{"id":"381fad3f-4365-4124-b131-a9f5d854c8f0","arxiv_id":"2505.08243","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Reasoning pre-training and dropout let robot policies benefit from chain-of-thought training without run-time reasoning, preserving most of the performance gains at standard VLA inference speeds.","lead":"This paper tests why robot chain-of-thought reasoning helps robot policies and finds that learning to produce reasoning steps improves the policy's internal representations, even when the reasoning is not generated at test time. The authors introduce two lightweight training recipes, reasoning pre-training and reasoning dropout, that match or approach full reasoning performance while running about 3x faster at inference.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Missing control for non-semantic token dropout leaves the claim that reasoning content, not generic regularization, drives the gains unsupported.","rationale":"The reader's weakest assumption identifies the same gap I consider most load-bearing: no condition isolates the semantic content of reasoning tokens from generic dropout regularization. The paper's other controls are informative but not decisive. Thinking tokens (Sec. 6.5) vary token quantity without dropout or generation loss, so they address Hypothesis 3 but not the dropout confound. Co-training (Sec. 6.3) removes action-to-reasoning attention but also changes data mixing and batch size. Reasoning pre-training (Sec. E.1) avoids dropout, but it has a compute mismatch (100k reasoning + 200k action steps vs. 200k action-only steps), so it cannot fully separate 'learning to generate reasonings' from 'more training on an auxiliary task.' The random-token dropout test I propose would settle the primary confound for the headline dropout recipe. I agree with the reader's CONDITIONAL verdict: the empirical recipes are well supported and the paper is transparent about limitations, but the central mechanistic claim should remain conditional on this control. Secondary concerns, such as pre-training compute matching and the provenance of the Bridge 'reasoning dropout' policy described in Appendix E.3, reinforce the same conclusion without changing it.","tokens_in":27972,"tokens_out":15181,"duration_ms":157553,"concrete_test":"Re-run the LIBERO-90 reasoning dropout condition with the reasoning tokens replaced by an equal number of tokens drawn from a fixed set of non-semantic tokens (e.g., repeated '.' or random BPE tokens), under the identical random-prefix-dropout schedule and token-level generation loss, keeping MiniVLA, 200k steps, batch size, and VQ-VAE action decoding unchanged. Evaluate with the token sequence absent at test time. If this random-token dropout policy reaches within about 1-2 points of the reported 89.4% on standard LIBERO-90 and the Perturb variants, the gain is generic dropout or auxiliary sequence prediction rather than the semantic content of embodied reasoning, and the abstract's causal claim must be weakened. If it remains near the 82.0% standard-VLA baseline, the semantic-content interpretation is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the semantic content of embodied reasoning—not just extra tokens, compute, or stochastic masking—improves VLA representations, and that attending to those reasonings leverages the improvement. The paper's most effective recipe, reasoning dropout (Sec. 5, E.3), differs from the standard VLA baseline in two confounded ways: the model is trained to generate and attend to semantic reasoning tokens, and it is trained with random prefix dropout. The thinking-token condition (Sec. 6.5) controls only for token count and compute; it does not apply the same dropout schedule or token-level generation loss, so it cannot rule out generic dropout regularization. With only 3,917 LIBERO training trajectories, stochastic masking of context tokens could plausibly improve robustness by ordinary regularization, matching the +7.4-point gain (89.4% vs 82.0%) without any reasoning-specific representation. The authors' own Limitations section (Sec. 9) concedes that the learning-dynamics mechanism is not investigated and describes their explanations as 'intuitive speculation'; no representation probing is reported. Thus the causal interpretation in the abstract is underdetermined: the practical recipes may be valid, but if random-token dropout reproduces the gain, the stated mechanism—the paper's main scientific contribution—would be wrong.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper investigates why embodied chain-of-thought (ECoT) reasoning improves vision-language-action (VLA) policies and then proposes two lightweight training recipes, \"reasoning pre-training\" and \"reasoning dropout,\" that retain most of the benefits of test-time reasoning while maintaining fast inference. The authors hypothesize three mechanisms: better representation learning, an implicit learning curriculum, and increased effective expressivity. They isolate these by testing reasoning pre-training, co-training, dropout, scaffolding, and non-semantic thinking tokens. On LIBERO-90, reasoning dropout reaches 89.4% on the standard split and 76.4% average across variants, versus 82.0%/67.4% for the standard VLA baseline; reasoning pre-training reaches 87.1%/72.8%. On Bridge real-robot tasks, both recipes outperform the standard VLA baseline on aggregate, with reasoning pre-training at 69.4% versus 50.5% for the baseline. The paper concludes that learning to generate