{"id":"c431ea54-d546-4c0e-91de-ca0c1ae973ef","arxiv_id":"2607.17563","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"FlexiGrad selectively removes conflicting and reinforces agreeing gradient components between hierarchy levels, improving multi-granularity accuracy on CUB, FGVC-Aircraft and Stanford Cars.","lead":"The paper proposes FlexiGrad, a way to adjust gradients during training so that coarse and fine labels in image recognition don't fight each other. It reports accuracy gains on three common bird/aircraft/car datasets with no added parameters.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Algorithm 1 and §3.5 define different update rules; §3.6's sum-descent proof is invalid. Without code, the reported gains cannot be attributed to a single, well-defined method.","rationale":"The reader's verdict is CONDITIONAL, and our concern supports keeping the paper conditional: the missing code and seeds are important, but the algorithm/text mismatch is an internal inconsistency that strengthens the conditionality. The reader already flagged that §3.6 proves per-task descent only, which matches our second point. However, the reader's weakest assumption focused on optimizer sensitivity and unreported training details, whereas we identify the unresolved algorithm specification as the more fundamental issue. A concrete reproducibility check—comparing the two algorithm variants—would settle whether the reported gains come from a well-defined FlexiGrad update or from an artifact of the pseudocode's asymmetric ordering. We do not escalate to REJECT because either variant might still yield the improvements, but the paper must disambiguate the method and provide code before the central claim can be fully assessed.","tokens_in":9781,"tokens_out":12448,"duration_ms":100721,"concrete_test":"Implement both the asymmetric Algorithm 1 (j>i only) and the symmetric §3.5 version (all j≠i) on CUB-200-2011 with ResNet50, SGD, and 100 epochs as reported. Compare the average multi-granularity accuracy against the claimed 89.19% and against each other. If the two variants differ by more than noise or neither matches the reported value, the paper's method is not uniquely defined and the central claim is not attributable to a specific algorithm. Additionally, run a 2-task toy search to check whether any FlexiGrad update yields (Σ g_FG^i)·(Σ g_i) < 0, falsifying the §3.6 descent claim.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim rests on a specific gradient-modulation rule, but the paper describes two incompatible versions. Algorithm 1 (lines 3–15) updates g_i only against tasks j with k>i, leaving the last task's gradient unmodified. Section 3.5 states each task interacts with 'all other tasks j≠i', which would be symmetric. These two rules are not equivalent, and the reported numbers do not clarify which was used. If the asymmetric version was used, the improvements may depend on the arbitrary ordering of tasks (e.g., the species gradient is never modulated, yet species accuracy improves most). If the symmetric version was used, the pseudocode is wrong. Either way, the method is not uniquely specified, so the empirical claim cannot be reproduced or independently verified.\n\nAdditionally, the theoretical justification in §3.6 is logically unsound. The paper argues that because each corrected gradient satisfies g_FG^i · g_i ≥ 0 (Eq. 9–10), the sum Σ g_FG^i is a 'principled surrogate descent for L_total'. This does not follow: the dot product of the sum with the total gradient involves cross-terms g_FG^i · g_j that are unexamined. A simple two-task example (e.g., g1=(1,0), g2=(-2,0), with g_FG^1=g1 and g_FG^2=0) has each nonnegative but the sum has negative dot with g1+g2. Thus the proof in §3.6 is invalid as written, weakening the claimed theoretical motivation.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FlexiGrad, a parameter-free gradient-modulation method for hierarchical fine-grained classification. Given task gradients from coarse and fine classifiers, FlexiGrad measures pairwise cosine similarity, removes the conflicting component when gradients disagree, and mixes the gradient with a projected component using weight w_ij=(1+cosθ_ij)/2 when they partially agree. The corrected per-task gradients are summed and used in place of the raw multi-task gradient. Experiments on CUB-200-2011, FGVC-Aircraft, and Stanford Cars with ResNet-50, Swin-B, ViT-B/16, and several FGVC models report consistent accuracy gains over Vanilla single, PCGrad, and FGoN, with ablations and visualizations supporting the claimed behavior.","tokens_in":10140,"tokens_out":16447,"duration_ms":123327,"significance":"If the exact update is specified and the empirical results are