diff --git a/codex_reviews/reviews/06GH83hDIv.json b/codex_reviews/reviews/06GH83hDIv.json new file mode 100644 index 0000000000000000000000000000000000000000..bc6a72d5012c7928013262c6cb14d458efd162d7 --- /dev/null +++ b/codex_reviews/reviews/06GH83hDIv.json @@ -0,0 +1,58 @@ +{ + "paper_id": "06GH83hDIv", + "title": "Auction-Based Regulation for Artificial Intelligence", + "model": "gpt-5.6-sol", + "reasoning_effort": "xhigh", + "created_at": "2026-08-09T17:19:55.973234+00:00", + "paper_md_chars": 81902, + "metadata": { + "pdf_time_verification_status": "verified_arxiv_version_before_review", + "review_time_revision_source": "arxiv", + "arxiv_version_id": "2410.01871v1", + "markdown_source": "arxiv" + }, + "review": { + "summary": "The paper proposes SIRA, a regulatory contest in which models must meet a minimum safety threshold to be deployed, are randomly paired, and receive an additional premium when safer than their paired opponent. It models safety investment as an all-pay bid, derives purported equilibrium bidding functions for Uniform and Beta value distributions, and reports synthetic simulations suggesting increased participation and spending relative to threshold-only regulation. The idea is interesting and clearly motivated, but the central equilibrium analysis does not correctly account for endogenous participation, so the main theoretical and empirical claims are not established.", + "strengths": [ + "The paper identifies an important problem and offers a concrete mechanism rather than only qualitative policy recommendations.", + "The reserve-threshold baseline and the distinction between deployment value, premium value, and sunk safety investment provide an intuitive starting model.", + "The manuscript is generally organized, with an algorithm, notation table, closed-form special cases, proofs, and simulations over multiple threshold values.", + "The authors explicitly discuss important limitations concerning safety evaluation, regulator access, and heterogeneous evaluation data.", + "Random pairwise comparison is operationally simple and guarantees a bounded number of premium recipients relative to rewarding every qualifying participant." + ], + "weaknesses": [ + "The principal equilibrium derivation ignores endogenous selection into the auction. Opponents are drawn only from agents whose equilibrium utility is positive, so the relevant premium-value CDF must be conditioned on that endogenous entry event. The appendices instead condition only on V_i >= p_epsilon. This is necessary but not sufficient for entry: utility also depends on lambda_i, v_i^d, the winning probability, and the bid. Consequently, the derived F_v, bidding functions, utilities, and participation rates are generally not an equilibrium of Algorithm 1.", + "The paper describes an incomplete-information setting but claims ordinary Nash equilibria. What is required is a Bayesian Nash equilibrium jointly specifying type-dependent entry and bidding. The proof only gives the familiar interior bidding condition for a two-player all-pay contest under an assumed opponent distribution; it does not solve the entry-distribution fixed point.", + "Theorem 2's strict claim b_i^* > p_epsilon is false as stated for arbitrary distributions. The formula equals p_epsilon when v_i^p=0 and can also fail to be strict for distributions with gaps or atoms. A participant with zero premium value but v_i^d > p_epsilon rationally bids exactly the threshold. The theorem therefore needs atomlessness/support assumptions and an almost-sure qualification.", + "The stated evaluation model does not justify the deterministic comparison used in the game. IID agent and regulator data imply neither identical finite-sample safety measurements nor that b_i > b_j guarantees the regulator ranks model i higher. Evaluation noise, sample size, and rank reversals are absent from the utility and proofs.", + "The experiments simulate the assumed value distributions and evaluate the paper's own closed-form strategies. They do not test actual AI models or safety metrics and cannot independently validate an equilibrium whose opponent distribution is misspecified. No uncertainty estimates, random seeds, or numerical tables are supplied for the reported 15% and 20% improvements.", + "A 20% increase in average bid is repeatedly described as a 20% increase in safety, but M is unspecified and may be nonlinear. Thus no quantitative safety improvement follows from the reported spending increase.", + "Several boundary and consistency issues remain: ties are absent from the utility; bids are capped at one without analyzing the resulting pooling or boundary best responses; strict monotonicity of M does not ensure that M^{-1}(b) exists for every b in [0,1]; and Reserve Thresholding has multiple optima when v_i^d=p_epsilon.", + "There are material formula and notation inconsistencies. For example, the second branch of the Uniform utility uses ln(2p_epsilon) in the corollary statement but ln(2v_i^p) in its proof. The proof also alternates between bids, types, and reports in expressions such as F_v(b_i), obscuring the best-response argument.", + "The mechanism does not establish the advertised budget practicality. The regulator's actual transfer or cost is never modeled, private v_i^p is an agent valuation rather than a budgeted payment, and paying one winner per pair need not satisfy any fixed budget. Pairwise winners are also not the globally top-performing half, contrary to the motivating description.", + "Important strategic concerns—including collusion, multiple identities, manipulation of the entrant pool, heterogeneous safety costs, and gaming of the scalar safety metric—are not analyzed." + ], + "questions": [ + "Can the authors solve the joint entry-and-bidding Bayesian equilibrium, with the opponent distribution conditioned on the actual positive-utility entry rule, and then recompute all participation results?", + "What precisely is the premium supplied by the regulator, what does it cost, and what formal budget constraint does the mechanism satisfy?", + "How do the equilibrium and safety guarantee change when regulator evaluations are noisy finite-sample estimates and safer-in-expectation models can lose a comparison?", + "What safety-to-cost function M was used to convert bids into the claimed percentage safety improvements? If no M was instantiated, why are bid increases reported as safety increases?", + "How are equality cases, odd numbers of participants, capped bids, and ties resolved, and do these rules preserve the proposed equilibrium?" + ], + "limitations": [ + "The model reduces multidimensional AI safety to one scalar metric and assumes a common deterministic cost-to-safety map across all firms.", + "All evidence is synthetic and distributional; there is no experiment with trained models, regulator evaluations, or realistic safety-investment costs.", + "Agent and regulator data are assumed IID, while distribution shift and evaluation design are deferred to future work.", + "The paper studies neither regulator welfare nor social welfare, so increased investment and participation are not shown to be efficient or socially beneficial.", + "The practical enforcement, audit, prize-design, and adversarial-resilience requirements of the proposed regulatory system remain unspecified." + ], + "soundness": 3, + "presentation": 5, + "contribution": 4, + "overall_score": 3, + "confidence": 4, + "ethical_concerns": "The work has a safety-positive objective, but deploying the mechanism as policy could create false assurance if a scalar benchmark is treated as comprehensive safety. Relative pairwise rewards may encourage benchmark gaming, strategic entry, collusion, or firm splitting, while a cost-based contest may advantage well-resourced incumbents. These risks and the distributional effects of regulator compensation require explicit analysis.", + "recommendation": "reject", + "review_text": "This is a well-motivated attempt to bring mechanism design to AI regulation, and the reserve-plus-pairwise-premium construction is intuitive. However, the central guarantee is not currently sound. Because only positive-utility agents enter and are paired, the opponent premium-value distribution must be conditioned on endogenous entry. The appendices condition only on V_i >= p_epsilon, even though many such types have negative utility depending on lambda_i and therefore do not enter. This invalidates the supplied CDFs as opponent distributions and, consequently, the claimed bidding equilibria, utilities, participation comparisons, and simulations. The appropriate object is a joint Bayesian entry-and-bidding equilibrium.\n\nAdditional gaps reinforce this concern. IID evaluation data do not make regulator rankings deterministic, the strict above-threshold claim has counterexamples at zero premium value, and an unspecified nonlinear M does not permit a 20% bid increase to be interpreted as a 20% safety increase. The experiments are Monte Carlo evaluations of the analytical model rather than independent or model-level evidence, and practical budget feasibility is not formalized. Internal formula inconsistencies, including ln(2p_epsilon) versus ln(2v_i^p) in the Uniform utility, also impede verification. The framing is promising, but establishing the contribution requires a corrected endogenous-entry equilibrium, noisy evaluation analysis, a defined transfer budget, and evidence that distinguishes spending from actual safety." + } +} diff --git a/codex_reviews/reviews/070DFUdNh7.json b/codex_reviews/reviews/070DFUdNh7.json new file mode 100644 index 0000000000000000000000000000000000000000..6745b7c982af58d802ff99d7af338a875e8b853d --- /dev/null +++ b/codex_reviews/reviews/070DFUdNh7.json @@ -0,0 +1,64 @@ +{ + "paper_id": "070DFUdNh7", + "title": "GraphGPT: Graph Learning with Generative Pre-trained Transformers", + "model": "gpt-5.6-sol", + "reasoning_effort": "xhigh", + "created_at": "2026-08-09T17:19:19.066494+00:00", + "paper_md_chars": 39981, + "metadata": { + "pdf_time_verification_status": "verified_revision_cdate_before_review", + "review_time_revision_source": null, + "arxiv_version_id": null, + "markdown_source": "dataset" + }, + "review": { + "summary": "GraphGPT serializes graphs or sampled subgraphs as token sequences using Eulerian/postman traversals, pre-trains a causal decoder with next-token prediction, and fine-tunes task-specific representations for graph, edge, and node prediction. The serialization-plus-generative-pretraining idea is intuitive and potentially useful, and the reported ogbl-ppa result is strong. However, several central claims are not established by the supplied text. In particular, reversibility is underspecified after edge duplication, the link-prediction setup may expose the target edge, pretraining split hygiene is unclear, scaling comparisons are confounded, and the claimed graph-level SOTA and demonstrated generation capability are unsupported by the tables.", + "strengths": [ + "The paper proposes a simple unifying interface for graph-, edge-, and node-level tasks using an otherwise standard causal transformer decoder.", + "Eulerian traversal is an interesting way to expose every edge while retaining local adjacency in the serialized sequence; random traversal sampling also provides a plausible source of augmentation.", + "The study covers three task granularities and includes ablations for pretraining, node re-indexing, and node identity encoding.", + "The pretraining ablation reports sizeable improvements on every evaluated task, including PCQM4Mv2 validation MAE from 0.1086 to 0.0875 and ogbl-ppa test HR@100 from 41.28 to 55.56 for the stated configurations.", + "The reported ogbl-ppa result is empirically strong: GraphGPT-large obtains 67.15±1.36 test HR@100 versus 63.22±1.74 for the strongest listed baseline.", + "The authors report multiple model scales, reaching roughly 400M parameters, and show monotonically improving results within each displayed GraphGPT family." + ], + "weaknesses": [ + "The claimed lossless and reversible transformation is not fully specified. Eulerization duplicates existing edges, but the text defines no token that distinguishes duplicated traversal edges from original parallel edges. Recovering adjacent token pairs therefore reconstructs the Eulerized multigraph, not necessarily the original graph. The treatment of duplicated edge attributes and directed edges is also unclear.", + "The link-prediction protocol has a potentially severe leakage issue. Positive fine-tuning examples are described as subgraphs induced around positive edges, but the paper never states that the target edge is removed before serialization. If it remains in the input, positive and negative examples can be distinguished by directly observing the queried adjacency.", + "Pretraining data hygiene cannot be verified. The paper reports up to 50B tokens but does not say whether graph-level validation/test molecules or held-out validation/test edges were excluded from pretraining. This is especially important because unique node identities are explicitly encoded.", + "The scaling evidence does not justify the causal claims about over-smoothing or over-squashing. Increasing transformer size while observing better accuracy does not demonstrate that these phenomena have been overcome, and no depth-, compute-, or parameter-matched GNN/graph-transformer scaling experiment is presented.", + "Some scaling comparisons are confounded by training budget. On ogbl-ppa, mini/base/large receive 25.6B/39B/50B pretraining tokens, respectively, so improvements cannot be attributed solely to parameter count. Compute, optimization steps, and FLOPs are not reported.", + "The claim of graph-level SOTA is contradicted by the presented results. On ogbg-molpcba, GraphGPT-large reaches 0.2722 test AP while the best listed baseline reaches 0.3031. On PCQM4Mv2, GraphGPT has only a validation result of 0.0875, slightly worse than the listed Graphformer value of 0.0864, and no test result.", + "The claimed graph-generation capability is not experimentally evaluated. There are no measurements of syntactic validity, graph validity, uniqueness, novelty, attribute consistency, or distributional fidelity.", + "Statistical reporting is inadequate. The number of independent runs is not given for several ± entries. For ogbn-proteins, the reported mean and variance are computed from three consecutive, validation-selected epochs rather than independent seeds, so they do not measure run-to-run uncertainty and are affected by selection.", + "The method remains order-sensitive: a causal transformer with positional information is not inherently invariant to graph isomorphisms or Eulerian-trail choice. Re-indexing and random paths may encourage robustness, but no invariance or traversal-sensitivity experiment is supplied.", + "Comparisons are not controlled for compute or pretraining exposure. GraphGPT uses tens of billions of tokens and up to hundreds of millions of parameters, while many listed baselines are small leaderboard models. Thus the experiments do not isolate the benefit of Eulerian serialization from model scale and training compute.", + "Node identity encoding is inherently transductive and can encourage memorization. The paper does not evaluate unseen-node transfer, unseen-graph transfer for the large-graph setting, or sensitivity to changing the identity partition.", + "Critical implementation information is deferred to appendices that are absent from the supplied text, including tokenizer construction, sampler parameters, optimization, architecture details, and graph-specific processing. Reproducibility therefore cannot be assessed from the submission provided." + ], + "questions": [ + "Are target edges removed from ogbl-ppa subgraphs before serialization during both pretraining and supervised fine-tuning? Please specify the exact graph available for train, validation, and test examples.", + "Which splits are used to construct the 16B–51.2B pretraining-token streams? Are validation/test molecules, held-out edges, labels, or evaluation-only graph structure ever exposed during pretraining?", + "How does decoding distinguish Eulerization duplicates from original parallel edges, and how are duplicated attributes and directed edges handled so that the original graph is uniquely recoverable?", + "Can the authors report independent-seed results, including the number of runs, for every task rather than treating adjacent epochs as replicates?", + "What happens under matched pretraining tokens, optimization steps, FLOPs, and parameter counts? In particular, does the apparent scaling trend remain when data and compute are held fixed?", + "How sensitive are predictions to the sampled Eulerian trail, and how does this serialization compare with DFS/BFS walks, random walks, edge-list serialization, and non-generatively pretrained decoder baselines?", + "Can the claimed generation capability be evaluated quantitatively, including valid decoding and removal of artificial jump or duplicated edges?" + ], + "limitations": [ + "Postman augmentation can substantially lengthen sequences, while transformer attention is quadratic in sequence length.", + "For large graphs, context-window limits force very small sampled neighborhoods; the paper itself reports roughly 10 nodes and 20 edges for ogbn-proteins at context length 256.", + "Unique node identifiers are tied to a fixed graph and do not directly support unseen nodes or transfer to a differently indexed graph.", + "Random Eulerian paths create multiple order-dependent representations of the same graph without guaranteed prediction invariance.", + "The approach requires very large pretraining streams and models, but wall-clock cost, hardware, memory, and energy consumption are not reported.", + "Generation and longer-context variants are discussed but not validated in the supplied experiments." + ], + "soundness": 4, + "presentation": 6, + "contribution": 6, + "overall_score": 4, + "confidence": 4, + "ethical_concerns": "No immediate human-subject or privacy issue is evident from the supplied text. However, the compute and environmental costs of training 400M+ parameter models on tens of billions of tokens are not documented. The paper also applies predictions to protein associations and functions without discussing the risks of treating benchmark predictions as biologically or clinically validated conclusions.", + "recommendation": "weak_reject", + "review_text": "The core idea—turning graphs into causal token sequences through Eulerian/postman traversals—is appealing, and the paper offers a coherent decoder-based interface across task granularities. The pretraining ablations and reported ogbl-ppa performance make the direction worth pursuing.\n\nThe current evidence is nevertheless insufficient for the main claims. Most importantly, the edge-prediction experiment does not state whether the queried positive edge is removed from the serialized subgraph. Because serialization explicitly reveals edges, leaving the target edge present would create a direct shortcut. The paper must also disclose precisely which graph structure and dataset splits enter pretraining; otherwise leakage from held-out molecules or edges cannot be ruled out.\n\nThere is also a gap in the method's foundational reversibility argument. A sequence reconstructs the Eulerized multigraph, but the text provides no mechanism for distinguishing duplicated postman edges from original multiplicities. The claim needs a formal encoding/decoding algorithm covering multigraphs, edge attributes, disconnected components, and directed graphs. The assertion of optimality should likewise specify the cost being optimized, since minimizing duplicated edges need not minimize token length when attributes have different tokenized lengths.\n\nThe empirical conclusions are overstated. The supplied tables do not establish graph-level SOTA, and graph generation is asserted without any generation experiment. Monotonic performance across larger transformers does not prove that over-smoothing or over-squashing has been overcome. Moreover, ogbl-ppa model size is confounded with different pretraining-token budgets, and leaderboard comparisons are not compute-matched. Finally, uncertainty estimates need independent runs; averaging three validation-selected adjacent epochs on ogbn-proteins is not an adequate substitute.\n\nOverall, this is a promising and reasonably well-presented concept, but resolving leakage questions, formalizing reversibility, adding controlled serialization and scaling baselines, and correcting the statistical and SOTA claims are necessary before the reported conclusions can be considered reliable." + } +} diff --git a/codex_reviews/reviews/0823rvTIhs.json b/codex_reviews/reviews/0823rvTIhs.json new file mode 100644 index 0000000000000000000000000000000000000000..5ffb93a8601b7d9ed593da405c36af867f45b2aa --- /dev/null +++ b/codex_reviews/reviews/0823rvTIhs.json @@ -0,0 +1,57 @@ +{ + "paper_id": "0823rvTIhs", + "title": "Weakly-Supervised Affordance Grounding Guided by Part-Level Semantic Priors", + "model": "gpt-5.6-sol", + "reasoning_effort": "xhigh", + "created_at": "2026-08-09T17:21:20.140344+00:00", + "paper_md_chars": 38556, + "metadata": { + "pdf_time_verification_status": "verified_revision_cdate_before_review", + "review_time_revision_source": null, + "arxiv_version_id": null, + "markdown_source": "dataset" + }, + "review": { + "summary": "The paper proposes a pseudo-supervised alternative to CAM-based weakly supervised affordance grounding. A manually specified object–affordance-to-part mapping supplies text prompts to VLPart and SAM, producing dense pseudo-labels for egocentric images. The approach adds exocentric object-masked feature alignment, an occlusion-inspired pseudo-label refinement stage, a noun/part reasoning module, and stitching augmentation. Evaluation is described on AGD20K seen/unseen splits, with two qualitative robotic grasping examples.", + "strengths": [ + "The affordance-to-semantic-part reformulation is intuitive and directly addresses the weak spatial precision of classification-based CAMs.", + "The method uses exocentric images in two complementary ways: object-confined feature alignment and occlusion-based label refinement.", + "Evaluation considers both seen and unseen object splits, and the paper reports averaging over five random seeds.", + "The stitching augmentation is simple and well motivated as a way to discourage ignoring the affordance query.", + "The proposed mapping and pseudo-label generators are reportedly unnecessary at inference time." + ], + "weaknesses": [ + "The numerical contents of Tables 1 and 2 are not present as readable text; they appear only as image references. Consequently, the claimed breakthrough, comparisons, ablation magnitudes, and consistency across metrics cannot be verified from the supplied paper text.", + "The reasoning-loss equation appears to swap its targets: the predicted part is compared with Enc_T(o), while the predicted object is compared with Enc_T(p). This contradicts both the module description and the variable names, leaving the implemented objective ambiguous.", + "The refinement objective is potentially degenerate. An empty or very small predicted affordance mask leaves nearly the entire egocentric object in its complement, which may already resemble the visible portion of the exocentric object. The paper acknowledges point-concentrated solutions but provides no area, entropy, or other constraint establishing that the separate phase and SAM post-processing resolve this issue.", + "The manual mapping P(o,a), a central source of part-level supervision, is not provided. Its construction choices, ambiguous or multi-part affordances, and failure cases therefore cannot be reproduced or audited. Calling the setting weakly supervised is defensible with respect to dense labels, but the additional manual semantic supervision should be characterized more carefully.", + "The open-vocabulary claim is weakly substantiated. Training still uses a fixed affordance classifier for the exocentric loss and a finite manual part taxonomy, while evidence for unseen affordance language consists only of a qualitative grasp example. There is no systematic evaluation of unseen verbs, synonyms, or compositional queries.", + "Important implementation details are missing, including the exact cross-modal fuser and decoder configuration, pseudo-mask-to-probability conversion for KL loss, SAM segment-selection threshold, refinement training schedule, frozen refinement encoder choice, and which encoder checkpoint is used for offline exocentric candidate selection.", + "The comparison may conflate the proposed learning ideas with access to strong external models and extra semantic priors. There is no readable controlled comparison isolating CLIP architecture, VLPart/SAM pseudo-labeling, manual part mappings, and the proposed modules under matched resources.", + "Although results are averaged over five runs, no standard deviations, confidence intervals, or per-class results are reported in the supplied text.", + "The robot experiment contains only two illustrative successes, without repeated trials, success rates, failure analysis, or a baseline. It does not establish reliable real-world generalization or safety." + ], + "questions": [ + "Is the reasoning-loss equation incorrect? Should f_pred-part be compared with Enc_T(p) and f_pred-obj with Enc_T(o)? Please state the objective actually implemented.", + "How is collapse prevented during label refinement, and what quantitative evidence shows that the refined masks are more accurate than the initial VLPart/SAM masks?", + "Can the complete P(o,a) mapping, pseudo-label normalization procedure, refinement schedule, SAM selection thresholds, and offline feature checkpoint be provided for reproducibility?", + "What are the exact Table 1 and Table 2 values, including run-to-run variance, and how does the method perform on a systematic unseen-affordance or paraphrase split?", + "How much improvement remains under resource-matched controls that separately introduce CLIP, VLPart/SAM pseudo-labels, manual part names, refinement, alignment, reasoning, and stitching?" + ], + "limitations": [ + "The method depends on the coverage and biases of CLIP, VLPart, and SAM, but their failure rates on the target dataset are not quantified.", + "A single manually selected semantic part may not represent diffuse, multimodal, context-dependent, or ambiguous affordance regions.", + "Scalability beyond roughly 100 object–affordance combinations is asserted rather than demonstrated; replacing the manual mapping with an LLM is only proposed.", + "Generalization is evaluated primarily across object categories, not rigorously across novel actions or free-form language.", + "Real-world evidence is anecdotal and limited to two objects and one grasp-selection pipeline." + ], + "soundness": 4, + "presentation": 5, + "contribution": 6, + "overall_score": 5, + "confidence": 4, + "ethical_concerns": "No major ethical concern is apparent from the supplied text. However, the two successful robot examples should not be interpreted as safety validation; deploying uncertain affordance predictions around people or fragile objects would require calibrated confidence, failure detection, and substantially broader testing.", + "recommendation": "weak_reject", + "review_text": "The central idea is promising: translating an object–affordance pair into a semantic part enables existing part-grounding and segmentation models to provide direct spatial supervision, avoiding the known limitations of CAMs. The object-masked exocentric alignment and stitching augmentation are also sensible additions, and the seen/unseen evaluation protocol is relevant.