reasonings improves VLA representations and that attending to reasonings during training is what matters, while the semantic content of thinking tokens is irrelevant. The two proposed recipes do not generate reasoning at test time and are about 3x faster than full ECoT.","tokens_in":28174,"tokens_out":4625,"duration_ms":48520,"significance":"If the central claim is correct, the paper contributes two practical and well-evaluated training strategies that give a large part of embodied CoT's performance benefit without the inference-time cost, which would be useful for real-time robot control. The empirical scale is a clear strength: 121,500 simulated trials and 444 real-robot trials, with shared evaluation seeds and reported standard errors. The thinking-token control is also a valuable negative result: it suggests that merely adding compute tokens is not enough, consistent with the paper's emphasis on semantically meaningful reasoning. However, the scientific mechanism claim is not yet fully established: the most effective recipe, reasoning dropout, is not compared against a non-semantic token-dropout control, and reasoning pre-training is not compute-matched to the baseline. The paper's own Limitations section (Sec. 9) acknowledges that the learning-dynamics mechanism is not investigated and describes the explanations as \"intuitive speculation.\" The practical recipes may still stand, but the causal interpretation in the abstract requires additional control experiments.","major_comments":[{"comment":"The reasoning-dropout recipe is confounded with generic dropout regularization. The thinking-token condition (E.5.2) inserts filler tokens without applying the same stochastic dropout schedule, and there is no control that drops an equivalent number of non-semantic tokens (e.g., random text tokens or image patches) on the same schedule. Because the LIBERO training set is only 3,917 trajectories, stochastic masking of context tokens could plausibly produce the observed +7.4-point average gain without any reasoning-specific representation change. This missing control leaves the abstract's central claim—that learning to generate reasonings improves VLA representations—underdetermined.","section":"Sec. 5 (Reasoning Dropout), E.3, Table 1"},{"comment":"Reasoning pre-training is not compute-matched to the standard VLA baseline: it uses 100k steps of reasoning pre-training followed by 200k steps of action tuning (300k total) versus 200k for the standard VLA. The +5.4% average improvement could therefore be due to additional gradient steps rather than to reasoning-specific representation learning. The co-training condition uses 200k steps with a doubled batch size, which changes the number of optimizer updates per sample. A compute-matched baseline (e.g., training a standard VLA for 300k steps) or a learning-curve analysis is needed to attribute the gain to the reasoning content.","section":"Sec. 6.1, E.1, Table 1"},{"comment":"The text says that both ECoT-Lite approaches \"improve significantly\" over the standard VLA on Bridge, but the reported standard errors do not support this for reasoning dropout: the aggregate is 60.4% ± 4.6% versus 50.5% ± 4.7% for the VLA baseline, a difference of about 1.5 combined standard errors. Additionally, the aggregate hides large per-task reversals (e.g., reasoning pre-training at 37.5% vs. 87.5% for a spatial task). The authors should either report a proper significance test with per-task paired comparisons or soften the significance claim.","section":"Sec. 6.2, Table 2"},{"comment":"The paper's learning-dynamics explanation for why pre-training outperforms co-training is explicitly labeled as \"intuitive speculation\" in the Limitations section, and the supporting illustration in Figure 7 is an abstract loss-landscape cartoon. Since the paper's central scientific contribution is the claim that reasoning changes the policy's representations, downstream success alone is only indirect evidence. Adding a direct representation probe (e.g., linear probing for object or subtask features in the intermediate layers after different trainings, or measuring representational similarity) would provide much stronger support for the mechanism and would be within the paper's scope.","section":"Sec. 9, A.1, Sec. 6.3"}],"minor_comments":[{"comment":"Typo: \"such as Molmo's synthetic reasonings are much ore verbose\" should be \"much more verbose.\"","section":"C.1"},{"comment":"In the Semantic Gen. row, \"Put the screw in the bowel\" should be \"bowl.\"","section":"Table 2"},{"comment":"The phrase \"Disabling reasonings does not see to affect the spatial generalization tasks\" should be \"does not seem to affect.\"","section":"A.2"},{"comment":"The abstract's claim of outperforming conventional VLAs on BridgeData V2 \"by 10-19%\" should be qualified with the reported standard errors, since the lower end of that range is within one standard error of the baseline for the dropout variant.","section":"Introduction / Abstract"},{"comment":"The description of how the reasoning dropout policy can be run with reasoning \"turned on\" at test time is useful, but the paper does not report the performance of the dropout-trained policy when reasoning is actually enabled at test time on LIBERO; adding this would directly quantify the value of test-time reasoning within the same model.","section":"E.3"}],"recommendation":"major_revision","confidential_remarks":"The empirical evaluation is extensive and the two proposed recipes are likely usable in practice, but the paper's mechanistic claim, which is the stated main scientific contribution, is not yet supported because of the missing non-semantic dropout control and the compute mismatch in pre-training. I do not believe this is fatally flawed; the missing controls are feasible to run and the authors could also soften the abstract if experiments are not possible. The Bridge significance claim should be corrected regardless. I would ask for a revised version with those additions or revisions before considering publication in a major venue."