA GPT-2-Med 0 0.88 0.63 0.57 0.70 1 0.77 0.78 0.64 0.75 2 0.83 0.64 0.76 0.80 3 0.84 0.61 0.74 0.85 4 0.72 0.67 0.81 0.78 5 0.74 0.63 0.83 0.74 6 0.74 0.62 0.75 0.71 7 0.78 0.60 0.69 0.73 8 0.78 0.66 0.63 0.68 9 0.77 0.78 0.54 0.73 10 0.75 0.88 0.53 0.71 11 0.72 0.79 0.49 0.73 Table 7: Fraction (%) of neurons with SSI>2SSI>2 at each layer. The denominator is dmlpd_mlp (3072 for GPT-2/BERT/ELECTRA, 4096 for GPT-2-Med). Pythia-1B’s layer 0 has the highest fraction (1.80%) due to its larger MLP producing more sense-differentiating neurons in absolute terms, but the fraction remains very small. B.4 SAE collision: 32k vs. 128k dictionary comparison Table 8 compares the two SAE dictionary sizes across all five probed layers. The collision ratios are strikingly similar despite a 4×4× difference in dictionary capacity, confirming that the lexical confound is not resolved by increasing the number of SAE features. 32k SAE 128k SAE Layer Active Blind Coll.% Active Blind Coll.% 0 33.6 10.4 31.7 31.5 9.8 32.6 2 27.8 7.4 29.1 26.4 7.6 31.3 5 21.3 4.3 24.1 20.0 4.4 25.7 8 18.6 2.7 17.7 15.2 2.6 22.4 11 18.4 3.9 28.9 15.2 2.9 25.9 Table 8: SAE collision analysis: 32k vs. 128k features. Active: mean features active per word (>>30% firing rate in either sense). Blind: mean sense-blind features (Cohen’s d<0.5d<0.5, fires for both senses). Coll.%: collision ratio (blind/active). Both SAEs show similar collision ratios at every layer, confirming that increasing dictionary size does not resolve the confound. B.5 Synonym subset validation The full RlexR_lex requires the SYN condition, which is available for 251 of 407 words (62%). We verify that this subset is unbiased by testing whether the core lexical confound—the per-word SL–PS cosine gap—differs between synonym-available and synonym-absent words. Subset comparisons. Table 9 reports Mann-Whitney U tests at five representative layers in each model. No test reaches significance (p<0.05p<0.05) in any model at any layer; most p-values exceed 0.20.2. This confirms that the two subsets are statistically indistinguishable on the metric that drives RlexR_lex. Model L0 Ln4 n4 Ln2 n2 L3n4 3n4 Ln-1 min p GPT-2 .43 .34 .45 .56 .37 .28 GPT-2-Med .50 .28 .38 .42 .28 .28 BERT-base .57 .31 .14 .18 .16 .14 ELECTRA .48 .33 .50 .53 .26 .26 Pythia-1B .62 .20 .33 .30 .16 .16 Pythia-6.9B .80 .43 .26 .13 .25 .13 Table 9: Mann-Whitney U test p-values comparing the per-word SL–PS cosine gap between synonym-available (n=251n=251) and synonym-absent (n=156n=156) words. No test reaches significance; the minimum p-value across all models and layers is 0.130.13 (Pythia-6.9B, layer 24). POS stratification. The synonym subset is POS-imbalanced: 69% verbs vs. 50% in the non-synonym set (χ2=13.8χ^2=13.8, p<0.001p<0.001). Nouns exhibit a larger SL–PS gap than verbs (p<0.005p<0.005 at all layers in GPT-2), reflecting stronger lexical identity effects for nouns. However, both POS categories show a consistently positive SL–PS gap at every layer in every model, confirming that the confound is universal; the POS imbalance affects magnitude but not the direction or existence of the effect. Other properties. Wu-Palmer similarity distributions are indistinguishable between subsets (KS p=0.98p=0.98, Mann-Whitney p=0.91p=0.91). Sentence counts per word are likewise non-significant (KS p=0.68p=0.68). Word length shows a small difference (6.3 vs. 5.9 characters; KS p<0.001p<0.001), but this does not correlate with the SL–PS gap at any layer. Appendix C Additional figures Figure 6: Jaccard overlap of active-neuron sets across layers (GPT-2). The condition ordering sl>ps>SYN>cl sl> ps>SYN> cl is consistent with the cosine similarity results, confirming that the lexical confound affects which neurons fire, not just how strongly they fire. Figure 7: Cross-sense (PS) cosine similarity per word per layer in GPT-2 (top 30 words by variance). Cooler colors indicate lower similarity (stronger sense-specific processing). Most words exhibit a U-shaped trajectory with a dip in mid-layers, though the depth and timing of the minimum varies substantially across words—sense separation is word-specific, not driven by a shared mechanism. Figure 8: RlexR_lex (solid) vs. Rlexno-synR_lex^no-syn (dashed) for all eight models. The alternative metric Rlexno-syn=(M¯ps−M¯cl)/(M¯sl−M¯cl)R_lex^no-syn=( M_ ps- M_ cl)/( M_ sl- M_ cl) does not require