{"id":"2bfa6218-2481-48f8-bfc1-5fded58c4ce2","arxiv_id":"2501.15065","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"TATR improves multi-task model merging by restricting task-vector additions to parameter dimensions that are orthogonal to task-specific loss gradients.","lead":"The paper introduces TATR, a training-free model merging method that masks out parameter dimensions where task vectors collide with task-loss gradients. It reports consistent multi-task accuracy gains over task arithmetic and related baselines on eight image classification datasets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No random-mask or magnitude-mask control: TATR's gains may not be attributable to gradient-orthogonality, and the first-order justification is self-undermined.","rationale":"The reader's CONDITIONAL verdict is appropriate. My pass does not move it: the missing control and the first-order tension both point to the same unresolved attribution question. If anything, the random-mask test is more decisive than the theoretical objection because it tests the causal mechanism directly. I keep CONDITIONAL rather than REJECT because the reported gains, if reproducible and robust to the control, would support the method, and the theory weakness alone does not falsify the empirical claim. I flag partial agreement: the reader locates the problem in the first-order Taylor justification; I locate it in the absence of a control that isolates the mask's selection criterion. They are complementary versions of the same load-bearing assumption, namely that |grad L (dot) Delta| specifically identifies conflict dimensions.","tokens_in":20894,"tokens_out":6995,"duration_ms":73615,"concrete_test":"Re-run the ViT-B/32 eight-task benchmark exactly as in Tables 1/3, with the same tau grid, but replace the TATR mask with (i) a random mask that zeroes the same tau fraction of dimensions (10 seeds) and (ii) a magnitude mask that zeroes the top-tau% dimensions by |Delta_i|. If either control reaches average accuracy within one standard error of TATR's 72.8%, the gradient-based trust-region selection is not the operative cause of the gains; if both remain near the 69.1% Task Arithmetic baseline, the gradient-specific mask is doing real work.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central causal claim is that the dimensions excluded by the trust-region criterion (Eq. 8, Algorithm 1) are precisely the conflict-inducing ones, so discarding them produces the observed gains. The paper compares TATR against no-mask Task Arithmetic and against global sensitivity variants (positive/negative/NTK/zero-shot, Table 6), but never against a control that discards the same fraction tau of dimensions randomly or by parameter magnitude. This control matters because the best tau is below 1% (Figure 3b, Section 6.3): TATR changes only a tiny fraction of parameters, and any small structured perturbation of the task sum could shift accuracy by a few points. Without such a control, the reported +3.7 points on ViT-B/32 cannot be attributed to gradient-orthogonality rather than to generic pruning. The theory also does not resolve this: Section 4 explicitly states that first-order Taylor expansion fails for large task vectors, yet Eq. 4/8 use the same first-order product (with absolute values) to rank dimensions. The sign collapse from signed to absolute product is a heuristic, not a consequence of the analysis. Hence the method's causal story is underdetermined by the present experiments, independently of whether the empirical gain is real.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies knowledge conflicts in Task Arithmetic (TA) model merging, defining a conflict as the degradation in one task's loss when another task's fine-tuned model is included in the merged model. The authors argue, via a first-order Taylor expansion around the pre-trained weights, that conflict originates mainly from task-vector components aligned with the gradient of the task-specific loss. They propose TATR, a training-free mask that keeps only parameter dimensions for which the accumulated product |\\nabla L_j(\\theta_{pre})| \\odot |\\Delta_i| is small, discarding the remaining dimensions before merging. TATR is evaluated as a standalone method and as a plug-in for Ties-Merging, AdaMerging, and Surgery on eight image-classification datasets with CLIP ViT-B/32, ViT-B/16, and ViT-L/14, reporting consistent average-accuracy gains (e.g., 72.8% vs. 69.1% for Task Arithmetic on ViT-B/32). The paper also includes a zero-shot variant that approximates the gradient by the task vector magnitude, loss-landscape visualizations, and ablations of positive, negative, NTK-based, and zero-shot sensitivity definitions.","tokens_in":21116,"tokens_out":5182,"duration_ms":49918,"significance":"If the empirical claims hold, TATR is a simple, inexpensive, plug-and-play masking scheme with consistent gains across multiple backbones and datasets. The one-shot and zero-shot variants make it particularly attractive for practical model merging, and the ablation in Table 6, where positive and negative masks fail catastrophically, provides meaningful evidence that the sign of the masking criterion matters. However, the central causal claim---that gradient-orthogonal dimensions are precisely the safe ones---is not yet convincingly established. The missing random-mask and magnitude-mask controls, together with the acknowledged failure of first-order Taylor approximation in the very regime where the mask is defined, leave