\n\nHowever, the current evidence and specification are insufficient to validate the strongest claims. Most importantly, the numerical contents of the main and ablation tables are unavailable in the provided text, so neither the claimed breakthrough nor the contribution of each module can be checked. Even setting that issue aside, the reasoning-loss equation contradicts its prose by apparently swapping object and part targets. The label-refinement loss also admits plausible trivial or weakly informative solutions, and the paper does not provide quantitative pseudo-label evaluation or enough procedural detail to show that SAM post-processing reliably fixes them.\n\nThe paper should also qualify its supervision and open-vocabulary claims. The method requires a manually defined mapping for every training object–affordance combination, which constitutes meaningful part-level semantic supervision even if it is not pixel-level annotation. The mapping is neither listed nor analyzed. Meanwhile, free-form inference is claimed, but the experiments described in text do not systematically test unseen verbs, synonyms, or compositional queries; two qualitative grasping examples are not sufficient.\n\nOverall, this is a useful and potentially impactful direction, but the supplied version leaves central quantitative claims unverifiable and contains methodological ambiguities that affect reproducibility and soundness. A revision with readable complete results, corrected equations, quantitative pseudo-label analysis, resource-matched controls, uncertainty estimates, and systematic open-vocabulary evaluation could materially strengthen the work." + } +} diff --git a/codex_reviews/reviews/09JVxsEZPf.json b/codex_reviews/reviews/09JVxsEZPf.json new file mode 100644 index 0000000000000000000000000000000000000000..c83c971ad411941a0f3c7a73c638da7bb79478ec --- /dev/null +++ b/codex_reviews/reviews/09JVxsEZPf.json @@ -0,0 +1,59 @@ +{ + "paper_id": "09JVxsEZPf", + "title": "Towards Comprehensive and Efficient Post Safety Alignment of Large Language Models via Safety Patching", + "model": "gpt-5.6-sol", + "reasoning_effort": "xhigh", + "created_at": "2026-08-09T17:20:37.557477+00:00", + "paper_md_chars": 38128, + "metadata": { + "pdf_time_verification_status": "verified_revision_cdate_before_review", + "review_time_revision_source": null, + "arxiv_version_id": null, + "markdown_source": "dataset" + }, + "review": { + "summary": "The paper proposes SafePatching, a post-safety-alignment framework that derives two parameter deltas from the same harmful dataset: a gradient-ascent patch intended to suppress harmful behavior and a gradient-descent patch intended to reduce excessive refusals. The deltas are sparsified and merged using SNIP-based importance masks. The framing around jointly optimizing safety, over-safety, and utility is valuable, and the evaluation covers several relevant dimensions. However, central mathematical and implementation details are ambiguous, the empirical evidence does not fully support several headline claims, and the evaluation lacks uncertainty estimates and sufficiently independent safety validation.", + "strengths": [ + "The paper identifies an important multi-objective problem: improving safety while reducing false refusals and preserving general capabilities. Evaluating all three dimensions is more informative than reporting safety alone.", + "Deriving complementary patches from a shared harmful dataset is an interesting and potentially data-efficient design. Applying a patch after derivation also avoids additional inference-time operations.", + "The evaluation includes two harmful-prompt datasets, two over-safety datasets, multiple utility benchmarks, a jailbreak setting, and a continual-alignment setting.", + "The component ablations are useful. In particular, the single-patch results clearly expose the tension between safety and over-safety, while the random-retention ablation suggests that dense patch merging damages utility.", + "The main 7B results show a meaningful practical trade-off: harmful rates fall from 31.73/19.64 to 3.27/7.87, refusal rates fall from 8.00/4.67 to 5.60/2.67, and MT-Bench remains close to the original model." + ], + "weaknesses": [ + "The core algorithm is not specified unambiguously. Equation defining I(a,b) contains only the asymmetric difference I_SE(a) minus I_OSM(b), although the prose discusses distinct important regions for both patches. It is unclear which masks are ultimately applied to the safety and over-safety patches, how their OSM-exclusive region is handled, and whether overlapping entries are removed from one or both patches.", + "The optimization convention for the safety patch is ambiguous. L_GA is positive log-likelihood: maximizing it would increase the probability of harmful answers, whereas minimizing it would suppress them. The paper says gradient ascent but does not explicitly state whether the displayed objective is maximized or minimized. This is a central reproducibility and correctness issue.", + "Essential implementation details are absent: learning rates, number of steps or epochs, optimizer, batch size, trainable parameter scope, exact training-set size and split, generation settings, p, a, b, alpha, beta, SNIP aggregation across examples, and random seeds. No variance across training runs or random masks is reported.", + "The claim of having “no extra training cost” is inaccurate as written. SafePatching trains both gradient-ascent and gradient-descent models and computes importance gradients. The paper reports only inference-time ratios for decoding baselines, with no training wall-clock time, GPU hours, peak memory, or storage comparison.", + "The claim that SafePatching enhances utility is overstated. Its aggregate utility decreases from 40.35 to 39.99, while MT-Bench increases only from 6.01 to 6.16. Without repeated evaluations or confidence intervals, the significance of the 0.15 MT-Bench difference cannot be established.", + "SafePatching does not uniformly outperform the baselines. Several baselines achieve lower harmful rates, while SafePatching obtains better refusal and utility results. This is a trade-off rather than strict superiority, and the paper provides neither a formal multi-objective metric nor a Pareto analysis to justify its overall ranking.", + "Safety judgments rely on one Longformer classifier, including for jailbreak ASR. The manuscript asserts that it performs identically to GPT-4 and human annotators but provides no within-paper validation, error analysis, or human safety assessment. Gradient ascent could produce incoherent outputs that receive a low harmfulness score without constituting appropriate safe responses.", + "The jailbreak experiment adapts the method using data induced by ReNeLLM and evaluates it under ReNeLLM again. The paper does not state whether prompts or transformations are held out, and it evaluates no unseen attack family, so the result may reflect attack-specific adaptation rather than general jailbreak robustness.", + "Although the implementation section says experiments use both LLaMA-2-Chat-7B and 13B, the supplied quantitative tables report only 7B. Evidence for cross-scale generality is therefore not available in the provided text.", + "Human evaluation of over-safety lacks annotation instructions, blinding, adjudication procedure, and inter-annotator agreement. MT-Bench evaluation similarly omits evaluator version and decoding details.", + "The continual-alignment section reports only averages over three time steps in its table. The unavailable numeric contents of the plotted curves prevent verification of retention of earlier safety improvements, which is the defining continual-learning claim.", + "Baseline comparisons are difficult to interpret because decoding methods do not receive the full utility evaluation, while training methods may use different data and compute. A matched-data, matched-compute comparison to simple refusal fine-tuning or preference-based safety tuning is missing." + ], + "questions": [ + "What are the exact masks applied to each patch? Please give separate set equations for the retained SE and OSM entries, including the treatment of their intersection, SE-only region, OSM-only region, and random remainder.", + "Is Equation 1 maximized or minimized in code? Please state the optimizer update explicitly and provide all training, mask-construction, merging, and generation hyperparameters.", + "Are the ReNeLLM evaluation prompts and transformations disjoint from those used to derive the patch, and how does the patched model perform against unseen jailbreak methods?", + "How stable are the results across random seeds and random-retention masks, and are the MT-Bench differences statistically meaningful?", + "Where are the stated 13B results, and do the same p, a, b, alpha, and beta values transfer across model scales and continual-alignment steps?", + "What are the end-to-end patch-construction costs, including both fine-tuning runs and SNIP computation, relative to every training-based baseline?" + ], + "limitations": [ + "This review is restricted to the supplied manuscript text and metadata; external citations and claims about prior evaluators cannot be independently verified.", + "Several referenced figures are not numerically represented in the provided Markdown, so the exact jailbreak and per-time-step continual results cannot be assessed.", + "No appendix, code, or supplementary implementation details are included in the supplied material." + ], + "soundness": 4, + "presentation": 5, + "contribution": 6, + "overall_score": 4, + "confidence": 4, + "ethical_concerns": "The work has a legitimate safety objective, but it deliberately trains an intermediate model to reproduce harmful answers and creates an over-safety patch that weakens internal safeguards. The manuscript gives only a content warning and does not discuss secure handling of harmful data, access controls for unsafe intermediate checkpoints and patches, release policy, researcher exposure, or misuse testing. These safeguards are especially important because the OSM patch alone raises harmful rates to 87.69% and 65.45%.", + "recommendation": "weak_reject", + "review_text": "SafePatching offers a useful framing of post-safety alignment as a three-way problem involving harmfulness, false refusals, and utility. Its complementary-patch construction is intuitive, and the breadth of evaluation and component ablations are positive. The reported 7B result is promising as a trade-off: it substantially lowers harmful rates while modestly improving refusal rates and approximately preserving utility.\n\nThe main obstacle is that the method cannot be reproduced reliably from the paper. The sign and optimization direction of the gradient-ascent objective are ambiguous, and the central difference-set masking procedure is defined only asymmetrically even though the narrative requires separate important regions for two patches. Most training, importance-scoring, merging, and decoding hyperparameters are missing. These are not minor presentation issues because different interpretations can reverse the safety update or materially change which unsafe parameters are merged.\n\nThe evidence also falls short of the strongest conclusions. Aggregate utility decreases slightly, and the small MT-Bench increase is reported without uncertainty. Other methods achieve greater safety enhancement, so “outperforms” requires a formal trade-off analysis. Reliance on one automatic harmfulness classifier leaves open whether gradient ascent creates genuinely safe responses or merely incoherent outputs. Reusing ReNeLLM for patch derivation and evaluation, without a stated held-out protocol or cross-attack tests, does not establish general jailbreak robustness. Claims of no additional training cost conflict with the two fine-tuning runs and SNIP computation, for which no cost measurements are supplied. Finally, the text claims experiments on both 7B and 13B backbones, but the provided quantitative evidence covers only 7B.\n\nI view the problem formulation and empirical direction as worthwhile, but the algorithm needs a precise specification and the claims need stronger, independently validated evidence. A revision should provide pseudocode and complete hyperparameters, multiple-seed statistics, human or diverse-model safety evaluation, unseen-attack testing, matched-compute baselines, end-to-end construction costs, and the stated cross-scale results." + } +} diff --git a/codex_reviews/reviews/0BBzwpLVpm.json b/codex_reviews/reviews/0BBzwpLVpm.json new file mode 100644 index 0000000000000000000000000000000000000000..a264bd3febe34a0ae0c534cc202cb753b0dbe3c9 --- /dev/null +++ b/codex_reviews/reviews/0BBzwpLVpm.json @@ -0,0 +1,59 @@ +{ + "paper_id": "0BBzwpLVpm", + "title": "Learning Identifiable Concepts for Compositional Image Generation", + "model": "gpt-5.6-sol", + "reasoning_effort": "xhigh", + "created_at": "2026-08-09T17:25:16.644959+00:00", + "paper_md_chars": 33861, + "metadata": { + "pdf_time_verification_status": "verified_revision_cdate_before_review", + "review_time_revision_source": "submission_note", + "arxiv_version_id": null, + "markdown_source": "dataset" + }, + "review": { + "summary": "The paper proposes a conditional StyleGAN2-ADA framework in which each binary attribute controls a separate stochastic latent block. A learnable sparse mask is intended to enforce a “minimal change principle,” while a causal-conditioning transformation splits child attributes according to parent values. The paper also claims component-wise identifiability and reports image-generation and editing experiments on MNIST4, Car9, Pizza13, and FFHQ. The problem is important and the proposed combination is intuitively appealing, but the central theoretical guarantee is not established by the supplied text, the causal interpretation is substantially under-justified, and essential implementation and evaluation details are missing. Consequently, the claims of identifiable concepts, causal disentanglement, data efficiency, and superior performance are not presently supported strongly enough.", + "strengths": [ + "The paper targets an important failure mode of conditional generators: changing one attribute can inadvertently change correlated attributes.", + "Separating attribute-specific stochastic blocks and penalizing their active dimensionality is a simple, potentially useful architectural bias for compositional generation.", + "The paper evaluates both generation and real-image editing and explicitly considers rare or withheld attribute combinations rather than only in-distribution quality.", + "The ablation described in the text attempts to distinguish the effects of sparsity and causal conditioning.", + "The authors openly report a failure on the more complex cat-and-ice-cream experiment and acknowledge limitations from the StyleGAN-based backbone and binary-label interface." + ], + "weaknesses": [ + "Theorem 1 is not justified by a proof in the supplied text. More importantly, it assumes only that the model matches the marginal image distribution, p_{G(Z)}=p_x. Marginal matching generally cannot rule out nonlinear transformations that mix latent concepts. The attribute variability in A4 can help only if an appropriate collection of label-conditional distributions or the joint distribution of images and labels is matched, but this is not stated in the theorem.", + "The assumptions behind the theorem do not closely match the claimed setting. A3 restricts every concept to one dimension, whereas the method learns multidimensional masked concepts. A4 requires label pairs differing only in one entry, which can fail precisely for dependent or absent combinations. Independence assumptions on the noise variables and the role or identifiability of the residual z_c are also not made sufficiently explicit.", + "The proposed causal transformation does not, by itself, recover an exogenous variable. Replacing E by indicators such as E|O=1 and E|O=0 creates parent-stratified conjunctions; it is not shown that these indicators isolate the SCM noise ε or yield an age-invariant eyeglasses concept. With multiple binary parents, enumerating parent configurations may also grow exponentially.", + "The causal-discovery procedure is underspecified. The paper suggests PC followed by manual orientation, but states no causal-sufficiency, faithfulness, sampling, significance-level, or background-knowledge assumptions. Vanilla observational discovery cannot generally distinguish directions within a Markov equivalence class or resolve latent common causes, despite the paper motivating the method using both causal influences and common causes.", + "The learnable mask mechanism is incomplete. The paper calls m_i binary but specifies only an L1 penalty, without explaining its parameterization, relaxation, estimator, or thresholding. If masks are continuous, their scale can be reduced while compensating in the generator weights; if they are discrete, the optimization method is necessary. Thus L1 norm alone does not establish minimal semantic influence.", + "There are several formal inconsistencies in the model description. The definitions z_i^deact=f_{i,1}(ε_i) and z_i^act=f_{i,0}(ε_i) reverse the earlier meanings of f_{i,1} and f_{i,0}. The adversarial fake term is written as 1-log D(x_fake|a), rather than a standard log(1-D) discriminator term or a clearly separated non-saturating generator loss; maximizing the displayed discriminator objective is unbounded as D(x_fake) approaches zero.", + "It is asserted, rather than demonstrated, that the shared f_null contains no concept information. Correlated attributes could permit one concept to be represented through another attribute block, while the residual block and generator introduce further representational non-uniqueness. No objective directly measures image-space locality or preservation of non-target concepts.", + "The quantitative entries of Tables 1 and 2 are not present as text, so the claimed improvements cannot be audited from the supplied paper. The experimental section also omits sample counts for evaluation, random seeds, uncertainty estimates, training budgets, mask sizes, sparsity coefficients, discovered causal edges, PC settings, and many architecture and optimization details.", + "The Out-FID protocol is unclear. The paper refers to labels that are unseen or rare during training and compares generated samples to test images, but it does not specify the exact held-out combinations, split construction, number of real examples per combination, or how FID is computed for combinations described elsewhere as nonexistent.", + "The FFHQ evaluation relies on attributes predicted by a pretrained CelebA classifier, but the classifier's accuracy and domain shift are not evaluated. For rare combinations, CLIP alignment does not by itself establish realism, identity preservation, or that non-target attributes remain fixed. Prompt construction, aggregation, and uncertainty are not reported.", + "Comparisons with GPT-4o, Meta AI, and Stable Diffusion 3 are anecdotal without prompts, sampling settings, number of trials, or a controlled quantitative protocol. Fairness of the learned-model baselines is likewise difficult to assess because training data, capacity, and compute are not described.", + "The claims of interpretability and data efficiency are not directly measured. There is no analysis of selected mask dimensions, concept leakage, sample-efficiency curves, intervention precision, identity preservation, or sensitivity to the sparsity coefficient and causal graph.", + "Presentation quality is weak: equations and captions contain typographical or corrupted text, Figure 4 is difficult to interpret from its caption, the results section refers to “Table 8” although the supplied item is Figure 8, and important notation changes meaning across sections." + ], + "questions": [ + "Can the authors state and prove an identifiability theorem based on matching p(x|A) or p(x,A), and explain why it excludes nonlinear mixing with other concepts and z_c? How does that theorem extend beyond one-dimensional concepts and remain applicable when A4 fails for absent combinations?", + "How exactly are binary masks trained? Please specify their parameterization, gradient estimator or relaxation, thresholding, initialization, mask dimension, λ_sparsity, and safeguards against mask-generator rescaling.", + "For a child with several parents, what exact transformed label vector is constructed and how is an intervention such as “add eyeglasses while preserving age” encoded at inference? What assumptions make these indicators equivalent to isolating exogenous noise, and how does the method avoid exponential growth in parent configurations?", + "Please provide the discovered FFHQ graph, manually oriented edges, causal-discovery settings, held-out combinations, per-condition sample counts, full numerical tables, uncertainty across seeds, and quantitative measures of realism, target-attribute success, identity preservation, and non-target-attribute leakage." + ], + "limitations": [ + "The authors acknowledge that the StyleGAN-based model fails on complex, non-curated compositions, illustrated by the cat-and-ice-cream experiment.", + "Control is restricted to binary attribute labels; text, masks, continuous attributes, and object-level spatial relationships are not supported.", + "The causal-conditioning procedure depends on reliable attribute labels, causal-graph estimation, and manual background knowledge.", + "The supplied paper text does not contain the proof of Theorem 1 or machine-readable numerical entries for the rasterized tables, limiting verification of the theoretical and empirical claims." + ], + "soundness": 3, + "presentation": 3, + "contribution": 5, + "overall_score": 3, + "confidence": 4, + "ethical_concerns": "The face experiments generate and edit sensitive demographic and appearance attributes using labels from an unvalidated pretrained classifier. Potential demographic bias, stereotyping, privacy or impersonation misuse, and classifier errors are not analyzed. These concerns do not preclude the research but merit an explicit discussion and subgroup evaluation.", + "recommendation": "reject", + "review_text": "The paper addresses a valuable question: how to prevent conditional image generators from carrying correlated, non-target information when composing attributes. Its main design—attribute-specific stochastic modules with sparse masks—provides a plausible inductive bias, and studying rare combinations is worthwhile.\n\nHowever, the paper's strongest claim, identifiable recovery of the true concepts, is not supported as stated. Theorem 1 requires only marginal image-distribution matching, which does not ordinarily eliminate nonlinear latent mixing. Its useful variability assumption concerns label interventions, so the theorem would need to formulate matching of the joint or all relevant conditional distributions. The one-dimensional assumption also differs materially from the multidimensional masked implementation, while A4 can be violated for exactly the dependent or nonexistent combinations motivating the paper. No proof is included in the supplied text, so these gaps cannot be resolved.\n\nThe causal-conditioning component is also conceptually incomplete. Conditioning or splitting an observed child label by parent values does not automatically identify its exogenous SCM noise; it may instead learn separate parent-specific versions of the attribute. Causal discovery from observational binary labels requires assumptions that are not stated, and manual orientation weakens the claim that the framework itself discovers causal structure. Handling latent common causes is particularly unclear.\n\nOn the implementation side, the paper does not explain how a supposedly binary mask is optimized. A continuous L1-penalized gate has a scale degeneracy with the downstream generator, whereas a discrete gate requires an estimator. The displayed adversarial loss and the activation/deactivation definitions are inconsistent. These are central details, not minor implementation choices.\n\nFinally, the empirical support is difficult to verify. The supplied tables do not expose their numerical values, and the text omits exact split definitions, evaluation sample sizes, seeds, uncertainty, mask statistics, graph details, and training budgets. FID or CLIP alone does not measure the central promise that only the requested concept changes. Controlled attribute-success, leakage, realism, reconstruction, and identity-preservation metrics are needed, together with sample-efficiency curves and graph/mask sensitivity analyses. The qualitative comparisons to external foundation models also lack reproducible prompts and sampling protocols.