reproducible, FlexiGrad would be a useful, simple, and architecture-agnostic contribution to hierarchical FGVC. It is genuinely parameter-free apart from a small epsilon, it does not alter the network structure, and the reported gains are consistent across datasets and backbones. The core idea of treating coarse-to-fine gradient interactions asymmetrically and continuously, rather than symmetrically projecting or blocking them, is well motivated. The paper also gives credit for reporting standard deviations over three runs and for ablating the two components. However, two load-bearing issues currently prevent acceptance: the algorithm is not uniquely specified, and the theoretical descent-direction claim is not proven.","major_comments":[{"comment":"Algorithm 1 (lines 3–15) defines a different update from the one described in §3.5. In the pseudocode the inner loop is over tasks with k > i, so the last task's gradient is never modified (g_FG^K = g_K) and the ordering of tasks matters. §3.5 states that each task 'sequentially applies the conflict or cooperation rule with all other tasks j ≠ i', which is symmetric. These two rules are not equivalent. The numbers in Tables 1–4 do not reveal which rule was used, and even the sequential version is ambiguous about whether the numerator in lines 8 and 12 uses the raw g_i or the already-updated g_FG^i. This makes the empirical claim unreproducible as stated. Please correct the pseudocode or the text, specify the exact rule, justify the ordering, and release code so the reported gains can be attributed to a single method.","section":"§3.5 / Algorithm 1"},{"comment":"The proof does not establish the stated conclusion. Eqs. (9)–(10) only show g_FG^i · g_i ≥ 0. To conclude that g_FG = Σ_i g_FG^i is a descent direction for L_total, one needs Σ_i g_FG^i · g_i + Σ_{i≠j} g_FG^i · g_j ≥ 0. The cross terms are not examined and can be negative. As a logical illustration, take g1=(1,0), g2=(-2,0), g_FG^1=g1, g_FG^2=0; then both per-task inner products are nonnegative but g_FG·(g1+g2) = -1. Therefore the surrogate-descent claim is not established. Provide a proof that controls the cross terms using the actual projection rule, or remove the theoretical claim and present §3.6 as intuition.","section":"§3.6, Eqs. (9)–(10)"}],"minor_comments":[{"comment":"When cos(θ_ij)=0, Eq. (7) gives g_coop_i = 0.5 g_i because the projection term vanishes. Orthogonal gradients are therefore halved, not 'reinforced'. This magnitude shrinkage is not discussed and may be part of why the method works; please analyze or at least acknowledge it.","section":"§3.4 / Algorithm 1"},{"comment":"The experiments are described as using SGD for 100 epochs, but the Figure 2 caption refers to 'the standard Adam optimizer' and 'Adam with PCGrad / FlexiGrad'. Please clarify which optimizer is used in the actual experiments and in the visualization.","section":"§4.1 / Fig. 2"},{"comment":"The epsilon in the denominator of the projection is never specified or ablated. Since the paper emphasizes that FlexiGrad is parameter-free, please state the value of epsilon and explain why it does not affect the qualitative behavior.","section":"Algorithm 1"},{"comment":"For reproducibility, please report random seeds, exact augmentation details, dataset split conventions, and learning-rate schedules. The statement 'standard augmentations' is too vague for a method whose reported gains are 0.5–1.0%.","section":"§4.1"},{"comment":"The accuracy-vs-epoch figure lacks axis labels and a legend entry for 'Vanilla single'. Please improve the figure captions so the visualization can be interpreted independently.","section":"Fig. 5"}],"recommendation":"major_revision","confidential_remarks":"The core idea is promising, but in its current form the method is not uniquely specified and the theoretical section overclaims. I would like to see a revised version with a corrected algorithm/theory and either code or a detailed implementation description. The gains over PCGrad are modest but consistent; once the ambiguities are resolved, this could be a useful contribution to the hierarchical FGVC literature."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I read FlexiGrad. The short version: it's PCGrad with an extra cooperative term, applied to hierarchical fine-grained classification. The empirical results are consistently positive (about +1% average accuracy) across three benchmarks and several backbones, and the ablations suggest both components contribute. But the paper has a load-bearing ambiguity: Algorithm 1 updates each task only against tasks with larger index, leaving the last gradient unmodified and making the update order-dependent, while §3.5 says each task interacts with all others j≠i. These are different rules, and the paper never says which one produced the numbers. That alone is enough to make the method not uniquely specified.