{"id":"b671a097-3c08-4396-bd03-f660b4f913e1","arxiv_id":"2605.24969","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":5.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":2,"one_line_summary":"OSDTW decomposes long-tailed recognition into head and tail tasks with shared encoder and derives a Fisher-based bias-variance proxy to select shared depth and task weights.","lead":"The paper introduces OSDTW, a framework that splits long-tailed recognition into head and tail tasks using a shared encoder plus task-specific decoders, then uses a three-stage training process and a second-order error expansion to pick optimal shared depth and task weights. A smart generalist might read it for a more systematic approach to handling severe class imbalance in machine learning without purely heuristic tuning.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"Block-diagonal Fisher approximation may fail to capture shared-encoder cross-task interactions, invalidating the second-order proxy.","rationale":"The reader's weakest_assumption already isolates the block-diagonal Fisher step as the critical unverified link; the abstract-only review correctly flags the lack of derivation or validation details. No stronger internal inconsistency is visible from the given text, so the concern does not move the verdict away from UNVERDICTED.","tokens_in":1788,"tokens_out":334,"duration_ms":13574,"concrete_test":"On a CIFAR-10-LT (imbalance 100) model trained to the three-stage pipeline, compute the full empirical Fisher at the final checkpoint; report the Frobenius norm of all off-block entries divided by the sum of the block-diagonal norms. If the ratio exceeds 0.15, re-run the hyper-parameter search with the full (non-block) second-order term and compare top-1 accuracy of the proxy-selected configuration versus the block-diagonal one.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that the block-diagonal Fisher approximation yields a usable second-order expansion of expected generalization error (encoder variance + bias + decoder variance) that reliably ranks shared-depth and task-weight choices. The factorized model is invoked to justify additive KL terms and block-diagonality, yet the shared encoder parameters couple head and tail gradients; any residual off-block Fisher entries (arising from class dependence or encoder features) would make the decomposed proxy inaccurate for hyper-parameter selection. The abstract provides no quantitative bound on the approximation error or regime where it holds.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper claims that long-tailed recognition can be improved by decomposing the problem into head and tail tasks via a shared encoder and task-specific decoders. Under a factorized model, the KL-divergence generalization error decomposes additively into task-wise terms. A three-stage pipeline (independent task training, weighted joint training, branch assembly) is proposed, and under a block-diagonal Fisher approximation a second-order expansion of expected generalization error is derived that decomposes into encoder variance, encoder bias, and decoder variance; this expansion serves as a computable proxy for selecting shared depth and task weights. Experiments on standard long-tailed benchmarks are reported to outperform strong baselines.","tokens_in":1937,"tokens_out":631,"duration_ms":23853,"significance":"If the block-diagonal Fisher-based proxy reliably ranks shared-depth and task-weight choices, the work would replace heuristic hyper-parameter selection with a principled, bias-variance grounded procedure. The explicit decomposition into encoder/decoder terms could also clarify the sources of the head-tail trade-off. The three-stage pipeline and factorized KL argument are technically interesting even if the proxy requires further validation.","major_comments":[{"comment":"The central claim that the second-order expansion supplies an independent, computable proxy for hyper-parameter selection rests on the block-diagonal Fisher approximation. Because the encoder is shared, head and tail gradients are coupled through the same parameters; any residual off-block entries in the true Fisher (arising from feature overlap or class dependence) would invalidate the additive variance/bias decomposition used for ranking. No quantitative bound on the approximation error or regime of validity is supplied in the derivation.","section":"Derivation of second-order expansion (under block-diagonal Fisher)"},{"comment":"The factorized model is invoked both to obtain additive KL terms and to justify block-diagonality. The mutual-exclusivity assumption between head and tail label groups is plausible for the output layer but does not automatically extend to the shared encoder features; residual statistical dependence would couple the encoder-variance term across tasks and undermine the proxy.","section":"Factorized model and task-wise objective"}],"minor_comments":[{"comment":"The abstract states that the proxy 'enables efficient hyper-parameter