\n\nOverall, the paper contains an interesting direction and potentially useful architectural heuristic, but major theoretical, causal, optimization, and evaluation issues must be resolved before its central claims can be considered reliable." + } +} diff --git a/codex_reviews/reviews/0CtIt485ew.json b/codex_reviews/reviews/0CtIt485ew.json new file mode 100644 index 0000000000000000000000000000000000000000..dfcdaf34fafa22d586b8eeced3a5960f41c6b670 --- /dev/null +++ b/codex_reviews/reviews/0CtIt485ew.json @@ -0,0 +1,59 @@ +{ + "paper_id": "0CtIt485ew", + "title": "Brain-inspired continual pre-trained learner via silent synaptic consolidation", + "model": "gpt-5.6-sol", + "reasoning_effort": "xhigh", + "created_at": "2026-08-09T17:27:34.017766+00:00", + "paper_md_chars": 73005, + "metadata": { + "pdf_time_verification_status": "verified_arxiv_version_before_review", + "review_time_revision_source": "arxiv", + "arxiv_version_id": "2410.05899v1", + "markdown_source": "arxiv" + }, + "review": { + "summary": "The paper proposes Artsy, a class-incremental learning framework that freezes a ViT-B/16 backbone, adds one AdaptFormer sub-network per task, and uses per-task MLP binary gates—described as artificial silent synapses—to select adapter features at inference. The reported CIFAR-100 and TinyImageNet accuracies exceed the strongest tabulated baseline, EASE, by roughly 0.9–2.1 percentage points. However, the learning protocol appears to require previous-task data for training every new gate, the task-routing mechanism is under-specified and potentially invalid for future tasks, and several equations, metrics, and dataset protocols are inconsistent. These issues prevent the principal empirical claims from being established under a well-defined and fair CIL setting.", + "strengths": [ + "The high-level modular design is understandable: a frozen pre-trained representation supplies stability, while task-specific adapters supply plasticity.", + "The evaluation includes two standard image-classification datasets, multiple incremental granularities, and comparisons with both conventional and pre-trained-model CIL methods.", + "The reported results are consistently better than the strongest tabulated baseline across all ten reported Average/Last accuracy entries.", + "The paper attempts to examine new-task learning and old-task retention separately and includes an ablation concerning features used by the gate.", + "The proposed use of learned routing between a frozen backbone and task-specific modules addresses a practically relevant problem in task-agnostic dynamic architectures." + ], + "weaknesses": [ + "The gate-training protocol appears to violate or materially alter the claimed CIL setting. Each new MLP gate is trained to classify all previous-task data, sum_{i=0}^{t-1} D^i, against current data D^t. The paper does not state that old raw data are stored, give a memory budget, or explain an alternative source of negative examples. If all previous data are retained, Artsy uses replay and is not directly comparable to rehearsal-free methods such as EASE under the described protocol.", + "The routing rule is not soundly specified for task-agnostic inference. Gate S_t is trained only to distinguish task t from tasks earlier than t. For an example belonging to a later task k>t, S_t has never received such samples as negatives, so its output is unconstrained. Multiple earlier gates may activate, yet the paper gives no normalization, exclusivity constraint, arbitration rule, or analysis of gate errors.", + "The mathematical definition is internally inconsistent. Equation 4 uses h + sum_i m_i h_t, repeating h_t rather than summing h_i, while h itself is undefined and apparently should be h_0. The classifier label space Y_b is also undefined. Algorithm 2 loops over t and invokes a classifier repeatedly but never defines which prediction is returned. Equations 3–4 therefore do not uniquely determine the implemented model.", + "The optimization problem is missing. Algorithm 1 says to optimize sums of networks but provides no sub-network loss, gate loss formula, prototype construction rule, classifier normalization, or alternating-training details. The gate threshold thr_t is not specified, even though it directly determines adapter activation.", + "The experimental protocols are contradictory. CIFAR-100 is said to use increments of 10 and 20 classes, whereas the table reports 10 and 20 steps; 20 classes per increment would yield five steps, not twenty. TinyImageNet is described as 100 base classes plus increments of 5, 10, or 20 classes, while the table labels these as 5, 10, and 20 steps. It is therefore unclear which task sequences produced the reported results.", + "The metric definition is ambiguous and likely incorrect: Avg_t is described as the mean through task t but is written as (1/T) sum_{t=1}^T Last_t. “Last accuracy” is alternately presented as performance at the final incremental stage, performance on the newest task, and a per-stage quantity. This ambiguity prevents precise interpretation of the table and plasticity/stability claims.", + "No number of runs, random seeds, variance, confidence intervals, or significance tests are reported. Consequently, the claim of “significant” improvement is unsupported, especially for the 0.93-point Average-accuracy improvement over EASE on CIFAR-100 with 10 steps. The prose also incorrectly reports this difference as 0.91 points.", + "The comparison lacks essential fairness details: exemplar budgets and old-data access are not reported; preprocessing and class orders are unspecified; baseline hyperparameters are absent; and the statement that all comparison methods were reimplemented is insufficient to verify matched protocols.", + "The ablation does not isolate the proposed contribution. “Good feature” and “bad feature” are undefined, quantitative values are not supplied in the text, and there is no comparison against no gating, always-on adapters, a conventional task router, Expert Gate/MoE-style routing, or an oracle task gate. Thus the evidence does not show that the silent-synapse formulation, rather than generic learned routing, produces the gains.", + "Artsy adds one adapter and one classifier per task, but parameter growth, storage, training cost, and inference FLOPs are not reported. This is particularly important because evaluating multiple gates and adapter features appears to scale with the number of tasks, undermining the stated efficiency motivation.", + "The biological mechanism is principally metaphorical. The implementation uses an SGD-trained MLP and hard threshold, without spike timing, local plasticity, receptor dynamics, or a computational counterpart to synaptic consolidation beyond routing. The paper does not empirically establish greater biological interpretability than other modular methods." + ], + "questions": [ + "Does training S_t access and retain every earlier raw dataset D^1,...,D^{t-1}? If so, what is the memory budget, and were baselines given the same replay data? If not, how are the negative gate examples obtained?", + "What are the exact inputs, hidden dimensions, objective, target balance, calibration procedure, and threshold for every S_t? How is the final prediction formed when zero or multiple gates activate?", + "How does S_t behave on data from tasks arriving after task t, which are absent from its training distribution? Please report the full gate-confusion matrix after the final task.", + "What is the corrected form of Equation 4, and what single output does Algorithm 2 return? Please also define h, Y_b, l, and the prototype construction and update process.", + "Are the table headings numbers of tasks or numbers of classes per increment? What base session and exact class counts are used for each CIFAR-100 and TinyImageNet experiment?", + "Can the authors report multiple-seed means and standard deviations under identical class orders and data-access budgets, together with no-gate, always-on, standard-router, and oracle-gate ablations?", + "What are the trainable parameter count, stored-data footprint, and training/inference FLOPs as the number of tasks grows?" + ], + "limitations": [ + "This review uses only the supplied paper text and metadata, as requested.", + "The numerical contents of referenced plot images are not present in the supplied text, so the claimed per-step curves and feature ablation cannot be independently checked.", + "No code, supplementary material, run-level results, or external sources were used; reproducibility and biological assertions can therefore only be assessed from the manuscript's descriptions." + ], + "soundness": 3, + "presentation": 4, + "contribution": 4, + "overall_score": 3, + "confidence": 5, + "ethical_concerns": "The paper extrapolates from an artificial feature-routing ablation to dementia and brain lesions, and suggests applications to neurodegenerative-condition simulation, without biological experiments or validation. These statements should be removed or sharply qualified because the presented ML evidence cannot support biomedical causal conclusions.", + "recommendation": "reject", + "review_text": "Artsy combines a frozen ViT-B/16, task-specific AdaptFormer modules, prototype classification, and learned binary gates framed as silent synapses. The tabulated gains over EASE are potentially interesting and consistent across the evaluated settings. Nevertheless, the current manuscript does not establish them under a coherent CIL protocol. Most critically, each new gate is said to train on all prior datasets as negatives, but old-data retention and its budget are never disclosed. This may give Artsy replay access that competing rehearsal-free methods do not receive. Independently, a gate trained to distinguish its task from only earlier tasks has unconstrained behavior on future-task examples, so the proposed collection of independent gates does not provide a well-defined task-agnostic router.\n\nThe method cannot be reproduced from the formal description. Equation 4 appears to contain an indexing error and undefined symbols; the testing algorithm generates multiple classifier calls without specifying the returned prediction; no adapter optimization loss is given; and gate thresholds and prototype updates are omitted. The dataset schedules and accuracy definitions are also contradictory. Finally, the paper provides no multi-run uncertainty, resource accounting, or direct ablation against standard routing alternatives. The biological narrative does not remedy these technical gaps: an SGD-trained MLP gate with a threshold does not by itself implement spike-timing-dependent plasticity or demonstrate synaptic consolidation. A substantially revised submission would need a precisely defined no-replay or budget-matched protocol, corrected equations and metrics, future-task-safe routing, complete implementation details, uncertainty estimates, routing baselines, and complexity analysis." + } +} diff --git a/codex_reviews/reviews/0CvJYiOo2b.json b/codex_reviews/reviews/0CvJYiOo2b.json new file mode 100644 index 0000000000000000000000000000000000000000..2e8ffa8d78e7f0f1b5e2bdab8001a9516eeac53a --- /dev/null +++ b/codex_reviews/reviews/0CvJYiOo2b.json @@ -0,0 +1,64 @@ +{ + "paper_id": "0CvJYiOo2b", + "title": "Revisiting PCA for Time Series Reduction in Temporal Dimension", + "model": "gpt-5.6-sol", + "reasoning_effort": "xhigh", + "created_at": "2026-08-09T17:27:29.709219+00:00", + "paper_md_chars": 35795, + "metadata": { + "pdf_time_verification_status": "verified_revision_cdate_before_review", + "review_time_revision_source": "submission_note", + "arxiv_version_id": null, + "markdown_source": "dataset" + }, + "review": { + "summary": "The paper proposes fitting PCA on training-set windows and using the resulting component scores as shorter inputs to time-series models. The idea is simple and potentially useful, and the evaluation spans classification, forecasting, and extrinsic regression. However, several central theoretical claims about information preservation are mathematically incorrect, and the experiments lack the controls, uncertainty estimates, and timing methodology needed to establish accuracy preservation or end-to-end efficiency reliably.", + "strengths": [ + "The method is simple, model-agnostic, and explicitly fits PCA only on the training split before applying it to validation and test data.", + "The empirical scope includes linear, Transformer, CNN, and RNN architectures across three task categories.", + "The paper considers both predictive performance and systems metrics such as training time, inference time, and GPU memory.", + "Comparisons with shortening, downsampling, and learned linear/1D-CNN reduction layers are directionally useful.", + "The paper identifies a practical and underexplored question: whether fixed-window temporal inputs can be compressed before model training." + ], + "weaknesses": [ + "The claimed theoretical justification is unsound. A truncated PCA projection does not generally preserve per-window means, sums, peaks, skewness, or kurtosis. Even an invertible orthogonal rotation need not preserve coordinate-wise marginal higher moments, and truncation discards additional information.", + "PCA is not inherently a denoiser: low-variance components may contain predictive signal, while high-variance components may contain nuisance variation. A smoother inverse reconstruction is not evidence that discarded components are noise.", + "The assertion that reversing trends or changing periodic amplitudes should not impair prediction is generally false unless the targets and data-generating relationship are transformed consistently. In forecasting, trend direction can be directly predictive of future values.", + "There is no formal theory despite repeated references to theoretical analysis: no predictive-risk bound, sufficient-statistic result, reconstruction-to-task-error connection, or assumptions under which temporal PCA preserves forecasting information.", + "Accuracy comparisons rely on point estimates and counts of cases in which PCA wins. No seeds, standard deviations, confidence intervals, or significance tests are reported, so claims of no performance loss cannot be distinguished from experimental noise.", + "Component counts such as 16, 48, and 96 appear manually fixed. The paper does not report explained variance, validation-based selection, or sensitivity curves showing the accuracy-efficiency tradeoff as k varies.", + "The efficiency evaluation is insufficiently specified: it is unclear whether PCA fitting and per-sample transformation are included, whether times are end-to-end, what hardware and batch sizes are used, whether GPU synchronization and warm-up are applied, and whether memory means peak allocated memory or average utilization. Normalized averages across tasks obscure absolute costs.", + "The evaluation is largely univariate: TSC keeps only the last channel, forecasting uses only oil temperature, and the proposed multivariate extension is restricted to channel-independent models. This weakens the broad generalization claims.", + "The PCA scores are ordered by variance rather than time, yet sequential models are apparently asked to treat component index as a sequence position. The consequences of this semantic change and the use of positional encodings are not analyzed.", + "The patch-based procedure is ambiguous about whether one PCA basis is shared across all patches or separate bases are fitted by patch position. These alternatives have different statistical meaning and computational cost.", + "Method notation is inconsistent: the paper defines D as containing inputs and targets but then describes applying PCA to D_train; eigenvalues are indexed to m although the covariance matrix is L by L; and it says k is smaller than m when the relevant original dimension is L.", + "The baseline set is limited. Fixed transforms and temporal aggregation methods such as random projections, Fourier/DCT or wavelet truncation, pooling/PAA, and compute-matched shorter-token models are absent. Learned reduction baselines also require clearer parameter and optimization matching.", + "The baseline Linear forecasting results are described as adapted from prior work, making comparability with newly run PCA experiments uncertain unless identical preprocessing, splits, implementations, and hyperparameters are demonstrated." + ], + "questions": [ + "What exact mathematical property supports the statements that truncated PCA preserves window means, sums, peaks, skewness, and kurtosis?", + "How is k selected without using test performance, and what are the explained-variance and accuracy-versus-k curves for every task?", + "Do reported training and inference times include PCA fitting, host-to-device transfer, and PCA transformation? What are the absolute times, hardware, batch sizes, warm-up protocol, and number of repetitions?", + "Are dataset splits created before sliding-window extraction, preventing overlapping train/test windows or other leakage?", + "For PatchTST, is PCA fitted once over all patches, separately for each patch position, or separately for each sample?", + "How do results change under multiple seeds, and are the observed accuracy differences statistically significant?", + "Why are multivariate datasets reduced to one channel, and does temporal PCA remain effective on standard multivariate settings with cross-channel dependence?", + "How does PCA compare at the same reduced dimension and compute budget with random projection, DCT/Fourier truncation, pooling/PAA, and smaller native models?", + "How are positional encodings handled when principal-component index replaces temporal position?", + "Are all non-PCA and PCA models retrained with independently tuned hyperparameters under exactly the same implementation and data pipeline?" + ], + "limitations": [ + "The supplied text does not expose the numerical cells of Tables 1-6, so exact effect sizes and possible large degradations cannot be independently checked from the provided prose.", + "The supplementary code is not included in the provided material, so implementation details and timing procedures cannot be verified.", + "The paper itself does not study irregular sampling, distribution shift, online adaptation, very long windows, or multivariate models with cross-channel interactions." + ], + "soundness": 3, + "presentation": 5, + "contribution": 4, + "overall_score": 4, + "confidence": 4, + "ethical_concerns": "No material ethical concerns are apparent from the provided paper text.", + "recommendation": "weak_reject", + "review_text": "This work evaluates a straightforward preprocessing strategy: fit PCA across temporal positions of fixed-length training windows, project each window to k component scores, and train standard TSA models on the shorter representation. The practical question is worthwhile, and the breadth across tasks and architectures is a positive feature.\n\nThe main obstacle is that the paper substantially overstates its theoretical basis. PCA optimizes rank-k linear reconstruction error under squared loss; it does not guarantee preservation of task-relevant information. In particular, the claims that the reduced scores preserve per-series means, sums, peaks, skewness, and kurtosis are false in general. Likewise, interpreting low-variance components as noise requires assumptions not supplied here. The trend-reversal argument is also invalid for forecasting because changing historical trend direction can change its relationship with the future target. These points affect the central explanation for why the method should work, rather than being minor presentation issues.\n\nThe empirical evidence is promising but not conclusive. The paper reports point estimates and win percentages without repeated trials or uncertainty. It does not explain how component counts were selected or report explained variance and sensitivity to k. The systems claims are based on normalized averages without enough detail to determine whether preprocessing cost is included or whether measurement is reproducible. The limited, largely univariate scope also does not support broad claims about time-series analysis generally. Moreover, interpreting variance-ordered PCA coefficients as temporal positions for Transformers, CNNs, and RNNs requires analysis, especially regarding positional encodings.\n\nA stronger paper would replace the incorrect preservation claims with a precise reconstruction objective and explicit assumptions connecting reconstruction to prediction; provide accuracy-efficiency curves over k; report repeated, end-to-end timing and memory measurements; test standard multivariate settings; and compare against stronger fixed-transform and compute-matched baselines. As written, the idea may be a useful engineering baseline, but the theoretical and empirical evidence is not yet sufficiently rigorous." + } +} diff --git a/codex_reviews/reviews/0JTwZ30qPH.json b/codex_reviews/reviews/0JTwZ30qPH.json new file mode 100644 index 0000000000000000000000000000000000000000..bb7413d1d1a0fed1383347a01ea44cfe8a13d049 --- /dev/null +++ b/codex_reviews/reviews/0JTwZ30qPH.json @@ -0,0 +1,62 @@ +{ + "paper_id": "0JTwZ30qPH", + "title": "Task-Oriented Multi-View Representation Learning", + "model": "gpt-5.6-sol", + "reasoning_effort": "xhigh", + "created_at": "2026-08-09T17:31:25.566080+00:00", + "paper_md_chars": 32375, + "metadata": { + "pdf_time_verification_status": "verified_revision_cdate_before_review", + "review_time_revision_source": "submission_note", + "arxiv_version_id": null, + "markdown_source": "dataset" + }, + "review": { + "summary": "The paper proposes TOMRL, a meta-learning wrapper for multi-view representation models. It embeds an episodic task using Fisher-information-like gradients, maps that embedding to FiLM scale/shift parameters, and modulates both view-specific and fused representations. The idea of making representations explicitly task-dependent is relevant, and the claimed plug-in compatibility is appealing. However, the mathematical construction, optimization procedure, and evaluation protocol are too underspecified to establish correctness or reproducibility. In particular, the proposed task embeddings are not well-defined for the general SVL/MVL losses claimed, the unsupervised task definition is ambiguous, and the experiments lack essential controls and statistical reporting in the supplied text.", + "strengths": [ + "Addresses a meaningful limitation of task-agnostic multi-view representations: different downstream tasks may require different features and fusion behavior.", + "The separation between view-specific modulation and shared fused-representation modulation is intuitive and permits useful component ablations.", + "The method is framed as an add-on to existing multi-view pipelines rather than requiring an entirely new backbone.", + "The paper considers both classification and clustering episodes and includes two directions of a cross-domain FashionMNIST-based experiment.", + "Ablations examine the two modulation locations, the view-index auxiliary loss, and several modulation mechanisms." + ], + "weaknesses": [ + "The Fisher information task embedding is not operationally specified. A full FIM is quadratic in the number of model parameters, yet no diagonal, block-diagonal, low-rank, normalization, dimensionality-reduction, or vectorization scheme is described. Consequently, its memory cost, input dimensionality to the meta-learner, and compatibility across different encoders are unclear.", + "Equation 4 assumes a probabilistic classifier p(y|x) for every view. Several advertised base methods instead use reconstruction, contrastive, clustering, or information-theoretic objectives. The assertion that arbitrary SVL losses can provide the same FIM is not justified, and the analogous entity-level FIM from an arbitrary MVL loss is not defined.", + "The unsupervised clustering episode is ambiguous and partly malformed: M is undefined, the support and query sets appear to be split by views rather than samples, their index ranges overlap at V', and the stated query consistency error is not defined. Sampling N classes for an allegedly unlabeled task may also use hidden labels, but the protocol does not explain whether or how labels are used.", + "The meta-learning objective does not clearly distinguish support adaptation from query evaluation. Equation 6 applies L_meta to an inner optimum but provides no support/query arguments, while Section 3.1.3 defines L_meta as the same base auxiliary loss plus view-index loss. It is therefore unclear what downstream task signal actually makes the learned modulation task-oriented.", + "No algorithm or sufficient implementation details are given for computing task embeddings, constructing the meta-learner, producing dimension-matched FiLM parameters, differentiating through the FIM and inner update, handling second-order derivatives, or applying the method at test time.", + "The comparisons confound task conditioning with added capacity and additional optimization. The paper states that TOMRL adds a meta-learning process after ordinary training, but it does not include matched controls such as extra fine-tuning without task conditioning, task-independent FiLM, random task embeddings, or parameter/compute-matched adapters.", + "The evaluation protocol omits train/validation/test class splits, episode sampling details, preprocessing, encoder architectures, meta-training duration, model-selection criteria, SVM hyperparameter selection, clustering assignment procedure, and whether test tasks influence training or tuning.", + "The text reports averages over 100 tasks but does not state standard deviations, confidence intervals, significance tests, or results across independent training seeds. Since the supplied tables are raster-image references, their exact numerical values and uncertainty cannot be verified from the provided text.", + "The cross-domain evaluation covers only two transformations of FashionMNIST in opposite directions. This is a narrow domain shift and does not strongly support broad cross-domain generalization.", + "The strong claims of consistent improvement, independence from auxiliary tasks and fusion strategies, robustness across task types, and compatibility with most existing methods exceed the evidence described. Ablations are limited to one dataset and one baseline, and no failure cases or computational costs are reported.", + "The exposition contains malformed equations and notation, including the output grouping in Equation 3, the view-index loss, and the optimization variables in Equation 2. These issues obscure the precise model being proposed." + ], + "questions": [ + "What exact approximation and representation of the FIM is passed to the meta-learner, what is its dimensionality, and what are its time and memory costs for each backbone?", + "How are z_t^v and Z_t defined when a baseline's SVL/MVL loss is not a normalized likelihood and provides no p(y|x)?", + "Please give a complete algorithm separating meta-training support updates, query losses, outer updates, and meta-test inference. Is differentiation through the FIM itself required?", + "For clustering episodes, what are M, V', the support/query split, the query consistency loss, and the procedure for resolving cluster-label permutations? Are ground-truth classes used to construct episodes?", + "Which loss directly encourages modulation to improve the downstream classification or clustering objective, rather than merely reducing the original representation-learning and view-index losses?", + "How do TOMRL gains compare with equal-compute extra fine-tuning, task-independent FiLM, random or pooled-gradient embeddings, and a model using only one of the two task embeddings?", + "What are the exact dataset splits, number of independent training runs, uncertainty estimates, parameter counts, runtime, and hyperparameter-selection procedures?", + "Can the authors demonstrate transfer between genuinely distinct datasets or modalities, rather than only NoisyFashion and EdgeFashion variants of the same underlying dataset?" + ], + "limitations": [ + "Exact effect sizes cannot be checked because the supplied paper text contains table images without transcribed numerical entries.", + "The method description is insufficient to determine whether the claimed general integration with arbitrary multi-view objectives is mathematically valid.", + "The experimental evidence is restricted to four relatively small or synthetic multi-view datasets, with cross-domain evaluation limited to FashionMNIST variants.", + "There is no reported computational or memory analysis despite the potentially prohibitive FIM and bi-level differentiation costs.", + "The related-work discussion alone does not substantiate the claim that this is the first exploration of task-oriented multi-view representations." + ], + "soundness": 3, + "presentation": 3, + "contribution": 5, + "overall_score": 3, + "confidence": 4, + "ethical_concerns": "No specific ethical concerns are apparent from the supplied paper. The datasets and tasks described are standard representation-learning benchmarks, although broader impacts and potential biases are not discussed.", + "recommendation": "reject", + "review_text": "TOMRL starts from a compelling observation: a single task-agnostic multi-view representation need not emphasize the information appropriate for every downstream task. The proposed response—derive a task embedding from gradients and use it to generate FiLM parameters for both view-specific and fused features—is conceptually reasonable and potentially useful as a modular extension to existing methods.