{"id":"2c4c0257-7960-48ae-b4db-cd1bd622538b","arxiv_id":"2507.18918","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"Sparse-autoencoder analysis of Gemma-2-2B shows that medium-to-low resource languages get up to 26% lower activations than English, and LoRA fine-tuning that explicitly minimizes the activation gap raises activations but does not consistently improve benchmark scores.","lead":"This paper measures neuron activation levels in a small multilingual language model across ten languages and finds that lower-resource languages such as Malayalam and Marathi activate far less than English in many layers. It then fine-tunes the model to make those activations look more like English and reports larger activations, but mixed and mostly small changes on standard benchmarks.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The English-selected SAE features used in §2.4 are not validated as language-neutral probes; without a per-language feature-selection control, the reported activation gaps and the fine-tuning results do not establish the central claim.","rationale":"The reader's rejection is justified by the unvalidated measurement and the inconsistent follow-up benchmark evidence. I focused on the measurement because it precedes everything else: if the SAE feature indices are English-specific, both the descriptive disparity and the fine-tuning target are defined relative to an English-biased dictionary, so the paper's central claim loses its empirical foundation. The paper's own Appendix A.1 acknowledges SAE features can be composed and possibly language-specific; the rebuttal there is an argument from the very data whose validity is in question. Appendix C.6 is decisive on the causal side: only one of three Malayalam benchmarks improves, and the abstract's phrase 'modest but consistent improvements' is contradicted by the appendix. A per-language feature-selection control would settle the issue inexpensively and should be a precondition for any claim about activation alignment. No ad hominem is intended; the issue is the evidence chain.","tokens_in":11167,"tokens_out":3935,"duration_ms":41860,"concrete_test":"Select features independently for each target language: for a random sample of 1,000 Gemma Scope features, compute each language's own top-activating phrases using the same 80%-of-max threshold, and measure mean activations at layers 0–25. If the high-/low-resource gap disappears or shrinks substantially, the English-selected indices were the source. As a control, also back-translate the top native phrases to English and check whether English activation on those same features is also low; if it is, the feature is a mismatch, not evidence of resource-level under-activation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that activation disparities are systematic and causally relevant. The load-bearing premise is the measurement: features whose indices were selected from English phrases (Section 2.1) are treated as language-neutral probes when activations of translated phrases are read (Section 2.4). This is not established. An SAE feature direction is fit to English co-occurrence statistics; a translated sentence can have low activation on that index simply because the feature's dictionary direction does not match the translated content, not because the language is under-represented. The paper provides no per-language feature selection, no control using each language's own top-activating phrases, and no translation-invariance check. Appendix A.1 even concedes SAEs learn composed features that may be language-specific, and Section 4 concedes Helsinki-NLP translation errors 'may introduce bias.' Thus the 26.27%/19.89% gaps and the r ≈ -0.95 correlations (Figures 9–11) could be artifacts of probe selection. The fine-tuning loss directly minimizes |u_l - v_l| on the same English-derived features, so the reported activation gains are the optimized objective, not independent confirmation. The only independent evidence is benchmark changes, and Appendix C.6 shows Malayalam declined on MMLU (-3.95) and HellaSwag (-2.41) while gaining 1.44 on ARC-C. Hence the abstract's 'modest but consistent improvements' and the conclusion that activation alignment is key are not supported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper investigates cross-linguistic activation disparities in Gemma-2-2B using the Gemma Scope sparse autoencoders. English phrases activating each SAE feature above 80% of its maximum are extracted from Neuronpedia and machine-translated into nine languages; activations of the translations are then read at the same English-selected feature indices across all 26 layers (§2.1–§2.5). The authors report that medium-to-low-resource languages receive up to 26.27% lower activations at layer 6 and a persisting 19.89% gap at layer 25 (Table 1), that embedding cosine similarity is high across languages despite large benchmark gaps on ARC-C, MMLU, and HellaSwag, and that the activation gap correlates strongly with benchmark accuracy (r ≈ -0.93 to -0.95; Figures 9–11). A LoRA fine-tuning