{"id":"948e452a-903a-4424-98fa-56d280e8d421","arxiv_id":"2608.06559","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A frozen in-context learning model, conditioned on bootstrap resamples of the interaction history and an arm-context feature map, selects actions and beats several contextual bandit baselines in cumulative regret.","lead":"This paper combines bootstrap resampling with a frozen pretrained tabular model to pick actions in contextual bandit problems. It reports lower cumulative regret than linear, kernel, and neural bandit baselines on several standard benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The abstract's unqualified claim that BC-ICL outperforms established baselines is contradicted by the paper's own Table 1: on MNIST, BC-ICL-TabICL (1936.8±187.3) is worse than several baselines, including BootstrapNN (1473.4±100.8) and LinUCB (1568.7±100.5).","rationale":"The reader's verdict is CONDITIONAL, and this stress-test supports that outcome without moving it. The reader's weakest_assumption focuses on pretraining overlap and inductive bias alignment; that is a legitimate external risk, but the manuscript provides no information to confirm or refute it. The more immediate and verifiable problem is the internal contradiction between the abstract and Table 1. BC-ICL-TabICL on MNIST does not outperform several baselines, and the error bars (187.3 vs. 126.1 for Linear TS) do not exclude a real deficit. The paper's own text in Section 3 correctly limits the claim to 'most datasets,' but the abstract does not. Since the central claim as advertised is the empirical superiority of the method, and since the paper's data contradict that claim for at least one variant on a standard benchmark, the claim needs qualification or removal. This does not reject the paper: BC-ICL-TabPFN wins on MNIST, and both variants win on most other datasets. The concern is addressable by rephrasing the abstract and main-claim statements, and by adding the paired significance analysis proposed above. The reader already flagged the MNIST overclaim in the rationale, so agreement is partial rather than full. The verdict remains CONDITIONAL because the paper should not be accepted with its current abstract; but the flaws are not terminal and do not warrant rejection.","tokens_in":12993,"tokens_out":6535,"duration_ms":58317,"concrete_test":"Reanalyze Table 1 by comparing, for each dataset and each BC-ICL variant, the reported mean regret against the best-performing baseline using the given standard errors (or re-run the 10 seeds to obtain paired per-round regrets for a paired test). Document every variant-dataset configuration where BC-ICL's mean is not lower than the best baseline. If any such configuration exists — MNIST/BC-ICL-TabICL being a clear candidate — the abstract's unqualified 'outperforming established baselines' must be revised to a qualified claim such as 'outperforms on most datasets' or 'outperforms when instantiated with TabPFN.'","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim, as stated in the abstract and in the reader's strongest_claim, is that BC-ICL 'outperforms established baselines' on standard contextual bandit suites. Table 1 does not support this universally. For MNIST, BC-ICL-TabICL has cumulative regret 1936.8±187.3, which is higher than Linear TS (1774.6±126.1), LinUCB (1568.7±100.5), NeuralUCB (1656.3±42.2), NeuralTS (1639.5±61.1), and BootstrapNN (1473.4±100.8). The paper's own Section 3 hedges to 'most datasets,' so the abstract overstates the evidence. This is an internal inconsistency, not a matter of external consensus: the manuscript's own numbers falsify the unqualified superiority claim. If the claim is intended to apply only to BC-ICL-TabPFN or only to 'most datasets,' that qualification must appear in the abstract. This concern is more load-bearing than the reader's leakage worry because it is demonstrable from the submitted data alone, whereas pretraining overlap is a plausible but unverified risk. The central claim's scope is exactly what the abstract advertises, and the evidence fails to cover one of the two proposed instantiations on a standard benchmark.