\n\nThe central difficulty is that the method is not specified precisely enough to evaluate its correctness. Equation 4 defines a full FIM over encoder parameters, but such a matrix is generally enormous and its approximation and conversion into a fixed-size meta-learner input are omitted. It also presupposes p(y|x), whereas the paper claims compatibility with arbitrary reconstruction, contrastive, clustering, and information-theoretic SVL/MVL objectives. The entity-level analogue is only asserted. Thus, the mechanism needed to support the main generality claim is undefined.\n\nThe episodic and optimization formulations have further ambiguities. The clustering task appears to split support and query data by view, uses an undefined M, overlaps the stated view ranges, and never defines the query consistency loss. Equation 6 does not identify separate support and query objectives, while L_meta is described as the original base loss plus a view-index loss. It is consequently unclear which outer-loop signal teaches the modulation to improve the actual downstream task. A complete training and inference algorithm is necessary.\n\nThe empirical design also does not isolate the proposed source of improvement. TOMRL adds parameters and an additional training phase, but there are no equal-compute fine-tuning or task-independent FiLM controls. Dataset splits, episode construction, architectures, meta-learner details, training duration, test-time adaptation, and model selection are insufficiently documented. The supplied text states that 100-task averages are reported but provides no uncertainty estimates or multiple-seed analysis. The cross-domain experiment is limited to two derived views of FashionMNIST, so it offers weak evidence for general domain transfer.\n\nOverall, the task-conditioning perspective is promising and the two-location modulation design is intuitive, but major gaps in the mathematical definition, experimental controls, and reproducibility prevent the current evidence from supporting the paper's claims." + } +} diff --git a/codex_reviews/reviews/0JcPJ0CLbx.json b/codex_reviews/reviews/0JcPJ0CLbx.json new file mode 100644 index 0000000000000000000000000000000000000000..472d46644895a352c1312d696c4bb6ed548a5f95 --- /dev/null +++ b/codex_reviews/reviews/0JcPJ0CLbx.json @@ -0,0 +1,64 @@ +{ + "paper_id": "0JcPJ0CLbx", + "title": "Revisiting MAE pre-training for 3D medical image segmentation", + "model": "gpt-5.6-sol", + "reasoning_effort": "xhigh", + "created_at": "2026-08-09T17:31:44.406674+00:00", + "paper_md_chars": 39901, + "metadata": { + "pdf_time_verification_status": "verified_revision_cdate_before_review", + "review_time_revision_source": "submission_note", + "arxiv_version_id": null, + "markdown_source": "dataset" + }, + "review": { + "summary": "The paper revisits masked-autoencoder pre-training for 3D MRI segmentation using a residual-encoder U-Net within nnU-Net. It pre-trains on 39,168 clinical MRI volumes and uses five development datasets to select masking, sparsification, and fine-tuning choices, followed by evaluation on additional datasets. The central claim is that the resulting Spark3D model consistently improves over training from scratch and several self-supervised baselines. The study asks an important question and adopts a substantially stronger development framework than many prior medical SSL papers. However, the supplied text does not include the numerical contents of any result tables, and several aspects of data separation, statistical analysis, baseline fairness, and reproducibility are insufficiently specified. Consequently, the headline improvements and state-of-the-art claim cannot be verified from the paper text.", + "strengths": [ + "The paper targets a practically important problem: whether self-supervised pre-training adds value over a strong, modern nnU-Net baseline rather than an outdated or poorly configured network.", + "The pre-training collection is unusually large for 3D medical imaging: 39,168 MRI scans from more than 9,000 patients and over 44 centers, with multiple scanners and MRI sequences.", + "The development protocol uses five downstream datasets spanning lesions, anatomy, and tumors, reducing the risk that design decisions are optimized for only one task.", + "The paper separates a fixed-configuration from-scratch control from dynamically planned nnU-Net. This is useful because it distinguishes benefits of initialization from benefits or costs of using a fixed preprocessing and architecture configuration.", + "The ablation scope is valuable: sparsification components, masking ratio, weight transfer, warm-up, learning rate, encoder freezing, pre-training duration, fine-tuning duration, low-data behavior, cross-center generalization, unseen modalities, and multi-channel inputs are all considered.", + "Reimplementing MG, VoCo, and VolumeFusion using the same pre-training data and backbone is potentially more informative than comparing against numbers reported under unrelated experimental settings." + ], + "weaknesses": [ + "The quantitative evidence is missing from the supplied paper text. All central tables are unresolved LaTeX inputs, such as `tables/test_dice`, `tables/sparse_ablation`, and `tables/ft_results_full`, while the figures are also unavailable. Thus, the claimed approximately 3-point gain, dataset-level consistency, NSD improvements, ablation effects, and ranking stability cannot be checked.", + "Dataset accounting is internally inconsistent. The paper defines five development datasets and eight test datasets, but repeatedly refers to 11 test datasets and an improvement on 10 of 11. D6–D12 provide seven fine-tuning test datasets, while D13 is a separate zero-shot generalization experiment. It is unclear exactly which datasets enter each aggregate result.", + "The headline comparison is ambiguous. The abstract and introduction claim approximately three Dice points over the strong nnU-Net baseline, whereas the Results section reports a two-point average gain over `No (Fixed)`. No explicit aggregate improvement over `No (Dyn.)` is stated in the available text, even though the conclusion claims consistent improvement over dynamic nnU-Net.", + "Potential data leakage is not adequately addressed. The paper does not state whether downstream subjects or scans were explicitly excluded from the proprietary pre-training collection, whether near-duplicate examinations were detected, or whether patients appearing at multiple centers were deduplicated.", + "The split unit is unspecified. Because the clinical corpus contains multiple scans per patient, image-level splitting could place examinations from the same patient in different partitions. The paper should state that all pre-training filtering and downstream train/validation/test splits are patient-level.", + "Statistical support is insufficiently described. D6–D12 use a single 80/20 split, each method is reportedly fine-tuned only once, and no repeated-seed variability, confidence intervals, or paired significance tests are specified. Bootstrapping is mentioned, but its sampling unit, number of replicates, handling of datasets and classes, and uncertainty intervals are not given.", + "Baseline fairness remains uncertain. S3D receives extensive tuning over five datasets, while competing SSL objectives use the same hyperparameters only `where possible` and appear not to receive comparable objective-specific tuning. The authors themselves acknowledge that VoCo and VolumeFusion hyperparameters may be suboptimal for this backbone, weakening conclusions that MAEs intrinsically dominate those methods.", + "The scope of the state-of-the-art claim is too broad. Only three SSL baselines are included in the final comparison, despite AMAES, HySparK, GL-MAE, and other methods being discussed. Using the same backbone is a strong controlled comparison, but it does not by itself establish overall state of the art.", + "Important reproducibility details are absent or underspecified: exact residual U-Net configuration and capacity, sparse-convolution and masked-normalization definitions, masking implementation at every resolution, decoder design, total downstream training schedule, inference procedure, modality classification, quality-control process, compute cost, seeds, and dataset sample counts.", + "The fixed 160^3 patch and 1 mm isotropic spacing can severely alter small structures and high-resolution scans, as acknowledged for D12. This makes the fixed baseline a weak comparator on some tasks. Pre-training should also be compared under each dataset's dynamically planned architecture or through a controlled initialization-compatible adaptation.", + "The proprietary pre-training dataset limits independent reproduction. The public code/model claim is useful, but an external repository cannot substitute for a complete description of the data composition and experimental protocol in the paper.", + "The clinical-corpus filtering rules are very coarse. File size, field of view, and spacing do not reliably remove scout, corrupted, or non-diagnostic scans, and the paper provides no audit of residual noise or filtering precision." + ], + "questions": [ + "Exactly which 11 datasets enter the main DSC/NSD averages, and how do they relate to D1–D13? Please reconcile the five-development/eight-test description with the repeated `11 test datasets` statement.", + "What are the actual per-dataset results, aggregate definitions, confidence intervals, and paired statistical comparisons? Are averages macro-averaged over classes, cases, folds, or datasets?", + "Were all splits performed at patient level, and were pre-training subjects, scans, or near-duplicates explicitly checked against every downstream dataset?", + "How many cases and foreground classes are used in each downstream train, validation, and test split? Were the random splits fixed before method development?", + "How is the bootstrap ranking computed—what is resampled, how many replicates are used, and are dependencies among classes and cases respected?", + "Were multiple random seeds evaluated for pre-training and fine-tuning? If not, how can gains of roughly 0.3 Dice points in some ablations be distinguished from stochastic variation?", + "How were baseline-specific hyperparameters selected? Were MG, VoCo, and VolumeFusion granted a tuning budget comparable to the S3D development process?", + "Does S3D improve the aggregate score over `No (Dyn.)` by approximately three points, or does that claim concern `No (Fixed)`? Please report both comparisons explicitly.", + "Can the authors initialize dynamically planned nnU-Net variants from the pre-trained model, or otherwise disentangle pre-training quality from the disadvantage imposed by fixed 1 mm preprocessing?", + "What ethics approval, consent or waiver, de-identification, data-governance, and demographic auditing procedures were used for the proprietary multi-center clinical corpus?" + ], + "limitations": [ + "This review is restricted to the supplied paper text. Numerical table contents, figure contents, references, appendices, and supplementary material were not provided.", + "The linked code and models were not accessed, so implementation details and reported results could not be independently checked.", + "Because dataset sizes and numerical results are missing, the magnitude, consistency, and statistical reliability of the reported gains cannot be assessed." + ], + "soundness": 4, + "presentation": 6, + "contribution": 7, + "overall_score": 5, + "confidence": 4, + "ethical_concerns": "The work uses a large proprietary clinical dataset from more than 9,000 patients and over 44 centers, but the supplied text does not describe ethics-board approval, consent or waiver, de-identification, data-use agreements, security controls, demographic composition, or bias analysis. It also does not establish that downstream subjects were absent from pre-training. These omissions should be addressed before the clinical-data methodology can be adequately assessed.", + "recommendation": "weak_reject", + "review_text": "This is a promising and potentially impactful empirical study. Its strongest aspect is the decision to evaluate SSL against a modern residual nnU-Net under a broad, validation-driven protocol, using substantially more pre-training data than is typical in 3D medical imaging. The controlled fixed-backbone comparisons and the breadth of ablations could provide useful guidance to the field.\n\nNevertheless, the supplied manuscript does not provide enough checkable evidence for its central conclusions. The numerical contents of every major result table are absent, making it impossible to verify the approximately 3-point improvement, the claimed consistency, or the ablation conclusions. There are also unresolved inconsistencies in the number of evaluated datasets and in whether the main improvement is measured against fixed or dynamically planned nnU-Net. The experimental protocol needs clearer guarantees against subject overlap, patient-level leakage, and overlap between the proprietary pre-training corpus and downstream datasets. Single fine-tuning runs on single 80/20 splits, without described uncertainty estimates, are insufficient for interpreting small differences. Finally, extensive S3D tuning is contrasted with baselines whose hyperparameters may be suboptimal for the common backbone, so the assertion that MAEs dominate alternative objectives is not yet supported.\n\nI would reconsider the score if a revision supplied complete numerical results, reconciled the dataset accounting, documented patient-level separation and ethics procedures, reported repeated-run or paired-bootstrap uncertainty, and gave competing methods a transparent and comparable tuning budget. As written in the provided text, the methodology is encouraging but the headline claims are not adequately verifiable." + } +} diff --git a/codex_reviews/reviews/0QePvFoqY6.json b/codex_reviews/reviews/0QePvFoqY6.json new file mode 100644 index 0000000000000000000000000000000000000000..ac03451eecaa9a747536479d598aec8eacbc1de4 --- /dev/null +++ b/codex_reviews/reviews/0QePvFoqY6.json @@ -0,0 +1,61 @@ +{ + "paper_id": "0QePvFoqY6", + "title": "IncEventGS: Pose-Free Gaussian Splatting from a Single Event Camera", + "model": "gpt-5.6-sol", + "reasoning_effort": "xhigh", + "created_at": "2026-08-09T17:38:00.147286+00:00", + "paper_md_chars": 56117, + "metadata": { + "pdf_time_verification_status": "verified_arxiv_version_before_review", + "review_time_revision_source": "arxiv", + "arxiv_version_id": "2410.08107v1", + "markdown_source": "arxiv" + }, + "review": { + "summary": "The paper proposes IncEventGS, an incremental monocular event-camera system that jointly estimates camera motion and a 3D Gaussian scene without externally supplied poses. Events are grouped into 50 ms chunks; each chunk uses an SE(3) trajectory interpolated between start and end poses. Differences between Gaussian-rendered log-intensity images supervise tracking and sliding-window bundle adjustment. Initialization jointly optimizes random Gaussians and poses, then uses a pretrained monocular depth model to reinitialize Gaussian centers. Reported experiments on five synthetic Replica scenes and five real TUM-VIE sequences show large NVS and ATE improvements over event-NeRF, E2VID-based, and DEVO baselines. The problem is important and the results are promising, but key formulation, evaluation, scale-alignment, and reproducibility details are missing or internally inconsistent.", + "strengths": [ + "Addresses a challenging and practically relevant setting: incremental dense reconstruction and motion estimation from a single event camera without externally provided camera poses.", + "The integration of a differentiable event-formation loss, continuous within-chunk SE(3) trajectories, Gaussian rendering, tracking, and sliding-window joint optimization is coherent at a high level.", + "The reported Replica NVS gains are substantial across PSNR, SSIM, and LPIPS rather than being confined to one metric or scene.", + "The paper evaluates trajectory estimation on both synthetic and real sequences and compares against an event-only visual-odometry baseline as well as an image-reconstruction pipeline.", + "The depth-initialization and event-window ablations indicate that bootstrapping and temporal support materially affect both reconstruction and tracking." + ], + "weaknesses": [ + "The metric-scale and ATE protocol is insufficiently specified. Monocular event reconstruction has a scale gauge, yet ATE is reported in centimeters without stating whether EVO uses rigid, scale, or Sim(3) alignment. The pretrained depth model may inject scale, but its depth convention, normalization, and alignment are not described. Consequently, the very small ATE numbers cannot be interpreted or verified from the paper.", + "The event-sampling description is internally inconsistent. The method first creates n_seg=100 segment-end timestamps and says the index of t_k is obtained by subtracting n_win, but n_win is sampled from bounds as large as 400k–500k. Such an index cannot refer to a list of 100 timestamps; the ablation instead suggests these values count events. This prevents faithful implementation.", + "The measured event image is not mathematically defined: E(x)=C{e_i} omits an explicit polarity sum and any handling of multiple events at a pixel. The formulation also assumes a fixed known contrast threshold, while the real-data value is selected empirically, with no sensitivity analysis or treatment of per-pixel threshold variation, noise, refractory effects, or hot pixels.", + "The SSIM objective is ambiguous. The system minimizes a positive SSIM term, which would be incorrect if SSIM denotes similarity; the prose calls it a structural dissimilarity loss but does not define the conversion. The sliding-window sums run from 0 to n_w while being divided by n_w, apparently including n_w+1 terms.", + "The NVS evaluation protocol does not identify training versus held-out views, how query views are selected, whether the evaluated timestamps contributed events to reconstruction, or how grayscale event-derived output is compared with ground truth. The linear color transformation can materially affect event-only intensity evaluation, but its fitting set and parameterization are not stated.", + "Baseline fairness is unclear. IncEventGS receives a strong pretrained monocular-depth prior, whereas no corresponding geometry prior is described for the baselines. The paper attributes gains specifically to Gaussian representation and bundle adjustment, but it does not isolate these factors from depth initialization, pose quality, implementation choices, or baseline tuning.", + "Real-data image quality is supported only by unaligned, cropped nearest RGB-camera references. These are explicitly not ground truth aligned with the event camera, so they cannot establish accurate novel-view synthesis. Moreover, the referenced image contents are not present in the supplied text and thus the qualitative claims cannot be independently assessed here.", + "The claimed efficiency is not substantiated with runtime, memory, Gaussian count, or throughput. Performing 200 tracking and 1500 mapping steps for every 50 ms chunk appears unlikely to be real time, but the paper gives no timing measurements.", + "Important system details are absent: initialization distribution and bounding box, pose initialization and propagation between chunks, enforcement of trajectory continuity, Gaussian insertion frequency and attributes, pruning criteria, depth-model scale handling, key optimizer settings, failure detection, and boundary handling for the sliding window.", + "There are several reporting inconsistencies: the depth ablation text reports improvement to 0.064 cm while its table reports 0.046 cm; the ablation scene or aggregation protocol is unspecified; no repeated-run variance is reported despite random initialization and stochastic sampling; and the projected covariance is declared 3x3 although J Sigma J^T is 2x2." + ], + "questions": [ + "Does ATE use no alignment, SE(3) alignment, or Sim(3) alignment, and exactly how are estimated trajectories converted to metric centimeters?", + "Are n_low and n_up counts of events, timestamp-segment indices, or another quantity? Please provide executable pseudocode for selecting t_k and t_{k+Delta t}.", + "What is the exact per-pixel definition of E(x), including polarity summation, repeated events, normalization, masking of inactive pixels, and treatment of contrast thresholds?", + "Is L_ssim equal to SSIM, 1-SSIM, or a different structural dissimilarity? Why do the written window sums contain n_w+1 terms?", + "What views constitute the NVS test set, and are all events or poses associated with those views excluded from optimization? How is the linear color transform fitted without test leakage?", + "Which pretrained depth model and checkpoint are used, and how are its affine/metric depth ambiguity, input normalization, and scale reconciled across chunks?", + "Can the authors report end-to-end processing time per 50 ms chunk, peak memory, Gaussian count, and trajectory/reconstruction accuracy under causal real-time operation?", + "Can the main comparisons be supplemented with repeated-run statistics and ablations for the depth prior, joint BA, tracking-only poses, trajectory continuity, chunk duration, and imperfect contrast thresholds across all scenes?" + ], + "limitations": [ + "Evaluation covers only five synthetic indoor scenes and five sequences from one real dataset, with no outdoor, dynamic-scene, extreme-lighting, or cross-sensor testing.", + "The model assumes a static scene, linear Lie-group interpolation within each 50 ms chunk, calibrated intrinsics, and a fixed contrast threshold; robustness when these assumptions fail is not evaluated.", + "Absolute appearance is underconstrained by brightness-change events, and the evaluation's fitted linear color transform obscures how much absolute intensity information is actually recovered.", + "The system depends critically on a pretrained monocular-depth prior, as shown by the large ATE degradation without it, limiting the strength of the claimed event-only geometric recovery.", + "Real-scene reconstruction quality lacks aligned quantitative ground truth, and no geometric map metric such as depth, point-cloud, or surface accuracy is reported." + ], + "soundness": 4, + "presentation": 5, + "contribution": 6, + "overall_score": 4, + "confidence": 4, + "ethical_concerns": "No material ethical concerns are apparent from the supplied paper text. The work uses synthetic Replica data and an established event-camera dataset; no sensitive personal-data use or harmful deployment is described.", + "recommendation": "weak_reject", + "review_text": "IncEventGS tackles an important and difficult problem, and its high-level construction is appealing: continuous within-chunk poses connect event accumulation to differentiable Gaussian rendering, while tracking and sliding-window mapping jointly recover motion and scene structure. The numerical gains reported on Replica are large, and the inclusion of real-sequence trajectory evaluation is valuable.\n\nHowever, the current paper does not specify several details needed to establish correctness. Most seriously, metric ATE is reported in centimeters without specifying trajectory alignment or resolving monocular scale. The depth prior may supply some scale information, but its output convention and scale handling are absent. The event-window sampling description is also impossible to reconcile as written: indices into 100 segment timestamps are offset by values of 400k–500k. The event accumulation equation omits the polarity sum, and the SSIM and sliding-window objectives contain ambiguities or apparent errors.\n\nThe experimental evidence also leaves major questions. The NVS split and possible overlap between evaluation views and optimized event data are unspecified. The fitted linear color transform is not described sufficiently to exclude leakage. Comparisons do not isolate the benefit of Gaussian representation or joint BA from the strong pretrained depth prior. Real-image evaluation uses unaligned nearest-camera references, and efficiency claims lack any runtime or memory results despite 1700 optimization steps per 50 ms chunk. Randomized optimization is reported without variance.