Sparse Autoencoders Find Partially Interpretable Features in Italian Small Language Models

Sparse Autoencoders Find Partially Interpretable Features in Italian Small Language Models Alessandro Bondielli 1,2,* , Lucia Passaro 1,2 and Alessandro Lenci 2 1 Department of Computer Science, University of Pisa 2 CoLing Lab, Department of Philology, Literature and Linguistics, University of Pisa Abstract Sparse Autoencoders (SAEs) have become a popular technique to identify interpretable concepts in Language Models. They have been successfully applied to several models of varying sizes, including both open and commercial ones, and have become one of the main avenues for interpretability research. A number of approaches have been proposed to extract latents from the model, as well as automatically provide natural language explanations for the concepts they supposedly represent. Despite these advances, little attention has been given to applying SAEs to Italian language models. This may be due to several factors: i) the small number of Italian models; i) the costs associated with leveraging SAEs, which includes the training itself, as well as the necessity to parse and assign an interpretation to a very large number of features. In this work, we present an initial step toward addressing this gap. We train a SAE on the residual stream of the Minerva-1B-base-v1.0model, for which we release the weights; we leverage an automated interpretability pipeline based on LLMs to evaluate both the quality of the latents, and provide explanations for some of them. We show that, albeit the approach shows several limitations, we find some concepts in the weights of the model. Keywords Mechanistic Interpretability, Sparse Autoencoders, Large Langauge Models, Italian, 1. Introduction The rise of Large Language Models (LLMs) have pro- foundly affected the landscape of Natural Language Pro- cessing (NLP). These models have demonstrated remark- able capabilities in many tasks, often achieving near- human performances and saturating benchmarks as soon as they are released. Nevertheless, many questions re- main about their internal workings: Whether and how they perform some form of reasoning [1], and to what extent their grasp of concepts through natural language approximates human conceptual understanding. The aim ofMechanistic Interpretability(MechIn- terp) is to address this pressing issue by attempting to reverse-engineer the learned representations and algo- rithms within their neural networks [2]. A promising technique within MechInterp is the use of sparse dictio- nary learning methods like Sparse Autoencoders (SAEs) [3]. The idea behind SAEs is similar to that of standard autoencoder. Autoencoders are unsupervised models that learn two functions: an encoding function, that projects the input data from an푛dimensional space into CLiC-it 2025: Eleventh Italian Conference on Computational Linguis- tics, September 24 — 26, 2025, Cagliari, Italy * Corresponding author. † These authors contributed equally. $alessandro.bondielli@unipi.it (A. Bondielli); lucia.passaro@unipi.it (L. Passaro); alessandro.lenci@unipi.it (A. Lenci) 0000-0003-3426-6643 (A. Bondielli); 0000-0003-4934-5344 (L. Passaro); 0000-0001-5790-43086 (A. Lenci) ©2025 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (C BY 4.0). a푘!=푛dimensional space; a decoding function, that should reconstruct the푘-dimensional data back into the original푛-dimensional one. Autoencoders are typically used for dimensionality reduction, i.e.,푘 << 푛. In the case of SAEs, instead,푘 >> 푛: the model is trained to project the input space into a much higher-dimensional (and thus sparser) one, and then project it back into the original dimensional space. In our context, SAEs are trained to reconstruct the internal activations of a lan- guage model’s residual stream by projecting them into a higher-dimensional latent space, while being constrained to use only a small number of “features” from a learned dictionary. Thissparsityconstraint encourages the SAE to learn a set ofmonosemanticfeatures, also referred to aslatents, that is, features each corresponding to a single, hopefully more interpretable concept [4]. This is in con- trast with apolysemanticrepresentation, which is typical of standard dense neural networks [5,6], in which sev- eral concepts are superimposed in the same activation patterns. SAEs allow to decompose model activations into a set of near-orthogonal, i.e., largely disentangled features that should be semantically coherent. Recent work has demonstrated the effectiveness of SAEs in uncovering meaningful features within both toy models [7] and large-scale commercial LMs, revealing representations for concepts ranging from concrete ob- jects to abstract ideas [8,9,10]. As noted in [9], several distinctive features have been identified in Claude-3.5- Sonnet – most notably, one corresponding to the “Golden Gate Bridge.” SAEs have also been applied successfully to smaller, English-centric models in the 1 to 10 Billion CEUR Workshop Proceedings ceur-ws.org ISSN 1613-0073 published 2025-11-30 parameter range [11]. This class of models is becom- ing more and more relevant, as research on Small Lan- guage Models (SLMs) [12] and Baby Language Models (BabyLMs) [13,14], that mitigate the costs of training and serving LLMs while attempting to retain most of their abilities, is a very active endeavour particularly in the