{"id":"552415ab-bb3c-476e-94f2-a562a95c2bd2","arxiv_id":"2504.12996","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A causal-tracing-guided method that freezes upper layers and retrains early MLP layers with a joint forget/retain loss achieved 2nd place in the SemEval-2025 Task 4 1B unlearning track.","lead":"Researchers devised an unlearning method that first locates where a language model stores sensitive facts and then retrains only those early layers to erase them. The approach placed 2nd on the 1B-parameter track of SemEval-2025 Task 4 while preserving most of the model's general knowledge.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The causal localization is established only on 125 Subtask-2 QA samples, with no per-subtask or 7B evidence, so the freeze-upper-layers design is the weakest link; the 7B utility collapse is consistent with localization failing at scale.","rationale":"The reader identified the same weakest assumption: causal localization on a narrow sample is the load-bearing step. I agree. The paper's contribution is specifically 'causal-informed layer optimization'; if the localization does not hold across subtasks and model sizes, the method is just an arbitrary choice of which layers to fine-tune, and the 7B failure is unexplained rather than a minor scaling issue. The empirical leaderboard result for 1B is real and should be credited, but it does not by itself validate the causal mechanism. I considered whether the internal tension in Tables 3-4 (MLP-only has low MIA and forget knowledge only 0.292, while MLP+MHSA scores higher) is the stronger concern; it is important, but it is entangled with the unstated configuration of the 1B submission, so the localization-transfer gap is cleaner and more directly testable. No adversarial reading is needed: the paper reports only one CMA experiment, with no per-subtask analysis, no noise scale, and no 7B causal evidence; the central claim overreaches that evidence. The proposed check would settle whether the concern lands. Until then, the reader's CONDITIONAL verdict remains appropriate; no verdict change is needed.","tokens_in":8877,"tokens_out":12375,"duration_ms":111410,"concrete_test":"Run the same CMA protocol, with the noise scale nu reported, on at least 50 public samples from Subtask 1, 50 from Subtask 3, and on OLMo-7B, and plot per-layer intervention scores; if the top-impact layers for any of these conditions are not within layers 0-5, the localization claim does not transfer and the freeze-upper-layers design has no mechanistic justification.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's mechanistic premise is that CMA on 125 synthetic QA samples from Subtask 2 (Section 3.1) shows OLMo layers 0-5 store the subject-attribute associations that need forgetting, and this justifies freezing all upper layers and editing only early MLP layers. That premise is not tested for Subtask 1 (creative documents), Subtask 3 (real documents), or for the 7B model; the noise scale nu is not reported, so the localization result cannot be reproduced or compared. Two consequences follow. First, if the associations to be forgotten in Subtasks 1 and 3 are encoded in later layers, those layers remain frozen and still contain the target information, so the forget loss on early layers must fight an incomplete path; this could explain the unexplained 46% MMLU drop in the 7B run (Section 4.1) rather than the paper's untested 'larger models require fewer update steps' explanation. Second, the public ablations in Tables 3-4 show the MLP-only configuration credited in the abstract has forget knowledge 0.292 and MIA score 0.010 (MIA accuracy 0.99) on the 7B model, while the better-scoring MLP+MHSA configuration is the one whose MMLU collapses; without knowing which configuration produced the 1B leaderboard result, the link between the causal claim and the reported success is not established. This is not a disagreement with consensus; it is a gap between the evidence presented (one subtask, one noise scale, one model size) and the general claim made.