the theoretical justification underdetermined. The contribution is therefore best viewed as a promising empirical heuristic whose mechanism requires further validation.","major_comments":[{"comment":"The reported gains are not contrasted with a control that discards the same fraction tau of dimensions randomly or by parameter magnitude. Since the selected tau is below 1% (Figure 3b), TATR alters only a tiny fraction of the model, and any small structured perturbation of the task-sum could shift accuracy by a few points. Without random-mask and magnitude-mask baselines matched to the same tau, the +3.7 point improvement over Task Arithmetic on ViT-B/32 cannot be attributed specifically to gradient-orthogonality rather than to generic sparsity. This control is essential for the paper's central claim.","section":"Section 6.2, Table 1; Algorithm 1"},{"comment":"The paper states that the first-order Taylor expansion fails for large task vectors because the negative component overshoots local optima, yet the masking criterion in Eq. (8) is precisely the first-order product |\\nabla L_j(\\theta_{pre})| \\odot |\\Delta_i|, with the sign collapsed by taking absolute values. The move from the signed scalar product in Eq. (4) to the absolute per-dimension product in Eq. (8) is an additional heuristic, not a consequence of the Taylor analysis. If first-order sensitivity is unreliable in the regime of real task vectors, the theory does not explain why the absolute first-order product should correctly rank conflict-inducing dimensions. Please either provide a theoretical justification for the absolute-value step or add an experiment that directly compares the predictions of the signed and absolute first-order criteria, e.g., by showing that the signed version fails precisely on dimensions that the absolute version correctly excludes.","section":"Section 4, Eq. (4) and Figure 2(b); Section 5, Eq. (8) and Algorithm 1"},{"comment":"The hyperparameter tau is described as 'tuned' over [0.1%, 0.2%, 0.5%, 1.0%, 2.0%, 5.0%], but the manuscript does not specify the selection protocol, in particular whether a validation split is used or whether the reported test accuracies are the best over the grid. If tau is selected on the test data, the headline numbers incorporate fitted selection and the 'training-free' characterization is weakened. Please state the selection criterion explicitly, report results for a fixed tau (e.g., tau = 0.5%) as a sensitivity check, and clarify the relationship between the choice of tau and the test-set accuracies shown in Figure 3(b).","section":"Section 6.1, Implementation details; Figure 3(b)"}],"minor_comments":[{"comment":"The positive and negative sensitivity variants are defined with an average over task pairs before thresholding; because positive and negative signed products can cancel in this average, the catastrophic failure of these variants does not by itself establish that every dimension's signed product is harmful. Please report, or at least discuss, the signed product without the averaging.","section":"Table 6"},{"comment":"Knowledge conflict is defined through loss differences, while all experimental results report accuracy. The connection between the two (e.g., using cross-entropy loss and noting that lower loss is expected to correspond to higher accuracy) should be stated explicitly.","section":"Definition 1, Eq. (3)"},{"comment":"The expectation is placed outside the absolute value in the gradient estimator, which is nonstandard for Fisher-type estimates; since |E[g]| <= E[|g|] and the two estimators can rank dimensions differently, this design choice should be justified or at least discussed.","section":"Eq. (6)"},{"comment":"The figure is not legible in the provided version, and the axis labels and layer indexing are unclear; the text's claims about periodic sensitivity and bias-layer sensitivity cannot be verified from the figure as rendered. Please redraw with clear axes and legends.","section":"Figure 4"},{"comment":"The text states that excluding less than 1% of parameters is sufficient, but it does not state whether the same tau is used across all datasets and backbones or how the value was chosen for the main tables; please make this explicit.","section":"Section 6.3, Figure 3(b)"},{"comment":"The relationship between TATR's trust-region mask and existing magnitude-based pruning in Ties-Merging is not discussed; given the small tau, a comparison with low-magnitude pruning at the same sparsity level would help the reader understand the novelty of the gradient-based criterion.","section":"Section 2.2, Related Work"}],"recommendation":"major_revision","confidential_remarks":"The empirical direction is promising, and the Table 6 ablation showing that positive and negative masks fail is a strong piece of evidence that the sign of the masking criterion matters. However, the missing random- and magnitude-mask controls and the self-undermined first-order justification are load-bearing. I would be willing to accept after the authors add the controls, clarify the tau-selection protocol, and either repair or substantially temper the theoretical claim. The zero-shot gradient approximation in Eq. (7) is quite bold and may attract criticism; the authors should calibrate their claims accordingly."