\n\nI view this as a promising system and a potentially useful contribution, but the missing alignment protocol, contradictory sampling definition, dependence on an underspecified depth prior, and incomplete evaluation currently prevent the central accuracy and pose-free reconstruction claims from being verified. A revision resolving these issues and providing causal runtime, held-out-view evaluation, repeated trials, and broader ablations could be substantially stronger." + } +} diff --git a/codex_reviews/reviews/0T8vCKa7yu.json b/codex_reviews/reviews/0T8vCKa7yu.json new file mode 100644 index 0000000000000000000000000000000000000000..3208ea2f4b223b0171dcb0d72fedd2c7879af55c --- /dev/null +++ b/codex_reviews/reviews/0T8vCKa7yu.json @@ -0,0 +1,61 @@ +{ + "paper_id": "0T8vCKa7yu", + "title": "LLM Compression with Convex Optimization—Part 1: Weight Quantization", + "model": "gpt-5.6-sol", + "reasoning_effort": "xhigh", + "created_at": "2026-08-09T17:42:28.710265+00:00", + "paper_md_chars": 38509, + "metadata": { + "pdf_time_verification_status": "verified_revision_cdate_before_review", + "review_time_revision_source": "submission_note", + "arxiv_version_id": null, + "markdown_source": "dataset" + }, + "review": { + "summary": "The paper proposes CVXQ, a post-training mixed-precision weight-quantization framework. It estimates layer or channel sensitivity using projected output Jacobians, allocates relaxed bit depths through a dual update derived from a high-rate rate–distortion approximation, then applies companding quantization, bias correction, and matrix partitioning or clustering. The direction is interesting, but the paper overstates the method as an exact convex solution: the original objective is nonconvex and discrete, while the actual optimization relies on local linearization, separability and distributional assumptions, followed by heuristic iteration and integer rounding. Important experimental evidence is also not verifiable from the supplied text because most tables are image-only and several central implementation and evaluation details are missing.", + "strengths": [ + "The paper formulates mixed-precision allocation explicitly as a rate-constrained optimization problem, giving a useful conceptual connection between LLM quantization and rate–distortion theory.", + "The proposed allocation considers both weight variance and output sensitivity, and it supports finer row, column, or clustered granularity rather than assigning one precision to an entire layer.", + "The combination of bit allocation, companding, bias correction, and pruning provides a reasonably broad post-training compression framework.", + "Experiments reportedly cover several OPT and Llama-2 scales, 3-bit and 4-bit settings, language modeling, and downstream question-answering tasks.", + "The paper discusses practical tradeoffs involving cluster size, metadata overhead, calibration cost, and quantized matrix-vector kernels." + ], + "weaknesses": [ + "The central characterization as convex optimization is not established for the stated objective. The distortion in Eq. (3) is a discrete, coupled, nonconvex function of all quantized weights. Convexity holds only for the surrogate terms proportional to 2^(-2B_n) when G_n and S_n are fixed; recomputing G_n after quantization does not provide a proof of convergence or global optimality.", + "The claim that the method solves the problem “exactly” is unsupported. Eq. (5) uses a high-rate approximation despite experiments emphasizing very low precisions, including 0–4 bits. It also neglects inter-layer cross terms and appears to require independence or isotropy assumptions to reduce Jacobian-weight interactions to G_n^2 S_n^2.", + "Continuous relaxation followed by independent integer rounding may violate the exact rate constraint and need not solve the integer allocation problem. No rounding or residual-budget correction algorithm, approximation guarantee, or optimality comparison against exact allocation on small instances is given.", + "The assumption H_1=...=H_N is consequential because the framework uses companding and layers or channels can have different distributions. Dropping H_n can change the bit allocation, but this approximation is not empirically tested.", + "Under the standard linear-layer convention, the stated bias correction appears to have the wrong sign: matching the mean output would require B_n^q = B_n + (Theta_n - Theta_n^q) Xbar_n, whereas the paper writes the reverse difference. The paper must clarify its tensor convention or correct the formula.", + "The stated model-size constraint counts only P_n B_n. Scale, mean, cluster-index, padding or packing, and corrected-bias costs are excluded from the optimization. Consequently, the claim of reaching any user-specified model size is not demonstrated, especially at fine grouping granularity.", + "Most numerical tables are embedded only as images in the supplied paper text, so their values cannot be independently checked here. The visible prose reports selected gains but is insufficient to verify all model-baseline comparisons, ablations, overheads, pruning rates, or downstream-task results.", + "The empirical protocol lacks important details: baseline versions and tuning, quantizer symmetry and packing, selection and rounding procedures, companding grid ranges, PCA dimension, token-subsampling policy, downstream prompting and shot settings, evaluation harness, repetitions, and variability. There are also inconsistent references—for example, the discussion attributes GSM8K results to Table 4(a), while Table 4 is captioned as pruning and overhead.", + "The final quantized checkpoint is selected using WikiText2 validation while WikiText2 test is the main reported result, but the exact model-selection protocol is unclear. Some ablations instead use C4 test or validation, making the split conventions difficult to follow.", + "Scalability and efficiency claims are only weakly supported in the supplied text. The paper claims minutes to hours and applicability to hundreds of billions of parameters, but gives no complete optimization-time or peak-memory table. A single matrix-vector benchmark for an OPT-175B-shaped matrix does not establish end-to-end inference performance or successful quantization of a 175B model.", + "The abstract states that a reference implementation “can be obtained from” without supplying a location. Algorithm 1 also contains inconsistent notation and apparent hyperparameter swaps between the gradient EMA and dual update, substantially limiting reproducibility." + ], + "questions": [ + "For precisely which surrogate objective is CVXQ convex, and what convergence or optimality result applies when G_n is recomputed after every quantization update?", + "Can the authors derive Eq. (5) while explicitly stating the high-rate, independence, zero-mean-error, and cross-layer assumptions, then measure the approximation error at 0–4 bits?", + "How are fractional bit depths converted to integers while satisfying the exact total-bit constraint? How does this allocation compare with an exact dynamic-programming or integer-programming solution on small models?", + "Is the sign in the bias-correction formula intentional under a nonstandard layer convention? Please derive the correction from the layer equation.", + "Are metadata, corrected biases, alignment, and packing included in each reported average bit rate? Can the authors report total serialized model sizes for CVXQ and every baseline?", + "What PCA rank is used, how is it computed from calibration data, and how does the projected Jacobian-norm estimate correlate with the unprojected sensitivity?", + "What are the complete downstream evaluation settings, and are improvements consistent across multiple calibration samples or random seeds?", + "Were models at the claimed hundred-billion-parameter scale actually quantized end to end? If so, what were wall-clock time, peak host and device memory, hardware, and end-to-end inference throughput?", + "Where is the promised reference implementation, and what are the precise companding inverse, parameter-search grids, and numerical update schedules needed to reproduce the results?" + ], + "limitations": [ + "This review is restricted to the supplied paper text. Several tables and figures are represented only as images without machine-readable numerical content, so many reported results cannot be verified.", + "The paper itself focuses on weight-only post-training quantization; its claims about activation quantization are deferred to another work and are not validated here.", + "The available evidence does not establish behavior outside the reported OPT and Llama-2 families, calibration distribution robustness, or hardware-general end-to-end speedups." + ], + "soundness": 4, + "presentation": 3, + "contribution": 5, + "overall_score": 4, + "confidence": 4, + "ethical_concerns": "No direct ethical concern is apparent. The paper motivates compression partly through environmental benefits, but it does not quantify energy use or account for the optimization and calibration cost, so that benefit remains unverified.", + "recommendation": "weak_reject", + "review_text": "CVXQ offers an appealing rate–distortion interpretation of mixed-precision quantization and combines sensitivity-based bit allocation with companding and fine-grained grouping. These are potentially useful ideas, and the reported breadth across model scales is promising. However, the main technical framing is substantially stronger than the evidence permits. The original rate-constrained quantization problem is discrete, coupled, and nonconvex. The closed-form bit allocation is convex only for a separable high-rate surrogate with fixed sensitivity statistics, while the implemented procedure repeatedly changes those statistics and rounds the solution. Thus neither exact solution of Eq. (3) nor global optimality has been shown. The derivation additionally suppresses cross-layer interactions and distribution-dependent coefficients, and its validity at the low bit depths of interest is not evaluated. The apparent sign reversal in the bias correction is another potentially serious correctness issue.\n\nThe empirical section does not currently resolve these concerns. Numerical tables are unavailable in the supplied text except as images, and the prose alone cannot substantiate the full comparisons. Total serialized sizes, integer-budget handling, reproducible baseline settings, statistical variability, and end-to-end quantization and inference costs are not adequately specified. Claims of arbitrary target sizes and hundred-billion-parameter scalability therefore remain unverified. A stronger revision would precisely delimit the convex surrogate, remove or prove the exactness claim, validate its approximations against exact small-scale allocation, correct or derive the bias update, and provide complete reproducible results using total storage and end-to-end runtime. In its present form, the conceptual contribution is interesting but the soundness and empirical substantiation are insufficient for acceptance." + } +} diff --git a/codex_reviews/reviews/0d1gQI114C.json b/codex_reviews/reviews/0d1gQI114C.json new file mode 100644 index 0000000000000000000000000000000000000000..b9e72c4f4e80ff9380d74295c282a616df7546a2 --- /dev/null +++ b/codex_reviews/reviews/0d1gQI114C.json @@ -0,0 +1,55 @@ +{ + "paper_id": "0d1gQI114C", + "title": "LiDAR-PTQ: Post-Training Quantization for Point Cloud 3D Object Detection", + "model": "gpt-5.6-sol", + "reasoning_effort": "xhigh", + "created_at": "2026-08-09T17:54:19.908381+00:00", + "paper_md_chars": 43810, + "metadata": { + "pdf_time_verification_status": "verified_revision_cdate_before_review", + "review_time_revision_source": "submission_note", + "arxiv_version_id": null, + "markdown_source": "dataset" + }, + "review": { + "summary": "The paper introduces LiDAR-PTQ, a post-training INT8 quantization pipeline for LiDAR-based 3D detectors. It combines max-min calibration with grid search, a pseudo-label-based task-level distillation loss (TGPL), and adaptive weight rounding. On Waymo, the reported mAPH gaps from FP32 are small for CenterPoint-Pillar (60.12 versus 60.32), CenterPoint-Voxel (65.18 versus 65.25), and FSD (70.73 versus 70.94). The empirical results are promising, but important algorithmic and deployment details are missing or internally inconsistent, limiting reproducibility and the strength of the causal and efficiency claims.", + "strengths": [ + "The paper addresses a practically important deployment problem and motivates why highly sparse LiDAR feature distributions may behave differently from dense image features under calibration.", + "The component-wise ablation is informative: CenterPoint-Pillar mAPH improves from 52.91 with max-min calibration to 58.39 with grid search, 59.40 with TGPL, and 60.12 with adaptive rounding.", + "Results cover a dense pillar model, a sparse-convolution voxel model, and the fully sparse FSD detector. All three reportedly retain accuracy within 0.21 mAPH of their FP32 counterparts.", + "The method uses only 256 unlabeled calibration frames, or 0.16% of the stated Waymo training set, and compares calibration cost against several PTQ baselines and QAT.", + "The paper includes measurements on an edge-oriented Jetson AGX Orin rather than reporting only theoretical operation counts." + ], + "weaknesses": [ + "Adaptive rounding is not technically specified well enough to be valid as written. Equation 13 applies a hard round to (x+θ)/s, whose gradient with respect to θ is zero almost everywhere. The paper does not define a differentiable relaxation, straight-through estimator, constraint mechanism, hardening schedule, or rounding regularizer. Consequently, it is unclear how θ is actually optimized.", + "The quantizer definition is internally inconsistent. The paper claims signed symmetric quantization while retaining and optimizing nonzero zero-points. Its scale formula, (x_max-x_min)/(2^b-1), also does not cleanly match the stated signed range [-128,127] when x_min=-x_max. Per-tensor versus per-channel quantization is not specified.", + "The claimed root cause is not established causally. The evidence shows that entropy calibration performs poorly and that degradation grows with distance, but it does not directly measure foreground clipping or test entropy calibration after excluding zero entries. Moreover, the proposed “sparsity-based” calibration is conventional max-min calibration plus grid search and does not explicitly use a sparsity mask or statistic.", + "Several essential hyperparameters and procedures are deferred to unavailable supplementary material: grid-search α, β, and T; optimization iteration count; λ1 and λ2; the regression weight α; optimizer and batch schedule; pseudo-label refresh policy; and details of propagating gradients through the remaining network. The reported results therefore cannot be reproduced from the paper text.", + "All tables label the models as 8/8 even though the first and last layers remain in full precision. The fraction of FP32 operations and parameters is not reported, so the models are mixed precision rather than uniformly INT8. BN folding, unsupported operators, sparse-convolution precision, quantization granularity, and accumulator precision are also unspecified.", + "The deployment claim is insufficiently documented. Latency lacks units and details about batch size, warm-up, runtime, power mode, included preprocessing/postprocessing, and variance. The reported speedups are approximately 2.95x for Pillar and 2.50x for Voxel, so an unqualified 3x claim does not hold for both models.", + "Experiments use one dataset and no repeated calibration draws or error bars. Since the final gaps to FP32 are only 0.07–0.21 mAPH, calibration-subset and run-to-run variability are important. Baseline and QAT reimplementation settings are not described sufficiently to verify comparison fairness.", + "The contribution is largely a combination of established max-min/grid-search calibration, pseudo-label task distillation, and AdaRound-style reconstruction. TGPL is useful but specialized to CenterPoint-like outputs, making the claimed generality broader than the demonstrated formulation." + ], + "questions": [ + "What differentiable parameterization, estimator, constraints, and hardening schedule are used to optimize θ through the hard rounding operation in Equation 13?", + "Are weights and activations quantized per tensor or per channel, and how can learned zero-points be reconciled with the claimed signed symmetric quantizer?", + "What are the values of grid-search α, β, and T, optimization T, λ1, λ2, and the TGPL regression weight, and are pseudo-labels generated once or refreshed during layer-wise optimization?", + "Which operations remain FP32 on the Jetson, and do the latency figures include voxelization, data transfer, decoding, and NMS? Please also provide batch size, units, runtime configuration, warm-up protocol, and variability.", + "How sensitive are the near-FP32 results to the random choice and size of the 256-frame calibration set? Do they persist across multiple seeds, datasets, and detector heads not based on CenterPoint-style heatmaps?" + ], + "limitations": [ + "Evaluation is restricted to Waymo and primarily to CenterPoint variants, with FSD as one additional architecture.", + "Only 8-bit weight/activation results are reported; the paper explicitly leaves lower-bit quantization for future work.", + "The method adds layer-wise optimization cost relative to simpler PTQ calibration, taking 2.12–2.75 GPU-hours in the reported setup.", + "The TGPL construction assumes detector-specific heatmap classification and box-regression outputs, so adaptation to other heads is not demonstrated." + ], + "soundness": 6, + "presentation": 5, + "contribution": 6, + "overall_score": 6, + "confidence": 4, + "ethical_concerns": "No direct ethical violation is apparent. However, deployment in autonomous systems is safety-critical, and the evaluation does not analyze rare classes, adverse conditions, calibration shift, or statistically significant changes in long-range recall. Near-equal aggregate mAPH alone is insufficient evidence of deployment safety.", + "recommendation": "weak_accept", + "review_text": "LiDAR-PTQ presents a practically relevant PTQ pipeline with unusually strong reported accuracy retention. The progression from 52.91 mAPH under max-min calibration to 60.12 after all components is compelling, and similar near-FP32 results on CenterPoint-Voxel and FSD suggest that the approach is not confined to one implementation. The use of a small unlabeled calibration set and on-device latency measurements further improves the practical relevance.\n\nThe central concern is that the method cannot be reconstructed faithfully from the paper text. In particular, adaptive rounding is expressed using a nondifferentiable hard-round operator without specifying the relaxation or gradient estimator required to optimize θ. The signed-symmetric quantizer is also inconsistent with learned zero-points and the stated scale equation. Important training and quantization hyperparameters are absent. These are substantive methodological issues rather than cosmetic omissions.\n\nThe empirical conclusions also need tighter qualification. The “sparsity-based” explanation is plausible but supported mainly by correlation, not an intervention isolating zero-valued features or foreground clipping. The models retain FP32 first and last layers, yet are uniformly labeled 8/8. Likewise, the latency table gives a 2.50x speedup for the voxel model, so the blanket 3x claim is overstated. No repeated calibration subsets or uncertainty estimates are provided despite accuracy differences of only a few tenths of a point. Overall, the empirical gains and practical problem justify a weak-accept assessment, but the quantizer definition, rounding optimization, reproducibility details, and deployment reporting should be clarified." + } +} diff --git a/codex_reviews/reviews/0gVatTOgEv.json b/codex_reviews/reviews/0gVatTOgEv.json new file mode 100644 index 0000000000000000000000000000000000000000..85db6e6dfb38dad737d8d826ffdee6ceb55c7fab --- /dev/null +++ b/codex_reviews/reviews/0gVatTOgEv.json @@ -0,0 +1,66 @@ +{ + "paper_id": "0gVatTOgEv", + "title": "Glider: Global and Local Instruction-Driven Expert Router", + "model": "gpt-5.6-sol", + "reasoning_effort": "xhigh", + "created_at": "2026-08-09T18:01:49.136692+00:00", + "paper_md_chars": 64488, + "metadata": { + "pdf_time_verification_status": "verified_arxiv_version_before_review", + "review_time_revision_source": "arxiv", + "arxiv_version_id": "2410.07172v1", + "markdown_source": "arxiv" + }, + "review": { + "summary": "The paper proposes GLIDER, a post-hoc router over independently trained LoRA experts that combines a Phatgoose-style token/module-level local router with a global router based on LLM-generated task descriptions and embedding similarity. On the T0 expert pool, GLIDER improves T0 held-in aggregate performance from 61.42 for Phatgoose to 68.04, near the reported best-single-expert oracle of 69.60. Held-out gains are much smaller and inconsistent across settings. The central idea is promising, but the evaluation protocol is underspecified and appears to give GLIDER three labeled demonstrations at inference, contradicting zero-shot and arbitrary-query claims and making comparisons with zero-shot routers difficult to interpret. Test-set hyperparameter selection, missing uncertainty, limited component ablations, and several mathematical and numerical inconsistencies further weaken the evidence.", + "strengths": [ + "The paper studies an important practical tension in routing independently trained experts: preserving specialization on held-in tasks while retaining compositional generalization to held-out tasks.", + "The contributor-aggregator formulation is useful and clearly separates local expert training from post-hoc aggregation.", + "Combining query-level semantic guidance with token- and module-level routing is intuitive. The global component can stabilize expert identity, while the local component can still compose experts at finer granularity.", + "The reported T0 held-in improvement is substantial in absolute terms: 68.04 versus 61.42 for Phatgoose, approaching the reported 69.60 best-single-expert oracle.", + "The evaluation covers pools of 36 T0 and 166 FLAN experts and several held-out benchmark collections, which is a meaningful experimental scale.", + "The alpha and top-k studies provide some evidence about the specialization/generalization trade-off, and the routing heatmaps offer a useful qualitative view of selected experts." + ], + "weaknesses": [ + "The inference protocol is the largest concern. The global router requires three input-output examples of similar queries. These are labeled demonstrations and implicitly reveal substantial task information. The paper does not explain where they come from, how similarity is determined without task identity, whether they are disjoint from evaluated examples, or how this mechanism handles a genuinely isolated arbitrary user query.", + "The supplied prompt template contains only the three input-output demonstrations and has no field for the current query u. Thus the query-dependent vector q_u cannot be reproduced as described; the global decision may instead be a task-level decision determined entirely by demonstrations.", + "The paper calls the held-out setting zero-shot in the introduction, but Table 1 explicitly categorizes GLIDER as few-shot. Comparisons against zero-shot Phatgoose, Arrow, and merging therefore do not isolate the benefit of the routing architecture. LoRA Hub is also few-shot, but the paper does not establish matched demonstrations, supervision, or optimization budgets.", + "The threshold and scaling protocol risks test-set tuning. The paper chooses p=0.8, beta=3, and a high global weight using distributions and performance on the reported T0-HI, T0-HO, BBH, and BBL evaluation collections. No separate validation tasks or pre-specified calibration procedure are described.", + "The main causal claim about LLM semantic reasoning is not isolated. Missing baselines include global-only and local-only routing, embedding the raw query or original task instruction without GPT-4, embedding demonstrations directly, an input-only prompt, a non-LLM task summarizer, and sensitivity to the three sampled demonstrations, LLM randomness, embedding model, and number of examples.", + "No per-task results, variance across seeds or demonstration samples, confidence intervals, or significance tests are provided. This is important because most held-out improvements over Phatgoose are small: 0.88 on T0-HO, 0.16 on T0 BBL, and 0.32 on FLAN BBL, while GLIDER is 0.53 worse on FLAN BBH.", + "The evaluation metrics and aggregation procedure are not defined. The benchmark collections contain heterogeneous tasks, but the paper reports only a generic aggregate 'performance' without specifying per-task metrics, normalization, macro/micro averaging, decoding, prompt aggregation, or sample counts.", + "Several internal inconsistencies impede reproducibility. The alpha formula with gamma=100 and beta=3 yields 103 above the threshold, while the text says the high alpha is 100. Top-2 BBH is 35.29 in the main and alpha tables but 35.39 in the routing table. The global-routing caption says 35 T0 experts, whereas the experimental setup says 36.", + "The final mixture equation uses weights w_k without defining whether they are softmax probabilities or whether they are renormalized over selected experts. It also omits the LoRA scale used earlier. The local-router description standardizes the activation but then writes the cosine similarity using the unstandardized symbol.", + "The 'Oracle Expert' is not an upper bound on mixtures: it is the best individual expert for a dataset, and a routed composition can legitimately outperform it. Calling it an upper bound is logically inconsistent with the reported results that exceed it.", + "The claim that GLIDER supports heterogeneous expert-module architectures is neither demonstrated nor clearly compatible with the equations, which assume aligned module locations and compatible matrices so that expert updates can be summed.", + "Scalability is not evaluated. GLIDER must score all experts at every routed module and invoke a proprietary LLM to create a task description at inference. Latency, API cost, memory, throughput, caching assumptions, and behavior as the expert pool grows are missing." + ], + "questions": [ + "Exactly how are the three inference-time input-output demonstrations obtained for each query, and how are they selected as similar without knowing the task identity? Are they fixed per dataset, retrieved automatically, or supplied by the user?", + "Is the current query included in the GPT-4 prompt? If so, where does it enter the provided template? If not, why is the resulting representation denoted q_u and described as per-query routing?", + "Are all demonstrations disjoint from evaluation examples, and can the authors rule out selecting the current example or otherwise leaking evaluation labels?", + "What validation collection was used to select p=0.8, gamma, beta, and top-k? Can the complete selection procedure be reported without using the final evaluation tasks?", + "How does GLIDER compare with global-only routing, Phatgoose plus direct query/instruction embeddings, and a task-description router that does not use GPT-4 or labeled outputs?", + "What is the variability over random selections of the three demonstrations and over repeated LLM generations? Do the small held-out gains remain statistically reliable?", + "What exactly are w_k in the final module equation, and are top-k probabilities renormalized? Is the high-threshold alpha 100, 103, or implemented differently?", + "What metrics, task weights, decoding rules, and prompt-aggregation procedures produce each reported benchmark score?", + "Can results be reported for FLAN held-in tasks and per individual task, rather than only aggregate held-out scores?", + "What are inference latency, external API cost, memory use, and throughput relative to Phatgoose for 36 and 166 experts?" + ], + "limitations": [ + "The evidence is limited to a T5.1.1-XL base model and rank-16 LoRA experts; transfer to decoder-only LLMs, other PEFT modules, heterogeneous architectures, or larger expert pools is untested.", + "The method depends on GPT-4 Turbo and nomic-embed-text-v1.5, so behavior may be model-version dependent and cannot be reconstructed from the mathematical description alone.", + "Only T0 held-in performance is reported; specialization within the 166-expert FLAN pool is not evaluated.", + "The method assumes access to three labeled examples at inference and therefore does not directly cover isolated queries without demonstrations.", + "The supplied text does not provide sufficient implementation detail to verify the code-availability claim or reproduce sampling, generation, and evaluation exactly." + ], + "soundness": 4, + "presentation": 6, + "contribution": 5, + "overall_score": 4, + "confidence": 4, + "ethical_concerns": "The method sends contributor data examples, their outputs, and potentially user-query context to an external proprietary LLM API. This creates possible privacy, confidentiality, data-governance, and consent concerns, especially under the decentralized contributor framing, but the paper does not discuss them. It also omits the cost and environmental implications of an LLM call at inference and the reproducibility risks of relying on a changing hosted model.", + "recommendation": "weak_reject", + "review_text": "GLIDER combines a Phatgoose-style local router with an LLM-assisted global semantic router. The motivation is compelling: token-level decisions can lose task-level context, so a global signal may improve retrieval of a held-in expert without eliminating local composition. The T0 held-in result supports this possibility, with 68.04 versus 61.42 for Phatgoose and 69.60 for the reported best-single-expert oracle. The contributor-aggregator framing and experiments with up to 166 experts are also useful.