synonym data and can be computed for all 407 words. The qualitatively identical trajectories confirm that our findings do not depend on synonym coverage (62% of words, validated by subset invariance test; see Appendix B.5). Figure 9: Mean magnitude divergence (Δ=1|A∩B|∑i∈A∩B|ai−bi| = 1|A∩ B| _i∈ A∩ B|a_i-b_i|) by condition and layer. Unlike cosine similarity (which shows a U-shaped trajectory), magnitude divergence increases monotonically with depth, reflecting growing activation scale differences as representations become more specialized. The condition ordering is reversed—cl>SYN>ps>sl cl>SYN> ps> sl—because more dissimilar conditions produce larger activation differences among their shared neurons. Figure 10: Condition means at three representative layers in GPT-2 (layers 0, 6, 11). At layer 0, ps and sl are close while SYN and cl are distant. By layer 6, the gap between ps and sl widens. At layer 11, all conditions retain the same ordering with moderate separation. Figure 11: RlexR_lex (lexical contribution ratio) across layers for all nine models with 95% bootstrap CIs. The lexical contribution decreases with depth but never reaches zero in any model. Larger models (LLaMA-2-70B, cyan) show sustained Rlex≈0.35R_lex≈ 0.35–0.450.45 across 80 layers. Appendix D Supplementary analyses Sense accuracy after ablation. On GPT-2, clean (no ablation) sense accuracy averages 73% (measured via diagnostic token probabilities on held-out sentences). After ablating sense-selective neurons, accuracy drops by 8 percentage points; after ablating sense-blind neurons, it drops by <<1 p (indistinguishable from random variation), confirming that sense-selective neurons are causally necessary for the model’s sense-discriminating behavior. Concrete causal example. For the word bank, sense-A (financial institution) and sense-B (geographic land) yield distinct diagnostic tokens (loan, account for sense A; river, shore for sense B). Ablating sense-A-selective neurons increases perplexity on financial-institution sentences while leaving river-bank sentences unchanged; ablating sense-blind neurons has negligible impact on either. Cross-model scaling details. Pythia-1B (>>5× GPT-2’s parameters) shows a qualitatively identical trajectory. Pythia-6.9B confirms the confound persists at the scale where modern interpretability work typically operates. Pythia-12B and LLaMA-2-13B (gated SiLU MLP, a fourth architecture family) maintain the strict condition ordering at every layer; LLaMA-2-13B achieves the lowest absolute similarities but the same pattern, with RlexR_lex ranging from 0.910.91 (layer 0) to 0.250.25 (layer 38). The U-shaped trajectory. All conditions exhibit a U-shaped similarity trajectory—high in early layers, lowest in mid-layers, partially recovering in final layers. This is consistent with prior observations that late-layer representations reconverge toward the output vocabulary space (nostalgebraist, 2020; Geva et al., 2022), possibly partially undoing sense-specific processing. LIS dose-response. Table 10 shows the full dose-response from projecting out the top-k LIS components. The PS–SYN gap closes monotonically; PS similarity rises slightly (removing between-word variance makes within-word representations more uniform) while SYN drops (shared meaning partly encoded along word-form directions). The gap is the informative quantity. PS–SYN gap RlexR_lex Δ Δ GPT-2 (d=3072, layer-averaged) Baseline 0.119 0.65 — — k=10k=10 (0.3%) 0.087 0.52 ++0.008 −-0.032 k=20k=20 (0.7%) 0.047 0.41 ++0.019 −-0.054 k=50k=50 (1.6%) −-0.028 <<0 ++0.031 −-0.116 Pythia-1B (d=8192, layer-averaged) Baseline 0.097 0.59 — — k=10k=10 (0.1%) 0.062 0.47 ++0.016 −-0.019 k=20k=20 (0.2%) 0.033 0.29 ++0.025 −-0.040 k=50k=50 (0.6%) −-0.032 <<0 ++0.037 −-0.092 Table 10: LIS dose-response ablation. Removing k=20k=20 components (≤ 1% of dimensions) reduces RlexR_lex by 37–51%. At k=50k=50, the PS–SYN gap reverses—the confound is eliminated. Per-word collateral damage. The collateral-damage experiment (Section 5.5) ablates sense-blind vs. sense-selective neurons for individual words at layer 6. Across 10 polysemous words tested, sense-blind ablation consistently produces low specificity (both senses equally affected), while sense-selective ablation is 55–72×72× more targeted. The full per-word results are included in the released code. Unsupervised LIS