{"id":"fecc1f5c-786f-4d27-8125-03a76c94bc97","arxiv_id":"2506.01074","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"In Llama-based code models, programming languages are represented through an English-like intermediate token space, with language-specific neurons concentrated in bottom layers and exclusive PL neurons in top layers; closely aligned languages such as C# and Java resist neuron identification.","lead":"Researchers probed how Llama-based language models represent seven programming languages and English during code translation tasks. They found that English-like tokens dominate internal intermediate layers, while language-specific neurons appear mostly in early and late layers, with closely aligned languages like C# and Java defying neuron identification.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"English-pivot claim may be inflated by lexical overlap and shared English surface text in the super-parallel dataset; logit-lens classification in §3.1 lacks a control for code-corpus token frequency.","rationale":"The reader's verdict is CONDITIONAL, and the reader's weakest assumption is that keyword-set membership reflects the language the model is thinking in, with lexical overlap potentially inflating English-token probabilities. My stress-test reaches the same conclusion and identifies it as the single most load-bearing concern because the abstract's first major claim—'concept space is closer to English'—is directly supported by Figure 2, which uses exactly this classification. The concern is concrete: the keyword sets are built in §2.1 from overlapping vocabularies, and the super-parallel dataset shares English comments and identifiers across all seven PLs, creating a surface-level confound that the paper's 'English (Excl. PL Keywords)' curve does not fully control for. The proposed check—rerunning the logit-lens analysis with comments removed, identifiers neutralized, and token frequencies matched—would settle whether the English pivot is a genuine representational phenomenon or an artifact of data and lexical overlap. Because this control is currently missing, the conditional verdict is appropriate; the paper's other contributions, especially the neuron-layer distribution and the C#/Java difficulty, are plausible and would not be invalidated by this test, but the headline English-centric finding requires the additional evidence before the paper should be accepted unconditionally.","tokens_in":16450,"tokens_out":5491,"duration_ms":65158,"concrete_test":"Recompute the Figure 2 probability/rank curves after preprocessing the 581 super-parallel snippets to (a) strip all comments, (b) replace variable names and string literals with neutral placeholders such as var0/var1, and (c) exclude any English token whose frequency in the GitHub-Code English subset (§2.1) exceeds a matched threshold. If the second-half rise in the 'English (Excl. PL Keywords)' curve persists, the English-pivot claim survives; if it flattens or disappears, the effect is driven by surface overlap and shared English text in the dataset.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that the model's concept space is closer to English rests on logit-lens decoding (§2.3, Figure 2), where the top-10 decoded tokens at each layer are classified as English or PL by keyword-set membership (§2.1). This classification does not establish that the model 'thinks in English.' Many PL keywords are English words (if, for, while, return), so the English and PL categories overlap by construction. Moreover, the super-parallel snippets (§2.1) share English comments and variable names across all seven PLs, so English tokens are present in the input surface itself. The high English-token probabilities in the second half of layers could therefore reflect (i) the unembedding prior over frequent tokens in code-heavy training and (ii) the model tracking the shared English text, rather than a representational pivot. The 'English (Excl. PL Keywords)' curve in Figure 2 removes only the exact keyword overlap; it does not remove English tokens that are common in comments or identifiers, nor does it frequency-normalize against code-corpus token distributions. Since the abstract's first finding is 'concept space is closer to English,' this confound is load-bearing. The Limitations section concedes that a precise English-to-PL token dictionary is missing, so the present evidence cannot distinguish genuine English-centric representation from lexical artifacts. The neuron-level findings are less affected by this confound, but the paper's headline result about an English pivot is directly at risk.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies how two Llama-family models (CodeLlama 7B and Llama 3.1 8B) represent programming languages (PLs) and English in their intermediate layers. Using a super-parallel dataset of code snippets in seven PLs, the authors perform few-shot translation between 21 PL pairs and apply logit lens to measure how often English and PL keyword tokens appear among top decoded tokens at each layer. They report that English tokens, including PL keywords, dominate intermediate representations in the second half of the layers. They also apply MEXA to measure cross-lingual alignment among PLs and LAPE to identify language-specific and exclusive neurons, finding that exclusive neurons tend to appear in top layers and that highly aligned PLs such as C# and Java make language-specific neuron identification difficult. The paper concludes with implications for multilingual code model design and releases code.","tokens_in":16664,"tokens_out":5620,"duration_ms":56001,"significance":"If the findings hold, this is a useful extension of the 'English pivot' hypothesis to the code domain and a first systematic neuron-level analysis of multiple programming languages. The paper leverages established interpretability methods, uses two models, builds a parallel dataset across seven PLs, and makes code available. The observed difference between bottom-layer shared neurons and