{"id":"9d5cb0d6-b4d4-44f1-a431-56e0446e3d9e","arxiv_id":"2412.20637","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"KNE selects top-attribution neurons across all layers and updates only those weights, claiming improved knowledge editing accuracy on Llama-2-7B and GPT-J-6B.","lead":"This paper proposes KNE, a method for editing factual knowledge in large language models by identifying and updating a small set of 'knowledge neurons' across layers using gradient attribution scores. The authors report high edit success and competitive portability and locality on Llama-2-7B and GPT-J-6B, but the paper lacks code, error bars, and contains a dimension mismatch in its core update equation.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The attribution-based localization in KNE is never validated against a same-size random-parameter control; without that control, the Table 1 results do not establish that the knowledge neuronal ensemble, rather than the restricted gradient update itself, drives the gains.","rationale":"The reader's weakest_assumption identifies the same load-bearing point: the method assumes that the gradient-attribution score in Eq. (5) identifies parameters whose update is sufficient for the edit, and the paper provides no causal validation. My read agrees with that assessment and adds a concrete way to test it. The paper's own ablations in Section 4.5 vary parameter count, not parameter identity, so they do not establish that the attributed neurons matter. A random-parameter control is the minimal experiment that would separate 'localization works' from 'any sufficiently large gradient-constrained FFN update works.' Since the manuscript also lacks specified hyperparameters, a consistent implementation is impossible as written, and Eq. (10) as printed is dimensionally inconsistent: assigning an n-by-d1 matrix to d2 columns of Delta W cannot be type-correct. These issues reinforce the rejection: the central claim is not currently supported, not because it is necessarily false, but because the necessary control and specification are absent. For these reasons, the reader's REJECT verdict remains appropriate, and my read does not change it.","tokens_in":12462,"tokens_out":3596,"duration_ms":38431,"concrete_test":"Run KNE exactly as described on Llama-2-7b-chat with WikiData counterfact, but draw M_kne uniformly at random (same n per layer) instead of selecting the top gradient-attribution scores, using the same m, alpha, p, and update rule; repeat over 10 random seeds. Compare Edit Succ, Portability, Locality, and Fluency against the Table 1 KNE row. If the random-ensemble mean is within about 2 points of KNE on Edit Succ and Portability, then the attribution localization is not load-bearing; if it is substantially worse, the localization assumption survives this check.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that KNE's gradient-attribution 'knowledge neuronal ensemble' (Eq. 5, Section 3.2.2) is what drives the Table 1 results is unsupported because no experiment contrasts the attributed ensemble with a same-size, same-procedure random ensemble. The edit step (Eqs. 10-11) is a gradient update restricted to a selected slice of FFN weights; any slice of the same cardinality could in principle improve Edit Success and Portability. The paper's Section 4.5 ablation varies only the number of edited parameters, not whether the selected parameters are the attributed ones. Without a random or identity control, the high Edit Success and Portability numbers are equally explained by 'update enough FFN parameters with gradient descent' as by 'the attributed neurons encode the fact.' Because the paper's novelty is precisely the localization, this missing control is load-bearing. The companion reproducibility defects (Eq. 10 is dimensionally inconsistent and alpha, m, and p are unspecified) mean the claimed numbers cannot currently be reproduced, which reinforces that the central claim is unverified.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Knowledge Neuronal Ensemble (KNE), a locate-then-edit method for knowledge editing in LLMs. KNE computes token-level gradient attribution scores for FFN parameters (Eq. 5), selects the highest-scoring parameters across layers into a 'knowledge neuronal ensemble' (Eqs. 6-7), and updates only those parameters via a scaled gradient update (Eqs. 8-11). The authors evaluate KNE on ZsRE, WikiData counterfact, and WikiData recent using Llama-2-7b-chat and GPT-J-6B, reporting high Edit Success and competitive Portability/Locality in Table 1 and Table 2. The paper also includes layer-wise storage analyses, batch-editing experiments, and an ablation varying the number of edited parameters. A brief limitations section acknowledges the need for better knowledge-set selection and for theoretical grounding of the layer-choice findings.","tokens_in":12681,"tokens_out":6870,"duration_ms":68884,"significance":"If its empirical claims were reproducible, KNE would be a practically useful and simple baseline for batch knowledge editing, since it edits roughly 1% of parameters and reports strong portability. The paper formulates knowledge editing as constrained optimization and extends token-level gradient attribution to multiple tokens, which is a sensible direction. However, the manuscript is currently not reproducible: the update rule in Eq. (10) has a dimensional inconsistency, the hyperparameters p, m, and alpha