\n\nThe theory in §3.6 is also invalid. The authors show that each corrected gradient has non-negative dot with its own raw gradient, and then claim the sum is a surrogate descent for L_total. That step needs each g_FG_i to have non-negative dot with every other g_j, or something stronger. Cross terms are unexamined. A simple counterexample (g1=(1,0), g2=(-2,0), with g_FG^2=0) has each non-negative but the sum ascending. So the theoretical motivation is wrong as written.\n\nWhat's actually new: the cooperative weighting w=(1+cos)/2 is a small but genuine variant of PCGrad, and the application to hierarchical FGVC with hierarchical labels is reasonable. The plug-in to four state-of-the-art FGVC models is a nice touch. The visualizations are suggestive, though not quantitative proof.\n\nOther soft spots: no code, no seeds, no full training details. Given the method's dependence on the exact update rule, that's a reproducibility problem. Also, the claimed gains are modest (~1%); not a breakthrough, but if the method is well-defined and reproducible, it's useful for this niche.\n\nI'd send it to peer review rather than desk reject, because the idea is simple and the experiments are broad. But a referee should push hard on the algorithm inconsistency and the descent proof. As is, I wouldn't accept it without major revision and code release.\n\nIf you work on gradient surgery or hierarchical classification, it's worth a quick read as a cautionary example of a promising idea that needs rigor. I'd probably not cite it in its current form.","headline":"PCGrad plus a cosine-weighted cooperation branch; consistent but modest gains, yet the update rule is ambiguous and the theory is a non sequitur.","tokens_in":10639,"tokens_out":6207,"would_cite":false,"duration_ms":51472,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"FlexiGrad, a parameter-free gradient-modulation rule, removes only the conflicting coarse–fine component and reinforces partial agreement, lifting multi-granularity accuracy on three fine-grained benchmarks.","keywords":["multi-granularity classification","fine-grained visual classification","hierarchical labels","gradient conflict","gradient surgery","parameter-free","coarse-to-fine learning","multi-task learning"],"falsifier":"On the three reported benchmarks, train with the same schedule but replace the cosine-based weights with a fixed constant w=0.5 (or with randomly shuffled pairwise cosines). If the accuracy gain over vanilla joint training matches FlexiGrad's, the alignment-sensitivity of the rule is not what drives the improvement. A second check: train with plain SGD instead of Adam; the theory only certifies descent for the aggregate gradient, so if the gain vanishes under SGD, the mechanism's causal claim would be in doubt.","tokens_in":9658,"feed_emoji":"🎯","tokens_out":11819,"duration_ms":87241,"temperature":0.7,"pith_summary":"This paper argues that the instability of joint coarse-and-fine training comes from gradient conflicts on the shared backbone, not from the labels themselves. It introduces FlexiGrad, a parameter-free backpropagation-time rule: when gradients from two granularity levels point in opposite directions, it subtracts the projection of one onto the other; when they partially agree, it blends them with a linear weight (1+cos θ)/2. The resulting update is a surrogate descent direction for the total objective, and experiments on CUB-200-2011, FGVC-Aircraft and Stanford Cars report consistent accuracy gains at every granularity, with the largest gains on the finest level and on hard taxonomic subtrees. The practical payoff is a drop-in optimisation fix that works across CNN and transformer backbones without modifying the architecture.","feed_headline":"One gradient rule lifts accuracy at every taxonomic level","feed_subtitle":"Removing only the conflicting gradient part lifts accuracy on birds, aircraft and cars without architecture changes.","key_machinery":"The central mechanism is the pairwise cosine similarity cosθ_ij between the gradients of two granularity tasks. When the cosine is negative, FlexiGrad subtracts the projection of one gradient onto the other, removing only the conflicting component; when it is positive, it blends the two gradients with a smooth linear weight (1+cos θ_ij)/2, reinforcing the shared direction while preserving each task's individual component. The corrected gradients are summed to form the final parameter update, so the method intervenes purely at backpropagation time and needs no architecture changes or extra parameters.","core_discovery":"The