open-source/open-weights community. Two key limitations remain for the applicability of SAEs to achieve interpretability. First, the computational cost of training a SAE. Given their nature, the internal layer of a SAE has to be a number of times larger than the size of residual stream, and thus the context window, of the target LM. The number of parameters of a SAE scales with a factor of the context size of the model, mul- tiplied by the number ofhookpointsin the models where activations are collected (e.g., after every transformer block/layer). Thus, the larger the target LM, the bigger and the more computationally expensive the SAE. Second, and most importantly, SAEs output a large number of features, that have then to be interpreted in some way. While the literature has not reached a con- sensus on what is the best practice, a popular method to address this is to leverage another LLM to provide expla- nations for the features based on examples of which to- kens (and respective contexts) they fired on. For example, if the feature푓 푖 fired on 10 tokens, the explainer model is fed with these tokens, their contexts and the request to find a common property among them. In most works, commercial LLMs with hundred of billions of parameters are successfully used for this task [9,10]. However, re- searchers have also shown that smaller and cheaper LMs can be leveraged effectively as well [15]. The vast majority of efforts regarding the use of SAEs for interpretability has been done on English-centric LMs[9,10,11]. In addition to this, several efforts have been made in the direction of finding universal features that apply across models and languages [16,17]. How- ever, models primarily trained on languages other than English have received less attention. In this work, we aim to provide an early evaluation on the feasibility of using SAEs to interpret models trained to be natively Italian. In the interest of main- taining a limited computational cost, we chose to use theMinerva-1B-base-v1.0from the Minerva model family [18]. We trained a SAE on the residual stream of every layer of the model using an Italian split of mC4 [19]. Then, we collected feature activations for the Italian dump of Wikipedia [20], and attempt to explain them and score explanations automatically using an LLM, fol- lowing [15]. Our contributions are the following: • Wetrain and release a Sparse Autoencoder onMinerva-1B-base-v1.0. We make the Autoencoder weights available to the research community via HuggingFace. 1 • We collect feature activations from a rela- tively large collection of Italian data, and pro- vide aquantitative and qualitative evalu- ation on the explanations using an auto- interpretability pipeline. We show that SAE are promising for finding concepts in Italian SLMs, but auto-interpretability pipelines shows several limitations for Italian. • We report on thechallenges and lessons learned on training and using SAEs, especially in computationally constrained settings. This paper is organised as follows: In Section 2 we detail the training procedure of the SAE; Section 3 pro- vides an overview of the auto-interpretability pipeline we employ; in Section 4 we present and discuss the obtained results; finally, Section 5 draws some conclusions and highlights future works. 2. SAE Training In the following, we detail the data and procedure used to train the SAE on theMinerva-1B-base-v1.0SLM. We trained the SAE on the residual stream of the model, with hookpoints on the outputs of each attention block. For our experiments we used the Sparsify library from EleutherAI, 2 which is built to roughly follow the train- ing recipe presented in [10] for a GPT-4 SAE. It trains a푘-Sparse Autoencoder [21]. The autoencoder uses a TopK activation function that allows for direct control over the number of active latents. Specifically, it only keeps the푘largest latents and assign zero to the rest. Authors in [10] argue that this eliminates the need for the L1 penalty, which biases activations toward zero and is only a rough proxy for L0, and supports any activa- tion function. They also show that it outperforms ReLU autoencoders in sparsity-reconstruction tradeoffs and en- hances monosemanticity as small activations are clamped to zero. Recipe.A full breakdown of the most relevant param- eters selected for training is presented in Table 1. The parameters were chosen following recipes for similar sized models, e.g. [11]. The expansion factor controls the size of the hidden layer, and is a multiplier over the model context size. In our case, an expansion factor of 32 yields a hidden layer of size2,048×32 = 65,536 parameters. 1 https://huggingface.co/alessandrobondielli/sae-Minerva-1B-32x The model can be used with the Sparsify and Delphi libraries for interpretabilty. 2 https://github.com/EleutherAI/sparsify ParameterValue ActivationTopK Expansion Factor 32 k32 Multi TopKFalse TranscodeFalse Batch Size16 Loss FunctionFraction of Variance Unexplained (FVU) OptimizerSignum Table 1 Parameters for the SAE training. Data.As for the training data, we chose to use mC4 [22]. Specifically, we consider the “tiny” split of the clean_mc4_itdataset [19]. It includes 6 Billion tokens (4 Billion words). The choice of the dataset was made on the basis that it is relatively large, especially for the Italian language, and it includes a variety of different texts. The data was not included in the training set for Minerva-1B-base-v1.0. We chose to use 6 Billion tokens following recent literature on training SAEs for similar-sized models [11]. Setup.We trained our model on a single Nvidia A100 with 80 GB VRAM. A full training run required 200 GPU hours, which roughly equates to 8 days. The final model, that we callsae-Minerva-1B-32x, occupies around 40 GB of disk space including hookpoints to all layers. The final model is available on HuggingFace 3 and can be loaded and used with Sparsify. 