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes BC-ICL, a contextual-bandit policy that at each round draws a bootstrap resample of the interaction history, conditions a frozen pretrained tabular in-context-learning (ICL) model on that resample, scores all arms through a multiplicative arm-context feature map, and plays the highest-scoring arm. The method is instantiated with TabPFN and TabICL and evaluated on eight UCI/MNIST classification-to-bandit benchmarks against linear, kernel, and neural baselines. The authors report strong cumulative-regret results on most datasets, ablation studies showing that bootstrap exploration beats greedy and sampling, and an arm-context feature-map ablation showing gains from multiplicative interactions. The paper also discusses computational cost and context-window strategies (FIFO and KNN). The central empirical claim, however, is stated too broadly in the abstract and introduction, because the paper's own Table 1 shows that BC-ICL-TabICL is worse than several baselines on MNIST.","tokens_in":13363,"tokens_out":3727,"duration_ms":34486,"significance":"If the empirical claims survive after necessary qualification, the paper makes a useful and simple contribution: it shows that a frozen pretrained tabular ICL model can be converted into a randomized bandit policy through bootstrap-conditioned action selection, avoiding online gradient training and relying on the model's pretrained inductive bias for exploration. The paper contains a clear method description, extensive benchmark comparisons, ablations for exploration strategy and arm-context encoding, and a practical runtime analysis with context-window selection. The main weaknesses are an overstated abstract claim contradicted by Table 1 on MNIST, the absence of code and seed-level reproducibility details, the lack of sensitivity analysis for the random projection matrix P, and the unaddressed possibility that benchmark datasets overlap with the ICL pretraining corpora. These issues are fixable but currently block acceptance.","major_comments":[{"comment":"The unqualified claim that BC-ICL 'outperforms established baselines' is contradicted by the paper's own results. On MNIST, BC-ICL-TabICL has cumulative regret 1936.8±187.3, which is worse than Linear TS (1774.6±126.1), LinUCB (1568.7±100.5), NeuralUCB (1656.3±42.2), NeuralTS (1639.5±61.1), and BootstrapNN (1473.4±100.8). Since MNIST is one of the eight reported suites and TabICL is one of the two proposed instantiations, the claim as written is internally inconsistent with Table 1. Please qualify the claim in the abstract and in the contribution list (e.g., 'on most datasets' or 'for the TabPFN instantiation'), and explicitly acknowledge the MNIST TabICL result in the main-text discussion.","section":"Abstract and §3, Table 1"},{"comment":"The random projection matrix P is a free parameter of the proposed multiplicative feature map, but it is never varied. The evidence for the benefit of multiplicative encoding (Figure 1, right) is based on a single draw of P, so the reported 30–40% improvement could be an artifact of that draw. Please report results across multiple draws of P or provide an argument that the regret is invariant to P; at minimum, state why the fixed random projection is not a load-bearing choice.","section":"§3, Implementation Details; §2, arm-context feature map"},{"comment":"The evaluation may be partly circular if TabPFN's or TabICL's pretraining corpora included the benchmark datasets, several of which are standard UCI sets (Adult, Covertype, Mushroom, Letter, Shuttle, Isolet, MagicTelescope). The paper does not discuss this possibility. Please add a statement on pretraining-data overlap with the evaluation benchmarks, or evaluate on datasets released after the models' pretraining cutoff; without this, the 'pretrained inductive bias' explanation for the observed gains is not fully identified.","section":"§3, Datasets; §4, Conclusion"},{"comment":"The theoretical paragraph states that 'one can derive a regret decomposition' into stability and representation terms, but the paper provides no formal statement, no precise stability assumption, and no proof. Because the paper's contribution is empirical, this paragraph as written overstates the theoretical content. Either remove it or formalize the decomposition with explicit assumptions and a precise claim.","section":"§4, Conclusion"}],"minor_comments":[{"comment":"The contribution list says BC-ICL 'outperforms linear, kernelized, and neural contextual bandit baselines' without qualification; this should be aligned with the actual results, as noted in the major comments.","section":"§1, Contributions"},{"comment":"No code or data release is mentioned, and the text reports '10 seeds' without specifying how seeds were generated or how dataset shuffling was performed. A reproducibility statement with seeds, exact hyperparameter values, and dataset split details would be needed for the empirical claims to be verifiable.","section":"§3, Implementation