are never specified, and no code is provided. More importantly, the central claim that the gradient-attribution ensemble is what drives the gains is not causally validated against a random-parameter control. The significance of the reported numbers is therefore conditional on substantial revisions.","major_comments":[{"comment":"The update rule as written cannot be executed: Eq. (8) defines W_kne as an n×d1 matrix, Eq. (9) defines Delta W as d2×d1, and Eq. (10) assigns Delta W[:, M_kne] = W_kne. Since M_kne has length n and its entries are stated to be natural numbers less than d2, the left-hand side has shape d2×n, so the assignment requires W_kne to be d2×n, not n×d1. If the intent is to select rows, the assignment should instead be Delta W[M_kne, :] = W_kne. This dimensional inconsistency must be corrected, and the indexing convention (row vs. column neurons, and 0- vs. 1-based indexing) must be stated explicitly.","section":"§3.2.3, Eqs. (8)–(10)"},{"comment":"The central claim that the gradient-attribution ensemble drives the reported gains is not tested. Section 4.5 varies only the number of edited parameters; it does not compare the attributed ensemble with a same-size random subset of FFN parameters or with any identity/control selection. Without such a control, the high Edit Success and Portability figures in Table 1 are equally compatible with the hypothesis that any sufficiently large restricted gradient update on FFN weights improves these metrics. Because the novelty of KNE is precisely the attribution-based localization, this missing control is load-bearing for the paper's main conclusion.","section":"§3.2.2, Eqs. (6)–(7), and §4.5"},{"comment":"The hyperparameters p (selection quantile in Eq. (6)), m (Riemann steps in Eq. (5)), and alpha (scaling factor in Eq. (11)) are never specified, and no code or repository is provided. The subset-selection procedure in Section 4.4, where 200 localized knowledge points outperform full-dataset localization, is also not described. These omissions make the numerical results in Tables 1 and 2 unreproducible, and the paper reports no error bars or significance tests to support the word 'significantly' in the abstract.","section":"§4.1, Appendix A, and Eq. (11)"},{"comment":"The quantile notation is self-contradictory: the text says KNE selects the top 1−p% of neurons, but Eq. (6) defines t_p = Quantile_{1−p}(...), and the following sentence says this quantile function 'calculates the value corresponding to the top p%'. This ambiguity makes it impossible to determine the fraction of parameters actually edited; please clarify the intended selection rule and report the fraction used in each experiment.","section":"§3.2.2, Eq. (6)"},{"comment":"The abstract claims KNE 'achieves, or even exceeds, the performance of the best baseline methods in portability and locality metrics,' but Table 1 contradicts this for locality on WikiData recent: KNE obtains 37.58, well below ROME (66.2), MEMIT (64.78), and FT-L (63.7). The claim should be revised to per-dataset and per-metric accuracy, and the large locality drop on this dataset deserves a specific explanation.","section":"Abstract and Table 1"}],"minor_comments":[{"comment":"The phrase 'Knowledge Neurona’l Ensemble' contains a typographical error ('Neurona’l') that should be corrected to 'Neuronal'.","section":"§5, Conclusion"},{"comment":"The symbol m is overloaded: it denotes the number of Riemann steps in Eq. (5) and the number of neurons in a layer in Definition 3. Please use distinct symbols for these two quantities.","section":"§3, notation"},{"comment":"Figure 2 is referenced as illustrating layer-wise findings, but the text does not walk through panels (a)–(d); adding a sentence per panel would make the conclusions in Section 4.3 easier to verify. The appendix figures B.3–B.5 would also benefit from explicit descriptions of the plotted axes and error bars, if any.","section":"Figure 2 and Appendix B"},{"comment":"Table 2 reports KNE results on two models but does not include baseline comparisons for GPT-J-6B, so it only shows KNE's own consistency across models and cannot support cross-model claims about relative performance.","section":"Table 2"},{"comment":"The paper states that baseline results were sourced from the EasyEdit GitHub repository, but no repository version, commit, or experiment configuration is given; please provide exact dataset splits and baseline hyperparameter details for reproducibility.","section":"Appendix A.3"}],"recommendation":"major_revision","confidential_remarks":"The paper is not ready for publication in its current form because the method cannot be run as written and the key attribution claim is untested against a random-parameter control. These issues are fixable in principle, but the editors should require a corrected update rule, full hyperparameters, the 200-item selection procedure, and a random/identity control experiment before sending the paper out for another round of review. I would also encourage the authors to release code, since the current text alone does not permit verification of the reported numbers."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is an incremental but reasonable locate-then-edit paper, and the most interesting thing in it is not the method but the empirical observation about which FFN layers you edit. As written, though, no one can reproduce it: Eq. (10) uses column indexing where row indexing is needed, and p, alpha, and m are never specified.