{"id":"2a7dd94b-dfba-4a41-9105-edfe3ac74333","arxiv_id":"2607.23970","paper_version":2,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"Unlearned models usually connect to their originals by smooth low-loss paths, and the smoothness of that path can predict how hard the unlearning task was.","lead":"Machine unlearning claims to erase a model's memory of your data without retraining. This paper maps the geometry of that erasure, finding that unlearned models often sit in smooth low-loss valleys connected to the original model, and that this shape predicts how well the erasure holds.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"MCU path definition is unstated, making the central connectivity claim and difficulty correlation protocol-dependent.","rationale":"The reader's weakest_assumption correctly identifies that the definition of a low-loss path in unlearning is ambiguous because the forget loss is intentionally high at the unlearned model. The reported correlation between MCU smoothness and unlearning difficulty may be tautological if both are derived from the same loss terms. Since the full text is unavailable and the abstract provides no protocol for the path construction, the strongest claim cannot be verified. This is a genuine load-bearing concern, but it does not move the verdict from UNVERDICTED; it reinforces the reader's assessment. My agreement with the reader is full on this point, and the recommended verdict remains UNCHANGED.","tokens_in":829,"tokens_out":1805,"duration_ms":29114,"concrete_test":"Pick one experimental setting from the paper (e.g., curriculum learning). For a fixed pair (original model, unlearned model), recompute MCU paths under three alternative path definitions: (i) retain-set loss only, (ii) forget-set loss only, (iii) the weighted loss used by the unlearning method. Then test whether basin membership (linear vs. nonlinear connectivity) and the MCU-smoothness/unlearning-difficulty correlation are invariant to this choice. Independently measure unlearning difficulty via a loss-agnostic metric such as membership-inference attack success rate. If the results change qualitatively across path definitions or the correlation vanishes, the abstract's central claim is protocol-dependent; if they persist, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract's central claim that unlearned models lie in connected basins with smooth retain/forget behavior presupposes a specific definition of a \"low-loss path.\" In standard mode connectivity, paths are low-loss with respect to the training objective. But unlearning deliberately raises the forget loss at the unlearned model, so no path connecting original and unlearned models can be low-loss on the full objective (original + forget). The authors must have chosen a sub-objective (e.g., retain loss only) or a weighted combination, but the abstract does not specify this. Consequently, basin membership, path smoothness, and any MCU-based conclusions are conditional on that unstated choice. Additionally, if \"MCU smoothness\" is measured via the same loss terms used to quantify unlearning difficulty (e.g., forget loss magnitude or its gradient), then the reported correlation between MCU smoothness and unlearning difficulty may be partially tautological. Without the path-loss protocol, the strongest claim is not independently checkable from the abstract alone.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes 'mode connectivity in unlearning' (MCU), a framework for studying the loss-landscape geometry of machine unlearning. Based on the abstract, the authors claim that, across various training settings (curriculum learning, second-order optimization, and different unlearning methods), many unlearned models lie in connected basins with smooth retain/forget behavior; that training dynamics can move solutions into different basins; that models within the same basin can differ substantially on privacy metrics; that unlearning progresses nonlinearly from the original model; that linear connectivity indicates most approximate unlearning methods are mechanistically distinct from retraining; that MCU-based ensembling improves generalization and robustness to relearning attacks; and that MCU smoothness correlates with unlearning difficulty. The abstract presents these as empirical findings, but no full text, datasets, protocols, or numerical evidence are available for review.","tokens_in":1089,"tokens_out":1502,"duration_ms":24130,"significance":"If the claims hold, this would be a novel and potentially useful geometric perspective on machine unlearning: basin membership and path smoothness could inform privacy behavior, ensembling across connected models could improve robustness, and MCU smoothness could serve as a proxy for unlearning difficulty. The proposed framework is falsifiable and testable. However, because this review is based solely on the abstract, the significance cannot be assessed beyond the plausibility of the claims; the central definitions and evidence are not yet visible.","major_comments":[{"comment":"The abstract does not specify what loss function defines a 'low-loss path' or a 'connected basin' in the unlearning setting. Unlearning deliberately raises the forget loss at the unlearned model, so a path that is low-loss on the full objective cannot exist between the original and unlearned models. The authors must have chosen a sub-objective (e.g., retain loss only), a weighted combination, or a path constrained to a particular subspace, but this choice is not stated. Every subsequent claim about basin membership, smooth retain/forget behavior, and linear connectivity is conditional on this unstated protocol.","section":"Abstract (central definition