3. Auto-Interpretation of Features For finding and explaining latents of the SAE models, we use the auto interpretability pipeline proposed in [15]. It is implemented via the Delphi library from EleutherAI. 4 The library includes tools for generating and scoring text explanations for SAE. The auto intepretability pipeline has three main steps: 1. Activations are collected from a text dataset. 2.AnExplainerLLM is shown activating contexts and is asked to provide interpretations in natural language for them. 3. AScorerLLM is tasked to distinguish between activating and non activating contexts of a fea- ture, as a binary classifier. This is achieved by asking the model, given several sequences and an intepretation, whether each of the sequences activates the SAE latent with that interpretation. 3 https://huggingface.co/alessandrobondielli/sae-Minerva-1B-32x 4 https://github.com/EleutherAI/delphi In the following we detail our implementation of the pipeline. Collecting Activations.As for the text dataset, we chose to use 20 Million tokens from the Italian subset of the November 2023 Wikipedia dump [20] available on HuggingFace. 5 The choice of Wikipedia as our test dataset rather than a sample of the SAE training data (clean_mc4_it) was made with the purpose of increas- ing the probability of finding concepts specific to the Italian language and culture, that could have been left out from a relatively small sample of a web-based dataset. We created equal-sized batches from the texts, shuffled them, and then collected their token-level activations. We collected the activations at three hookpoints, namely at layers 2, 8 and 14. We did so with the aim of under- standing whether there is any difference in the features found near the beginning, middle, or near the end of the residual stream. In the following we use the hookpoint notation to refer to layers, namelyLayer.푥. Generating Explanations.As for the explanation generation step, we followed the same procedure as [15]. We showed the Explainer LLM 40 examples of the activat- ing tokens and their contexts. We used a context length of 32 tokens. The activating token can be in any of the 32 positions, but is highlighted as"« token »". We show an example of explanation generation in Figure 1. To limit the computational cost, we attempted to gen- erate explanations only for a sample of 2,000 latents se- lected from the pool of 65k. Latents with less than 40 examples were skipped. We used the number of latents with enough examples at each hookpoint in the residual stream to highlight their differences. The chosen model to generate explanations is Meta-Llama-3.1-8B-Instruct-AWQ-INT4, 6 a quan- tized version ofMeta-Llama-3.1-8B-Instruct[23]. We prompted the model both in English and Italian. For the English prompt, we used the one provided in [15] for the zero-shot experiment. The Italian version is a direct translation of the English prompt. The translation was made semi-automatically: first, the prompts were translated with Gemini-2.5 Pro. 7 Then, the translated prompt was manually revised to ensure its quality. 8 Scoring Explanations.Finally, we scored the explana- tions. We employed a binary classification method. For each explanation, the model was shown five examples of sentences, where each had equal probability of being associated with the latent. The model was then asked 5 https://huggingface.co/datasets/wikimedia/wikipedia 6 https://huggingface.co/hugging-quants/Meta-Llama-3. 1-8B-Instruct-AWQ-INT4 7 https://gemini.google.com, accessed in June 2025. 8 See Appendix A for the prompts in both languages. Figure 1:Explanation Generation with examples. Activating tokens are marked as« token ». In the Figure we highlight them also inbold red. to decide, for each example, whether it corresponded to the explanation, and output a list of of decisions. If the output did not match a list of decision, it was assigned None. The output was then compared with the ground truth provided by the activations. The model for scoring was the same one used to generate explanations. As for the prompt and its translation in Italian, we followed the same translation procedure as well. We evaluated the quality of explanations with accuracy. Specifically, we considered aper-sample accuracy(i.e., how many out of the five examples the scorer model got right) and the average accuracy across across latents for the same hookpoint. We acknowledge that our choice of using a multilin- gual, relatively small, and quantized LLMs for generating and scoring explanations is far from ideal, and it is not an adequate substitute neither for human evaluation nor for more performing LLMs. The choice of a multilingual model rather than an Italian-only one was made due to the current lack of such models with open weights, high performances and capability to follow instructions. This choice led also to prompting the model both in English and Italian; this