procedure with the loss f(u_l, v_l) = |u_l - v_l| + α||u_l - u_orig,l||² targeted at layer 20 is then reported to yield activation gains of up to 87.69% (Malayalam) with roughly 91% English activation retention, and the abstract claims 'modest but consistent' benchmark improvements. The central claim is that SAE activation values can diagnose and partially correct cross-linguistic disparities, with activation alignment as a key factor in multilingual performance.","tokens_in":11450,"tokens_out":16286,"duration_ms":149802,"significance":"If the measured disparities were validated, the paper's descriptive finding—that embedding-level similarity coexists with large SAE-activation gaps and large benchmark gaps—would be a useful contribution to multilingual interpretability, and the openly documented, layer-wise activation tables (Table 1) would support follow-up work. The paper makes good use of public resources (Gemma Scope, Neuronpedia, OPUS-MT) and reports full per-language benchmark and activation numbers, and it deserves credit for explicitly documenting in Appendix C.6 that post-fine-tuning benchmark results are inconsistent rather than suppressing them. The falsifiable core—that embedding similarity does not guarantee equal task performance in multilingual LLMs—is clearly and testably stated. However, the significance claimed in the abstract is not currently established: the activation measurement rests on unvalidated English-derived probes, the headline activation gains are the direct optimum of the fine-tuning loss, and the abstract's 'consistent improvements' claim is contradicted by the paper's own Table 9.","major_comments":[{"comment":"The activation measurements assume that feature indices selected from English phrases (top-activated phrases above 80% of each feature's maximum, §2.1) are language-neutral probes when activations of machine-translated phrases are read at those same indices (§2.4). The paper supplies no per-language feature-selection control, no translation-invariance check, and no analysis using each language's own top-activating phrases. Appendix A.1 explicitly concedes that SAEs learn composed features that may be language-specific, and Section 4 concedes that Helsinki-NLP translation errors may introduce bias in cross-lingual analysis. Under these conditions, the headline gaps (26.27% at layer 6 and 19.89% at layer 25 in Table 1) and the correlations in Figures 9–11 could reflect a mismatch between the English-derived probe directions and the translated content rather than a systematic language-level activation deficit. Because the activation-gap measurement is the paper's central diagnostic contribution, a per-language validation of the probes is required before the gaps can be interpreted as evidence of under-representation.","section":"§2.1, §2.4, Appendix A.1"},{"comment":"The reported activation gains after fine-tuning are the direct optimum of the training objective, so they cannot serve as independent evidence for the paper's intervention claim. The loss f(u_l, v_l) = |u_l - v_l| + α||u_l - u_orig,l||² in §2.6 is exactly the activation gap that Table 5 and Figure 4 report as the main fine-tuning result; with target layer 20 and fine-tuned layers 0–20 (Appendix B.1), the 87.69% Malayalam and 86.32% Hindi gains at layer 20 measure how well gradient descent minimized the training loss, and the fine-tuning dataset is the same as the analysis dataset described in §2.1. These numbers therefore confirm the optimizer's behavior, not the value of activation alignment. The paper should report held-out activation alignment (on phrases and layers not used in training) and should present any layer-20 gain as an optimization sanity check rather than as a discovered effect.","section":"§2.6, Appendix B.1, §3.3, Table 5"},{"comment":"The abstract's claim of 'modest but consistent improvements' is contradicted by the paper's own data. For Malayalam, Table 9 shows +1.44 points on ARC-C but -3.95 points on MMLU and -2.41 points on HellaSwag after fine-tuning, so two of three benchmarks decline and the improvements are not consistent. Appendix C.6 acknowledges this inconsistency and Section 5 states that the effect on task performance 'remains limited,' but the abstract and the §3.3 framing ('demonstrating the positive impact of activation alignment on downstream tasks') do not reflect that qualifier. In addition, the pre-fine-tuning MMLU score for Malayalam is 29.44% in Table 4 but 29.05% in Table 9, an internal inconsistency in the only table that supports the causal claim. As written, the evidence does not support the statement that activation alignment is a key factor in enhancing multilingual performance.","section":"Abstract, §3.3, Table 9, Appendix C.6"},{"comment":"The correlations between activation gap and benchmark accuracy (r ≈ -0.95, -0.93, and -0.95 in Figures 9–11) are computed over nine or ten languages and are reported without confidence intervals, p-values, or any check that the relationship survives control for the probe-selection issue in §2.1–§2.4. Because the same