top-layer exclusive neurons, and the association between cross-lingual alignment and the difficulty of isolating language-specific neurons, are potentially valuable insights. However, the headline claims currently rest on a keyword-based classification prone to lexical confounds and on aggregate curves without statistical support, so the results should be treated with caution until these issues are addressed.","major_comments":[{"comment":"The classification of decoded tokens into 'English' versus 'PL' categories is based on keyword-set membership, and many PL keywords are ordinary English words (e.g., if, for, while, return). The 'English (Excl. PL Keywords)' curve removes only the exact keyword overlap; it does not remove English words that appear in comments or identifiers, which the super-parallel dataset shares across all seven PLs by construction (§2.1). Without a frequency-normalized control against code-corpus token distributions or a control using a non-English natural language, the high English-token probabilities in the second half of layers could reflect lexical overlap and the input surface text rather than an English-centric concept space. The Limitations section explicitly concedes that a precise English-to-PL token dictionary is missing, so this confound is load-bearing for the paper's central claim.","section":"§2.1, §2.3, Figure 2"},{"comment":"No error bars, confidence intervals, or significance tests are reported for any empirical quantity in the paper. All conclusions are based on aggregated curves over 581 snippets and 42 translation directions; comparisons such as 'English keywords maintain a high and similar 1/rank value of 0.4' or 'C# achieves the best alignment overall' are presented without variance or significance information. This makes it impossible to assess whether the reported differences between conditions are robust, which is a load-bearing issue for the central empirical claims.","section":"§3.1–§3.3, Figures 2–7"},{"comment":"The claim that 'neurons exclusive to each PL tend to appear in the top layers' is not quantitatively established. The text states that 'most language-specific neurons ... are selected from the bottom layers' while 'those that are exclusive to a specific PL are predominantly selected from the top layers,' but Figure 6 is described as showing only exclusive neurons, making the two statements difficult to reconcile. No per-layer counts or statistical comparison of exclusive versus shared neuron distributions are provided, so the abstract's second major finding is not supported by the evidence as presented.","section":"§3.3, Figure 6"},{"comment":"The conclusion that 'identifying language-specific neurons is not feasible' for C# and Java is stronger than the evidence warrants. The results show that with the LAPE method and the chosen hyperparameters (ν=400, τ=0.95), the PPL gap between primary and other languages is smaller for these PLs. This does not demonstrate infeasibility in general; it only indicates that this particular method and setting fail to isolate effective neurons. The claim should be restricted to LAPE under these hyperparameters, or additional identification methods should be tested.","section":"§3.3"},{"comment":"The MEXA alignment scores are computed on the same super-parallel snippets that, as stated in §2.1, share variable names and English comments across all seven PLs. The high alignment of C# and Java with many other languages may therefore be inflated by surface-level lexical overlap rather than representational alignment. This also weakens the subsequent link between alignment and the difficulty of neuron identification. A control that removes comments or uses renamed variables would help ensure the alignment scores reflect semantic alignment rather than shared surface tokens.","section":"§2.4, Figure 3"}],"minor_comments":[{"comment":"The definition of the 'probability' and '1/rank' curves is ambiguous: it is not stated whether the values are sums over all tokens in a keyword category or the maximum over those tokens. This should be clarified for reproducibility.","section":"§2.3, Figure 2"},{"comment":"The table uses the symbol '✓?' without explanation, making the row for non-English/PL neuron identification unclear.","section":"Table 1"},{"comment":"The dataset name appears inconsistently as both 'GeeksForGeeks' and 'GeeksForGeeks'; the spelling should be unified.","section":"§2.1"},{"comment":"The text contains a typo: 'rougly' should be 'roughly'.","section":"§3.1"},{"comment":"The heading 'Acknowlegments' is misspelled and should be 'Acknowledgments'.","section":"Acknowledgements"},{"comment":"The FFN equation is missing explicit definitions for the subscript notation in φ(h̃^(ℓ)W1^(ℓ))_r; including these would improve readability.","section":"§2.5"}],"recommendation":"major_revision","confidential_remarks":"The paper's novelty is moderate: it applies existing interpretability tools to a new domain (multiple programming languages) rather than introducing a new method. The central risk is the keyword-based confound in the English-pivot claim; if the authors add control analyses (e.g., removing comments from prompts, frequency-normalizing token probabilities, or comparing against a non-English natural-language pivot), the paper would be substantially stronger. The lack of error bars and significance testing is a general weakness that should be addressed before publication. I would not recommend rejection because the underlying questions are timely and the neuron-level findings, if properly quantified, could be a real contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Things to know: this is the first paper to map concept space and neuron sharing across multiple programming languages, and the top-layer concentration of exclusive PL neurons is the solid part. The English-pivot finding is interesting but the evidence is softer than the abstract suggests.