Details"},{"comment":"The legend in the extracted text appears garbled (e.g., 'TabICL (©mult)' and 'TabICL (©one ¡ hot)'); the figure should be checked for correct rendering of the subscript and one-hot labels.","section":"Figure 1, right panel"},{"comment":"The notation eDt for the bootstrap resample is introduced but not used consistently in the subsequent text and in the description of the arm-selection rule; please clarify that the resample is drawn independently at each round and specify how ties in argmax are broken.","section":"§2, Method"},{"comment":"The text says that reducing TabICL's n_estimators from 32 to 8 'does not degrade performance,' but no table or figure is provided to support this claim; please report the comparison or soften the statement.","section":"Appendix B.2"},{"comment":"The FIFO context-selection description says the current context is added to the conditioning set with bootstrap probability p=0.8, which is a different mechanism from the bootstrap resample described in Section 2; please reconcile the two procedures.","section":"Appendix B.4"}],"recommendation":"major_revision","confidential_remarks":"The paper is clearly written and the core idea is simple and potentially useful for practitioners. The main blocker is the mismatch between the abstract's unqualified superiority claim and the MNIST/TabICL results in Table 1. In addition, the lack of sensitivity analysis for the random projection and the absence of reproducibility details would likely be raised by other referees. The pretraining-overlap concern is real but not verifiable from the manuscript; asking the authors to address it directly is reasonable. If the authors qualify the claims, add the P-sensitivity experiment, and provide code/seeds, the paper could be acceptable for a workshop venue."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper is worth a look. The core idea—bootstrap-resample the accumulated history, feed each resample to a frozen tabular ICL model, and play the argmax—is simple and I haven't seen it before. The multiplicative arm-context feature map is a useful small contribution. Empirically, BC-ICL (especially with TabPFN) gets lower cumulative regret than linear, kernel, and neural baselines on six of eight datasets, and on Mushroom the gains are huge.\n\nThe paper does several things right. The ablation against greedy and sampling is honest: it shows bootstrap adds something beyond the model's own predictive uncertainty. The runtime analysis with FIFO/KNN context selection is practical. The writing is clear about the method's limits in the last section.\n\nBut the abstract goes further than the data. Table 1 shows BC-ICL-TabICL on MNIST has regret 1936.8, worse than BootstrapNN (1473.4), NeuralTS (1639.5), LinUCB (1568.7), and Linear TS (1774.6). That's from the paper's own numbers. So the blanket claim \"outperforming established baselines under a strict online protocol\" is false for one of the two instantiations on a standard benchmark. The stress-test note is right: this is a demonstrable internal inconsistency, more concrete than the leakage worry. Fixing it is just a matter of rewording the abstract.\n\nOther soft spots are standard but real: no code, no seed details, and no discussion of whether the ICL pretraining distribution overlaps the evaluation datasets. The random projection P is not varied, so its sensitivity is unknown. These are not fatal, but they are exactly what a reviewer should ask for.\n\nThe theory section is speculative, but it is labeled as future work, so I don't hold that against it.\n\nVerdict: this deserves a serious referee. The method is novel, the experiments are mostly convincing, and the flaws are fixable. I would want to see a revised version with an honest abstract, code (or at least seeds), and a leakage check before trusting the broadest claims. For a reading group on exploration with foundation models, this is a good pick.","headline":"A genuinely new combination that mostly works, but the abstract overclaims and the paper needs code and leakage checks before the headline claim is trusted.","tokens_in":13879,"tokens_out":2474,"would_cite":true,"duration_ms":22471,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A frozen, pretrained tabular in-context-learning model can be turned into a low-regret contextual bandit policy by resampling the interaction history and playing the highest-scored arm, beating linear, kernelized, and neural baselines on…","keywords":["contextual bandits","in-context learning","tabular foundation models","bootstrap