{"id":"5723a701-0a5f-44ad-8fe4-076160511c65","arxiv_id":"2505.07293","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"AttentionInfluence uses retrieval-head masking loss differences in a 1.3B model to select 73B tokens that improve a 7B model's reasoning benchmarks by 1.4 to 3.5pp.","lead":"This paper proposes a training-free method for picking pretraining data: a small language model scores documents by how much its loss rises when important attention heads are masked, and the top-ranked documents are added to the training mix. The authors report gains of 1.4 to 3.5 percentage points on reasoning and knowledge benchmarks for a 7B model, though the experimental design does not isolate selection quality from simple data repetition.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline gains may be caused by repeated exposure of the selected 73B subset, not by reasoning-specific selection; the missing random-upsampling control is the decisive test.","rationale":"The reader's weakest-assumption diagnosis is correct and is the most load-bearing concern. The paper's own ablations in Appendix D validate that masking retrieval heads degrades the 1.3B selector on reasoning benchmarks, but they do not validate that the resulting ranking of pretraining documents transfers to a 7B model. The design in Section 5.1 confounds the choice of which documents to emphasize with the act of emphasizing a subset at all, and duplication or upsampling of a chosen subset is a known mechanism by which pretraining performance can improve. The mirror-effect analysis is correlational and cannot disambiguate selection quality from repetition. This is an internal experimental-control risk rather than an outside-consensus disagreement, and it is testable at a fraction of the original pretraining cost by running the random-upsampling control. Secondary issues such as unreleased code, single-run evaluations, and lack of error bars would be less important if the control were run. The paper has substantial positive features, including large-scale experiments and diverse qualitative analyses, but the central causal claim remains conditional. Since the reader already assigned a conditional verdict, my stress-test pass does not change that verdict.","tokens_in":27648,"tokens_out":3396,"duration_ms":36179,"concrete_test":"Train a 7B control model with the identical 1T-token budget on the full SmolLM-Corpus plus a randomly selected 73.1B-token subset, repeating that subset exactly as the AttentionInfluence setup repeats its selected subset, using the same optimizer, WSD schedule, and evaluation harness; report MMLU, MMLU-Pro, AGIEval-en, GSM8K, and HumanEval. If the random-upsampling control closes more than half of the reported gaps, the selection signal is not responsible for the headline improvements.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The experimental design does not separate the selection signal from token-level upsampling. In Section 5.1, the AttentionInfluence model is trained on the full 241B SmolLM-Corpus plus a repeated 73.1B selected subset, while the baseline is trained on the 241B corpus alone; both consume 1T tokens. The treatment therefore has two simultaneous changes: selected documents receive roughly twice the exposure of unselected documents, and the whole corpus is traversed fewer times because the token denominator is larger. Repeated exposure to a subset can improve benchmark performance independently of why the subset was chosen, so the reported +1.4 to +3.5pp gains in Table 1 could be produced by random upsampling at the same budget. The paper never runs a random-upsampling control, and the 'mirror effect' in Section 5.2 is a post-hoc correlation that does not rule out duplication as the active ingredient. Equation (5) may still measure retrieval-head sensitivity, but the central claim that AttentionInfluence identifies high-quality reasoning data is only supported if its gains exceed those from repeating an unselected 73B slice.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes AttentionInfluence, a training-free data-selection score that uses a small pretrained 1.3B model to rank documents by the relative increase in loss when top-5% retrieval heads are masked (Eq. 5). The method is applied to the SmolLM corpus, selecting the top 20% of samples (~73.1B tokens). A 7B model is pretrained on the full SmolLM corpus plus this selected subset for 1T tokens (the treatment), and compared with a baseline 7B model trained on the SmolLM corpus alone for 1T tokens. The authors report consistent gains of 1.4--3.5pp on MMLU, MMLU-Pro, AGIEval-en, GSM8K, and HumanEval, along with a 'mirror effect' between masking-induced degradation on the 1.3B selector and downstream gains in the 7B model. They also provide analyses of data diversity, word-frequency overlap with FineWeb-Edu, clustering distributions, and GPT-4o-based education/reasoning scores.","tokens_in":27888,"tokens_out":3655,"duration_ms":39628,"significance":"If the reported gains were attributable to the AttentionInfluence selection signal rather than to the experimental design, the paper would make a useful contribution: it proposes a supervision-free, training-free scoring mechanism that is cheap to compute and connects mechanistic interpretability