was done to assess its explanation/scor- ing capabilities both in its “native” language, albeit on data from another language, and on Italian, in order to limit potential biases in the interpretation of results from using only one or the other language. As for the choice of a medium-sized quantized model, this was made in the interest of limiting the computational costs of our experiments, i.e., both in terms of the memory footprint of the model, and of the overall GPU hours. Using larger (including non-quantized variants) models would have drastically increased both the need of resources and over- all time of the experiments. Nonetheless, we argue that our choice represents a lower-cost alternative to using much larger and costlier models, that could prove es- pecially useful to provide some early insights into the quality of the latents found by the SAE, and of the model being interpreted. Authors in [15] estimate a cost in the order of hun- dreds or thousand of dollars for explaining and scoring 100k latents with larger or commercial models; our exper- iments, in contrast, can be easily replicated on a single GPU. In our case, generating and scoring explanations for 2,000 latents at three different hookpoints, in two different languages, took 0.5 GPU hours each on a single Nvidia A100, for a grand total of 3 GPU hours. Given the size of the model used, the experiments could be also replicated on much less performing hardware as well, provided a trade-off on GPU hours. 4. Results and Discussion In the following, we present our results. First, we show a quantitative evaluation of the extracted latents, and the performances of the generation and scoring pipeline, both with Italian and English prompts. To explore the results in greater depth, we also perform a qualitative evaluation. We consider explanations that received high- est scores by the scorer model. We use the results to discuss the feasibility of the proposed approach on Ital- ian SLM, as well as potential shortcomings. 4.1. Quantitative Evaluation The core of our quantitative analysis is based on the results we obtained using the Delphi library, with the configuration presented in Section 3. Quality of the Latents.To evaluate the quality of the latents obtained via the SAE encoding, several metrics can be used. Recall that we collected latent activations us- ing 20 Million tokens from the Italian subset of Wikipedia. Note also that here we are not yet using prompts, so we do not distinguish between Italian and English. Table 2 provide several common metrics used to evalu- ate the quality of the extracted latents at each hookpoint. First, we look at fraction of alive latents. A latent is con- sidered alive if at least one input token in the dataset made it fire. With the exception of Layer.8, the other two have much smaller fractions of alive latents than it is typical for SAEs (see for examples results reported in Figure 2:Accuracy distribution with Italian prompts.Figure 3:Accuracy distribution with English prompts. MetricLayer.2Layer.8Layer.14 Fraction of latents alive (%)72.0295.1684.65 Latents fired>1% of the time (%)0.270.450.38 Latents fired>10% of the time (%)0.060.000.01 Weak single-token latents (%)9.932.202.77 Strong single-token latents (%)12.400.550.47 Table 2 Latent activity statistics across selected layers [10] and [11]). This may be the results of several factors. On the SAE side, we could hypothesize an overcomplete latent space for the evaluation data, i.e. a too broad latent space for encoding the evaluation data. Recall in fact that we used mC4 to train the SAE, and evaluated it on Wikipedia, which may present less variety in terms of texts. On the Language Model side, we could hypothesize that the latent space of the analyzed model is very anisotropic at both earliest and latest layers, while more isotropic near the middle of the stack. This however is in direct contrast with works such as [24], and thus re- quires a more in-depth analysis, that we leave to future works. Another interesting aspect to consider are weak and strong single-token latents, that is latents that fire on a specific token only. Weak ones are those for which the token in question makes many other latents fire; strong ones are cases where the token preferentially activates the specific latent. We observe thatLayer.2is heavily biased towards single token latents. This may indicate that earliest layers sill leverage the embedding represen- tation quite strongly. Finally, we see that latents that fired either more than one or 10% of the times are less and less as we move towards the residual stream. These latents may be used to store single-token concepts of words such as function ones. Quality of the Explanations.To evaluate the quality of explanations, we consider the results of the explana- tion generation and scoring pipeline. Specifically, for each latent, we compute the accuracy at distinguishing between sequences that activate and do not activate the latent. Figures 2 and 3 show respectively the distribution of Accuracy for the scorer model using Italian and En- glish prompts for each hookpoint in the residual stream. We observe that, in both cases, there are significant differences both in distribution and averages for the three hookpoints. We also observe that explanations for latents extracted from later layers seem to be easier to score cor- rectly for the scorer model. This may