resampling","exploration","regret minimization","arm-context interactions","Thompson sampling"],"falsifier":"Run BC-ICL on a synthetic or held-out bandit task deliberately outside the backbone's pretraining distribution, such as one with class labels permuted on a dataset the in-context model never saw, and compare cumulative regret against BootstrapNN; if the advantage vanishes or reverses, the reported gains come from prior-task alignment rather than from the bootstrap mechanism.","tokens_in":12827,"feed_emoji":"🎰","tokens_out":9496,"duration_ms":79515,"temperature":0.7,"pith_summary":"The paper asks whether a pretrained tabular foundation model, designed for in-context prediction, can serve as the reward model of an online contextual bandit. It argues yes, introducing BC-ICL: each round the algorithm resamples the observed interaction history, conditions a frozen in-context learning predictor on that resample, scores every candidate arm, and plays the arm with the highest sampled score. This turns the model's pretrained inductive bias into a decision rule, with exploration coming from bootstrap-induced variation rather than from training an ensemble or estimating uncertainty online. On eight standard contextual bandit suites converted from classification datasets, the paper reports that BC-ICL, instantiated with the TabPFN or TabICL backbones, achieves lower cumulative regret than linear, kernelized, and neural baselines, with especially large early-round gains on small-data problems. If the result holds, cold-start personalization could inherit the sample efficiency of large tabular pretraining without fragile uncertainty estimates.","feed_headline":"Neural bandits beaten by frozen tabular AI plus bootstrap resampling","feed_subtitle":"BC-ICL resamples history, scores every arm with a frozen in-context model, and cuts regret on standard bandit suites.","key_machinery":"The mechanism is randomized conditioning of a frozen backbone. At round $t$, BC-ICL draws a bootstrap resample $\\tilde{D}_t$ of the interaction history, computes the feature map $\\Phi_{\\mathrm{mult}}(x,a) = [x; e_a; (x^\\top P) \\odot e_a]$ for every candidate arm, feeds these to the frozen in-context learning predictor $M(\\cdot|\\tilde{D}_t)$, and plays the arm with the largest predicted score. The shared random projection $P$ makes bootstrap perturbations propagate through a common context representation, so scores for all arms move together, which the paper identifies as the shared-exploration mechanism that avoids isolated-arm bootstrap failure. The named object is the multiplicative arm-context feature map: arm identity enters through a fixed embedding $e_a$ and arm–context interactions enter through elementwise multiplication with the shared projected context, and the paper argues this is what lets the shared projection help all arms simultaneously.","core_discovery":"On the paper's own terms, the central discovery is that data-level bootstrapping of the conditioning history, not the in-context model's internal predictive uncertainty, is what makes frozen pretrained predictors explore effectively. The ablation shows that greedy selection and sampling from the model's softmax incur substantially higher regret on hard multiclass datasets, while bootstrap resampling reduces regret by 5–19%, and on Mushroom BC-ICL-TabICL reaches 37.7 regret versus 252.3 for NeuralTS. A second element is the multiplicative arm-context feature map, which the paper shows is worth a roughly 30–40% regret reduction over a one-hot arm encoding on ISOLET, independent of the ICL backbone. The paper presents these results as evidence that a pretrained in-context model's inductive biases are directly usable for online decision making under a strict online protocol.","pith_inferences":["Editorial inference: because BC-ICL only needs a score for each (context, arm) pair, the same bootstrap-conditioning recipe should transfer to frozen models beyond the two tested backbones, and to offline evaluation settings where resampling logged data is already routine.","Editorial inference: the regret decomposition the paper sketches implies a clear boundary condition—when the in-context predictor's representation error on the bandit's reward class is large, the bootstrap concentrates on the wrong arm and regret is dominated by that error; formalizing this rate would tell practitioners when not to use BC-ICL.","Editorial inference: the paper does not address possible overlap between the backbones' pretraining data and the evaluation datasets, so part of the reported gap could reflect memorized