\n\nHowever, the central experimental comparison is not presently sound enough to support the claims. At inference, GLIDER is given three similar input-output pairs and sends them to GPT-4 Turbo to infer a task description. This is labeled few-shot information, as Table 1 acknowledges, yet the introduction describes the held-out result as zero-shot and compares it prominently with zero-shot routers. More fundamentally, the paper never explains how those examples are found for an arbitrary query, whether task identity is used, or whether the demonstrations are disjoint from test examples. The supplied prompt does not even contain the current query, despite the method defining a query-dependent vector q_u. Under the written procedure, the global router appears to infer a task from demonstrations rather than route based on the current query. This distinction is central because demonstrations can act as an implicit task label.\n\nThe evidence that LLM reasoning itself is responsible for the improvement is also insufficient. There is no global-only/local-only decomposition and no comparison with direct embeddings of the raw query, task instruction, or demonstrations. Input-only versus input-output prompts, random demonstration sensitivity, repeated LLM generations, and alternative embedding or instruction-generation mechanisms are not evaluated. Consequently, the results could reflect access to labeled task examples or a conventional embedding-retrieval signal rather than the claimed semantic reasoning advantage.\n\nThe hyperparameter protocol raises an additional generalization concern. The 0.8 threshold and global scaling are motivated using the reported evaluation-task score distributions, and alpha performance is examined on the same T0-HI, T0-HO, BBH, and BBL collections. No independent validation-task protocol is described. The threshold may therefore encode the held-in/held-out structure of these benchmarks rather than transfer to new tasks or expert pools.\n\nReporting is too incomplete to assess the smaller held-out effects. Relative to Phatgoose, the aggregate gains are only 0.88 on T0-HO, 0.16 on T0 BBL, and 0.32 on FLAN BBL, while GLIDER is 0.53 lower on FLAN BBH. There are no variances, significance tests, per-task tables, sample counts, or definitions of the heterogeneous-task aggregate metrics. The paper also does not report FLAN held-in performance. The qualitative routing examples are encouraging but cannot substitute for these analyses.\n\nFinally, several inconsistencies should be fixed: the stated alpha parameters yield 103 rather than the claimed 100 in the high-similarity branch; Top-2 BBH is reported as both 35.29 and 35.39; the T0 pool is described as both 35 and 36 experts; and the final mixture weights are undefined. The best individual expert is not an upper bound on routed mixtures, as the paper's own results demonstrate. The heterogeneous-architecture claim and practical scalability are likewise unsupported.\n\nOverall, the multi-scale routing idea is promising and the held-in gain is noteworthy, but the inference supervision, baseline fairness, validation protocol, and reproducibility issues prevent a reliable conclusion. A revised study with a precisely defined leakage-free demonstration protocol, matched few-shot baselines, independent task-level validation, essential component baselines, uncertainty estimates, and efficiency measurements would materially strengthen the work." + } +} diff --git a/codex_reviews/reviews/0nxocR2qx4.json b/codex_reviews/reviews/0nxocR2qx4.json new file mode 100644 index 0000000000000000000000000000000000000000..599b8a74102936832b6a3553f9c8252d9572dcaf --- /dev/null +++ b/codex_reviews/reviews/0nxocR2qx4.json @@ -0,0 +1,61 @@ +{ + "paper_id": "0nxocR2qx4", + "title": "ROPO: Robust Preference Optimization for Large Language Models", + "model": "gpt-5.6-sol", + "reasoning_effort": "xhigh", + "created_at": "2026-08-09T18:14:03.498912+00:00", + "paper_md_chars": 89207, + "metadata": { + "pdf_time_verification_status": "verified_arxiv_version_before_review", + "review_time_revision_source": "arxiv", + "arxiv_version_id": "2404.04102v2", + "markdown_source": "arxiv" + }, + "review": { + "summary": "ROPO combines a conservative DPO-derived loss, iterative loss-based sample filtering, and model-guided rejection sampling to improve preference optimization under label noise. The central intuition—that DPO assigns large gradients to strongly contradictory examples and can therefore amplify mislabeled preferences—is useful, and the reported gains are large across three datasets and two 7B model families. However, the principal theoretical guarantees rely on unstated realizability assumptions, do not apply to the actual combined ROPO loss, and conflate stochastic but valid Bradley–Terry outcomes with corrupted labels. Experimental conclusions are also weakened by single-run GPT-4 evaluation, incomplete implementation details, and evaluation inconsistencies.", + "strengths": [ + "The paper identifies a concrete failure mode of DPO: its gradient weight approaches its maximum when the observed preference strongly contradicts the model's implicit reward margin, which can amplify corrupted labels.", + "ROPO presents a coherent practical pipeline combining conservative gradients, loss-based filtering, and data regeneration without requiring a separate teacher or reward model.", + "The empirical study covers three datasets, two 7B model families, several noise levels, four relevant preference-optimization baselines, component ablations, loss distributions, and performance across training epochs.", + "The reported robustness improvements are substantial, especially on TL;DR; for example, Llama-2-7B ROPO remains near 79% across all three artificial-noise levels while DPO declines from 56.8% to 35.2%.", + "The paper explicitly acknowledges the risk of using a fixed filtering ratio when the true noise rate is unknown." + ], + "weaknesses": [ + "The population theorems implicitly optimize each pairwise probability independently. A shared finite-capacity LLM need not realize the pointwise Bayes optimum for every query-response pair. No realizability or model-class assumption justifies moving from minimization over θ to unconstrained minimization over each probability.", + "Theorem 3 claims that the optimal ROPO regularizer produces probabilities exactly equal to 0 or 1. This conflicts with the stated compact parameter space and an ordinary finite-logit softmax model, for which the induced sigmoid preference probability is strictly between 0 and 1. Thus the result does not hold for the stated model without substantial additional assumptions or a limiting argument.", + "The robustness result is proved only for ℓ_reg, not for the actual loss ℓ_ropo, which contains a positive DPO cross-entropy component. The minimizer of this combined loss will generally depend on the corrupted-label probability, so the paper has not established that its deployed loss is noise-tolerant.", + "The distinguishability analysis assumes that every clean observed label c satisfies P*(c)>1/2. This does not follow from the paper's own Bradley–Terry model: a valid label may be a minority stochastic outcome. Such an unflipped but valid observation can receive high loss and be filtered. The analysis therefore distinguishes Bayes-majority labels from their opposites rather than clean samples from corrupted samples in general.", + "The convergence argument only shows that exact alternating minimization produces a bounded, non-increasing sequence of objective values. It does not establish convergence of parameters or weights, and it does not cover approximate SGD, changing reference models, rejection sampling, or the changing dataset. The claim of a provably convergent training process is too strong.", + "The proof of Theorem 1 contains an invalid exchange construction: setting w'_j=1 and w'_l=w_j+w_l−1 can make w'_l negative, and the proof also states that the new weights sum to 1 rather than N_rho. The theorem itself is repairable through a valid mass-transfer argument, but the supplied proof is incorrect.", + "The assertion that symmetric and asymmetric noise are equivalent after swapping response positions is unsupported. Swapping the encoding does not eliminate unequal class-conditional flip rates or prompt-, response-, and annotator-dependent corruption. Experiments cover only artificial independent symmetric flips, despite broader motivation involving ambiguity and real annotation errors.", + "Evaluation relies on one GPT-4 referee with no repeated seeds, confidence intervals, significance tests, human validation, judge-version details, decoding details, or position-bias controls. Several reported differences are small enough that uncertainty is material.", + "Table 1 is labeled as comparison against SFT targets, but dialogue models are described as being compared with text-davinci-003. Moreover, the 0%-noise Mistral UFB DPO number is imported from a prior Zephyr result rather than produced in the same experimental protocol.", + "Important reproduction details are missing, including the precise alternating algorithm, reference-model update schedule, whether training continues or restarts after filtering, validation splits used for hyperparameter selection, and the generation cost of rejection sampling. The claimed computational efficiency is not supported by runtime, memory, or generation-token measurements.", + "Rejection sampling assumes without verification that both original responses are preferable to every newly generated response and selects the resulting pair using the same model's loss. This can create incorrect labels and confirmation bias. The ablation also contradicts the statement that all components improve performance: adding rejection sampling lowers the 40%-noise UFB result from 88.07 to 87.70.", + "The conclusion that gains on nominally clean data demonstrate substantial inherent label noise is not identified by the experiment. Improvements could instead arise from regularization, example selection, optimization differences, or evaluator preferences." + ], + "questions": [ + "Can the authors state explicit realizability assumptions and redo Theorems 2–4 for the constrained shared LLM model class, including the fact that finite-logit preference probabilities cannot attain 0 or 1?", + "Is the full combined loss ℓ_ropo noise-tolerant under symmetric corruption? If not, can its deviation from the clean optimum be bounded as a function of α and η?", + "How are stochastic Bradley–Terry minority outcomes treated? Can filtering precision and recall be evaluated under a noise definition that distinguishes genuine corruption from valid preference variability?", + "What is the exact training algorithm, including initialization, reference-model updates, filtering permanence, rejection-sampling timing, and whether optimizer state is preserved between iterations?", + "Were results averaged over training seeds and multiple judge orderings? What are the confidence intervals, judge version and settings, and outcomes under human or an independent evaluator?", + "How much additional generation time and compute does rejection sampling require, and how often are its assumed preference labels actually correct?", + "Which validation data and metric were used to choose learning rate, β, α, ρ, and noise-specific smoothing parameters without tuning on the reported evaluation sets?" + ], + "limitations": [ + "The evidence is limited to synthetic independent symmetric label flips and does not test class-conditional, instance-dependent, ambiguous, or naturally occurring annotation noise.", + "Only two 7B model families and three datasets are evaluated, all primarily through a single model-based judge.", + "ROPO requires a fixed filtering ratio and may discard difficult, rare, or minority-preference examples rather than genuinely corrupted examples.", + "No computational-cost comparison, safety evaluation, or analysis of demographic and viewpoint selection effects is provided." + ], + "soundness": 4, + "presentation": 6, + "contribution": 6, + "overall_score": 4, + "confidence": 4, + "ethical_concerns": "No direct ethical violation is apparent, but loss-based filtering can systematically remove minority or difficult preferences, and the self-generated rejection-sampling labels may reinforce the model's existing biases. The paper itself motivates the method with occupational discrimination, yet it provides no fairness or safety evaluation and treats these risks only briefly.", + "recommendation": "weak_reject", + "review_text": "ROPO is a promising attempt to make preference optimization robust to corrupted pairwise labels. Its main technical idea is to supplement DPO with a bounded symmetric regularizer, producing a gradient weight that decreases after the model-label disagreement becomes sufficiently large. It then removes high-loss samples and regenerates pairs for discarded prompts. The gradient-based motivation is intuitive, and the empirical gains—particularly on TL;DR—are large enough to merit attention.\n\nThe main obstacle is that the theory does not establish the claimed guarantees for the implemented method. Theorems 2–4 treat each pairwise probability as an independently selectable scalar, although the actual probabilities are coupled through one finite-capacity LLM. Theorem 3 additionally requires exact probabilities of 0 or 1, which are not attainable by an ordinary finite-logit model under the stated compact parameter space. Most importantly, noise tolerance is shown only for ℓ_reg, whereas ROPO optimizes a positive combination of ℓ_reg and the explicitly noise-sensitive DPO loss. Dominance of the regularizer is not equivalent to invariance of the combined optimum.\n\nThe clean/noisy distinction also needs reconsideration. Under the paper's Bradley–Terry formulation, an uncorrupted label is a stochastic realization and need not be the more probable label. The proofs instead assume that a clean label always has probability above one half. Consequently, high-loss filtering may remove valid minority preferences, and the theory does not prove clean/noisy separation as defined in the problem setup. The symmetric-noise experiments do not resolve this issue, nor do they support the broader assertion that asymmetric noise is equivalent after response swapping.\n\nThe experiments are broad but lack uncertainty estimates and controlled evaluation details. Dialogue results are described as comparisons with text-davinci-003 while Table 1 calls the comparator an SFT target, and one central DPO number is imported from an earlier Zephyr evaluation. No multiple seeds, judge-order controls, human evaluation, or confidence intervals are reported. The rejection-sampling labels are assumed rather than verified, its cost is unreported, and its ablation slightly hurts the 40%-noise result despite the claim that every component helps.\n\nOverall, the conservative-gradient idea and empirical pattern are interesting, but correcting the theoretical scope, fully specifying the algorithm, and providing controlled and statistically supported evaluation are necessary before the central robustness claims are convincing." + } +} diff --git a/codex_reviews/reviews/0owAtTCOlU.json b/codex_reviews/reviews/0owAtTCOlU.json new file mode 100644 index 0000000000000000000000000000000000000000..466b2b868955312c7aabec59eee771c0b5ad6f53 --- /dev/null +++ b/codex_reviews/reviews/0owAtTCOlU.json @@ -0,0 +1,61 @@ +{ + "paper_id": "0owAtTCOlU", + "title": "GRIC: General Representation and Informative Content for Enhanced Out-of-Distribution Detection", + "model": "gpt-5.6-sol", + "reasoning_effort": "xhigh", + "created_at": "2026-08-09T18:09:48.857073+00:00", + "paper_md_chars": 39311, + "metadata": { + "pdf_time_verification_status": "verified_revision_cdate_before_review", + "review_time_revision_source": "submission_note", + "arxiv_version_id": null, + "markdown_source": "dataset" + }, + "review": { + "summary": "GRIC combines two post-hoc modifications to CLIP-based OOD detection: it estimates PCA on a small ID reference set, ranks embedding coordinates using PCA loadings, masks selected coordinates at inference, and augments class prompts with LLM-generated superclasses. OOD confidence is the maximum softmax-normalized cosine similarity to the augmented class prompts. The manuscript reports substantial FPR95 improvements, especially on ImageNet-1k and MS-COCO, but the central interpretation of PCA coordinates as class-specific features is not established, several essential implementation details are missing or mathematically inconsistent, and the claimed zero-shot/label-free characterization conflicts with stratified ID sampling.", + "strengths": [ + "The proposed detector is conceptually simple, post-hoc, and does not require optimizing CLIP parameters.", + "The evaluation reportedly covers several ID datasets, four common OOD test sets, and two CLIP model sizes.", + "The ablations attempt to isolate image masking and informative prompts, include reverse masking of non-PCA features, and examine reduced ID reference-set sizes.", + "The paper reports both threshold-specific FPR95 and threshold-independent AUROC rather than relying on a single detection metric.", + "The manuscript eventually acknowledges that PCA requires ID examples and therefore calls the method few-shot, although this conflicts with claims elsewhere." + ], + "weaknesses": [ + "The key technical premise is unsupported. PCA on pooled, unlabeled ID embeddings identifies directions of total variance, not class-specific variation. Separating class-specific from shared structure would normally require, at minimum, a between-class versus within-class analysis. High variance can also arise from pose, background, illumination, or genuinely shared ID properties.", + "The proposed feature statistic is not clearly a variance statistic. Equation 1 appears to average absolute PCA loadings across components without weighting them by explained variance. It also divides by r even though the number of fitted components need not equal r. Thus the connection between the computed quantity and the claimed importance or specificity of an embedding coordinate is unclear.", + "The coordinate-masking operation is basis-dependent: an arbitrary orthogonal rotation of an otherwise equivalent embedding space could change the selected coordinates and the detector. The paper neither analyzes this issue nor compares coordinate masking against removing a PCA subspace through an explicit projection.", + "The algorithm is not reproducibly specified. The definition of k mixes a rank in the sorted list with an original feature index in SI; the stated variance-normalized scaling is not defined; epsilon values are given without a selection protocol; and it is unclear exactly how many components are retained or how embeddings are centered and normalized before PCA.", + "The scoring implementation is ambiguous. Equation 6 omits a temperature or CLIP logit scale, while modifying only image coordinates may alter cross-modal geometry. The paper does not clearly say whether text embeddings receive the same mask/scaling, whether embeddings are renormalized, or whether masked or original features are used for reported ID classification.", + "The paper alternately calls GRIC zero-shot, training-free, label-free, and few-shot. It uses 12, 16, or 25 ID samples per class, and selecting a fixed number per class appears to require class membership information even if those labels are not used by PCA. This data and supervision requirement should be stated consistently and compared fairly with baselines.", + "The LLM component is insufficiently documented. The exact instruction, generated superclass lists, number of hierarchy levels, decoding configuration, model snapshot, handling of ambiguous labels, and variation across generations are absent. The claim that the method is cost-free is also inaccurate when it depends on LLM inference.", + "The abstract claims that LLMs simulate potential OOD scenarios, but the described method only generates ID superclass taxonomies. No simulated OOD samples or labels appear in the algorithm, so this claimed mechanism is not supported by the method presented.", + "Important empirical controls are missing: random coordinate masks of matched size, masks based on coordinate variance without PCA, low- versus high-variance PCA subspace removal, sensitivity to k and epsilon, robustness to the sampled ID subset, alternative non-LLM taxonomies, and tests on another vision-language model family.", + "The evaluation protocol is incomplete. Exact ImageNet-10/20/100 class lists and splits are not provided; Stanford-Cars appears in the results discussion but not the dataset list; the text says six ID datasets while discussing eight; and the manuscript does not establish whether hyperparameters or prompts were selected without inspecting OOD results.", + "Although ten runs are mentioned, the supplied text provides no means, standard deviations, confidence intervals, or significance analysis. ID classification accuracy is listed as a metric and later mentioned as an experiment, but corresponding results are absent from the supplied manuscript. Consequently, the assertion that OOD gains are achieved while maintaining ID accuracy cannot be verified." + ], + "questions": [ + "What is the exact mathematical definition of the feature score, including PCA centering, normalization, number of components, eigenvalue weighting, and the corrected rule for selecting k?", + "What evidence shows that the selected coordinates encode class-specific rather than shared or nuisance variation? Can the authors report within-class/between-class covariance analyses and comparisons with random or variance-matched masks?", + "Why mask embedding coordinates instead of projecting away the estimated PCA subspace, and how stable is the result under rotations or different random ID reference sets?", + "Are the same mask and variance scaling applied to text embeddings? Are both modalities renormalized, and is a temperature or CLIP logit scale used in Equation 6?", + "How are the 12, 16, or 25 examples per class selected without labels? What are the total ID sample counts, and is any OOD validation data used to choose epsilon, k, prompts, or other settings?", + "Can the authors release the exact LLM instruction, raw superclass outputs for every dataset, model version and decoding settings, and results across multiple generations or manually defined taxonomies?", + "Can the paper provide complete mean-plus-uncertainty results for all runs, exact dataset class lists and splits, ID classification accuracy before and after masking, and computational cost for PCA and LLM prompting?", + "What operation in GRIC corresponds to the abstract's claim that the LLM simulates potential OOD scenarios?" + ], + "limitations": [ + "This review is restricted to the supplied manuscript. Several tables and Algorithm 1 are image-only in the provided markdown, so not every numerical entry or pseudocode line can be independently checked.", + "The supplied paper references Section B and additional experiments that are not included, limiting verification of sensitivity analyses and implementation details.", + "As presented, GRIC depends on a fixed ID reference set, dataset-specific thresholds, and externally generated taxonomies; behavior under domain shift within an ID class or changing label sets is not evaluated.", + "Evidence is concentrated on CLIP variants, so the claim that the method adapts to other contrastive vision-language models remains unverified." + ], + "soundness": 3, + "presentation": 3, + "contribution": 4, + "overall_score": 3, + "confidence": 4, + "ethical_concerns": "No major human-subject, privacy, or dual-use concern is evident from the supplied text. However, the real-world autonomous-driving anecdote is not substantiated within the paper and should not be used to imply safety benefits without deployment-oriented evidence. The provenance, reproducibility, and cost of the GPT-generated taxonomy should also be disclosed.", + "recommendation": "reject", + "review_text": "The paper proposes a potentially useful and inexpensive post-hoc recipe, and the reported FPR95 gains are large enough to warrant investigation. Nevertheless, the current manuscript does not establish the mechanism on which GRIC is based. Pooled PCA estimates total variance, whereas the paper repeatedly interprets selected embedding coordinates as class-specific and the remaining coordinates as general. The stated mean-variance statistic instead appears to average absolute loadings, without explained-variance weighting, and coordinate masking is inherently basis-dependent. These issues make the central interpretation technically unconvincing even if the empirical heuristic ultimately works.