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a genuinely useful empirical paper with a half-explained trick. TATR masks out parameter dimensions where the product of absolute gradient and absolute task-vector component is large, then does task arithmetic on the rest. It works: +3.7 average points on ViT-B/32 over plain task arithmetic, +0.8 on ViT-L/14, and it ports cleanly to Ties-Merging and AdaMerging. The positive/negative/orthogonal decomposition is a real conceptual contribution, and the ablation showing that merging the components aligned with gradient descent collapses performance is a nice counterintuitive result. I trust the accuracy numbers.\n\nThe soft spots, in order. First, there is no control that discards the same fraction of dimensions randomly or by raw magnitude. Since the best tau is below 1%, the mask touches very few parameters, and a small generic prune could easily shift accuracy by a point or two. Without that control, the causal claim that gradient orthogonality is what matters is underdetermined. The paper does compare against NTK and zero-shot sensitivity variants, which helps, but it doesn't close the gap. Second, the theory is internally strained: Section 4 argues that first-order Taylor expansion fails for large task vectors, then Equation 8 uses exactly that first-order product, with absolute values, to build the mask. The absolute value is a reasonable heuristic and the empirical curves make sense, but it is not a consequence of the analysis. Third, tau is tuned on the test set and no error bars are reported. That is common in this literature, but it means the headline gains should be read as optimistic bounds. No code is provided either.\n\nWho is this for? Anyone who wants a cheap, training-free plug-in to squeeze an extra couple of points out of task arithmetic. The paper deserves a serious referee: the idea is new, the experiments are mostly well done, and the flaws are addressable rather than fatal. I can't fully endorse the theoretical story, and I'd want the random-mask control before treating the mechanism as established, but the empirical result merits engagement.","headline":"A training-free mask on task vectors that buys a few accuracy points, with a plausible but under-tested story about why it works.","tokens_in":21655,"tokens_out":2335,"would_cite":false,"duration_ms":22649,"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 shows that knowledge conflicts in model merging come from task-vector components aligned with task-loss gradients, and that a trust-region mask fixing this improves merged-model accuracy.","keywords":["model merging","task arithmetic","knowledge conflict","trust region","task vector","gradient orthogonality","training-free","multi-task learning"],"falsifier":"Compare, for each task pair, the true change in task j's loss when task i's vector is added against the first-order prediction $\\langle \\nabla L_j(\\theta_{\\text{pre}}), \\Delta_i \\rangle$; TATR's mask is only justified if small predicted products correspond to small true loss changes. Alternatively, apply TATR to a small, heavily trained model with little parameter redundancy where near-orthogonal gradient directions are rare, and check whether the accuracy gains over Task Arithmetic disappear or reverse.","tokens_in":20655,"feed_emoji":"🧩","tokens_out":4576,"duration_ms":36874,"temperature":0.7,"pith_summary":"When multiple fine-tuned models are combined by adding their task vectors to a pre-trained model, the tasks can interfere, degrading performance. The paper defines this as 'knowledge conflict' and traces it to task-vector components that align with the gradient of each task's loss at the pre-trained weights. It proposes Task Arithmetic in the Trust Region (TATR), a training-free method that merges only the parameter dimensions where the product |gradient × task vector| is small, i.e., directions nearly orthogonal to the gradient. On eight image-classification datasets, TATR raises Task Arithmetic's average accuracy from 69.1% to 72.8% with a ViT-B/32 backbone, and also improves Ties-Merging, AdaMerging, and Surgery when used as a plug-in. The central claim is that discarding gradient-aligned components, rather than keeping gradient-descent ones, is what resolves cross-task interference.","feed_headline":"Masking gradient-aligned weights improves multi-task model merging","feed_subtitle":"A training-free trust-region mask lifts Task Arithmetic from 69.1% to 72.8% average accuracy across eight datasets.","key_machinery":"The central object is the trust-region mask defined by Eq. (5): a parameter dimension $n$ belongs to the trust region when $\\sum_{i \\neq j} |\\nabla_\\theta L_j(\\theta_{\\text{pre}})[n] \\cdot \\Delta_i[n]| < \\epsilon$, with $\\epsilon$ set by a proportion threshold $\\tau$ on the ranked sensitivity. The mask is applied by element-wise multiplication with the task vectors before adding them to the pre-trained weights (Algorithm 1). The mechanism exploits overparameterization: many parameter directions are near-orthogonal to task gradients, so a small mask can discard the conflicting aligned components without losing task-specific knowledge. Gradients are approximated from small exemplar sets or, in the zero-shot variant, by the task vectors themselves.","core_discovery":"Knowledge conflicts in task arithmetic are primarily caused by components of task vectors that align with the gradient of task-specific losses at the pre-trained model; even gradient-descent-aligned components harm performance because the first-order approximation overshoots local optima. Therefore the paper defines a trust region of parameter dimensions where the absolute product $|\\nabla_\\theta L_j(\\theta_{\\text{pre}})[n] \\cdot \\Delta_i[n]|$ is below a threshold, and merges only within it. TATR uses per-task exemplar samples to estimate gradients, ranks dimension sensitivity, masks away a small proportion of parameters (typically under 1%), and can be layered on top of existing task-arithmetic variants. Empirically this yields consistent accuracy gains across eight datasets and three backbones, with the largest gains when combined with test-time adaptation methods.","pith_inferences":["A natural extension is to apply the trust-region mask iteratively, recomputing gradients at the merged point; if the first-order assumption is weak, this could strengthen the mask beyond the paper's single-shot version.","The orthogonal-directions argument connects to parameter-level redundancy results; one could test TATR on smaller, less overparameterized models where orthogonal directions are scarce, which would clarify whether the mechanism is the stated orthogonality or just sparsification.","The method's success with one-shot exemplar sets suggests gradient estimation is robust, so a practical extension would be to use the mask for continual learning or model editing where only a few samples per task are available."],"forward_implications":["If the trust-region diagnosis is right, any task-arithmetic-style merge can be improved by masking gradient-aligned dimensions, without retraining.","The mask also reduces interference for test-time adaptation methods like AdaMerging and Surgery, so those methods can train their coefficients in a less conflicting parameter space.","The drop in accuracy from masking only ~1% of parameters suggests conflicts are concentrated in a small fraction of dimensions.","The zero-shot variant, using task vectors as gradient surrogates, offers a fully training-free path to conflict reduction when exemplars are unavailable."],"supporting_citations":[{"why":"Introduces task vectors and task arithmetic, the base method TATR modifies and the experimental protocol used.","marker":"Ilharco et al., 2023b"},{"why":"Ties-Merging, a baseline and a plug-in target whose pruning and sign operations TATR complements.","marker":"Yadav et al., 2023"},{"why":"AdaMerging, a test-time adaptation baseline and plug-in target; TATR's largest gains come with it.","marker":"Yang et al., 2024b"},{"why":"Representation Surgery, another test-time adaptation baseline and plug-in target.","marker":"Yang et al., 2024a"},{"why":"Provides the Fisher-information-style motivation for placing the expectation outside the absolute value in gradient estimation.","marker":"Wasserman, 2013"},{"why":"Tangent-space task arithmetic, an alternative attempt to reduce task-vector interference that frames the conflict problem TATR positions against.","marker":"Ortiz-Jimenez et al., 2023"}],"fun_headline_variants":["Task Arithmetic trust region lifts merging accuracy","Training-free mask resolves knowledge conflicts in model merging","Guard gradient-aligned weights to improve task merging","TATR: safe merging by masking sensitive parameters","Trust-region merging beats plain Task Arithmetic"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The trust-region mask is built from first-order gradient×task-vector products at the pre-trained weights, yet the paper itself shows that first-order Taylor expansions fail for full-size task vectors; if those products mis-rank the true cross-task interference, the mask discards the wrong dimensions.","fun_headline_variants_meta":{"raw":{"variants":["Task Arithmetic trust region lifts merging accuracy","Training-free mask resolves knowledge conflicts in model merging","Guard gradient-aligned weights to improve task merging","TATR: safe merging by masking sensitive parameters","Trust-region merging beats plain Task Arithmetic"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000173,"raw_usage":{"total_tokens":1314,"prompt_tokens":1014,"completion_tokens":300,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":630,"completion_tokens_details":{"reasoning_tokens":231}},"tokens_in":630,"tokens_out":300,"duration_ms":3331,"temperature":1.0,"reasoning_tokens":231,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T14:38:38.384102+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compare, for each task pair, the true change in task j's loss when task i's vector is added against the first-order prediction $\\langle \\nabla L_j(\\theta_{\\text{pre}}), \\Delta_i \\rangle$; TATR's mask is only justified if small predicted products correspond to small true loss changes. Alternatively, apply TATR to a small, heavily trained model with little parameter redundancy where near-orthogonal gradient directions are rare, and check whether the accuracy gains over Task Arithmetic disappear or reverse.","supporting_citations":[{"cited_title":"TIES -merging: Resolving interference when merging models","cited_arxiv_id":null,"evidence_quote":"Ties-Merging, a baseline and a plug-in target whose pruning and sign operations TATR complements."},{"cited_title":"All of statistics: A concise course in statistical inference, 2013","cited_arxiv_id":null,"evidence_quote":"Provides the Fisher-information-style motivation for placing the expectation outside the absolute value in gradient estimation."},{"cited_title":"Task arithmetic in the tangent space: Improved editing of pre-trained models","cited_arxiv_id":null,"evidence_quote":"Tangent-space task arithmetic, an alternative attempt to reduce task-vector interference that frames the conflict problem TATR positions against."}],"review_version":1}