prior knowledge; a held-out or label-permuted evaluation would separate transfer from memory.","Editorial inference: the resampling weights of the bootstrap set the exploration distribution, so downweighting stale interactions or annealing the resample size would give a tunable exploration-exploitation continuum the paper leaves unexplored."],"forward_implications":["A frozen pretrained in-context model can act as a drop-in reward model for contextual bandits, so online personalization systems could gain sample efficiency without per-task network training or hand-tuned uncertainty estimates.","Exploration in this setting comes from randomizing the conditioning set, not from posterior or predictive uncertainty; the ablation's greedy-versus-bootstrap gap implies that algorithms built on frozen predictors should randomize the data they condition on.","The multiplicative arm-context feature map is load-bearing: on ISOLET it yields roughly 30–40% lower regret than a one-hot arm encoding, so feature construction matters as much as the backbone choice.","KNN-based context selection keeps regret close to full-history BC-ICL while bounding per-round cost, which makes the method usable for horizons far beyond the 10,000-round benchmarks.","On simple problems, greedy selection is nearly as good, so the practical benefit of bootstrap exploration concentrates in hard, high-dimensional cold-start settings."],"supporting_citations":[{"why":"Provides TabPFN, one of the two frozen pretrained in-context learning backbones used to instantiate BC-ICL.","marker":"Hollmann et al., 2025"},{"why":"Provides TabICL, the second in-context learning backbone, giving an alternative instantiation of BC-ICL.","marker":"Qu et al., 2025"},{"why":"Supplies the benchmark protocol of converting classification datasets into contextual bandits and the block-diagonal context representation used for neural baselines.","marker":"Riquelme et al., 2018"},{"why":"Defines NeuralUCB and the neural bandit training setup that serves as the strongest non-bootstrap baseline family in the comparison.","marker":"Zhou et al., 2020"},{"why":"Introduces bootstrap ensembles for exploration, the idea BC-ICL adapts by resampling the conditioning set of a frozen model.","marker":"Osband et al., 2016"},{"why":"Proposes bootstrap Thompson Sampling, the direct antecedent of data-resampling-based exploration in bandits.","marker":"Eckles & Kaptein, 2014"},{"why":"Documents the linear-regret failure of naive bootstrapping in isolated-arm bandits, motivating the shared arm-context structure BC-ICL uses.","marker":"Kveton et al., 2019"},{"why":"Frames personalized recommendation as a contextual bandit and establishes the global-policy setting the paper builds on.","marker":"Li et al., 2010"}],"fun_headline_variants":["Bootstrap beats softmax for frozen tabular bandit exploration","Frozen ICL bandits explore via bootstrap, not softmax sampling","Arm-context conditioning lifts frozen tabular bandits by 30%","Resampling history, not model uncertainty, drives bandit exploration","Bootstrap resampling cuts regret in frozen tabular bandits"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method only works if the pretrained model's inductive bias is aligned with the bandit's reward function; on a mismatched task, the bootstrap cannot fix the bias.","fun_headline_variants_meta":{"raw":{"variants":["Bootstrap beats softmax for frozen tabular bandit exploration","Frozen ICL bandits explore via bootstrap, not softmax sampling","Arm-context conditioning lifts frozen tabular bandits by 30%","Resampling history, not model uncertainty, drives bandit exploration","Bootstrap resampling cuts regret in frozen tabular bandits"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000735,"raw_usage":{"total_tokens":3252,"prompt_tokens":880,"completion_tokens":2372,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":496,"completion_tokens_details":{"reasoning_tokens":2282}},"tokens_in":496,"tokens_out":2372,"duration_ms":15874,"temperature":1.0,"reasoning_tokens":2282,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T04:17:44.719373+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run BC-ICL on a synthetic or held-out bandit task deliberately outside the backbone's pretraining distribution, such as one with class labels permuted on a dataset the in-context model never saw, and compare cumulative regret against BootstrapNN; if the advantage vanishes or reverses, the reported gains come from prior-task alignment rather than from the bootstrap mechanism.","supporting_citations":[],"review_version":1}