English-selected probe set determines the activation values and the language grouping, these correlations cannot be distinguished from a selection artifact, and the very high r-values are implausible for such small samples unless the underlying variable is nearly a monotone transform of benchmark difficulty. The authors should report uncertainty quantification and a per-language feature-control analysis before using these correlations as evidence of a systematic link between activation magnitude and performance.","section":"§3.2, Figures 9–11"}],"minor_comments":[{"comment":"Tables 4 and 9 report the pre-fine-tuning Malayalam MMLU accuracy as 29.44% and 29.05%, respectively; the discrepancy should be resolved.","section":"Tables 4 and 9"},{"comment":"Section 2.2 samples 1,000 feature indices (indices 16i for i = 0..999), while the caption of Figure 7 refers to 4,000 feature indices; the sampling description should be reconciled.","section":"§2.2, Figure 7"},{"comment":"Sections 2.5 and 2.6 do not say whether A_l denotes SAE latent activations or residual activations, whether values are mean-pooled over tokens, or how A_high and A_med-low are aggregated over feature indices; the measurement definitions should be stated explicitly for reproducibility.","section":"§2.5, §2.6"},{"comment":"Appendix B.1 reports 4,000 fine-tuning samples, two iterations, and α = 1.00 but gives no learning rate, LoRA rank, or number of seeds; given that the benchmark deltas are on the order of 1–4 points, standard errors or multiple runs are needed to rule out noise.","section":"Appendix B.1"},{"comment":"Okapi is cited via Lai et al. (2023), which describes instruction-tuned models; the reference does not clearly identify an 'evaluation framework,' so the citation should be corrected or replaced with the actual benchmark source.","section":"References, §2.3"},{"comment":"Table 5's caption says 'across ten languages' but the table lists nine non-English languages, and Section 3.3 claims gains 'across all ten languages' while reporting English retention separately; the counting and the meaning of 'retention' should be clarified.","section":"Table 5, §3.3"}],"recommendation":"reject","confidential_remarks":"The abstract overstates the paper's findings relative to its own appendix: the 'consistent improvements' claim in the abstract conflicts with Table 9, and the headline activation gains are the optimized training objective. The authors' careful conclusion in Section 5 is closer to the evidence, and a resubmission would need to bring the abstract in line with it and add the per-language probe validation. The contribution is also closer to a negative result (activation alignment alone does not reliably improve multilingual benchmarks) than the framing suggests, so the novelty claim should be recalibrated. I would also note that the paper's self-reported limitations in Sections 4 and A.1 directly undermine the central measurement, which is unusual and should be weighed carefully."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things worth knowing about arXiv:2507.18918. First, the descriptive core—layer-wise SAE activation ratios for ten languages on Gemma-2-2B—is new and worth a look. Second, the causal claim built on top of it does not hold; the fine-tuning evidence is circular and the benchmark evidence contradicts the abstract.\n\nWhat the paper does well: it uses Gemma Scope properly, reports per-layer activation gaps (peak 26.27% at layer 6, persistent 19.89% at layer 25), and shows embedding similarity does not predict benchmark performance. Those tables and figures are useful. The appendix is also honest about translation noise, the single-model scope, and the fact that benchmark improvements were mixed.\n\nWhere it falls down. The measurement probe is the biggest issue. Features are selected from English phrases (Section 2.1) and then applied to translations without any per-language feature-selection control or translation-invariance check. A low activation on an English-derived feature index could simply mean the feature direction doesn't match the translated content. That makes the 26% gap, and the r ≈ -0.95 correlations, potentially artifacts of probe selection. Second, the fine-tuning loss directly minimizes |u_l - v_l| on those same features, so the reported 87.69% activation increase for Malayalam is the optimizer converging, not a finding. Third, the abstract says 'modest but consistent improvements,' but Appendix C.6 shows Malayalam dropped 3.95 points on MMLU and 2.41 on HellaSwag while gaining 1.44 on ARC-C. That is inconsistent, and it undercuts the 'activation alignment is key' conclusion. Minor stuff: the GitHub link is missing, LoRA hyperparameters are thin (2 iterations, 4,000 samples, layer 20 only), and the feature-index count is inconsistent—§2.2 says 1,000 sampled indices, Figure 7 says 4,000.\n\nThis is not a reliable demonstration of its central claim as written. But the descriptive measurements could become useful if the probe validity is fixed. If the authors added controls—per-language feature selection, random feature baselines, translation invariance checks—the paper would be worth another look.