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a system for the SemEval-2025 Task 4 targeted-unlearning challenge. The method has two stages: (1) causal mediation analysis (CMA) on 125 synthetic question-answering samples from Subtask 2 to locate factual associations in the early transformer layers (0-5) of OLMo 1B and 7B models, and (2) constrained optimization that freezes layers above 5 and trains only the MLP modules in layers 0-5 with a joint loss consisting of cross-entropy on forget-set outputs plus an adaptively weighted retain-set cross-entropy term. The authors report second place in the 1B track with final score 0.652, task aggregate 0.973, and 88% of baseline MMLU, while the 7B run reaches task aggregate 0.964 but suffers a 46% MMLU drop. The central mechanistic claim is that early-layer MLPs function as stores of subject-attribute associations and that training only those layers achieves effective unlearning with preserved utility.","tokens_in":9211,"tokens_out":4215,"duration_ms":40830,"significance":"If the causal-localization claim held across all three subtasks and both model sizes, the paper would make a useful practical contribution: it would show that unlearning can be reduced to updating a small, mechanistically motivated parameter subset, and the competitive leaderboard result provides external validation of the overall pipeline. The paper also releases code and reports externally measured scores, which are concrete strengths. However, the significance as stated is currently limited by the narrow evidence base for the mechanistic claim, the lack of a reproducible noise scale for the CMA, and an internal inconsistency between the paper's recommended configuration and its own tables. The work is best read as a competition system description with an interesting but not yet fully supported mechanistic hypothesis.","major_comments":[{"comment":"The causal localization that justifies freezing all layers above 5 rests on 125 synthetic QA samples from Subtask 2 only. The Gaussian noise scale ν used for the corruption step is never reported, so the experiment cannot be reproduced or compared with other CMA studies. More importantly, no per-subtask evidence is provided: Subtask 1 (creative documents) and Subtask 3 (real documents) may encode the target information in different layer ranges or in attention rather than MLP pathways, and the paper gives no reason to assume the Subtask-2 QA localization transfers. Please report ν, add per-subtask heatmaps or at least per-subtask localization summaries, and either provide 7B localization evidence or explicitly limit the mechanistic claim to the tested setting.","section":"3.1, Figure 1"},{"comment":"The paper concludes that 'training only MLP layers is the most effective strategy' and repeats in the abstract and conclusion that MLP layers 0-5 can 'effectively remove information without causing much loss in model utility.' This is contradicted by the paper's own numbers. In Table 4, the 0-5 MLP+MHSA configuration dominates 0-5 MLP on final score (0.467 vs 0.353), task aggregate (0.775 vs 0.572), and MIA score (0.217 vs 0.010). In Table 3, MLP-only leaves forget-set knowledge at 0.292 and an MIA accuracy near 0.99, meaning the forget set remains almost perfectly identifiable. The MLP-only configuration is the best only on MMLU preservation, not on unlearning efficacy or privacy. No table or text identifies which configuration produced the submitted 1B leaderboard result, so the connection between the causal claim and the reported success is not established. Please clarify which configuration was submitted and reconcile the 'MLP-only' recommendation with the reported metrics.","section":"4.2, Tables 3-4"},{"comment":"The 46% MMLU drop in the 7B model is attributed in Section 4.1 to 'overfitting on this expanded corpus' and to the suggestion that 'larger models require fewer update steps,' but no learning curves, epoch sweeps, or other evidence are provided to support this explanation. An equally consistent explanation is that the causal localization does not transfer to the 7B model or to the non-QA subtasks, so freezing upper layers prevents the removal of associations stored there and forces the optimization into an inconsistent regime. This matters because the conclusion draws a general scaling lesson ('layer freezing thresholds must scale non-linearly with model depth') from this single unexplained drop. Please provide the missing evidence or explicitly acknowledge that the scaling behavior is currently unexplained.","section":"4.1, 5"}],"minor_comments":[{"comment":"The abstract contains a duplicated phrase: 'with a with a final score of 0.652.'","section":"Abstract / 1"},{"comment":"'the contents of theretain-set' is a typo; also, the distinction between public and private test sets should be stated more clearly because some numbers in Section 4.1 appear to come from the private evaluation while others come from public ablations.","section":"2.2.1"},{"comment":"The citation 'Vig et al., 2004' is incorrect; the Causal Mediation Analysis paper by Vig et al. is from 2020.","section":"3.1"},{"comment":"The adaptive α formula in Section 3.2 is not numbered, making it awkward to reference; please number the equations. Also, the hyperparameters a=0.3, b=6, c=0.8, α_min=1.2, α_max=2.8 