indicate thatcon- cepts identified in later layers are, on average, more easily interpretable by an LLM. The accuracy scores obtained using the Italian prompt are generally higher than those for the English one, with average scores rang- ing from 0.64 to 0.69; the English ones, in contrast, range from 0.55 to 0.62. However, these results in isolation can- not be taken as a direct indication that explanations in Italian are better than English ones. It may as well be the result of poorer and broader explanations provided by the Explainer model. We also plot the aggregate confusion matrices over all the predictions of both prompts. The confusion matrices are shown in Figure 4. While the model prompted in Italian seem to fare better in all metrics except for True Positives, we also see that the number of times the model was not able to follow instructions and provide a predic- tion with the Italian prompt is three times higher than with the English one. This may be further indication that the Explainer/Scorer model used struggles with Italian. 4.2. Qualitative Evaluation To dig deeper into the quality of the explanations, we di- rectly looked at them and provide examples of seemingly good and bad explanations. Specifically, we sampleed from the 50 explanations that received highest scores by the Scorer, both in English and Italian. As for the Italian explanations, we immediately ob- served that a large fractions of them suffer fromDegen- erate Repetition[25]: The model starts to generate the same token or sequence of tokens over and over. On (a) Italian Prompt. (b) English Prompt. Figure 4:Confusion Matrices for the Scorer model on both the Italian and English prompt. the contrary, English ones does not suffer from this is- sue. However, if we look at the quality of explanations, aside from repetitions, we observe that at least some of the Italian ones are quite relevant to the examples, and while sometimes slightly missing the mark, they high- light some interesting aspects of the tokens that fire the latent. Among these, we can clearly see thatLayer.2is mostly represented by single token latents: the token “ale” as part of “federale” (federal), in several contexts, or the token “letto”, as both a noun (bed) and a verb (read). Layer.14latents on the other had appear to represent more abstract concepts. For example, we see latents firing on the final number of a year date, and a very interesting latent firing on the concept ofcompetition(see Fig.??). Layer.8explanations are generally more confusing and less interesting. Examples are reported in Figure 5 with the relative explanation, cut to avoid showing repetitions. As for the English explanations on the other hand, we observed that most of them actually miss the mark. In fact, they often provide an explanation related to the con- texts, rather than the firing tokens. This may be due to the fact that, while it is specified in the prompt, we use Italian texts as examples but instructions and expected outputs are in English. Neverhteless, we observe an inter- esting trend: most explanations, at all layers, that actually focus on the firing tokens refer to functional aspects of the text, including punctuation marks, special charac- ters, and functional words. For example, Latent 1818 of Layer.14is explained as “Prepositions and conjunctions used to connect words or phrases in Italian text, such as "a", "di", "nel", "in", "su", "da", "al", "nei", "all", "sulle", "col" [...]”. This is in contrast with what we observed for Italian explanations. 4.3. Discussion of Key Findings In the following, we highlight some of the key aspects that emerged from the experiments. SAEs can find partially interpretable features in Italian Small Language Models.First, we observe that using a SAE we are able to extract features that somewhat align to interpretable concepts, despite some limitations that we can mostly attribute to the quality of the training data, both for the original model and the SAE, and to the limitations of the auto-interpretability pipeline (see below). It is possible that leveraging a dataset more attuned with the Italian culture would yield better results in finding relevant latents. Different behaviours in the residual stream.We observed some relevant differences in the quality and types of latents that are properly identified in various points of the residual stream. In general, we observed that latents obtained from earlier in the stream are more relevant to single tokens and grammatical aspects of the language, while latents in later points of the stream show a slight tendency towards more abstract conceptualiza- tions. Auto-interpretability is promising, but currently shows limitations for Italian.Auto-interpretability pipelines are definitely a promising approach for simpli- fying and reducing the costs of finding explanations for latents of SAEs. Our experiment suggest in fact that this is a low-cost alternative that is nonetheless able to de- liver some interesting results. Nevertheless, we observed two main limitations that we can argue are actually two sides of the same coin. On the one side, the Explainer model showed some limitations in understanding the task and providing coherent texts for the explanations, while the Scorer model performed quite poorly in the binary classificationt task. This is especially true in the case of language mixing, i.e. when the model is prompted in its “main” language, i.e. English, but has to work on another Figure 5:Examples of explanations for latents in Italian. language, in this case Italian. On the other side, the size of the model used in our experiments could severely limit its performances. Thus, both issues could be solved either by leveraging a stronger Italian-centric model as the Explainer/Score, or by using a generally larger and better performing model. However, as for the first solution, there are currently no models on par with English ones in the 7-15B parameters range, wich whould allow for reducing the cost. As for the second solution, this would dramatically increase the costs, both computational and monetary. 