of MCU)"},{"comment":"The claimed correlation between 'MCU smoothness' and 'unlearning difficulty' is at risk of being partly tautological if the two quantities are computed from the same underlying loss terms, gradients, or path integrals. The abstract does not state how 'difficulty' is measured (e.g., retrain distance, relearn resistance, or forget-loss magnitude). If the difficulty metric and the smoothness metric share components, the reported correlation would not be informative. This needs an explicit independent definition.","section":"Abstract (MCU smoothness as a difficulty proxy)"},{"comment":"The abstract claims that MCU is 'evaluated across a range of settings' and reports qualitative findings ('many unlearned models lie in connected basins,' 'MCU-based ensembling can improve generalization'), but no datasets, baselines, error bars, or negative controls are mentioned. Without these, the central empirical claims are unevidenced at the abstract level. The reader cannot judge whether the findings are robust across architectures, data modalities, or unlearning methods, or whether they depend on particular hyperparameters.","section":"Abstract (empirical scope and evidence)"}],"minor_comments":[{"comment":"The term 'MCU smoothness' is introduced without a formal definition; the abstract should at least sketch what quantity is being measured.","section":"Abstract (terminology)"},{"comment":"The phrase 'linear connectivity suggests that most approximate unlearning methods are mechanistically distinct from retraining' is vague. The reader cannot tell what threshold of nonlinearity counts as 'mechanistically distinct' or how this is derived.","section":"Abstract (mechanical distinctness claim)"},{"comment":"The claim that 'unlearning progresses nonlinearly' needs a concrete notion of progression (e.g., interpolation parameter, training time, or loss trajectory). Without that, the statement is not checkable.","section":"Abstract (nonlinear progression)"}],"recommendation":"uncertain","confidential_remarks":"This review is severely constrained because only the abstract was provided. The core issues are not defects visible in the full text but rather missing definitions and missing evidence that the abstract alone cannot resolve. I would need the full paper to determine whether the path-loss protocol is well specified and whether the difficulty correlation is non-tautological. Given the journal's standards, I would lean toward requesting the full manuscript before any accept/reject decision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a sensible first pass at bringing mode connectivity to unlearning, and the empirical checklist in the abstract is genuinely interesting. But the paper is not evaluable from the abstract because the central quantity, MCU, is not defined. Specifically, unlearning deliberately raises forget loss, so a 'low-loss path' connecting original and unlearned models cannot be low-loss on the full objective. The authors must be using some sub-objective or weighted sum, but they don't say which. That omission is not cosmetic—basin membership, path smoothness, and the reported correlation between MCU smoothness and unlearning difficulty all change with that choice.\n\nWhat is new: I don't know of a prior mode-connectivity study in unlearning, so the framing is novel. The claims about nonlinear trajectories and linear connectivity being evidence against retraining equivalence are testable and would be useful if supported. The ensembling idea is a practical payoff.\n\nWhat's good: the authors clearly know the mode-connectivity literature and are asking the right questions. They also seem aware that different training dynamics produce different basins, which suggests they are not just force-fitting connectivity.\n\nSoft spots: the tautology risk is real. If 'unlearning difficulty' is measured by forget loss magnitude or gradient norm, and MCU smoothness is measured along a path in the same loss terms, then the correlation may be built in. The abstract doesn't show that difficulty is measured independently (e.g., by retrain distance or relearn resistance). I can't tell from the abstract whether the full paper addresses this. If it doesn't, the central claim is vacuous. If it does, this could be a solid paper.\n\nAlso, no datasets, baselines, or error bars at this level, so I can't verify any of the empirical statements. That's normal for an abstract, but it means my confidence is low.\n\nBottom line: this deserves a careful referee only if the full paper defines the path-loss protocol and shows that the difficulty metric is independent of the path quantities. If those are missing, desk reject. Otherwise send it out.","headline":"Nice idea, plausible claims, but the abstract leaves the definition of 'low-loss path' in unlearning unspecified, and that is the load-bearing choice.","tokens_in":1521,"tokens_out":2064,"would_cite":false,"duration_ms":30857,"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":"Machine unlearning can be understood through mode connectivity, the smooth low-loss paths connecting unlearned models in parameter space.","keywords":["machine unlearning","mode connectivity","loss landscape","parameter space","privacy metrics","ensembling","relearning attacks","model basins"],"falsifier":"One concrete check: fix an unlearning method and compute MCU using retain loss only versus forget loss only versus a weighted sum; if the reported correlations with unlearning difficulty or privacy change sign or vanish under the choice, the central claim is protocol-dependent. Another check: train the same unlearning method from different seeds and see whether the 'same basin' conclusion holds under a stricter path