\n\nThe novelty is real. Prior work looked at Python-only (Wu et al.) or monolingual neurons (Haider et al.); here they run logit lens over 21 PL pairs, MEXA over 7 PLs, and LAPE over 11 PLs + English, on two Llama models. The pattern that language-specific neurons sit in bottom layers while exclusive PL neurons concentrate in top layers is new and worth building on. The C#/Java case, where identification of language-specific neurons is hard, fits with the alignment measures and is a nice observation.\n\nMethod choice is sound: logit lens rather than tuned lens, LAPE with threshold sensitivity acknowledged, code link provided. The paper is honest about its limitations, including the missing keyword dictionary.\n\nSoft spots. The English-pivot claim rests on classifying decoded tokens by keyword lists. Since many PL keywords are English words and the super-parallel snippets share English comments and variable names, the high English probabilities in the middle layers could partly reflect the model tracking shared surface text or the unembedding prior. The 'English (Excl. PL Keywords)' curve removes one confound but not comments, identifiers, or token frequency. That said, the layer-wise dynamics—English keywords rising in the second half then falling as the expected token takes over—are not trivially explained by surface text alone, so I wouldn't call the claim dead. It needs a control with non-English comments or frequency-matched baselines.\n\nThere are no error bars or significance tests anywhere. The 581-snippet super-parallel set is small, and the 'not feasible' for C#/Java is an overstatement of a smaller PPL gap. Those are fixable in revision.\n\nNet: the neuron findings are the most robust contribution; the English-pivot finding is plausible but under-supported. The paper deserves a serious referee; a good referee will ask for controls on the token classification and some measure of variance.","headline":"First cross-PL map of concept space and neuron sharing; the neuron-layer findings are solid, the English-pivot claim needs tighter controls.","tokens_in":17300,"tokens_out":3008,"would_cite":true,"duration_ms":31344,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Llama code models translate programming languages by passing through an English-like concept space, with language-exclusive neurons only near the output.","keywords":["code language models","programming languages","multilingual code translation","logit lens","language-specific neurons","cross-lingual alignment","concept space","model interpretability"],"falsifier":"Translate between two programming languages whose keywords share no English words, or remove all tokens that are simultaneously English words and PL keywords from the keyword sets, and check whether the intermediate-layer English probability peak persists. If it disappears, the English pivot is an artifact of keyword overlap; if it remains for non-overlapping keywords, the concept-space claim is supported.","tokens_in":16189,"feed_emoji":"🧠","tokens_out":9609,"duration_ms":83035,"temperature":0.7,"pith_summary":"This paper tries to establish that programming languages are not stored as isolated systems inside Llama-based code models. During few-shot translation between languages such as Java and Rust, the model's intermediate layers first assign high probability to English tokens and programming-language keywords before committing to the output language, suggesting an English-like conceptual hub. It also claims that neurons in these models follow a layered division of labor: neurons that respond to one language are concentrated in the bottom layers, while neurons exclusive to a single programming language appear in the top layers. For closely aligned languages such as C# and Java, no effective language-specific neurons can be found, because their representations are shared too broadly. If these claims hold, multilingual code generation can be understood as routing through a shared concept space, which has direct consequences for how code models are adapted, compressed, and interpreted.","feed_headline":"English is the hidden pivot for code translation in Llama models","feed_subtitle":"Intermediate layers favor English tokens, and language-exclusive neurons sit only near the output, explaining how code concepts are shared.","key_machinery":"The main instrument is the logit lens: at an intermediate layer, the model's unembedding matrix is applied to the hidden state to turn it into token probabilities, revealing which languages' tokens are 'thinkable' before the final prediction. For neuron analysis, the paper uses language activation probability entropy (LAPE), which labels a neuron as language-specific when its activation probability concentrates on one or two languages, and then deactivates those neurons to measure perplexity changes. A third method, MEXA, computes cross-lingual alignment by counting how often parallel code snippets are nearest neighbours in embedding space, identifying which languages sit close to the central concept space.","core_discovery":"The central claim is that in CodeLlama 7B and Llama 3.1 8B, the internal 'concept space' for programming is English-centric. When the authors decode intermediate-layer embeddings with the logit lens during a few-shot translation task over 21 language pairs, token probabilities and rank scores for English and programming-language keywords rise sharply in the second half of the layers, overtaking the expected output token before the final layer reasserts the correct target token. This indicates that the model passes through an English-like representation even when both the source and target are programming languages. A second claim concerns neuron organization: applying an entropy-based selector to neuron activations over eleven programming languages and English, the authors find that language-specific neurons cluster in the bottom layers, while neurons exclusive to a single programming language concentrate in the top layers near the output. For highly aligned languages such as C# and Java, deactivating their 'specific' neurons also degrades other languages, so their