\n\nThe empirical claims are also difficult to verify. Essential details of feature scoring, selection of k, variance scaling, cross-modal masking, score temperature, LLM prompting, dataset construction, and hyperparameter selection are missing or inconsistent. The zero-shot and label-free framing conflicts with collecting a fixed number of ID examples per class, while the manuscript itself later calls the method few-shot. Ten runs are mentioned without uncertainty estimates, and the promised ID-accuracy evidence is absent from the supplied text. The abstract's claim that LLMs simulate OOD scenarios is not reflected in the described algorithm. I would require a corrected and complete specification, substantially stronger controls, exact reproducibility artifacts, and evidence that gains are not simply caused by dataset-specific feature deletion or prompt choices before the reported performance could support the paper's broad conclusions." + } +} diff --git a/codex_reviews/reviews/0sU4myabw1.json b/codex_reviews/reviews/0sU4myabw1.json new file mode 100644 index 0000000000000000000000000000000000000000..8e3dd21064b0cc60cd001de9fbbee54f8ee6cb25 --- /dev/null +++ b/codex_reviews/reviews/0sU4myabw1.json @@ -0,0 +1,65 @@ +{ + "paper_id": "0sU4myabw1", + "title": "RapidDock: Unlocking Proteome-scale Molecular Docking", + "model": "gpt-5.6-sol", + "reasoning_effort": "xhigh", + "created_at": "2026-08-09T18:20:08.512691+00:00", + "paper_md_chars": 48999, + "metadata": { + "pdf_time_verification_status": "verified_arxiv_version_before_review", + "review_time_revision_source": "arxiv", + "arxiv_version_id": "2411.00004v1", + "markdown_source": "arxiv" + }, + "review": { + "summary": "RapidDock is a 60M-parameter transformer for blind protein–ligand docking. It combines atom/residue tokens, ESM-2 features, charge embeddings, and learned pairwise-distance attention biases; predicts ligand–ligand, protein–protein, and ligand–protein distances; and reconstructs ligand coordinates with L-BFGS. The reported top-1 success rates are 52.1% on Posebusters and 44.0% on DockGen, with average runtimes of 0.03–0.05 seconds. The speed–accuracy tradeoff is promising, but important ambiguities in the loss and evaluation protocol prevent the central practical claims from being established convincingly.", + "strengths": [ + "The paper addresses an important bottleneck: reducing blind-docking latency enough to support much larger screens.", + "The architecture is relatively simple and computationally attractive. Pairwise distance biases provide a natural way to expose 3D information to a standard transformer without coordinate-frame dependence.", + "The reported accuracy is competitive with the included fast baselines, while the measured runtime advantage over DiffDock-L and NeuralPLexer is substantial.", + "The paper includes useful ablations for ESM-2 embeddings, folding pretraining, and attention scaling. The respective Posebusters decreases from 52.1% to 42.6%, 46.4%, and 48.6% provide evidence that all three components contribute.", + "Several implementation details are reported, including model dimensions, training schedules, distance discretization, and L-BFGS parameters. Reconstruction time is stated to be included in the reported inference time.", + "The authors recognize relevant limitations, including the absence of confidence estimates and the inability to predict binding strength or distinguish binders from non-binders." + ], + "weaknesses": [ + "The apostructure evaluation is not end-to-end. Although apostructures provide the model's input distance embeddings, the paper explicitly uses holostructure coordinates for ligand reconstruction. The reconstruction objective and initialization directly depend on protein coordinates, so bound-state pocket geometry still enters the final predicted pose. Consequently, the reported 51.7% and 42.3% results do not establish practical docking from an apostructure alone.", + "The ligand loss is internally inconsistent as written. Earlier, D^l denotes the rigid input matrix, with flexible distances set to -1. The loss later compares nonnegative Euclidean predictions against D^l “as defined above.” Taken literally, every flexible pair is trained toward an impossible distance of -1 and is retained by the less-than-20-angstrom mask. Presumably the intended target is the complete crystallographic ligand distance matrix, but this is not specified.", + "Evaluation-set construction is insufficiently documented. The method removes ligands above 128 heavy atoms and entries for which CCD matching, conformer generation, or force-field processing fails, but the paper gives neither retained example counts nor evidence that every baseline is evaluated on the exact same retained complexes. This prevents verification of the accuracy comparisons and may induce selection bias.", + "The proteome-wide experiment validates throughput only. It reports runtime for 12 ligands against 19,222 proteins but provides no biological validation, enrichment analysis, known-target recovery, or negative controls. Because the model neither predicts binding nor ranks calibrated interaction likelihoods, this experiment does not demonstrate useful proteome-scale screening.", + "Only symmetry-corrected ligand RMSD is reported. There are no PoseBusters-style physical-validity metrics, clash statistics, bond/angle errors, stereochemistry checks, or strain measurements. This is especially important because coordinates are recovered by unconstrained optimization over predicted distances and the input representation does not clearly encode bonds or chirality.", + "Runtime accounting remains incomplete. It is unclear whether ESM-2 feature computation and generation of 96 ligand conformers are included, what batching was used, and how memory or latency scales with sequence length. The AlphaFold-3 comparison is additionally based on multiplying a reported 16-GPU runtime by 16, which is not a direct single-GPU measurement and assumes linear equivalence.", + "No uncertainty estimates, repeated runs, confidence intervals, or seed sensitivity are reported. Differences such as the 0.4-point apostructure change therefore cannot be interpreted statistically.", + "The summed ligand, protein, and docking losses are not normalized by their number of active pairs. Their relative weighting consequently varies with protein and ligand size, potentially allowing the O(P^2) protein term to dominate. This design choice is neither justified nor ablated.", + "The full-attention architecture has quadratic dependence on the combined protein and ligand length. A maximum training sequence length of 2,760 is reported, but the handling of longer proteins in the claimed whole-proteome run is not explained.", + "Reproducibility is limited by the absence of a stated RapidDock code, checkpoint, processed split, or accession list. The assertion that the test benchmarks were not in training cannot be independently checked from the supplied paper text.", + "The terminology is imprecise: the paper calls the distance representation “equivariant, i.e. invariant,” although invariance and equivariance are different properties. The symbol H is also used both for the number of attention heads and for a 16-dimensional output space, despite the model having four heads." + ], + "questions": [ + "What are the Posebusters and DockGen results when apostructure coordinates—not holostructure coordinates—are used throughout both prediction and L-BFGS reconstruction?", + "Does the ligand-loss target use the complete crystallographic ligand distance matrix or the masked rigid input D^l? Please correct the notation and explain how flexible pairs are supervised.", + "How many examples remain in each benchmark after every filtering step, and were RapidDock, DiffDock-L, and NeuralPLexer evaluated on precisely the same complexes and atom mappings?", + "How are molecular automorphisms enumerated for the symmetry-invariant loss, and what is the worst-case computational cost of considering “all possible permutations”?", + "Do the reported runtimes include ESM-2 inference, ligand conformer generation, preprocessing, data transfer, and reconstruction? What batch size and precision were used?", + "How does RapidDock perform on physical-validity tests, including steric clashes, bond geometry, chirality, and ligand strain?", + "How are proteins longer than the 2,760-token training limit handled in the 19,222-protein experiment, and what are runtime and memory as functions of protein length?", + "What evidence supports the usefulness of the proteome-wide outputs—for example, recovery or ranking of established targets relative to decoy proteins?", + "Were training/test overlaps checked at the complex, ligand, protein-sequence, or structural-domain level, and can exact split identifiers be provided?" + ], + "limitations": [ + "The model currently provides no confidence score.", + "It does not predict binding affinity or distinguish binders from non-binders, limiting its immediate use for screening.", + "It requires a protein 3D structure as input and does not yet dock directly from sequence.", + "The evaluation covers two pose-prediction benchmarks but does not establish performance on physical validity, target ranking, or prospective screening.", + "The practical apostructure setting is not evaluated end-to-end because reconstruction uses holostructure coordinates.", + "Quadratic full attention and an unexplained maximum sequence length may constrain applicability to long proteins." + ], + "soundness": 4, + "presentation": 6, + "contribution": 6, + "overall_score": 4, + "confidence": 4, + "ethical_concerns": "No direct ethical concern is evident. However, claims about identifying toxic interactions or enabling drug-safety screening should be qualified carefully because the presented model predicts poses for supplied pairs but does not determine binding, affinity, or calibrated confidence.", + "recommendation": "weak_reject", + "review_text": "This paper presents a potentially valuable engineering direction: replacing iterative generative docking with a distance-aware transformer followed by short coordinate reconstruction. The reported combination of 44.0–52.1% RMSD success and 0.03–0.05-second inference is promising, and the ablations provide some support for the proposed representation and pretraining choices.\n\nMy primary concern is that the paper's strongest practical validation is not end-to-end. In the apostructure experiment, the model consumes apostructure-derived distance embeddings but the final ligand coordinates are reconstructed using holostructure protein coordinates. Since both the L-BFGS objective and initialization use those coordinates, the final RMSD still benefits from bound-state geometry. The small reported degradation therefore cannot substantiate the claim that RapidDock is nearly unaffected when only computationally folded structures are available. A valid test must use the apostructure coordinates through reconstruction as well.\n\nA second major issue is the ambiguous ligand objective. D^l is defined as a partially observed rigid-distance matrix whose flexible entries equal -1, yet the stated ligand loss uses this same D^l as its target. Literal implementation would train nonnegative predicted distances toward -1 for flexible pairs. The method is not reproducible until the authors distinguish the input mask from the complete crystallographic target and specify the exact symmetry-permutation procedure.\n\nThe empirical comparison also lacks essential accounting. Because the preprocessing filters failed graph matches, failed conformer generation, force-field failures, and ligands above 128 atoms, the paper must report retained benchmark sizes and demonstrate that every method is scored on an identical subset. It should also report statistical variability and clarify whether ESM-2 computation and 96-conformer preparation are included in latency. The extrapolated AlphaFold-3 runtime should not be treated as equivalent to a direct single-GPU benchmark.\n\nFinally, RMSD alone is insufficient for an unconstrained distance-reconstruction method. Physical validity, clashes, stereochemistry, and ligand strain should be evaluated. The proteome experiment demonstrates computational throughput but not screening utility: without affinity, non-binding examples, confidence, or target-ranking validation, docking every pair does not establish that meaningful interactions can be identified.\n\nOverall, the architecture and reported speed are interesting, but the current evidence does not yet validate the central apostructure and proteome-scale claims. Corrected end-to-end evaluation, exact common-subset reporting, clarification of the loss, and physical-validity results could materially strengthen the work." + } +} diff --git a/codex_reviews/reviews/0tAn34IkXI.json b/codex_reviews/reviews/0tAn34IkXI.json new file mode 100644 index 0000000000000000000000000000000000000000..faf31d137012a74d91b5cbd030f864edf92d7ccb --- /dev/null +++ b/codex_reviews/reviews/0tAn34IkXI.json @@ -0,0 +1,60 @@ +{ + "paper_id": "0tAn34IkXI", + "title": "Flat Posterior Does Matter For Bayesian Model Averaging", + "model": "gpt-5.6-sol", + "reasoning_effort": "xhigh", + "created_at": "2026-08-09T18:21:08.502536+00:00", + "paper_md_chars": 184764, + "metadata": { + "pdf_time_verification_status": "verified_arxiv_version_before_review", + "review_time_revision_source": "arxiv", + "arxiv_version_id": "2406.15664v3", + "markdown_source": "arxiv" + }, + "review": { + "summary": "The paper argues that approximate Bayesian neural networks may sample sharp solutions, proposes SA-BMA—a KL/Fisher-constrained sharpness-aware optimizer over posterior parameters—and combines it with subnetwork Bayesian transfer learning. It reports broad image-classification experiments spanning VI, SGLD-based MCMC, SWAG, few-shot transfer, corruptions, and Hessian-based flatness analyses. The empirical results are promising, but the central theoretical claims contain fundamental gaps, and several experimental choices prevent attributing the gains specifically to the proposed posterior-space geometry.", + "strengths": [ + "The paper studies an important question: how posterior geometry interacts with the quality of Bayesian model averaging.", + "The experimental scope is broad, covering multiple approximate-inference frameworks, CNN and ViT backbones, scratch training, few-shot transfer, calibration, NLL, and corruption robustness.", + "SA-BMA is evaluated against several deterministic and Bayesian sharpness-aware baselines, and the paper reports repeated runs and extensive hyperparameter tables.", + "The reported gains can be substantial, such as SA-BMA (SWAG) improving CIFAR100 RN18 accuracy from 59.48 to 63.63 and SA-BMA (VI) improving RN18 CIFAR10 10-shot accuracy from 57.29 for MOPED to 64.98.", + "The appendices attempt to document the algorithm, derivations, flatness measurements, and experimental details rather than leaving the method at a purely conceptual level." + ], + "weaknesses": [ + "Theorem 1 is not established. Weyl's inequality bounds the eigenvalues of the matrix average (1/M)ΣH(w_m), but the theorem interprets this as the Hessian H(w_WA) at w_WA=(1/M)Σw_m. In general H((1/M)Σw_m) is not equal to (1/M)ΣH(w_m). Moreover, closeness of the BMA and weight-averaged losses up to O(Δ) does not imply closeness of their Hessians; function-value proximity alone gives no derivative bound. Consequently, the claimed BMA flatness bound and its interpretation are unsupported.", + "The local KL expansion is incomplete. The standard second-order expansion contains a factor of 1/2 and higher-order remainder; the derivation retains a product of first-order terms while omitting other second-order terms. More seriously, Algorithm 1 replaces F^{-1} with ggᵀ/||g||⁴ from one score vector g. This is rank one. Unless perturbations are explicitly restricted to span(g), the Fisher trust-region problem is unbounded in its nullspace whenever the loss gradient has a nullspace component. The closed-form perturbation assumes an invertible positive-definite Fisher and therefore is not justified for the implemented approximation.", + "The objective and algorithm do not match. Equations defining SA-BMA contain β KL(q_θ||p), so their gradient should include β∇KL, but the stated update and Algorithm 1 use only the perturbed empirical-loss gradient. The omission is especially consequential for the VI and transfer-learning interpretations.", + "The claimed reductions to FSAM and natural gradient are not valid as written. The normalizing evidence called Z may be independent of w but generally depends on θ, so its θ-gradient cannot simply be discarded. A deterministic posterior is also a delta distribution for which the stated Fisher argument is not directly valid. The natural-gradient result uses η_SA-BMA proportional to the matrix F^{-1}; that is a matrix preconditioner, not an ordinary learning rate, and effectively inserts the desired natural gradient by definition.", + "The method is not consistently defined across inference frameworks. The derivation assumes an explicit variational distribution q_θ, but for SGLD the paper sets β=0 and adds noise without defining the corresponding θ-distribution or Fisher geometry. For SWAG, it is unclear whether θ denotes current weights, trajectory statistics, or the final Gaussian parameters. Thus the same theoretical objective cannot be verified for the MCMC and SWAG variants from the paper text.", + "The empirical evidence does not establish that flatness is necessary for BMA. Sorting 30 posterior samples by Hessian eigenvalue and cumulatively ensembling them is a post-hoc selection procedure, not standard Monte Carlo BMA. Because sharpness is already correlated with individual-model error, adding progressively worse individual models can explain the observed curves without identifying a distinct BMA mechanism. A matched analysis of ensemble gain, predictive diversity, and individual accuracy is missing.", + "The flatness metric is an average of per-sample weight-space Hessian eigenvalues, which is coordinate- and scale-dependent and is not itself the curvature of the posterior predictive. This is particularly problematic when comparing models with normalization layers and different stochastic parameterizations. The paper does not establish a mathematical connection between this evaluation metric and the KL/Fisher neighborhood optimized by SA-BMA.", + "Several major confounders lack ablations: nearly eliminating the KL term for SA-BMA VI with β=1e-8 while VI/MOPED baselines search β from 1e-2 to 1; training only normalization and classifier parameters; last-layer reinitialization; choice of prior construction; early stopping; and potentially different posterior sample counts. Matched controls are needed to identify which component causes the gains.", + "The reproducibility description is internally inconsistent. For example, the scratch-training text gives the SGD learning-rate search as [1e-4, 1e-3, 1e-2], while the tables report selected values of 5e-2 and 1e-1. The stated SA-BMA (VI) γ grid excludes the reported 0.2 setting for RN18/CIFAR100, and the stated SA-BMA (SWAG) grid likewise excludes the reported 0.2 setting.", + "Some claims overstate the tables. SA-BMA (SWAG) worsens RN18/CIFAR100 ECE from 0.030 to 0.052, SA-BMA (VI) worsens RN18/CIFAR100 10-shot ECE relative to MOPED from 0.047 to 0.071, and SA-BMA is not best in RN18/CIFAR100C NLL at severities 4 and 5. Thus improvement is not consistent across all reported uncertainty or robustness metrics." + ], + "questions": [ + "Can the authors reformulate Theorem 1 with explicit assumptions that relate H(w_WA) to the average of H(w_m), and provide a derivative-level result connecting weight averaging to BMA rather than only an O(Δ) loss-value statement?", + "How is the rank-one Fisher approximation used without making the trust-region maximization unbounded in its nullspace? Is the perturbation explicitly restricted to the score-vector span in the implementation?", + "What exactly are θ and q_θ for the SGLD and SWAG versions, and how is the posterior-parameter score used to compute their Fisher matrices?", + "Why is the KL gradient absent from the update rule and Algorithm 1, and why is β fixed to 1e-8 for SA-BMA VI while substantially larger values are used for the VI/MOPED baselines?", + "What validation split and selection criterion were used for grid search and early stopping in the few-shot setting, what does each ± value represent, and was the same number of BMA samples used for every method?", + "Can the authors provide matched ablations for posterior-space versus output-space perturbations, β, last-layer reinitialization, trainable subnetwork, prior construction, and identical computational budgets?", + "How do predictive diversity and ensemble improvement over the mean individual-model performance vary with sharpness after controlling for individual accuracy?" + ], + "limitations": [ + "The experimental evaluation is restricted to vision classification and predominantly to normalization/classifier subnetworks; effectiveness for full-network Bayesian inference or other modalities is not demonstrated.", + "The transfer-learning scheme assumes access to suitable pretrained weights, as the paper itself acknowledges.", + "Only three repetitions are reported for most experiments, and the text does not clearly identify whether error bars are standard deviations, standard errors, or confidence intervals.", + "The efficiency analysis counts trainable parameters but does not report wall-clock time, peak memory, or the cost of the additional gradient/Fisher computations.", + "The paper text does not provide enough information to verify the implementation details behind the figures, validation protocol, or posterior sampling fairness." + ], + "soundness": 3, + "presentation": 5, + "contribution": 5, + "overall_score": 4, + "confidence": 4, + "ethical_concerns": "No substantive ethical concerns are apparent from the paper text.", + "recommendation": "reject", + "review_text": "This paper presents an interesting empirical hypothesis and an ambitious set of experiments, but its principal theoretical justification is unsound. The most serious issue is Theorem 1: the proof bounds the eigenvalues of an average of Hessian matrices and then treats that matrix as the Hessian at the averaged weights. These quantities are generally different. The subsequent BMA connection relies only on closeness of loss values, which cannot establish closeness of Hessians. The proposed Fisher trust region also becomes problematic in the actual algorithm, where the inverse Fisher is replaced by the pseudoinverse of a single rank-one score outer product. The resulting constraint has a large nullspace, making the stated maximization unbounded unless an unstated restriction is imposed. The FSAM and natural-gradient reductions contain additional invalid steps, including dropping a θ-dependent normalizer and calling a matrix-valued preconditioner a learning rate.\n\nThe empirical results are broad and sometimes strong, so the paper contains potentially useful evidence for a sharpness-aware Bayesian training heuristic. However, the experiments do not isolate that mechanism. The progressive flatness-ordered ensembles conflate individual-model quality with ensemble behavior, while SA-BMA changes β, trainable parameters, layer initialization, prior construction, and early-stopping behavior. The near-zero β used for SA-BMA VI is a particularly important mismatch with the VI baselines. Hyperparameter descriptions also conflict with reported selected values. Finally, some consistency claims are contradicted by the ECE and corruption-NLL tables.