detection. To test whether the LIS can be approximated without sense annotations, we replace WordNet synonyms with nearest neighbors from GPT-2’s embedding layer (n=393n=393 pairs). The resulting subspace overlaps with the WordNet-based LIS at 0.190.19–0.250.25 (77–10×10× above a random baseline of 0.0260.026). This partial recovery suggests the subspace is identifiable without labels, though a fully unsupervised method remains open. Appendix E Attention representations We apply the same four-condition decomposition to attention head outputs (concatenated per-head output vectors at the target word position). The ordering sl>ps>SYN>cl sl> ps>SYN> cl holds in attention representations at every layer in all models, with RlexR_lex values comparable to MLP values in early layers. Since the residual stream sums attention and MLP outputs, the confound propagates to SAEs trained on resid_post (Section 5.3). Appendix F Reproducibility details Compute. All experiments were run on a university HPC cluster. Models up to 1B parameters used NVIDIA V100 GPUs (16 GB VRAM); Pythia-6.9B required an A100 GPU (40 GB VRAM); Pythia-12B and LLaMA-2-13B each required an NVIDIA GH200 GPU (96 GB VRAM). Decomposition experiments took approximately 2–4 hours per model on a single GPU (6–8 hours for the 12B and 13B models). The SAE collision experiments took approximately 20 minutes (32k SAE) and 30 minutes (128k SAE) on a V100. Software. Models were loaded using HuggingFace Transformers (v4.36+) in float32 precision. SAE experiments used TransformerLens (v1.6+) and SAE Lens (v3.0+). Random seed was fixed at 42 for all experiments. Bootstrap confidence intervals used 10,000 resamples stratified by word. Metric definitions (formal). For completeness, we provide the formal definitions of all metrics. Cosine similarity: cos(,)=⋅‖‖cos(a,b)= a·b\|a\|\;\|b\| Jaccard overlap: A neuron is active if its absolute activation exceeds the per-sentence median. J=|A∩B|/|A∪B|J=|A∩ B|/|A∪ B|. Magnitude divergence: Δ=1|A∩B|∑i∈A∩B|ai−bi| = 1|A∩ B| _i∈ A∩ B|a_i-b_i| Sense Selectivity Index: SSIi=|μi(A)−μi(B)|/sipooledSSI_i=| _i^(A)- _i^(B)|/s_i^pooled where sipooled=((nA−1)sA2+(nB−1)sB2)/(nA+nB−2)s_i^pooled= ((n_A-1)s_A^2+(n_B-1)s_B^2)/(n_A+n_B-2). Neurons with SSI>2SSI>2 are classified as sense-selective. F.1 Modern corpus validation (CoarseWSD-20) A potential concern is that SemCor draws on the 1960s Brown Corpus. To rule out the possibility that our findings are artifacts of vintage text, we replicate the full four-condition decomposition on CoarseWSD-20 (Loureiro et al., 2022), a dataset of 20 polysemous nouns with sense-annotated Wikipedia sentences. We construct all four conditions—SL, PS, SYN, and CL—by pairing CoarseWSD-20 sense labels with WordNet synonyms whose sentences are drawn from the same Wikipedia corpus (different words’ entries), ensuring the SYN condition uses modern text throughout. We run the decomposition on GPT-2, BERT-base, and Pythia-1B. Model SL PS SYN CL PS>>SYN GPT-2 .90–.60 .89–.47 .31–.45 .26–.45 p<.001p<.001 ✓ BERT-base .70–.58 .54–.30 .09–.32 .05–.35 p<.01p<.01 ✓ Pythia-1B .99–.46 .99–.27 .09–.28 .09–.26 p<.001p<.001 ✓ Table 11: Modern corpus validation on CoarseWSD-20 (Wikipedia sentences, 20 nouns). Ranges show cosine similarity across layers (layer 0 to final). The critical PS >> SYN gap—indicating lexical form drives more overlap than shared meaning—is statistically significant at every layer in all three models (Wilcoxon signed-rank, Holm-Bonferroni corrected). RlexR_lex bootstrap 95% CI remains above zero everywhere. The results replicate the SemCor findings on modern text. The critical PS >> SYN gap is significant at every layer in all three models, confirming that lexical form is a stronger driver of MLP neuron overlap than semantic similarity even on Wikipedia text. GPT-2’s RlexR_lex ranges from 0.980.98 (layer 0) to 0.570.57 (layer 11), closely matching the SemCor trajectory (0.740.74–0.290.29). Pythia-1B shows an even stronger confound (Rlex≥0.41R_lex≥ 0.41 at all layers). The full ordering sl>ps>SYN>cl sl> ps>SYN> cl holds at the majority of layers; at a few late layers, SYN and CL converge (both approach 0.30.3–0.40.4), consistent with the SemCor pattern where all conditions converge at depth. The lexical confound is a property of how transformers represent polysemous words, not an artifact of the SemCor corpus.