paper's central claim is that the naive sum of coarse and fine task gradients is suboptimal because it treats all cross-level interactions as equally compatible. FlexiGrad replaces it with a hierarchy-aware update: for each pair of task gradients, on conflict (negative cosine) the projection of one onto the other is removed; on partial agreement the gradient is re-weighted as a convex combination of itself and its projection, using w_ij = (1+cos θ_ij)/2. The paper shows each corrected gradient keeps a non-negative inner product with its original, so the sum remains a descent direction for the total loss. Empirically, on three benchmarks and across four base models, this yields higher mul","pith_inferences":["The same pairwise alignment rule should extend to hierarchies deeper than two levels and to any ordered set of tasks, since the procedure is already sequential and pairwise; a testable variant would apply it to kingdom–phylum–class–order taxonomies and check whether per-level gains compound.","The paper's descent-direction guarantee covers only the aggregate gradient, so the method's benefit under adaptive optimisers could partly be a learning-rate interaction; running FlexiGrad with plain SGD under the same schedule would isolate the mechanism's contribution.","Because the weight w=(1+cos θ)/2 assumes gradients are on comparable scales, per-task loss scaling could change the outcome; a whitened-gradient variant would test whether normalisation is needed for very different loss magnitudes."],"forward_implications":["Joint training with FlexiGrad reports higher accuracy at every granularity than vanilla joint training, with the largest absolute gains at the finest label level (species or model).","FlexiGrad is a drop-in backpropagation module: no architectural change, no extra parameters, and only a small training-time overhead (about 7% in the reported CUB setting).","The gains persist across backbone families: the paper reports improvements for both CNN-based and transformer-based fine-grained models when FlexiGrad is plugged in.","The improvement is largest precisely where coarse-to-fine interference is strongest: hard families and subtrees show the biggest species-level gains, up to about 19 percentage points on CUB in the paper's analysis.","Because the corrected update is a surrogate descent direction for the total loss, the method can be combined with any optimiser and any hierarchical head setup without changing the objective."],"fun_headline_variants":["Remove conflicting gradients only, lift accuracy at every level","One gradient rule to settle hierarchy conflicts in fine-grained tasks","FlexiGrad: adaptive gradient modulation for stable coarse-to-fine learning","Stop gradient clashes: new method boosts bird, car, and aircraft recognition","Smooth hierarchy-aware gradients improve multi-granularity classification"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that coarse and fine supervision are best treated as two separate tasks whose harmful interactions can be repaired by pairwise gradient projections; if the reported gains actually come from Adam-specific dynamics or from an accidental regularisation effect of the reweighting, the 'hierarchical gradient conflict' explanation would not be the cause.","fun_headline_variants_meta":{"raw":{"variants":["Remove conflicting gradients only, lift accuracy at every level","One gradient rule to settle hierarchy conflicts in fine-grained tasks","FlexiGrad: adaptive gradient modulation for stable coarse-to-fine learning","Stop gradient clashes: new method boosts bird, car, and aircraft recognition","Smooth hierarchy-aware gradients improve multi-granularity classification"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000235,"raw_usage":{"total_tokens":1308,"prompt_tokens":683,"completion_tokens":625,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":427,"completion_tokens_details":{"reasoning_tokens":537}},"tokens_in":427,"tokens_out":625,"duration_ms":5885,"temperature":1.0,"reasoning_tokens":537,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T17:35:23.789704+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On the three reported benchmarks, train with the same schedule but replace the cosine-based weights with a fixed constant w=0.5 (or with randomly shuffled pairwise cosines). If the accuracy gain over vanilla joint training matches FlexiGrad's, the alignment-sensitivity of the rule is not what drives the improvement. A second check: train with plain SGD instead of Adam; the theory only certifies descent for the aggregate gradient, so if the gain vanishes under SGD, the mechanism's causal claim would be in doubt.","supporting_citations":[],"review_version":1}