are reported only in Appendix A, but no sensitivity analysis is given; a small grid over α_max or b would help the reader judge robustness.","section":"3.2"},{"comment":"The abbreviations 'Reg.' and 'Know.' are not defined in the captions; please spell out 'Regurgitation' and 'Knowledge' and clarify the direction of each score (higher/lower is better).","section":"4.2, Tables 3-4"},{"comment":"The sentence 'training different set of parameters for 8 epochs shows where that by training only MLP layers...' contains a grammatical error ('shows where that').","section":"4.2"},{"comment":"The conclusion asserts that 'output token cross-entropy provides a more surgical intervention than full-sequence loss calculations,' but no experiment in the paper compares these two loss variants; please either add the comparison or remove the claim.","section":"5"}],"recommendation":"major_revision","confidential_remarks":"The paper is a competition system description, and the leaderboard result is the strongest asset. The main risk is overinterpretation: the causal mechanism is asserted as a general property of OLMo models on the basis of a small single-subtask experiment, and the paper's own ablation tables partially contradict the recommended configuration. The authors should be asked to either add per-subtask and per-scale evidence or substantially soften the mechanistic claims. If the per-subtask ablations are infeasible, the paper could still be acceptable as a system-report contribution, but not with the current abstract and conclusion wording."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, this is a competent SemEval system report, and the 1B leaderboard result is real: second place, task aggregate 0.973, externally scored. Second, the paper's distinctive claim—that OLMo stores subject-attribute associations in layers 0–5, so you can freeze everything else and just edit those MLPs—is genuinely interesting but much thinner than the abstract suggests.\n\nWhat it does well: it ships code, gives a clear ablation of layer groups (all layers vs. 0–5 MLP vs. MHSA), and is honest about the 7B MMLU collapse. The localization figure for OLMo 1B/7B is a concrete empirical artifact that isn't in the cited literature, as far as I know. The adaptive-alpha joint loss is a reasonable engineering choice, though not a conceptual advance.\n\nSoft spots, in proportion. The causal mediation analysis is run on 125 synthetic QA samples from Subtask 2, with no reported noise scale. That's the load-bearing evidence for freezing upper layers, but there's no per-subtask or per-size evidence that the localization transfers. The 7B run losing 46% MMLU is consistent with the localization failing at scale, and the paper's explanation ('larger models require fewer update steps') is an untested guess. There's also a mismatch: the abstract credits MLP-only updates for the 1B success, while the public 7B ablations show the MLP-only config has forget knowledge 0.292 and MIA accuracy 0.99—so it leaves the target data in. It's not clear which configuration produced the 1B leaderboard submission, and the paper should say so. The alpha hyperparameters were tuned on the public set, so the forget-set improvement is partly an optimized objective rather than a prediction. These are real concerns, but they don't sink the 1B empirical result, which is externally evaluated.\n\nFor whom: anyone working on unlearning or model editing will want to look at the layer-localization figure and the ablation tables. It's a legitimate system paper with a specific discovery. I'd send it to a serious referee, with instructions to push on the causal evidence and the configuration mismatch. It doesn't need a desk reject.","headline":"A solid shared-task report with a genuinely interesting layer-localization finding for OLMo, but the causal claim is thinner than the abstract implies and the 7B failure is unexplained.","tokens_in":9731,"tokens_out":2513,"would_cite":false,"duration_ms":22213,"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":"The paper claims that in OLMo models the facts targeted for unlearning are stored in the MLP modules of transformer layers 0–5, and that freezing all upper layers while training only those early MLPs with a joint forget/retain loss…","keywords":["machine unlearning","causal mediation analysis","knowledge localization","large language models","OLMo","targeted unlearning","privacy","layer freezing"],"falsifier":"Run the same unlearning pipeline but edit only layers 0–5 while keeping upper layers frozen, and measure task aggregate separately per subtask (Subtask 1 creative documents, Subtask 2 PII, Subtask 3 real documents). If