\n\nWhat's actually new: combining gradient attribution (Dai et al. 2022) with a masked gradient update restricted to the top-scoring FFN parameters, and the finding that editing mapping layers (gate_proj/up_proj) gives better locality than editing value layers. The partial-localization experiment (1/4 of the dataset suffices) is also a useful datapoint. Those are worth taking seriously.\n\nThe soft spots are proportional. The dimension mismatch is likely a typo—change Delta W[:, Mkne] to Delta W[Mkne, :]—but it's still a sign the manuscript wasn't checked before posting. Missing hyperparameters and missing code/data make the headline numbers unverifiable. Single-point results with no error bars are common in this subfield but they do weaken any claim of 'significantly improves.'\n\nThe stress-test concern is the one I'd push hardest: the paper never contrasts the attributed ensemble against a random same-size set of parameters. Since the update is just gradient descent restricted to a slice of FFN weights, any slice of the same cardinality might produce similar edit success and portability. Without that control, Table 1 does not establish that the localization is doing the work. That's not fatal to the method—the empirical pattern might survive—but it means the central claim is unverified, and the ablation in 4.5 only varies the count, not the selection.\n\nOn circularity: I don't think that's a problem here. Portability and locality are measured on held-out inputs, not on the attribution set.\n\nWho's this for: people working on locate-then-edit knowledge editing. They'll get value from the mapping-layer result and the batch-editing exploration. I'd send it to a serious referee, but only with a request for code, hyperparameters, corrected equations, and the random-selection control. If a revision provides those, it could be a solid workshop or second-tier conference paper.","headline":"A plausible locate-and-edit variant that can't be reproduced as written, but the mapping-layer finding and the missing random-control critique are the two things to argue about.","tokens_in":13213,"tokens_out":3180,"would_cite":false,"duration_ms":31493,"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 gradient-attribution scores can pick a sparse set of neurons—around 1% of parameters—whose update edits a factual association in a large language model while keeping unrelated knowledge intact.","keywords":["knowledge editing","large language models","knowledge neuronal ensemble","gradient attribution","parameter localization","locate-then-edit","feed-forward networks","batch editing"],"falsifier":"Run the KNE edit with the same number of neurons chosen from the lowest attribution scores instead of the highest; if edit success and locality stay roughly unchanged, the attribution ranking is not doing the causal work. A second test is to compare an edit that updates only neurons above the threshold against an edit that updates the same count of parameter positions chosen at random, where the claim predicts a decisive gap.","tokens_in":12246,"feed_emoji":"🧠","tokens_out":4978,"duration_ms":48272,"temperature":0.7,"pith_summary":"The paper proposes Knowledge Neuronal Ensemble (KNE), a locate-then-edit method for changing a large language model's factual knowledge without retraining the whole model. KNE computes a gradient-attribution score for every neuron in every layer, keeps the top-scoring ~1% of neurons as a 'knowledge neuronal ensemble,' and backpropagates through only those parameters. The paper claims this yields higher edit success than five baselines on ZsRE, WikiData counterfact, and WikiData recent, with portability and locality that match or exceed the best baselines on most benchmarks. A secondary claim is that factual knowledge is not confined to FFN value layers; editing mapping layers gives better locality, and localizing only a quarter of the facts can suffice.","feed_headline":"Gradient scores pick the 1% of neurons to edit LLM facts","feed_subtitle":"A locate-then-edit method claims 97-99% edit success across three benchmarks while updating only ~1% of parameters.","key_machinery":"The load-bearing object is the 'knowledge neuronal ensemble' (KNE): the set of neurons whose token-level gradient-attribution score, summed over answer tokens and approximated by a Riemann sum, lies above the top-1-p% quantile across all layers. The paper uses these scores to build a sparse update matrix $W_{\\text{kne}}$, maps it into a zero matrix $\\Delta W$ at the selected indices, and updates the weights with $W \\leftarrow W + \\frac{\\alpha}{\\sqrt{n}} \\Delta W$. The dynamic-interaction claim rests on computing losses and gradients only over this ensemble and backpropagating through it.","core_discovery":"The central claim is that a small set of neurons selected by gradient-attribution scores, rather than a fixed layer, can act as a distributed ensemble that carries a fact, and that updating only this ensemble with gradients flowing across layers edits the fact while preserving unrelated knowledge. On the three benchmarks the paper reports Edit Success of roughly 97-99%, Portability of 53-63%, and Locality that is competitive with or better than the baselines on two of the three datasets. The paper also argues that knowledge storage is more distributed than the key-value memory view: editing mapping layers (gate and up