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two takeaways. First, the ECoT-Lite recipes are genuinely useful: reasoning dropout gives you most of ECoT's generalization at standard VLA inference speed, and reasoning pre-training is a solid alternative when you have unpaired reasoning data. The simulated evaluation is thorough — 121,500 trials with shared seeds and standard errors — and the thinking-token negative result is a good counterpoint to the 'extra compute' explanation. The practical recipes, at least in LIBERO, work.\n\nSecond, the paper's central claim — that semantic content of reasoning, not just extra tokens or stochastic masking, drives the gains — is plausible but not proven. The authors don't run the obvious control: reasoning dropout with non-semantic tokens dropped at the same rate. With ~3.9k LIBERO trajectories, ordinary regularization from random context dropout could explain part of the +7.4 point gain. Thinking tokens don't apply dropout, so they can't rule that out. Pre-training helps, which supports the representation story, so this is a gap in the mechanism claim rather than a disproof.\n\nA bigger issue I noticed: the Bridge results labeled 'Reasoning Dropout' don't actually use the dropout training recipe. Per Appendix E.3, they use the released ECoT policy from Zawalski et al. and simply omit reasoning at inference. That's a fine probe of test-time reasoning, but it doesn't validate the training recipe on real robots. The real-world validation of dropout rests on LIBERO alone. The compute mismatch for pre-training (300k vs 200k steps) is a lesser-but-real concern; at least the authors disclose it.\n\nWho this is for: anyone training VLAs or working on robot reasoning. The recipes are easy to adopt, and the paper's guidance on when to use which variant is clear. The mechanism section is hypothesis-generating, not conclusive, and should be read that way.\n\nRecommendation: send it to peer review. A serious referee should push for the non-semantic dropout control and for a properly trained dropout policy in the real-world evaluation. With those, the causal story would be much stronger. Even without them, the empirical recipes and the LIBERO results justify publication, so this is a clear conditional accept rather than a reject.","headline":"Useful recipes and solid LIBERO ablations; the mechanism claim is under-tested and the Bridge dropout numbers come from a model not trained with the dropout recipe.","tokens_in":28719,"tokens_out":4700,"would_cite":true,"duration_ms":47831,"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":"Robot reasoning pays off in training, not at slow inference.","keywords":["embodied chain-of-thought reasoning","vision-language-action models","representation learning","reasoning dropout","reasoning pre-training","robot policy generalization","LIBERO-90","real-robot manipulation"],"falsifier":"Train the same architecture with the reasoning-dropout protocol but replace the dropped reasoning steps with an equal number of random text tokens at matched sequence length. If the perturbed-split success rates stay near 70-76%, the semantic content of reasoning is not load-bearing and the effect is generic regularization; if performance drops back toward the standard-VLA baseline, the reasoning content itself is doing the work.","tokens_in":27781,"feed_emoji":"🤖","tokens_out":5543,"duration_ms":54730,"temperature":0.7,"pith_summary":"The paper asks why chain-of-thought reasoning improves vision-language-action (VLA) robot policies and tests three candidate mechanisms: better representations, a learning curriculum, and extra compute. Its answer is that the main benefit is representational: training a VLA to generate reasoning traces such as plans, bounding boxes, and motion rationales reshapes its internal features, and the policy exploits those features best when actions are trained to attend to them. The paper introduces two lightweight recipes, reasoning pre-training and reasoning dropout, that capture most of this benefit without generating any reasoning at test time. If right, robot policies can gain the generalization advantages of embodied reasoning at the same speed as a standard VLA, a roughly threefold inference speedup over full chain-of-thought control.","feed_headline":"Robot reasoning pays off in training, not at slow inference","feed_subtitle":"Two fast VLA policies that skip test-time reasoning match or beat full chain-of-thought on LIBERO-90.","key_machinery":"The central machinery is a pair of training-time interventions that decouple learning to reason from reasoning at inference: reasoning pre-training, which supervises the VLM on embodied reasoning traces alone and then fine-tunes on actions, and reasoning dropout, which trains the VLA to predict actions while randomly masking or keeping the reasoning tokens in context. Combined with the embodied reasoning annotations themselves (plans, subtask rationales, bounding