findings to data curation. The manuscript is generally careful in reporting hyperparameters, evaluation settings, and ablations, and it includes extensive appendices with case studies and implementation details. However, the central experimental comparison is confounded, and the paper does not currently provide the control needed to support its main claim that AttentionInfluence identifies high-quality reasoning-intensive data.","major_comments":[{"comment":"The main comparison is confounded by token-level upsampling. The treatment model is trained on the full SmolLM corpus (241B tokens) plus the selected 73.1B subset, while the baseline is trained on the corpus alone; both consume 1T tokens. This means the treated model gives roughly twice the exposure to the selected documents and fewer exposures to the unselected documents, relative to the baseline. Any subset, including a random 73.1B slice, could produce gains under this protocol simply because of repeated exposure and reduced coverage of the rest of the corpus. The paper never runs a random-upsampling control at the same budget. Without such a control, the reported +1.4 to +3.5pp gains in Table 1 do not establish that the AttentionInfluence score, rather than the duplication protocol, is responsible for the improvement. Please add a training run that upsamples a randomly chosen 73.1B subset of the corpus in exactly the same way, and report the same benchmark table for that control.","section":"§5.1, Table 1"},{"comment":"The claim that a larger relative loss increase when retrieval heads are masked measures 'reasoning intensity' is not tested against simpler alternative mechanisms. A high (L_ref - L_base)/L_base value could also flag out-of-distribution, high-surprisal, or repetitive text, since such texts typically have high loss and may be particularly sensitive to masking any important component. The paper provides no comparison with selection by base loss, perplexity, or random ranking, so the mechanism underlying the score is unidentified. Please add controls that select the same 73.1B budget by high base loss (or high perplexity) and by random ranking; if the high-loss control matches the reported gains, the reasoning-specific interpretation in Section 5.2 is not supported.","section":"§4.2, Eq. (5)"},{"comment":"The 'mirror effect' is presented as post-hoc evidence, but it is a selected correlation rather than a predictive test. The paper identifies tasks where masking retrieval heads degrades the 1.3B selector and then observes gains on those same tasks in the 7B model, but no a-priori list of tasks is defined and no rank correlation across all evaluated benchmarks is reported. Moreover, because the mirror effect is computed on the same benchmarks used to select the discussion points, it cannot rule out the possibility that the gains come from upsampling generally useful documents. To make this evidence load-bearing, specify the full benchmark list in advance, report the correlation between masking-induced degradation and downstream improvement across all benchmarks, and show that a random-up sampling control does not exhibit a similar pattern.","section":"§5.2, Appendix D"}],"minor_comments":[{"comment":"The caption and text refer to the 7B model trained on selected data as 'AttentionInfluence-1.3B', which is confusing because the selector is 1.3B but the trained model is 7B; please use distinct names such as 'AI-1.3B-selected' for the trained model.","section":"§5.2, Figure 1"},{"comment":"The masking operation is described only as setting attention weights to equal weights 1/L; please clarify whether the mask is applied to the attention logits before softmax or to the normalized attention probabilities, since this affects the resulting reference model and the score in Eq. (5).","section":"Appendix C"},{"comment":"The pretraining configuration says 'the amount of training tokens is 1 TB', which conflates bytes and tokens; this should read 1T tokens, and the same correction applies to Table 7 entries such as '1.5TB' and '9TB' for the pretrained 1.3B and 7B models.","section":"Appendix E"},{"comment":"There is a typo in the figure caption: 'TirvialQA' should be 'TriviaQA', and the heading 'Increasing Parameter Size of AttentionInfluce' in Section 5.2 should be 'AttentionInfluence'.","section":"§5.2, Figure 9"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe thing to know: AttentionInfluence is a genuinely new, cheap recipe for data selection—mask retrieval heads to make a weak reference model and rank pretraining docs by relative loss increase—but the main experiment doesn't isolate the selection signal from token duplication. The treatment trains on the full SmolLM corpus plus the selected 73B subset, while the baseline trains on the corpus alone, both for 1T tokens. So selected docs get roughly twice the exposure of unselected docs in the same run, and more exposure than the baseline's docs. Without a random-upsampling control, the +1.4 to +3.5pp gains could be mere repetition of a 73B slice. I think that concern is the right one, and it is load-bearing for the central claim.