\n\nMy recommendation: if this lands on your desk, send it to referees only if you expect them to demand those controls. The topic is timely and the tooling is appropriate, but this version needs major revision before the conclusions can be trusted.","headline":"Descriptive SAE comparison is new and worth a look, but the causal claim collapses under probe-selection circularity and benchmark evidence that contradicts the abstract.","tokens_in":12027,"tokens_out":4228,"would_cite":false,"duration_ms":42580,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that sparse-autoencoder features reveal systematic under-activation of medium-to-low resource languages in Gemma-2-2B, and that fine-tuning to align those activations with English reduces the gap but only modestly…","keywords":["sparse autoencoders","multilingual LLMs","activation disparities","low-resource languages","activation-aware fine-tuning","LoRA","Gemma-2-2B","cross-lingual benchmarks"],"falsifier":"Select SAE features independently from non-English phrases using the same 80% activation threshold and measure activation of English translations on those indices. If the cross-linguistic activation gap vanishes or becomes symmetric, the reported disparity is an artifact of English-anchored feature selection rather than a model-level representational deficit. A second check: fine-tune with the activation-alignment loss but freeze all English-retention terms and evaluate MMLU; if MMLU still declines, the loss itself, not knowledge drift, is causing the drop.","tokens_in":10919,"feed_emoji":"🌐","tokens_out":6752,"duration_ms":60430,"temperature":0.7,"pith_summary":"This paper tries to establish that multilingual language models can look language-neutral at the embedding level while still being systematically weaker for medium-to-low resource languages inside the model. Using sparse-autoencoder features of Gemma-2-2B across all 26 layers and 10 languages, it reports that non-English languages receive lower feature activations, up to 26.27% lower in early layers and a persistent gap near 20% in deep layers, and that these activation gaps track benchmark gaps on ARC-Challenge, MMLU, and HellaSwag. It then shows that fine-tuning with an activation-alignment loss raises activations for under-resourced languages by large percentages while keeping English activations mostly intact, but downstream accuracy improves only modestly and unevenly. If the paper is right, SAE activation values are a practical diagnostic for cross-linguistic inequality and a target for partial correction, not a guarantee of task-level gains.","feed_headline":"Low-resource languages fire LLM neurons up to 26% less, study finds","feed_subtitle":"Activation probes find the gap in every layer; alignment fine-tuning narrows it but task gains stay modest.","key_machinery":"The load-bearing object is the sparse autoencoder (SAE) feature: a decomposition that turns each layer's dense activation vector into a small set of sparse, interpretable unit features. The paper anchors each feature using English phrases whose activation exceeds 80% of the feature's maximum, then measures how strongly the same feature fires on translations of those phrases in nine languages. The quantity doing the work is the per-layer activation gap, the percentage by which the mean activation for medium-to-low resource languages falls below that for high-resource languages; the same gap feeds the fine-tuning loss, which drives target-language activations toward English activations under a stability regularizer.","core_discovery":"On the paper's own terms, the central discovery is that Gemma-2-2B systematically activates learned features less for medium-to-low resource languages than for English and other high-resource languages, despite near-equal residual embeddings. The gap peaks at layer 6 with 26.27% lower activation, narrows but never closes deeper in the network at 19.89% at layer 25, and is strongly negatively correlated with benchmark accuracy, with Pearson coefficients around -0.93 to -0.95 across ARC-Challenge, HellaSwag, and MMLU. Fine-tuning with a LoRA objective that penalizes the absolute difference between English and target-language activations while adding a stability term for original English activations produced activation gains of 87.69% for Malayalam and 86.32% for Hindi with roughly 91% English retention. The downstream effect was modest and inconsistent: ARC-Challenge for Malayalam rose 1.44 percentage points, while MMLU and HellaSwag declined, so the paper claims activation alignment is a key factor but not a sufficient one.","pith_inferences":["If the activation gap is causal rather than merely correlated, then fine-tuning only the early layers where the gap is largest should produce larger benchmark gains than aligning all layers, at lower cost; this is directly testable.","The same English-anchored SAE-probe protocol could be run on other model families; comparing gap sizes would show whether