forget-set scores for Subtask 1 or Subtask 3 stay near baseline (knowledge not removed) or retain-set scores collapse, the localization assumption is falsified. Alternatively, repeat the causal tracing on Subtask 1 and Subtask 3 samples; if restoring hidden states in layers above 5 recovers correct output for those subtasks, the early-layer claim does not transfer.","tokens_in":8682,"feed_emoji":"🧠","tokens_out":8302,"duration_ms":76344,"temperature":0.7,"pith_summary":"Large language models memorize sensitive facts, and retraining to forget them is prohibitive; this paper argues that in the OLMo family the facts targeted for forgetting are stored in the MLP modules of the earliest transformer layers (0–5), so selective unlearning can be achieved by freezing every layer above 5 and optimizing only those early MLPs. The optimization minimizes a joint loss that drives up cross-entropy on forget-set outputs while adaptively penalizing any rise in retain-set loss, converting unlearning into a localized memory edit instead of a full-model update. On the 1B model this recipe placed second in the task's private evaluation with a final score of 0.652, a task aggregate of 0.973, and roughly 88% of the baseline MMLU accuracy; on the 7B model it removed forget-set content about as well (0.964 task aggregate) but cut MMLU by 46%. If the causal localization is correct, the same two-stage recipe—locate the association, then edit its storage layers under a retain-preserving constraint—offers a cheap, surgical path to privacy compliance for deployed LLMs.","feed_headline":"Edit six layers, not all 32, to make LLMs forget","feed_subtitle":"Causal tracing finds stored facts in early MLP layers; a constrained loss erases them while keeping 88% of general accuracy.","key_machinery":"The carrying mechanism has two parts. First, causal tracing: add Gaussian noise to the subject token's embedding, run the model to see the prediction corrupt, then restore the original hidden state at candidate layer/token positions and measure whether the correct attribute prediction returns—the positions that restore prediction are judged causally responsible for storing the fact. Second, a layer-restricted joint loss $\\mathcal{L}_{joint}=-\\mathcal{L}^{CE}_{forget}+\\alpha\\,\\mathcal{L}^{CE}_{retain}$ applied only to MLP parameters of layers 0–5, with an adaptive weight $\\alpha=\\mathrm{clip}(a\\cdot b^{\\Delta L}+c,\\alpha_{min},\\alpha_{max})$ chosen so that retain-set drift is penalized exponentially (empirically $a=0.3$, $b=6$, $c=0.8$, $\\alpha_{min}=1.2$, $\\alpha_{max}=2.8$). Together they turn unlearning into a targeted edit of the storage site rather than a whole-model re-optimization: the trace decides which parameters to touch, and the joint loss decides how hard to push them.","core_discovery":"Using causal mediation analysis on 125 synthetic question-answering samples from the task's Subtask 2, the authors corrupt subject-token embeddings with Gaussian noise and then restore hidden states at each layer; restoring layers 0–5 of OLMo-1B and OLMo-7B recovers correct attribute predictions, identifying those early layers as the storage site for subject-attribute associations. Guided by that map, they freeze layers 6 through L and train only the MLP weights of layers 0–5 with the joint loss $\\mathcal{L}_{joint}=-\\mathcal{L}^{CE}_{forget}+\\alpha\\,\\mathcal{L}^{CE}_{retain}$, where $\\alpha$ grows exponentially whenever retain-set loss drifts above its epoch-0 baseline. Parameter studies (their Tables 3 and 4) show MLP-only editing removes forget-set knowledge with much less utility damage than editing both attention and MLP, or all layers. On the private test set the 1B submission reached a 0.652 final score with 0.973 task aggregate and forget-set knowledge retention reduced to 0.14 (an 86% reduction), while the 7B submission reached 0.964 task aggregate at the cost of MMLU falling from 0.509 to 0.275. The paper reads these results as evidence that causal-informed, layer-restricted editing is an efficient and precise unlearning paradigm.","pith_inferences":["Editorial: Causal tracing on one subtask is slim ground for a universal layer map; a cheap test, run the same tracing on Subtask 1 creative documents and Subtask 3 real documents, and if the recovery peak shifts above layer 5 the early-layer claim is subtask-specific.","Editorial: The 7B utility collapse suggests the localization itself may be scale-dependent; comparing traced layer maps between 1B and 7B on identical samples would separate a genuine mechanistic difference from an overfitting artifact.","Editorial: The adaptive $\\alpha$ rule is a manual trust-region controller; one