5. Conclusions and Future Works In this paper, we have shown that SAEs can partly un- cover interpretable concepts in Italian Small Language Models. Specifically, we did so by training a SAE model on the residual stream of theMinerva-1B-base-v1.0 SLM, and then applying an auto-interpretability pipeline to generate explanations for its latents. Our findings suggest that SAE can be used to this end, and that it exist a hierarchical representation within the model, with earlier layers showing more token-centric features and later layers more abstract concepts. As for the auto-interpretability pipeline, while promising for its low cost, underscored the need for better language- specific tools for Italian. Moving forward, we aim to explore several avenues. First, we plan to scale our experiments in two directions: on the one hand, we aim to train SAEs on larger Italian models, e.g. larger variants of Minerva as well as others; on the other hand, we observe that we need to improve the models used for auto-interpretability, in order ob- tain more reliable explanations. This could be achieved both by scaling them up substantially, and by tuning Italian-speaking models to the specific tasks of latent ex- planation and scoring. Second, we plan to leverage SAE and auto interpretability to address potential differences of representations in models pre-trained specifically on Italian data, e.g. Minerva and Velvet [26], and multilin- gual models that received only fine-tuning in Italian, like the LLaMAntino variants [27] and Cerbero [28]. Finally, we plan to explore the larger latent space to attempt to uncover features linked specifically to Italian-centric con- cepts, in addition to properties of the Italian Language. This work is an early first step in exploring inter- pretability research using Sparse Autoencoders for non- English-centric Language Models. Albeit limited in scope, we are optimistic that it may provide a relevant founda- tion for this yet under explored research area, both in terms of approach and the release of open models for the community. Limitations Our initial effort to interpret Italian SLMs using Sparse Autoencoders has several limitations. The choice of the smallerMinerva-1B-base-v1.0model, driven by computational constraints, means our findings might not generalize to larger Italian models. The SAE’s training data, while substantial for Italian, might not fully capture all linguistic nuances, potentially affecting the quality of learned features. Additionally, using different data to train and evaluate the SAE, while arguably not problem- atic in principle, may have introduced some unwanted biases. A key limitation stems from our cost-effective auto- interpretability pipeline, which relies on a relatively small, quantized multilingual LLM. This model strug- gled with generating coherent Italian explanations, often repeating itself, and performed poorly in scoring when mixing languages. This highlights the strong dependence of explanation quality on the explainer/scorer model’s capabilities, and the current lack of robust, affordable, Italian-specific tools. Finally, our analysis was based on a sample of 2000 latents across only three layers, not the entire SAE latent space. While insightful, this limited scope and subjec- tive qualitative assessment means we cannot yet claim a comprehensive understanding of the model’s internal workings. Acknowledgments This work has been supported by the PNRR MUR project PE0000013-FAIR (Spoke 1), funded by the European Com- mission under the NextGeneration EU programme, and the EU EIC project EMERGE (Grant No. 101070918). References [1] P. Shojaee, I. Mirzadeh, K. Alizadeh, M. Horton, S. Bengio, M. Farajtabar, The illusion of thinking: Understanding the strengths and limitations of rea- soning models via the lens of problem complexity, 2025. URL: https://ml-site.cdn-apple.com/papers/ the-illusion-of-thinking.pdf . [2]C. Olah, N. Cammarata, L. Schubert, G. Goh, M. Petrov, S. Carter, Zoom in: An introduction to circuits, Distill 5 (2020) e24. [3] B. A. Olshausen, D. J. Field, Sparse coding with an overcomplete basis set: A strategy employed by v1?, Vision research 37 (1997) 3311–3325. [4] H. Cunningham, A. Ewart, L. Riggs, R. Huben, L. Sharkey, Sparse autoencoders find highly interpretable features in language models, 2023.URL:https://arxiv.org/abs/2309.08600. arXiv:2309.08600. [5]N. Elhage, T. Hume, C. Olsson, N. Schiefer, T. Henighan, S. Kravec, Z. Hatfield-Dodds, R. Lasenby, D. Drain, C. Chen, R. Grosse, S. McCan- dlish, J. Kaplan, D. Amodei, M. Wattenberg, C. Olah, Toy models of superposition, Transformer Circuits Thread (2022). [6]A. Scherlis, K. Sachan, A. S. Jermyn, J. Benton, B. Shlegeris, Polysemanticity and capacity in neu- ral networks, 2025. URL: https://arxiv.org/abs/2210. 