\n\nWhat the paper does well: the combination is new and sensibly motivated. It takes retrieval-head detection from Wu et al., swaps in a masking-based weak model instead of training a strong reference model as in Mirror Influence/Rho-1, and uses the delta-loss as a quality score with per-domain normalization. That is a clean, supervision-free, training-free pipeline. The authors also ran a serious 7B pretraining budget (1T tokens, WSD) with many benchmarks, and they include a sanity check (Appendix D) showing masking retrieval heads hurts reasoning tasks more than masking random non-retrieval heads—that part supports the mechanism. The case studies and clustering analyses are useful but not central.\n\nSoft spots, in order: (1) the missing random-upsampling control; (2) no released code or selected-data indices, so the result is not independently checkable; (3) single-run evaluation with no error bars, so the 1-3pp deltas are not quantified against run-to-run noise; (4) the 'mirror effect' is a post hoc correlation and doesn't rule out duplication as the active ingredient. The 'reasoning intensity' interpretation also isn't fully pinned down—high delta-loss could partly reflect high-surprisal or OOD text—but the per-domain normalization and head-specific masking make that secondary, and the paper itself doesn't claim more than a heuristic.\n\nWho this is for: people working on pretraining data curation and anyone interested in using mechanistic interpretability for data selection. It deserves a serious referee: the method is novel, cheap, and the confound is fixable with one control run. I'd send it out, but I'd ask for the random-upsampling control, error bars or seeds, and code/data release before acceptance.","headline":"A novel, cheap data-selection recipe that is not yet proven: the headline gains are confounded with repeated exposure of the selected 73B subset, so the missing random-upsampling control is the decisive test.","tokens_in":28378,"tokens_out":3306,"would_cite":true,"duration_ms":32588,"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":"AttentionInfluence shows that masking a small model's retrieval heads yields a supervision-free score for selecting reasoning-intensive pretraining data.","keywords":["pretraining data selection","attention heads","retrieval heads","weak-to-strong generalization","influence measure","reasoning-intensive data","training-free data curation"],"falsifier":"Train a 7B model on a random 20% sample of the corpus (matched token count and length distribution) in addition to the full corpus, and compare with the AttentionInfluence-selected subset; if the random subset produces the same gains, the selection signal adds nothing. Alternatively, check whether the highest-scoring documents are simply the highest-perplexity documents for the intact 1.3B model; if so, the score reduces to a surprisal filter rather than a reasoning-specific signal.","tokens_in":27434,"feed_emoji":"🧠","tokens_out":6756,"duration_ms":56228,"temperature":0.7,"pith_summary":"AttentionInfluence proposes that the loss increase a small pretrained language model suffers when its retrieval attention heads are masked reveals how reasoning-intensive a pretraining document is. The paper uses this signal, computed on a 1.3B model, to pick the top 20% of the SmolLM corpus (about 73B tokens), then mixes those tokens back into the full corpus to pretrain a 7B model for 1T tokens. Against a baseline trained only on SmolLM, the selected-data model gains 1.4 to 3.5 percentage points on MMLU, MMLU-Pro, AGIEval-en, GSM8K, and HumanEval. The claim is that attention-head influence works as a weak-to-strong data selector: a small model, with no labels and no classifier training, can improve a larger model.","feed_headline":"Attention heads rank pretraining data; 7B model gains 3.5 points","feed_subtitle":"It scores documents by loss increase when retrieval heads are masked—no labels, no classifier training.","key_machinery":"The key machinery is the retrieval-head masking probe. First, the selector model's attention heads are scored on a synthetic key-passage retrieval task, and the top 5% become 'retrieval heads.' Then each candidate document receives an AttentionInfluence score: the relative increase in token-level cross-entropy loss when those heads are masked to uniform attention, compared with the intact model. This score is compared only within data domains, and a higher score is taken to mean higher reasoning intensity. The masking probe carries the entire selection: it is the only signal that ranks the 241B-token corpus.","core_discovery":"The paper's central claim is that the relative loss increase when the top 5% retrieval heads of a small pretrained model are masked—$(L_{\\mathrm{ref}} - L_{\\mathrm{base}})/L_{\\mathrm{base}}$—measures a document's reasoning intensity, and that ranking a corpus by this score and upsampling the top fraction yields better downstream reasoning and knowledge performance than training on the original corpus alone. The authors validate this on a 1.3B selector and a 7B target, reporting consistent gains across knowledge and reasoning benchmarks, and observe a mirror effect: the tasks where masking the selector's retrieval heads hurts most are the tasks where the 7B model trained on the selected data improves most. They interpret this as evidence that the masking probe isolates the reasoning-relevant content in the data.","pith_inferences":["The