search' but provides no concrete complexity comparison or timing results; a table or paragraph quantifying the cost of the three-stage pipeline versus grid search would strengthen the efficiency claim.","section":"Abstract and experimental section"},{"comment":"Notation for the task-wise objectives and the Fisher blocks should be introduced with explicit equation numbers in the main text rather than only in appendices, to make the bias-variance decomposition self-contained.","section":"Notation and preliminaries"}],"recommendation":"major_revision","confidential_remarks":"The manuscript's reliance on an unvalidated approximation for its main algorithmic contribution makes it a borderline fit for a methods-focused venue; if the authors can supply either a theoretical error bound or strong empirical evidence that the proxy ranking matches held-out performance, the paper would become substantially stronger."},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the detailed and constructive review. We address each major comment below and describe the revisions we will make.","responses":[{"response":"We acknowledge that the block-diagonal Fisher approximation is central to the additive decomposition and that no theoretical bound on the approximation error is derived. The shared encoder does induce parameter coupling, and off-block terms could in principle affect the ranking. While a general quantitative bound is technically difficult to obtain for deep networks, we will revise the manuscript to add an empirical analysis section that quantifies the magnitude of off-diagonal Fisher blocks on the benchmark datasets and discusses the observed regimes of validity. This will support the practical reliability of the proxy without claiming theoretical guarantees.","revision_made":"yes","referee_comment":"[Derivation of second-order expansion (under block-diagonal Fisher)] The central claim that the second-order expansion supplies an independent, computable proxy for hyper-parameter selection rests on the block-diagonal Fisher approximation. Because the encoder is shared, head and tail gradients are coupled through the same parameters; any residual off-block entries in the true Fisher (arising from feature overlap or class dependence) would invalidate the additive variance/bias decomposition used for ranking. No quantitative bound on the approximation error or regime of validity is supplied in the derivation."},{"response":"The factorized model p(y|x) = p_head(y_head|x) p_tail(y_tail|x) is introduced specifically to establish additive decomposition of the KL-based generalization error at the task level; it does not claim to remove all statistical dependence in the shared encoder features. The block-diagonal Fisher approximation is applied separately as a computational device for the bias-variance proxy. We will revise the text to clarify this distinction, explain how the three-stage pipeline (independent pre-training followed by weighted joint training) reduces feature-level coupling in practice, and note that the proxy remains an empirical tool whose utility is demonstrated by improved hyper-parameter selection over heuristics.","revision_made":"yes","referee_comment":"[Factorized model and task-wise objective] The factorized model is invoked both to obtain additive KL terms and to justify block-diagonality. The mutual-exclusivity assumption between head and tail label groups is plausible for the output layer but does not automatically extend to the shared encoder features; residual statistical dependence would couple the encoder-variance term across tasks and undermine the proxy."}],"tokens_in":1489,"tokens_out":507,"duration_ms":27769,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The paper splits long-tailed recognition into head and tail tasks that share an encoder but use separate decoders. They introduce a factorized model so the KL-based generalization error splits into additive task terms, then use a three-stage pipeline (independent training to get optima and Fisher, weighted joint training, then assembly) and a second-order expansion under block-diagonal Fisher to decompose error into encoder variance, bias, and decoder variance. That decomposition is meant to serve as a computable proxy for picking the shared depth and the task weights.\n\nWhat stands out is the attempt to replace heuristic choices with something derived from the bias-variance structure. The factorized model and the additive KL property are a clean way to handle the mutual exclusivity between head and tail groups. If the math holds, the proxy could cut down on trial-and-error for these hyperparameters.\n\nThe soft spot is exactly the block-diagonal Fisher assumption. Because the encoder is shared, head and tail gradients flow through the same features, so cross terms are likely present. The abstract gives no error bound or regime where the approximation stays accurate, and the stress-test concern about residual off-block entries from class dependence looks real. Without seeing the full derivation or the quantitative checks in the experiments, it is unclear whether the proxy actually ranks the choices reliably or just reproduces something fitted.