01892.arXiv:2210.01892. [7]E. Anders, C. Neo, J. Hoelscher-Obermaier, J. N. Howard, Sparse autoencoders find composed fea- tures in small toy models, https://shorturl.at/YOtYR, 2024. [8]T. Bricken, A. Templeton, J. Batson, B. Chen, A. Jermyn, T. Conerly, N. Turner, C. Anil, C. Deni- son, A. Askell, R. Lasenby, Y. Wu, S. Kravec, N. Schiefer, T. Maxwell, N. Joseph, Z. Hatfield- Dodds, A. Tamkin, K. Nguyen, B. McLean, J. E. Burke, T. Hume, S. Carter, T. Henighan, C. Olah, Towards monosemanticity: Decomposing language models with dictionary learning, Transformer Cir- cuits Thread (2023). https://transformer-circuits. pub/2023/monosemantic-features/index.html. [9] A. Templeton, T. Conerly, J. Marcus, J. Lind- sey, T. Bricken, B. Chen, A. Pearce, C. Citro, E. Ameisen, A. Jones, H. Cunningham, N. L. Turner, C. McDougall, M. MacDiarmid, C. D. Free- man, T. R. Sumers, E. Rees, J. Batson, A. Jermyn, S. Carter, C. Olah, T. Henighan, Scaling monose- manticity: Extracting interpretable features from claude 3 sonnet,Transformer Circuits Thread (2024). URL: https://transformer-circuits.pub/2024/ scaling-monosemanticity/index.html. [10] L. Gao, T. D. la Tour, H. Tillman, G. Goh, R. Troll, A. Radford, I. Sutskever, J. Leike, J. Wu, Scaling and evaluating sparse autoencoders, arXiv preprint arXiv:2406.04093 (2024). [11]T. Lieberum, S. Rajamanoharan, A. Conmy, L. Smith, N. Sonnerat, V. Varma, J. Kramar, A. Dragan, R. Shah, N. Nanda, Gemma scope: Open sparse autoencoders everywhere all at once on gemma 2, in: Y. Belinkov, N. Kim, J. Jumelet, H. Mo- hebbi, A. Mueller, H. Chen (Eds.), Proceedings of the 7th BlackboxNLP Workshop: Analyzing and Interpreting Neural Networks for NLP, Association for Computational Linguistics, Miami, Florida, US, 2024, p. 278–300. URL: https://aclanthology.org/ 2024.blackboxnlp-1.19/. doi:10.18653/v1/2024. blackboxnlp-1.19. [12] B. Yuan, C. Li, C. Zhang, X. Chen, Y. Liu, Y. Zhang, Y. Wu, Z. Wang, Y. Wang, Y. Cao, et al., Small language models: A survey of the state of the art, arXiv preprint arXiv:2407.01513 (2024). [13]M. Y. Hu, A. Mueller, C. Ross, A. Williams, T. Linzen, C. Zhuang, R. Cotterell, L. Choshen, A. Warstadt, E. G. Wilcox, Findings of the sec- ond BabyLM challenge: Sample-efficient pretrain- ing on developmentally plausible corpora, in: M. Y. Hu, A. Mueller, C. Ross, A. Williams, T. Linzen, C. Zhuang, L. Choshen, R. Cotterell, A. Warstadt, E. G. Wilcox (Eds.), The 2nd BabyLM Challenge at the 28th Conference on Computational Natural Language Learning, Association for Computational Linguistics, Miami, FL, USA, 2024, p. 1–21. URL: https://aclanthology.org/2024.conll-babylm.1/. [14]L. Capone, A. Bondielli, A. Lenci, ConcreteGPT: A baby GPT-2 based on lexical concreteness and cur- riculum learning, in: M. Y. Hu, A. Mueller, C. Ross, A. Williams, T. Linzen, C. Zhuang, L. Choshen, R. Cotterell, A. Warstadt, E. G. Wilcox (Eds.), The 2nd BabyLM Challenge at the 28th Conference on Computational Natural Language Learning, Asso- ciation for Computational Linguistics, Miami, FL, USA, 2024, p. 189–196. URL: https://aclanthology. org/2024.conll-babylm.16/. [15] G. Paulo, A. Mallen, C. Juang, N. Belrose, Auto- matically interpreting millions of features in large language models, arXiv preprint arXiv:2410.13928 (2024). [16]M. Lan, P. Torr, A. Meek, A. Khakzar, D. Krueger, F. Barez, Sparse autoencoders reveal universal fea- ture spaces across large language models, arXiv preprint arXiv:2410.06981 (2024). [17] J. Lindsey, W. Gurnee, E. Ameisen, B. Chen, A. Pearce, N. L. Turner, C. Citro, D. Abrahams, S. Carter, B. Hosmer, J. Marcus, M. Sklar, A. Tem- pleton, T. Bricken, C. McDougall, H. Cunningham, T. Henighan, A. Jermyn, A. Jones, A. Persic, Z. Qi, T. B. Thompson, S. Zimmerman, K. Rivoire, T. Con- erly, C. Olah, J. Batson, On the biology of a large language model,Transformer Circuits Thread (2025). URL: https://transformer-circuits.pub/2025/ attribution-graphs/biology.html. [18] R. Orlando, L. Moroni, P.-L. Huguet Cabot, S. Co- nia, E. Barba, S. Orlandini, G. Fiameni, R. Nav- igli,Minerva LLMs: The first family of large language models trained from scratch on Italian data, in: F. Dell’Orletta, A. Lenci, S. Montemagni, R. Sprugnoli (Eds.), Proceedings of the 10th Italian Conference on Computational Linguistics (CLiC- it 2024), CEUR Workshop Proceedings, Pisa, Italy, 2024, p. 707–719. URL: https://aclanthology.org/ 2024.clicit-1.77/. [19]G. Sarti, M. Nissim, IT5: Text-to-text pretraining for Italian language understanding and generation, in: N. Calzolari, M.-Y. Kan, V. Hoste, A. Lenci, S. Sakti, N. Xue (Eds.), Proceedings of the 2024 Joint In- ternational Conference on Computational Linguis- tics, Language Resources and Evaluation (LREC- COLING 2024), ELRA and ICCL, Torino, Italy, 2024, p. 9422–9433. URL: https://aclanthology.org/2024. lrec-main.823. [20]W. Foundation, Wikimedia downloads, ???? URL: https://dumps.wikimedia.org. [21]A. Makhzani, B. Frey, K-sparse autoencoders, arXiv preprint arXiv:1312.5663 (2013). [22]L. Xue, N. Constant, A. Roberts, M. Kale, R. Al- Rfou, A. Siddhant, A. Barua, C. Raffel,mT5: A massively multilingual pre-trained text-to-text transformer, in: Proceedings of the 2021 Con- ference of the North American Chapter of the Association for Computational Linguistics: Hu- man Language Technologies, Association for Com- putational Linguistics, Online, 