could replace the hand-set $a,b,c$ with an online estimate of retain-loss curvature and get the same protection with zero tuning.","Editorial: 'Unlearned' here is defined by output-level metrics; probing intermediate representations before and after editing would show whether the association is deleted from the weights or merely made unreachable, which matters for adversarial extraction."],"forward_implications":["Unlearning a fact costs one forward/backward pass on early layers instead of full-model retraining; the approach is computationally cheap enough to run per forget request.","General capabilities survive because upper layers are frozen: the 1B model kept about 88% of baseline MMLU while forget-set knowledge dropped 86%.","The two-stage recipe is model-agnostic in principle: any transformer can be traced first, then edited at its own storage layers, though the paper only demonstrates OLMo.","Task aggregate and MIA scores can simultaneously be strong (0.973 and 0.741 on 1B), suggesting that disrupting output-level regurgitation also weakens membership inference.","Scaling is not free: the 7B run shows the same six-layer edit erases the forget set but destroys 46% of MMLU, so the layer-freezing threshold must scale nonlinearly with model depth (the paper's own caution)."],"supporting_citations":[{"why":"Supplies the causal mediation analysis method used to locate where factual associations are stored.","marker":"Vig et al., 2004"},{"why":"Establishes that MLP modules in transformer layers store factual associations, the premise for editing only MLPs.","marker":"Geva et al., 2023"},{"why":"Shows how to locate and edit factual associations in GPT, the basis for restricting updates to MLP layers.","marker":"Meng et al., 2022"},{"why":"Provides the OLMo model family (1B and 7B) that is fine-tuned and edited in the experiments.","marker":"Groeneveld et al., 2024"},{"why":"Defines the task, dataset, private test set, and evaluation scores (final, task aggregate, MIA, MMLU).","marker":"Ramakrishna et al., 2025b"},{"why":"Provides the MMLU benchmark used to measure preservation of general ability after unlearning.","marker":"Hendrycks et al., 2021"},{"why":"NPO baseline that the joint-loss method is compared against in the leaderboard.","marker":"Zhang et al., 2024"}],"fun_headline_variants":["Edit just six early layers to make LLMs forget specific facts","Causal map reveals early layers store facts; edit only those","Freeze upper layers, hone lower ones for targeted unlearning","86% forget-set erased with 88% accuracy retained by layer-restricted edit"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the causal-tracing result from 125 synthetic question-answering samples—showing that the first six layers store the subject-attribute associations—holds for all three task subtasks and for the 7-billion-parameter model; if creative documents, real documents, or the larger model store the target knowledge in other layers, freezing everything above layer 5 has no justification.","fun_headline_variants_meta":{"raw":{"variants":["Edit just six early layers to make LLMs forget specific facts","Causal map reveals early layers store facts; edit only those","Freeze upper layers, hone lower ones for targeted unlearning","86% forget-set erased with 88% accuracy retained by layer-restricted edit"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000272,"raw_usage":{"total_tokens":1691,"prompt_tokens":1060,"completion_tokens":631,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":676,"completion_tokens_details":{"reasoning_tokens":556}},"tokens_in":676,"tokens_out":631,"duration_ms":6843,"temperature":1.0,"reasoning_tokens":556,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T12:16:43.626155+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same unlearning pipeline but edit only layers 0–5 while keeping upper layers frozen, and measure task aggregate separately per subtask (Subtask 1 creative documents, Subtask 2 PII, Subtask 3 real documents). If forget-set scores for Subtask 1 or Subtask 3 stay near baseline (knowledge not removed) or retain-set scores collapse, the localization assumption is falsified. Alternatively, repeat the causal tracing on Subtask 1 and Subtask 3 samples; if restoring hidden states in layers above 5 recovers correct output for those subtasks, the early-layer claim does not transfer.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Shows how to locate and edit factual associations in GPT, the basis for restricting updates to MLP layers."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the MMLU benchmark used to measure preservation of general ability after unlearning."}],"review_version":1}