paper does not include a random-upsampling control, so part of the observed gains could come from simply training on more tokens or on a longer, more diverse subset rather than from reasoning-specific selection; a matched random-20% baseline would isolate the mechanism.","Because the score is only compared within domains, a cross-domain normalization would be needed to use AttentionInfluence for global corpus reweighting; the current design avoids comparing math to general text.","The masking probe is a natural template for other capability-specific heads: replacing the retrieval task with, say, an induction-head probe could yield selectors for in-context learning or long-range coherence.","The mirror effect could be turned into a cheap diagnostic before large-scale runs, but its reliability across model families, tokenizers, and corpora has not been demonstrated."],"forward_implications":["A small pretrained model can act as a supervision-free data selector for a larger model, eliminating the need for human-labeled or LLM-labeled training data and classifier training.","The selected data is more balanced across content categories and longer than classifier-selected data, suggesting the signal captures procedural and reasoning content rather than just educational style.","The mirror effect offers a pre-training predictor: benchmarks that degrade under retrieval-head masking in the selector are the ones most likely to improve in the target model.","The approach scales with selector size: a 7B selector selects higher-scoring reasoning data than a 1.3B selector, improving downstream performance further.","Combining AttentionInfluence with an educational-value classifier yields comprehensive improvements on tasks requiring factual knowledge, reasoning, or both."],"supporting_citations":[{"why":"Identifies retrieval heads and defines the retrieval score that AttentionInfluence uses to select the top 5% of heads to mask.","marker":"[47]"},{"why":"Provides the mirrored-influence idea of using a weaker and stronger model pair to estimate data influence, which AttentionInfluence adapts by masking heads instead of training a reference model.","marker":"[22]"},{"why":"Shows that token-level loss gaps between a base and reference model can identify important training tokens, the precedent for the delta-loss scoring.","marker":"[26]"},{"why":"Supplies the key-passage retrieval task design that the synthetic 800-sample proxy for detecting retrieval heads is based on.","marker":"[34]"},{"why":"Defines the SmolLM corpus that is ranked and from which the top 20% is selected for the pretraining experiments.","marker":"[3]"},{"why":"Provides the FineWeb-Edu classifier baseline whose selected-data distribution, education scores, and reasoning scores are compared with AttentionInfluence's.","marker":"[32]"}],"fun_headline_variants":["Mask attention heads to pick reasoning data: 7B gains up to 3.5pp","No labels: small model's head-masking scores pretraining data","Loss change from head masking predicts data's reasoning value","Weak-to-strong: 1.3B head selector improves 7B performance"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that the loss jump a 1.3B model shows when its retrieval heads are masked measures a document's reasoning intensity, and that this signal transfers to improve a 7B model trained on the selected data.","fun_headline_variants_meta":{"raw":{"variants":["Mask attention heads to pick reasoning data: 7B gains up to 3.5pp","No labels: small model's head-masking scores pretraining data","Loss change from head masking predicts data's reasoning value","Weak-to-strong: 1.3B head selector improves 7B performance"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000518,"raw_usage":{"total_tokens":2524,"prompt_tokens":970,"completion_tokens":1554,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":586,"completion_tokens_details":{"reasoning_tokens":1472}},"tokens_in":586,"tokens_out":1554,"duration_ms":11683,"temperature":1.0,"reasoning_tokens":1472,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:19:57.896396+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train a 7B model on a random 20% sample of the corpus (matched token count and length distribution) in addition to the full corpus, and compare with the AttentionInfluence-selected subset; if the random subset produces the same gains, the selection signal adds nothing. Alternatively, check whether the highest-scoring documents are simply the highest-perplexity documents for the intact 1.3B model; if so, the score reduces to a surprisal filter rather than a reasoning-specific signal.","supporting_citations":[{"cited_title":"The mirrored influence hypothesis: Efficient data influence estimation by harnessing forward passes","cited_arxiv_id":null,"evidence_quote":"Provides the mirrored-influence idea of using a weaker and stronger model pair to estimate data influence, which AttentionInfluence adapts by masking heads instead of training a reference model."},{"cited_title":"CLongEval: A Chinese Benchmark for Evaluating Long-Context Large Language Models","cited_arxiv_id":"2403.03514","evidence_quote":"Supplies the key-passage retrieval task design that the synthetic 800-sample proxy for detecting retrieval heads is based on."}],"review_version":1}