\n\nThis is for groups already working on long-tailed or imbalanced vision tasks who want a more structured way to set sharing and weighting. A reader who follows bias-variance arguments in deep models will see the point of the decomposition even if they end up skeptical of the approximation.\n\nI would send it to peer review. The framing is worth a proper look even if the central proxy needs stronger validation.","headline":"OSDTW gives a Fisher-derived proxy for choosing shared depth and task weights in long-tailed nets, but the block-diagonal approximation is the part that needs checking.","tokens_in":2403,"tokens_out":424,"would_cite":false,"duration_ms":22606,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"OSDTW derives a bias-variance decomposition of generalization error to select shared encoder depth and task weights for long-tailed recognition.","keywords":["long-tailed recognition","task decomposition","shared encoder","task weighting","Fisher approximation","bias-variance decomposition","generalization error"],"falsifier":"On a standard long-tailed benchmark, compare test accuracy obtained with depth and weights chosen by minimizing the decomposed error terms against accuracy obtained with heuristic or grid-search choices; the proxy is falsified if the selected values do not improve or match the best heuristic performance.","tokens_in":2690,"feed_emoji":"📊","tokens_out":423,"duration_ms":27796,"temperature":0.7,"pith_summary":"The paper partitions long-tailed recognition into head and tail tasks implemented with a shared encoder and task-specific decoders. It introduces a factorized model so that the KL-based generalization error decomposes into task-wise terms up to a constant. Under a block-diagonal Fisher approximation, a second-order expansion of the expected generalization error is derived that splits into encoder variance, encoder bias, and decoder variance. This decomposition supplies a computable proxy for choosing the shared depth and task weights inside a three-stage training pipeline.","feed_headline":"Error decomposition selects shared depth for long-tailed nets","feed_subtitle":"Second-order expansion splits encoder and decoder variance to pick depth and weights without exhaustive search","key_machinery":"The factorized model for head and tail label groups, which lets the generalization error be written as the sum of task-wise terms, together with the block-diagonal Fisher approximation that produces the second-order bias-variance expansion.","core_discovery":"Under a block-diagonal Fisher approximation, the expected generalization error of the factorized model admits a computable second-order expansion that decomposes into encoder variance, encoder bias, and decoder variance; this bias-variance decomposition serves as a proxy for selecting the shared depth and task weights in the three-stage training pipeline.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Fisher approximation decomposes generalization error for depth selection","Second order expansion picks shared depth and weights in long tailed nets","Bias variance decomposition proxies optimal OSDTW hyperparameters","Task wise error terms guide encoder sharing in long tailed recognition","Decoupled training uses computable error expansion to set depth"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"The block-diagonal Fisher approximation is accurate enough that the second-order error expansion reliably ranks choices of shared depth and task weights.","fun_headline_variants_meta":{"raw":{"variants":["Fisher approximation decomposes generalization error for depth selection","Second order expansion picks shared depth and weights in long tailed nets","Bias variance decomposition proxies optimal OSDTW hyperparameters","Task wise error terms guide encoder sharing in long tailed recognition","Decoupled training uses computable error expansion to set depth"]},"model":"grok-4.3","cost_usd":0.004473,"raw_usage":{"total_tokens":2161,"prompt_tokens":690,"num_sources_used":0,"completion_tokens":77,"cost_in_usd_ticks":44728000,"prompt_tokens_details":{"text_tokens":690,"audio_tokens":0,"image_tokens":0,"cached_tokens":64},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":1394,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":690,"tokens_out":77,"duration_ms":5314,"temperature":1.0,"reasoning_tokens":1394,"cache_read_input_tokens":64,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-30T11:44:39.661944+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"On a standard long-tailed benchmark, compare test accuracy obtained with depth and weights chosen by minimizing the decomposed error terms against accuracy obtained with heuristic or grid-search choices; the proxy is falsified if the selected values do not improve or match the best heuristic performance.","supporting_citations":[],"review_version":1}