definition, such as Bézier curves with a low maximum loss along the path.","tokens_in":725,"feed_emoji":"🔗","tokens_out":3471,"duration_ms":43785,"temperature":0.7,"pith_summary":"The paper argues that the geometry of unlearning is informative: most unlearned models sit in connected basins, meaning a smooth path of low-loss models links the original and the unlearned state. Changes in training dynamics—curriculum, optimizer choice—can push solutions into separate basins, and connectivity across different unlearning methods reveals that approximate unlearning is mechanistically distinct from full retraining. If this is right, researchers can use basin membership and path smoothness as a proxy for unlearning difficulty and privacy behavior. The paper also shows that averaging models along the connecting path improves generalization and resistance to relearning attacks. A sympathetic reader would care because this turns unlearning from a black-box procedure into a geometric object with testable structure.","feed_headline":"Unlearned models cluster in connected low-loss basins","feed_subtitle":"Smooth paths between original and unlearned models predict unlearning difficulty and privacy behavior.","key_machinery":"Mode connectivity in unlearning (MCU), the central object, is the finding that two models in parameter space—here, a model trained on all data and its unlearned counterpart—can be joined by a path along which the loss stays low. The paper uses MCU to measure basin membership, path smoothness, and nonlinearity of the unlearning trajectory, and uses it as a tool for ensembling and as a predictor of unlearning difficulty.","core_discovery":"The paper introduces mode connectivity in unlearning (MCU) and reports that many unlearned models lie in connected basins with smooth retain/forget behavior. It finds that changes in training dynamics can move solutions into different basins, that models in the same basin can differ substantially on privacy metrics, and that unlearning progresses nonlinearly from the original model to the unlearned model. Linear connectivity across different unlearning methods indicates that approximate unlearning methods are mechanistically distinct from retraining. MCU-based ensembling improves generalization and robustness to relearning attacks, and MCU smoothness correlates with unlearning difficulty.","pith_inferences":["If MCU smoothness is a reliable proxy for unlearning difficulty, then cheap path-smoothing checks could replace expensive attacks as a first screening for unlearning quality.","The nonlinear progression from original to unlearned model hints that intermediate checkpoints along the path are not simple interpolations; this could matter for audits that try to trace what was removed.","A natural testable extension is to check whether MCU also predicts resistance to membership inference, not only relearning, and whether the correlation survives when the path is defined by retain loss only rather than a weighted combination.","Because the abstract does not specify the loss combination that defines a smooth low-loss path, MCU is best interpreted as a family of measures; the reported correlations may depend on that unspecified choice."],"forward_implications":["Unlearning difficulty can be read off the geometry: smoother MCU paths correlate with easier unlearning.","Ensembling models along an MCU path yields a final model that generalizes better and resists relearning attacks.","Approximate unlearning methods fall into basins distinct from retraining, so their outputs should not be treated as close to a retrained model.","Curriculum learning or second-order optimization can move unlearning into a different basin, changing privacy behavior.","Privacy metrics can vary substantially among models in the same basin, so basin membership alone does not determine privacy."],"fun_headline_variants":["Mode connectivity maps unlearning's hidden basins","Unlearning basins linked by smooth loss paths","Connectivity predicts unlearning difficulty and privacy","Why unlearned models share low-loss paths","Unlearning geometry: connected basins, distinct methods"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that a low-loss path between the original and unlearned model is well-defined and informative when the forget set's loss is deliberately raised; the notion of 'smoothness' and which loss combination defines the path are not specified.","fun_headline_variants_meta":{"raw":{"variants":["Mode connectivity maps unlearning's hidden basins","Unlearning basins linked by smooth loss paths","Connectivity predicts unlearning difficulty and privacy","Why unlearned models share low-loss paths","Unlearning geometry: connected basins, distinct methods"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000194,"raw_usage":{"total_tokens":1180,"prompt_tokens":720,"completion_tokens":460,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":464,"completion_tokens_details":{"reasoning_tokens":392}},"tokens_in":464,"tokens_out":460,"duration_ms":7761,"temperature":1.0,"reasoning_tokens":392,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T03:27:13.453939+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"One concrete check: fix an unlearning method and compute MCU using retain loss only versus forget loss only versus a weighted sum; if the reported correlations with unlearning difficulty or privacy change sign or vanish under the choice, the central claim is protocol-dependent. Another check: train the same unlearning method from different seeds and see whether the 'same basin' conclusion holds under a stricter path definition, such as Bézier curves with a low maximum loss along the path.","supporting_citations":[],"review_version":2}