projections) can match or beat editing value layers on locality and fluency.","pith_inferences":["Editorial inference: If the localization premise holds, KNE's attribution scores could be reused as a diagnostic probe, and the overlap between ensembles for semantically close facts could predict whether batch edits will conflict.","Editorial inference: The finding that only 200-300 localized facts suffice hints that similar facts share storage regions; an external test would be to perturb one region and measure whether all related facts change together.","Editorial inference: The method's reliance on answer-token-level gradients suggests it may extend to low-resource or multilingual settings where only a few correct answer tokens are available, though the paper does not test this."],"forward_implications":["If the reported numbers hold, a 7-billion-parameter model can have a single fact changed with roughly 1% of weights updated, making repeated edits cheap enough for ongoing deployment.","The partial-localization result implies that a model may not need every fact localized before editing; a representative subset of similar facts can locate the region, cutting localization cost by about 75%.","Batch editing is possible with modest degradation as batch size grows, which extends the method beyond single-fact editors.","The layer analysis suggests practitioners can trade edit success for locality by choosing which projection to edit, with mapping layers offering safer localized edits.","The method's low parameter footprint could make knowledge editing practical on resource-constrained or on-device models."],"supporting_citations":[{"why":"Supplies the formal definition of knowledge editing as a constrained optimization and the evaluation framing used throughout the paper.","marker":"[1]"},{"why":"Provides the token-level gradient-attribution method whose scores KNE thresholds to select its neuron ensemble.","marker":"[6]"},{"why":"Establishes the key-value memory view of FFN layers that KNE builds on and partially revises.","marker":"[7]"},{"why":"Supports the assumption that FFN layers build predictions by promoting concepts in the vocabulary space.","marker":"[8]"},{"why":"Serves as a locate-then-edit baseline using causal mediation and a rank-one update; KNE compares against it directly.","marker":"[22]"},{"why":"Serves as the batch-editing baseline that extends ROME; KNE claims better portability and locality on the tested datasets.","marker":"[23]"},{"why":"Provides the AdaLoRA parameter-efficient fine-tuning baseline used in the comparison tables.","marker":"[26]"}],"fun_headline_variants":["KNE edits LLM facts by targeting a tiny neuron ensemble","Gradient scores find the ~1% of neurons that encode a fact","Distributed neuron ensemble enables precise LLM fact edits","KNE: Editing LLM knowledge via ~1% neuron ensemble","Neuron ensemble selected by gradients yields 97-99% edit success"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything reported depends on the premise that the gradient-attribution score actually locates the parameters that causally carry the fact, so that editing the top-scoring ~1% is both sufficient to install the new fact and safe for other knowledge.","fun_headline_variants_meta":{"raw":{"variants":["KNE edits LLM facts by targeting a tiny neuron ensemble","Gradient scores find the ~1% of neurons that encode a fact","Distributed neuron ensemble enables precise LLM fact edits","KNE: Editing LLM knowledge via ~1% neuron ensemble","Neuron ensemble selected by gradients yields 97-99% edit success"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000779,"raw_usage":{"total_tokens":3404,"prompt_tokens":866,"completion_tokens":2538,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":482,"completion_tokens_details":{"reasoning_tokens":2449}},"tokens_in":482,"tokens_out":2538,"duration_ms":15904,"temperature":1.0,"reasoning_tokens":2449,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T23:14:37.103205+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the KNE edit with the same number of neurons chosen from the lowest attribution scores instead of the highest; if edit success and locality stay roughly unchanged, the attribution ranking is not doing the causal work. A second test is to compare an edit that updates only neurons above the threshold against an edit that updates the same count of parameter positions chosen at random, where the claim predicts a decisive gap.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the token-level gradient-attribution method whose scores KNE thresholds to select its neuron ensemble."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the key-value memory view of FFN layers that KNE builds on and partially revises."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supports the assumption that FFN layers build predictions by promoting concepts in the vocabulary space."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Serves as a locate-then-edit baseline using causal mediation and a rank-one update; KNE compares against it directly."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Serves as the batch-editing baseline that extends ROME; KNE claims better portability and locality on the tested datasets."},{"cited_title":"Zhang, M","cited_arxiv_id":null,"evidence_quote":"Provides the AdaLoRA parameter-efficient fine-tuning baseline used in the comparison tables."}],"review_version":1}