the disparity is architecture-driven or data-driven, without needing new benchmark suites.","A stricter test of measurement validity would select features independently from each language and compare feature sets; if the same features are not activated across languages, the gap is partly a probe artifact rather than pure under-representation.","The MMLU decline after alignment suggests that forcing activation toward English may overwrite factual knowledge; a knowledge-retention term in the loss could separate alignment effects from forgetting effects."],"forward_implications":["SAE activation magnitude can be used as a layer-wise diagnostic: a 20% or larger activation gap flags a language that is likely to underperform, before running expensive benchmarks.","Activation-aware LoRA fine-tuning is a viable cheap intervention: it raised Malayalam activations by 87.69% and Hindi by 86.32% while keeping English activations at about 91%, showing the gap is partially movable.","High embedding similarity is not sufficient evidence of equitable multilingual ability; activation-level measurements reveal disparities that cosine similarity hides.","Since deeper-layer gaps persist near 20%, shared representations do not automatically erase language-specific deficits; targeted early-layer interventions may be needed.","Activation alignment alone does not guarantee task gains: MMLU and HellaSwag declined for Malayalam, so the goal should be alignment plus knowledge preservation."],"supporting_citations":[{"why":"Supplies the open SAE decompositions of Gemma-2-2B that all layer-wise activation measurements rely on.","marker":"(Lieberum et al., 2024)"},{"why":"Provides the 16,384 interpretable neuron features and English activation records used to anchor the probes.","marker":"(Joseph Bloom and Chanin, 2024)"},{"why":"Provides the OPUS-MT translation models used to render English phrases into the nine target languages.","marker":"(Tiedemann and Thottingal, 2020)"},{"why":"Extends the OPUS-MT translation pipeline used for the medium-to-low-resource languages.","marker":"(Tiedemann et al., 2023)"},{"why":"Defines the Okapi evaluation framework and benchmark setup whose scores the activation gaps are compared against.","marker":"(Lai et al., 2023)"},{"why":"Introduces the Gemma-2-2B model whose 26 residual layers are the subject of the study.","marker":"(Riviere et al., 2024)"},{"why":"Supplies the LoRA fine-tuning method used for activation-aware adaptation.","marker":"(Hu et al., 2021)"},{"why":"Establishes sparse autoencoders as a tool for extracting interpretable neuron-level features.","marker":"(Cunningham et al., 2023)"},{"why":"Provides the cross-lingual representation-learning background that connects shared representations to activation behavior.","marker":"(Conneau et al., 2020)"}],"fun_headline_variants":["Sparse autoencoders reveal 26% activation gap in low-resource LLMs","LLMs under-activate for low-resource languages by up to 26%","Activation-aware LoRA lifts low-resource LLM neurons 87%, but tasks lag","Multilingual LLM parity: activation alignment helps but not enough","How sparse autoencoders pinpoint LLM language activation disparities"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument assumes that features selected from English phrases are language-neutral sensors, so a lower activation on a translated phrase means the model genuinely represents that content more weakly, rather than the feature simply matching English wording better.","fun_headline_variants_meta":{"raw":{"variants":["Sparse autoencoders reveal 26% activation gap in low-resource LLMs","LLMs under-activate for low-resource languages by up to 26%","Activation-aware LoRA lifts low-resource LLM neurons 87%, but tasks lag","Multilingual LLM parity: activation alignment helps but not enough","How sparse autoencoders pinpoint LLM language activation disparities"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00058,"raw_usage":{"total_tokens":2760,"prompt_tokens":1001,"completion_tokens":1759,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":617,"completion_tokens_details":{"reasoning_tokens":1659}},"tokens_in":617,"tokens_out":1759,"duration_ms":15591,"temperature":1.0,"reasoning_tokens":1659,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T18:05:41.198940+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Select SAE features independently from non-English phrases using the same 80% activation threshold and measure activation of English translations on those indices. If the cross-linguistic activation gap vanishes or becomes symmetric, the reported disparity is an artifact of English-anchored feature selection rather than a model-level representational deficit. A second check: fine-tune with the activation-alignment loss but freeze all English-retention terms and evaluate MMLU; if MMLU still declines, the loss itself, not knowledge drift, is causing the drop.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the OPUS-MT translation models used to render English phrases into the nine target languages."}],"review_version":1}