\n\nA substantially revised paper would need to correct or narrow the theoretical claims, define the method rigorously for each inference framework, align the algorithm with its objective, and add matched ablations and a clearly separated validation protocol. In its current form, the promising empirical observations are insufficient to support the claimed general theory of flat posteriors and BMA." + } +} diff --git a/codex_reviews/reviews/0tMcsHsHgQ.json b/codex_reviews/reviews/0tMcsHsHgQ.json new file mode 100644 index 0000000000000000000000000000000000000000..983071436c9bb70131d7260775857ad721df963f --- /dev/null +++ b/codex_reviews/reviews/0tMcsHsHgQ.json @@ -0,0 +1,56 @@ +{ + "paper_id": "0tMcsHsHgQ", + "title": "Towards Undistillable Models by Minimizing Conditional Mutual Information", + "model": "gpt-5.6-sol", + "reasoning_effort": "xhigh", + "created_at": "2026-08-09T18:14:57.991159+00:00", + "paper_md_chars": 32820, + "metadata": { + "pdf_time_verification_status": "verified_revision_cdate_before_review", + "review_time_revision_source": "submission_note", + "arxiv_version_id": null, + "markdown_source": "dataset" + }, + "review": { + "summary": "The paper proposes CMIM, a teacher-training objective that combines cross-entropy with a penalty on the maximum within-class conditional mutual information after power-transforming output probabilities. The within-class CMI is an interpretable measure of output dispersion, and the proposed alternating centroid formulation is plausible. However, the paper does not establish the central implication from low CMI to undistillability. Its universal definition cannot be demonstrated by a finite collection of attacks, the optimized temperature range is only finite and discretized, and the supplied experimental section lacks the numerical results and protocol required to verify its claims.", + "strengths": [ + "The defense is formulated through a clear, interpretable quantity: the expected KL divergence between each prediction and its class-conditional centroid.", + "Considering power-transformed outputs addresses the standard possibility that an attacker changes the KD temperature.", + "The variational centroid identity offers a potentially practical way to optimize the CMI regularizer.", + "The stated evaluation spans CIFAR-100, TinyImageNet, and ImageNet and considers both teacher utility and resistance to distillation.", + "The paper explicitly acknowledges the absence of a formal undistillability guarantee and the additional computational cost." + ], + "weaknesses": [ + "The central logical step is unsupported. Even zero within-class CMI only implies a constant target distribution for each class. Such class-specific targets can encode a learned label-smoothing matrix and may still improve a student over uniform label smoothing. If the targets become one-hot, KD becomes hard-label training, which is also not generally equivalent to LS training.", + "The definition of undistillability quantifies over every KD method and student, whereas the evidence covers only an unspecified finite set. The conclusion should therefore be limited to resistance under the evaluated attacks.", + "The claim of minimizing CMI over the entire temperature spectrum conflicts with the actual objective, which restricts alpha to [0, beta] and then uses a finite grid alpha_i = i beta/N, finite omega, and an unspecified beta. No approximation-error or sensitivity analysis is provided.", + "The method controls empirical CMI on training data, but there is no argument that concentration generalizes to unseen, augmented, or out-of-distribution queries available to a model-extraction attacker.", + "Theorem 4.1 is essentially a continuous log-sum-exp limit. Its regularity assumptions are not stated, and it provides neither a finite-omega nor a quadrature-error bound.", + "The supplied text does not expose the numerical contents of Tables 1-3, Algorithm 1, the tested KD methods, teacher/student architectures, query data and budgets, optimization settings, LS tuning, CMIM hyperparameters, or uncertainty estimates. Consequently, the main empirical claims and comparison fairness cannot be verified.", + "There are no reported adaptive attacks, such as hard-label extraction, alpha values outside the defense grid, learned calibration or nonlinear transformations, or attacks using additional query augmentation.", + "The explanation that concentrated clusters cause higher teacher accuracy is asserted from visualization rather than established through controlled ablations or causal evidence." + ], + "questions": [ + "What exact teacher/student architectures, KD losses, query datasets, query budgets, temperatures, and training schedules were used, and was each LS baseline independently tuned with the same data and compute?", + "What are beta, omega, N, lambda, and the class-specific batch sizes? How sensitive are teacher accuracy, student accuracy, and computational overhead to each parameter?", + "Can the authors provide a theorem or controlled experiment showing that a class-conditional constant soft target cannot outperform uniform label smoothing?", + "How does CMIM perform against adaptive attacks using hard pseudo-labels, alpha greater than beta, learned output calibration, nonlinear target transformations, and additional augmented or out-of-distribution queries?", + "Do low training-set CMI values persist on held-out and attacker-generated query distributions, and are results statistically significant across seeds?" + ], + "limitations": [ + "There is no formal connection between the proposed CMI objective and the paper's universal definition of undistillability.", + "The defense covers only a bounded, discretized family of scalar power transformations.", + "The experimental evidence in the supplied paper text is incomplete and cannot be quantitatively audited.", + "Generalization of the concentration property beyond the training distribution is unexamined.", + "The added training cost is acknowledged but not quantified or compared at equal compute." + ], + "soundness": 3, + "presentation": 3, + "contribution": 4, + "overall_score": 3, + "confidence": 4, + "ethical_concerns": "No acute ethical concern is apparent. The work is framed as IP protection, although making models resistant to distillation could also impede legitimate compression, interoperability, and auditing; these tradeoffs are not discussed.", + "recommendation": "reject", + "review_text": "CMIM is an interesting regularizer with a reasonable intuition: reduce the variation in a teacher's outputs within each ground-truth class, including after temperature-equivalent power transformations, so that the outputs expose less instance-specific information. The CMI-as-average-KL formulation is interpretable, and the auxiliary centroid distributions appear to give a workable optimization route.\n\nThe main problem is that concentration does not establish undistillability. If every example in class y receives the same distribution s_y, the teacher can still communicate class-dependent off-diagonal probabilities. Those targets constitute a learned, nonuniform label-smoothing scheme and can contain useful class-similarity information absent from the uniform LS baseline. Even exact one-hot outputs merely reduce the procedure to hard-label training, which is not generally equivalent to LS. Thus, the proposed trait is at most a heuristic, and no theorem or decisive experiment validates the crucial implication.\n\nThere is also a mismatch between the paper's broad claims and its objective. The method considers alpha only on a bounded interval and approximates the maximum using finite omega and a finite grid. It therefore does not minimize CMI across the entire temperature spectrum, much less across arbitrary logit-based KD transformations. The universal definition of undistillability cannot be established by testing finitely many attacks and students. The defensible claim would be resistance to the particular evaluated configurations.\n\nFinally, the supplied experimental record is insufficient to assess the empirical contribution. The tables are embedded as images without readable numerical contents, and the text omits the attack list, architectures, data/query assumptions, hyperparameters, baseline tuning, variability, and computational overhead. No adaptive attacks or generalization tests for held-out query distributions are reported. These omissions prevent verification of the claimed accuracy improvements and attack resistance. A substantially narrower claim, complete experimental protocol, adaptive evaluation, ablations, and either a theoretical link or explicit counterexample analysis are needed." + } +} diff --git a/codex_reviews/reviews/12gMsxpu4G.json b/codex_reviews/reviews/12gMsxpu4G.json new file mode 100644 index 0000000000000000000000000000000000000000..4955306e23bc67aef23035e993529baffce712c5 --- /dev/null +++ b/codex_reviews/reviews/12gMsxpu4G.json @@ -0,0 +1,56 @@ +{ + "paper_id": "12gMsxpu4G", + "title": "OCS+: Improving PTQ with Outlier Translation", + "model": "gpt-5.6-sol", + "reasoning_effort": "xhigh", + "created_at": "2026-08-09T18:26:17.758239+00:00", + "paper_md_chars": 28179, + "metadata": { + "pdf_time_verification_status": "verified_revision_cdate_before_review", + "review_time_revision_source": "submission_note", + "arxiv_version_id": null, + "markdown_source": "dataset" + }, + "review": { + "summary": "The paper proposes OCS+, an offline channel-expansion transformation that represents clipped activation residuals in duplicated channels. For nonnegative ReLU activations with a uniform quantizer, the core decomposition is intuitive and potentially useful: one channel retains the usual clipped value while another encodes activation above x_clip, and the following layer combines them. The reported prose suggests improvements over a reconstructed OCS baseline, especially for 2-bit lightweight networks. However, central theoretical claims are overstated, the claimed support for general nonlinearities is not established, and the supplied text omits enough experimental detail that the headline SOTA result cannot be verified.", + "strengths": [ + "The method addresses a concrete weakness of activation OCS: halving an entire selected channel coarsens quantization for its inliers.", + "For ReLU-based, unsigned, uniformly quantized structures, translating the residual above x_clip into an added channel is simple and can be implemented through structured changes to adjacent weights and biases.", + "The paper studies the channel-expansion ratio k and reports that OCS+ does not exhibit the degradation of OCS as more channels are selected.", + "The reported comparisons attempt to account for inference cost and to distinguish saved-outlier benefits from merely training wider networks." + ], + "weaknesses": [ + "The claimed exact one-bit gain is mathematically inaccurate under the stated quantizer. With x_u = 2^b-1, the construction covers integer levels 0 through 2x_u, giving 2x_u+1 = 2^(b+1)-1 levels, not 2^(b+1). Thus the stated 3-bit example yields 15 levels rather than 16.", + "The transformation is proved only after specializing to ReLU and a nonnegative unsigned quantizer. Shifting a preactivation before GeLU or H-Swish does not generally produce the translated post-activation residual required by Equation 7. The broad claims covering GeLU, H-Swish, ViTs, and other nonlinearities therefore remain unverified.", + "The paper frames the method as obtaining an unavailable bit under the same hardware bitwidth, but it stores the representation across additional b-bit channels and increases convolutional computation, weight storage, activation memory, and bandwidth. This is not equivalent to a free general-purpose (b+1)-bit activation representation.", + "The supplied text jumps from Section 3 to Section 5, lacks the promised Algorithm 1, and provides no complete experimental protocol. Calibration-set size, optimization schedule, quantization granularity, scale estimation, seeds, variance, hardware, batch size, and latency methodology are not reported here.", + "The abstract's 12.73-point SOTA improvement for W2A2 MobileNet-v2 cannot be verified from the supplied text. Most tables are image references without transcribed values, and no complete main baseline comparison is available.", + "The deployment argument is based primarily on fake-quantization equations. It is not shown that the translated biases, duplicated channels, quantizer sharing, and accumulation can be executed by an integer inference stack without unsupported operations or material conversion overhead.", + "The channel score—the sum of activations in [x_clip, 2x_clip]—is called a sensitivity criterion but contains no task-loss sensitivity information. Comparisons against random selection, gradient-based criteria, or simpler outlier-frequency criteria are absent.", + "The wider-network experiment in Table 8 is not a clean causal control: a separately trained wider model differs in both optimization and learned function from the structured duplicated OCS+ model. An equal-topology ablation with the translated residual branch disabled would be more informative.", + "The manuscript contains malformed equations and figures, inconsistent terminology, and numerous grammatical errors that materially hinder verification." + ], + "questions": [ + "How is the transformation modified for GeLU, H-Swish, signed activations, and two-sided outliers? Please provide an exact identity or restrict the claim to ReLU-like activations.", + "How do the authors reconcile the one-bit claim with the level count 2(2^b-1)+1 = 2^(b+1)-1?", + "Can the authors provide the complete experimental protocol and textual main-results table supporting the claimed 12.73-point improvement, including calibration data, hyperparameters, repetitions, and variance?", + "What are the end-to-end changes in FLOPs, parameter storage, activation memory, memory traffic, energy, and latency for each k, measured on named hardware at controlled batch sizes?", + "Can OCS+ be demonstrated using actual integer inference kernels rather than fake quantization, including the representation and quantization of the shifted biases?" + ], + "limitations": [ + "OCS+ applies only where adjacent layers permit structured channel duplication and recombination.", + "The presented derivation assumes nonnegative activations, uniform quantization, and effectively ReLU behavior.", + "Coverage is limited to outliers up to 2x_clip for the beta=2 construction; larger outliers remain clipped.", + "Fixed channel selection from calibration data may miss channels that become outliers under distribution shift.", + "Because numerical table contents are unavailable in the supplied text, this review can assess only the prose-reported results, not the full quantitative evidence." + ], + "soundness": 4, + "presentation": 3, + "contribution": 5, + "overall_score": 4, + "confidence": 4, + "ethical_concerns": "No material ethical concerns are apparent from the supplied paper text.", + "recommendation": "weak_reject", + "review_text": "OCS+ contains a useful core idea: for a ReLU activation and unsigned uniform quantization, a clipped value and a translated residual can be stored in separate channels and recombined by duplicating the following weights. This directly avoids OCS's coarsening of ordinary values in selected channels, and the reported low-bit results suggest that the idea may be practically valuable.\n\nTwo central theoretical statements nevertheless need correction. First, the construction provides 2^(b+1)-1 distinct levels, not the claimed 2^(b+1), because x_clip=(2^b-1)s and the represented range ends at 2x_clip. Second, the proof relies on ReLU: applying a bias shift before GeLU or H-Swish is not equivalent to translating their output. Consequently, the claimed generality to those activations and ViTs is unsupported by the derivation. The method also uses extra b-bit channels, so its comparison to obtaining a hardware bit should be qualified by complete compute, storage, bandwidth, and latency accounting.\n\nThe empirical evidence is insufficiently reproducible in the supplied text. Section 4 and Algorithm 1 are absent, experimental settings are not specified, and the headline SOTA improvement cannot be checked against a complete results table. The prose-reported comparisons with OCS are promising, but stronger equal-resource baselines, variance across runs, selection-criterion ablations, and real integer-kernel deployment are needed. I therefore view the current submission as an interesting but insufficiently validated and theoretically overstated contribution." + } +} diff --git a/codex_reviews/reviews/15dVqf7VXR.json b/codex_reviews/reviews/15dVqf7VXR.json new file mode 100644 index 0000000000000000000000000000000000000000..6e882d34748f46419575f3281dbe1dc3dfd319eb --- /dev/null +++ b/codex_reviews/reviews/15dVqf7VXR.json @@ -0,0 +1,56 @@ +{ + "paper_id": "15dVqf7VXR", + "title": "Learning with User-Level Local Differential Privacy", + "model": "gpt-5.6-sol", + "reasoning_effort": "xhigh", + "created_at": "2026-08-09T18:29:02.002516+00:00", + "paper_md_chars": 126035, + "metadata": { + "pdf_time_verification_status": "verified_arxiv_version_before_review", + "review_time_revision_source": "arxiv", + "arxiv_version_id": "2405.17079v1", + "markdown_source": "arxiv" + }, + "review": { + "summary": "The paper studies user-level local differential privacy for mean estimation, stochastic optimization, classification, and regression, proposing privacy-adaptive grouping, Kashin representations, and Hadamard transforms. The topic is important and the proposed unification is potentially valuable. However, several central upper and lower bounds are not established by the supplied arguments, and some theorem statements are false without additional parameter restrictions or saturation terms.", + "strengths": [ + "The distinction between item-level and user-level LDP is well motivated, especially the possibility that a user-level encoder can exploit all local samples before privatization.", + "The privacy-adaptive grouping strategy is an interesting attempt to cover small, medium, and large ε within one framework.", + "The one-dimensional two-stage construction has a clear privacy argument, and using mean estimation as a primitive across several learning tasks gives the paper a coherent structure.", + "The paper attempts both achievable rates and information-theoretic lower bounds and explicitly states several modeling limitations." + ], + "weaknesses": [ + "The main lower-bound argument misapplies Lemma 20 to product distributions. In classification, p_v^m and p_v'^m differ whenever any of the m samples falls in B_1, so the relevant set has probability 1-(1-l^d)^m, not l^d. Likewise, in the high-dimensional mean proof the corresponding probability is 1-(1-1/d)^m, not 1/d. The extra factors obtained from the incorrect substitutions are essential to the claimed rates, so the mean, classification, and regression minimax results are not proved.", + "The classification upper bound ignores that each Q_k estimator receives only n/K users. The stage-I failure probability should therefore involve exp(-c(n/K)ε²), followed by a union bound over K estimators, whereas the proof uses exp(-cnε²). Consequently, Theorem 10's condition n(ε²∧1) >= c(log m+log n) is insufficient for the stated algorithm. The same problem occurs with the 2K groups used for regression.", + "Several lower bounds are false as stated because they omit saturation. For example, D²/[nm(ε²∧1)] can exceed D² even though the constant-zero estimator has worst-case error at most D². The appendix's own construction retains a minimum with a constant-scale term and does not imply the unqualified theorem statement.", + "The high-dimensional lower-bound proofs only control privatized KL through (e^ε-1)², but then replace this by ε² and claim dependence on ε²∧ε for all privacy levels. This approximation is valid only for small ε; the asserted medium-privacy 1/ε lower bound is not derived.", + "The heavy-tailed analysis has substantive gaps: stage I naturally localizes the mean μ_T of the clipped user average rather than μ, the displayed bias decomposition incorrectly uses subtraction instead of addition, and the theorem and appendix contain inconsistent dependence on M_p. Thus the claim that user-level LDP is faster than item-level LDP for heavy tails is not verified by the supplied proof.", + "The regression algorithm returns H_K P/K rather than H_K hat(P)/K and predicts with the potentially unstable ratio hat(q)_k/hat(p)_k. The proof silently introduces clipping by T, assumes a bounded estimator on failure events, and reverses the definition of its good event. Without an explicit two-sided clipped or denominator-stabilized estimator, the risk bound does not follow.", + "The one-dimensional variance calculation treats second-stage outputs as independent despite their shared random interval [L,R]. A law-of-total-variance term for the randomness of the first stage is missing. This may be repairable using the successful-stage event, but it is not done.", + "The stochastic-optimization result defines a constrained domain Θ but performs unprojected updates and uses properties appropriate to an unconstrained stationary minimizer. Either Θ must be all of Euclidean space or a projection and corresponding analysis are needed. Globally bounded gradients are also incompatible with global strong convexity on an unbounded domain.", + "There are numerous additional inconsistencies, including an incorrect non-squared optimization rate in the contribution list, undefined binary-label support in classification, inconsistent symbols T and t_0, and an incorrect orthogonality assertion in the Kashin proof. These make the already technical arguments difficult to verify." + ], + "questions": [ + "Can the lower bounds be rederived using a valid divergence or total-variation bound for the full m-sample user dataset, with explicit regimes depending on m and the probability of the perturbed region?", + "What rates and sample-size conditions remain after using n/K users per Hadamard coefficient and union-bounding all stage-I failures in classification and regression?", + "What is the exact regression estimator, including denominator stabilization and two-sided clipping, and how is its error controlled on the bad event?", + "Can all lower-bound statements be revised to include saturation and can the claimed ε²∧ε dependence be proved beyond the small-ε regime?", + "For stochastic optimization, is Θ intended to be unconstrained, or should Algorithm 3 include projection?" + ], + "limitations": [ + "The work is entirely theoretical and provides no empirical evaluation of the proposed mechanisms or their constants.", + "The model assumes equal numbers of i.i.d. samples per user, excluding heterogeneous and non-i.i.d. federated data.", + "Classification and regression require a density uniformly bounded away from zero; regression additionally assumes bounded responses, and optimization assumes bounded gradients and strong convexity.", + "Communication, computation, and the construction cost of the transforms are not analyzed.", + "Several guarantees require the number of users to grow with dimension and log m, limiting applicability in small-user regimes." + ], + "soundness": 2, + "presentation": 4, + "contribution": 6, + "overall_score": 3, + "confidence": 5, + "ethical_concerns": "No direct ethical concern is apparent. However, because the work concerns formal privacy and utility guarantees, the mechanisms should not be presented as deployment-ready until the proof gaps are resolved.", + "recommendation": "reject", + "review_text": "The paper addresses an important and relatively underexplored problem, and its privacy-adaptive grouping perspective is promising. Nevertheless, I cannot support the paper in its current form because the principal minimax results and two of the main upper bounds contain central mathematical gaps. Most importantly, the generalized Pinsker argument is applied to m-fold product distributions using the probability of a one-sample differing set. The product distributions differ on the event that at least one local sample belongs to that set, whose probability is 1-(1-a)^m rather than a. This invalidates the key extra factors used in the mean, classification, and regression lower bounds. Independently, the classification and regression upper analyses overlook that localization is run with only n/K or n/(2K) users per coefficient, leading to substantially larger failure probabilities and stronger sample requirements than stated. The regression estimator and proof also disagree about clipping and the use of hat(P), while the optimization method does not handle its declared constraint set. Finally, the lower bounds omit necessary saturation and do not justify the claimed 1/ε behavior outside the small-ε regime. These issues affect the headline conclusions rather than only constants or exposition. A substantially revised analysis could still make the underlying algorithmic ideas valuable." + } +} diff --git a/codex_reviews/reviews/1CK45cqkEh.json b/codex_reviews/reviews/1CK45cqkEh.json new file mode 100644 index 0000000000000000000000000000000000000000..f750ba9c3fc6eacc226fc3becbe012c928a7f9a0 --- /dev/null +++ b/codex_reviews/reviews/1CK45cqkEh.json @@ -0,0 +1,61 @@ +{ + "paper_id": "1CK45cqkEh", + "title": "Unsupervised Order Learning", + "model": "gpt-5.6-sol", + "reasoning_effort": "xhigh", + "created_at": "2026-08-09T18:36:28.931210+00:00", + "paper_md_chars": 32395, + "metadata": { + "pdf_time_verification_status": "verified_revision_cdate_before_review", + "review_time_revision_source": "submission_note", + "arxiv_version_id": null, + "markdown_source": "dataset" + }, + "review": { + "summary": "The paper proposes Unsupervised Order Learning (UOL), which alternates an ordered variant of k-means with representation learning. Ordered k-means assigns each sample according to its squared distances from a centroid and its two neighboring centroids, while the network is trained using a straight-line loss and a deviation loss. The idea is relevant and potentially useful, but the supplied paper contains serious metric, optimization, identifiability, and reproducibility concerns that prevent the empirical and theoretical claims from being established convincingly.", + "strengths": [ + "The paper targets an interesting setting—discovering both clusters and their order without target-label supervision—that is less studied than nominal deep clustering.", + "The ordered k-means objective is intuitive: incorporating neighboring centroids explicitly encodes local adjacency rather than treating cluster identities as unrelated.", + "The centroid, assignment, and permutation perspectives provide a reasonably interpretable connection to ordinary k-means, with ordinary k-means recovered when alpha is zero.", + "The evaluation spans facial age, retinal-disease severity, and facial-expression data, and compares against several deep-clustering methods using the same stated VGG16 backbone.", + "The paper explicitly acknowledges reversal symmetry when evaluating an order, rather than treating the two orientations as intrinsically different.", + "The loss ablation and embedding visualizations are relevant analyses, although their numerical evidence cannot be fully inspected from the supplied text." + ], + "weaknesses": [ + "The stated SRCC equation is incorrect as written. It omits the square on the rank difference, uses k rather than the number of evaluated instances in the normalization, and applies a permutation-only formula to repeated class and cluster ranks. Consequently, the reported SRCC values cannot be validated without the actual implementation and a correct definition.", + "The learned ordering is not identifiable as the claimed semantic order. The objective favors a chain through clusters, but it does not establish that this chain corresponds to age, disease severity, or another intended latent factor rather than pose, acquisition conditions, identity, or demographic attributes. Reversal is only one ambiguity; alternative nuisance-driven chains can also be meaningful to the objective.", + "The straight-line and deviation losses do not enforce the biconditional order desideratum in Equation 7. They impose local geometric constraints around neighboring centroids but contain no explicit global separation, non-overlap, or monotonic-distance constraint for every i