2021, p. 483–498. URL: https://aclanthology.org/2021.naacl-main.41. doi:10.18653/v1/2021.naacl-main.41. [23]A. Grattafiori, A. Dubey, A. Jauhri, A. Pandey, A. Ka- dian, A. Al-Dahle, A. Letman, A. Mathur, A. Schel- ten, A. Vaughan, et al., The llama 3 herd of models, arXiv preprint arXiv:2407.21783 (2024). [24] A. Razzhigaev, M. Mikhalchuk, E. Goncharova, I. Os- eledets, D. Dimitrov, A. Kuznetsov, The shape of learning: Anisotropy and intrinsic dimensions in transformer-based models,in: Y. Graham, M. Purver (Eds.), Findings of the Association for Computational Linguistics: EACL 2024, Association for Computational Linguistics, St. Julian’s, Malta, 2024, p. 868–874. URL: https://aclanthology.org/ 2024.findings-eacl.58/. [25]H. Li, T. Lan, Z. Fu, D. Cai, L. Liu, N. Collier, T. Watanabe, Y. Su, Repetition in repetition out: towards understanding neural text degeneration from the data perspective, in: Proceedings of the 37th International Conference on Neural Informa- tion Processing Systems, 2023, p. 72888–72903. [26] A. Team, Almawave presents velvet: The sustain- able and high-performance italian ai, 2025. URL: https://w.almawave.com. [27] P. Basile, E. Musacchio, M. Polignano, L. Siciliani, G. Fiameni, G. Semeraro, Llamantino: Llama 2 mod- els for effective text generation in italian language, 2023.arXiv:2312.09993. [28]F. A. Galatolo, M. G. Cimino, Cerbero-7b: A leap for- ward in language-specific llms through enhanced chat corpus generation and evaluation,arXiv preprint arXiv:2311.15698 (2023). Explainer Prompt (Eng) You are a meticulous AI r e s e a r c h e r conducting an important i n v e s t i g a t i o n i n t o p a t t e r n s found in the I t a l i a n language . Your task i s to analyze t e x t and p ro v id e an e x p l a n a t i o n that thoroughly e n c a p s u l a t e s p o s s i b l e p a t t e r n s found in i t . G u i d e l i n e s : You w i l l be given a l i s t of t e x t examples in I t a l i a n on which s p e c i a l words are s e l e c t e d and between d e l i m i t e r s l i k e << t h i s > >. I f a sequence of c o n s e c u t i v e tokens a l l are important , the e n t i r e sequence of tokens w i l l be contained between d e l i m i t e r s << j u s t l i k e t h i s > >. How important each token i s f o r the behavior i s l i s t e d a f t e r each example in parentheses . −Try to produce a c o n c i s e f i n a l d e s c r i p t i o n . Simply d e s c r i b e the t e x t l a t e n t s that are common in the examples , and what p a t t e r n s you found . −I f the examples are uninformative , you don ’ t need to mention them . Don ’ t focus on g i v i n g examples of important tokens , but t r y to summarize the p a t t e r n s found in the examples . −Do not mention the marker tokens ( < < > >) in your e x p l a n a t i o n . −Do not make l i s t s of p o s s i b l e e x p l a n a t i o n s . Keep your e x p l a n a t i o n s s h o r t and c o n c i s e . −The l a s t l i n e of your response must be the formatted explanation , using [EXPLANATION ] : prompt Explainer Prompt (Ita) S e i un m e t i c o l o s o r i c e r c a t o r e di i n t e l l i g e n z a a r t i f i c i a l e che conduce un ’ importante indagine s u g l i schemi p r e s e n t i n e l l a l i n g u a i t a l i a n a . I l tuo compito e ’ a n a l i z z a r e i l t e s t o e f o r n i r e una s p i e g a z i o n e che racchiuda in modo e s a u r i e n t e i p o s s i b i l i schemi in esso r i s c o n t r a t i . Linee guida : Ti verra ’ f o r n i t o un elen co di esempi di t e s t o in i t a l i a n o in c u i p a r o l e s p e c i a l i sono s e l e z i o n a t e e i n s e r i t e t r a d e l i m i t a t o r i come << questo > >. Se una sequenza di token c o n s e c u t i v i e ’ t u t t a importante , l ’ i n t e r a sequenza di token sara ’ contenuta t r a d e l i m i t a t o r i << p r o p r i o come questo > >. L ’ importanza di c i a s c u n token per i l comportamento e ’ e l e n c a t a dopo ogni esempio t r a p a r e n t e s i . −Cerca di produrre una d e s c r i z i o n e f i n a l e c o n c i s a . D e s c r i v i semplicemente g l i elementi l a t e n t i d e l t e s t o comuni n e g l i esempi e g l i schemi che hai t r o v a t o . −Se g l i esempi non sono i n f o r m a t i v i , non e ’ n e c e s s a r i o m e n z i o n a r l i . Non c o n c e n t r a r t i s u l f o r n i r e esempi di token importanti , ma c e r c a di riassumere g l i schemi t r o v a t i n e g l i esempi . −Non menzionare i token marca tori ( < < > >) n e l l a tua s p i e g a z i o n e . −Non c r e a r e e l e n c h i di p o s s i b i l i s p i e g a z i o n i . Mantieni l e tue s p i e g a z i o n i b r e v i e c o n c i s e . −L ’ ultima r i g a d e l l a tua r i s p o s t a deve e s s e r e l a s p i e g a z i o n e formattata , usando [ SPIEGAZIONE ] : prompt Figure 6:Explainer prompts in English (original, from [15]), and Italian (translated). A. Explainer Prompts In Figure 6 we provide prompts fed to the Explainer model, both in English (original from [15]) and Italian (translation). Declaration on Generative AI During the preparation of this work, the author(s) used ChatGPT (OpenAI) and Grammarly in order to: Paraphrase and reword, Improve writing style, and Grammar and spelling check. After using these tool(s)/service(s), the author(s) reviewed and edited the content as needed and take(s) full responsibility for the publication’s content.