language-specific neurons are not identifiable.","pith_inferences":["If the English pivot appears during code translation, then generating code from non-English natural-language comments may also detour through English, meaning code models could inherit the same English-centric behavior documented for natural languages; the paper does not test this directly.","Because many programming keywords are ordinary English words, the high English-token probabilities may partly reflect lexical overlap rather than a true conceptual pivot; an experiment with non-overlapping keyword sets would separate these explanations.","The layer distribution of language-exclusive neurons suggests a concrete recipe for continual learning: freeze the bottom layers and add only top-layer neurons when teaching a model a new programming language; this is an architectural proposal implied but not stated by the paper.","The alignment results predict that the centrality of a programming language in concept space grows with the size of its keyword set and its prevalence in pretraining, so a synthetic language with a deliberately small keyword set should appear peripheral; this is testable with the paper's methods."],"forward_implications":["Because English and a few prominent programming languages like C++ and C# sit centrally in the concept space, multilingual code translation could be improved by routing through these hubs, minimizing the distance between source and target languages.","The layered neuron pattern—shared and general neurons in bottom layers, language-exclusive neurons in top layers—suggests architectures that share bottom layers across languages and swap only top layers could support new programming languages with minimal added parameters.","For closely aligned languages such as Java and C#, shared representations make parameter sharing or adapter-based multilingual support practical, since the model does not need separate neurons for each.","The failure to find language-specific neurons for C# and Java means interpretability tools that locate language-specific neurons in natural languages do not transfer cleanly to code languages.","The difference between CodeLlama 7B and Llama 3.1 8B implies that a code-heavy training recipe produces more shared, less language-specific neurons, so training data composition directly controls how separable code languages are internally."],"supporting_citations":[{"why":"Introduces the logit lens technique used to decode intermediate-layer embeddings into token probabilities.","marker":"Nostalgebraist, 2020"},{"why":"Shows that English acts as a pivot language in English-centric LLMs, the pattern this paper extends to programming languages.","marker":"Wendler et al., 2024"},{"why":"Provides the language activation probability entropy (LAPE) method used to identify language-specific neurons.","marker":"Tang et al., 2024"},{"why":"Supplies the MEXA cross-lingual alignment score used to measure how closely programming languages align in intermediate layers.","marker":"Kargaran et al., 2024"},{"why":"Supplies the super-parallel multilingual code dataset used for the few-shot translation task.","marker":"Zhu et al., 2022b"},{"why":"Provides the programming-language keyword lists used to classify decoded tokens.","marker":"Meyer and McCulloch, 2022"},{"why":"Provides the PanLex English lexicon used to classify English tokens.","marker":"Kamholz et al., 2014"},{"why":"Supplies the raw GitHub code dataset used for neuron activation analysis across eleven programming languages.","marker":"CodeParrot, 2022"}],"fun_headline_variants":["English is the hidden pivot for code translation in LLMs","Code LLMs use English as internal pivot, neurons reveal PLs near output","English-centric concept space found in code LLMs; PL neurons near top","LLMs translate code via English, language-specific neurons near output","Code LLMs think in English; unique neurons per language near output"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a token's membership in a keyword list tells you which language the model is 'thinking in,' but many programming-language keywords are ordinary English words, so the English-token signal could be inflated by lexical overlap.","fun_headline_variants_meta":{"raw":{"variants":["English is the hidden pivot for code translation in LLMs","Code LLMs use English as internal pivot, neurons reveal PLs near output","English-centric concept space found in code LLMs; PL neurons near top","LLMs translate code via English, language-specific neurons near output","Code LLMs think in English; unique neurons per language near output"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001056,"raw_usage":{"total_tokens":4448,"prompt_tokens":977,"completion_tokens":3471,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":593,"completion_tokens_details":{"reasoning_tokens":3380}},"tokens_in":593,"tokens_out":3471,"duration_ms":22995,"temperature":1.0,"reasoning_tokens":3380,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T11:51:44.591712+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Translate between two programming languages whose keywords share no English words, or remove all tokens that are simultaneously English words and PL keywords from the keyword sets, and check whether the intermediate-layer English probability peak persists. If it disappears, the English pivot is an artifact of keyword overlap; if it remains for non-overlapping keywords, the concept-space claim is supported.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the logit lens technique used to decode intermediate-layer embeddings into token probabilities."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the programming-language keyword lists used to classify decoded tokens."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the PanLex English lexicon used to classify English tokens."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the raw GitHub code dataset used for neuron activation analysis across eleven programming languages."}],"review_version":1}