boxes, gripper positions, and movement rationales), these recipes force the action head to rely on reasoning-shaped representations during training while removing the need to decode reasonings during deployment.","core_discovery":"The paper's central claim is that learning to generate reasonings leads to better VLA representations, while attending to the reasonings during training is what lets the policy leverage those features for improved action prediction. Experimentally, reasoning dropout and reasoning pre-training outperform a standard VLA on LIBERO-90 (89.4% and 87.1% versus 82.0% on the standard split) and on real-robot Bridge tasks (69.4% versus 50.5% for pre-training), matching or approaching full embodied chain-of-thought while running at the faster speed of a plain VLA. The paper finds only weak support for the learning-curriculum hypothesis and no support for the expressivity hypothesis, since adding non-semantic thinking tokens actually hurts performance.","pith_inferences":["If the mechanism is semantic representation learning, then scaling the diversity of reasoning data should widen performance gaps between the recipes and may make test-time reasoning increasingly valuable, a trend the paper's own LIBERO-versus-Bridge comparison already hints at.","A direct test of the mechanism would compare reasoning dropout against dropping matched-length random text tokens; if the gains are equivalent, the benefit is generic regularization rather than reasoning content.","The null result for thinking tokens suggests that extra inference compute without a meaningful training signal will not transfer to manipulation, and a further control could insert semantically meaningful but task-irrelevant tokens to isolate content from mere sequence length.","Reasoning pre-training may generalize as a recipe for adapting any VLM before action fine-tuning, effectively extending domain-adaptive pretraining to robot control."],"forward_implications":["Policies trained with reasoning pre-training or dropout can be deployed at the control rate of a standard VLA, not the 1-1.2 Hz of full embodied CoT, while matching or beating it on LIBERO-90.","Reasoning dropout reaches state-of-the-art LIBERO-90 performance (89.4%) without generating any test-time reasoning, indicating that in narrow benchmarks the reasoning traces can be fully internalized.","Reasoning pre-training lifts real-robot Bridge success from 50.5% to 69.4%, and because it does not need paired reasoning-action data, it can in principle learn from unpaired reasoning data from other embodiments.","The paper's practical prescription follows from its ablations: use full ECoT to maximize performance, reasoning dropout for narrow task domains, and reasoning pre-training for diverse domains or when unpaired reasoning data are available."],"supporting_citations":[{"why":"Supplies the embodied chain-of-thought formulation, the Bridge reasoning annotation dataset, and the training pipeline that the paper's recipes modify.","marker":"[14]"},{"why":"Supplies the MiniVLA backbone and VQ-VAE action-chunking scheme used for all LIBERO-90 experiments.","marker":"[31]"},{"why":"Supplies the OpenVLA backbone used for the real-robot Bridge evaluations.","marker":"[6]"},{"why":"Provides the LIBERO-90 benchmark and its 50-rollout-per-task demonstration dataset used for training and evaluation.","marker":"[20]"},{"why":"Provides the BridgeData V2 dataset and evaluation splits used for real-robot generalization tests.","marker":"[8]"},{"why":"Sets the previous state-of-the-art LIBERO-90 success rate that the paper's best policies are compared against.","marker":"[69]"},{"why":"Supplies the thinking-token approach used to implement and test the expressivity hypothesis.","marker":"[64]"},{"why":"Supports the domain-adaptive pre-training argument invoked to explain why reasoning pre-training improves representations before action fine-tuning.","marker":"[66]"}],"fun_headline_variants":["Reasoning helps robot learning, not inference","Skip test-time reasoning, keep training gains","Fast robot policies from training-time reasoning","Training reasoning beats test-time reasoning for robots","Two fast robot policies that use reasoning only in training"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The conclusion that reasoning improves policies through semantic representation learning rests on the untested assumption that the gains of reasoning dropout come from the meaning of the reasoning tokens rather than from the generic regularization effect of randomly dropping tokens during training.","fun_headline_variants_meta":{"raw":{"variants":["Reasoning helps robot learning, not inference","Skip test-time reasoning, keep training gains","Fast robot policies from training-time reasoning","Training reasoning beats test-time reasoning for robots","Two fast robot policies that use reasoning only in training"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000149,"raw_usage":{"total_tokens":1179,"prompt_tokens":917,"completion_tokens":262,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":533,"completion_tokens_details":{"reasoning_tokens":195}},"tokens_in":533,"tokens_out":262,"duration_ms":3221,"temperature":1.0,"reasoning_tokens":195,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:58:52.098846+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same architecture with the reasoning-dropout protocol but replace the dropped reasoning steps with an equal number of random text tokens at matched sequence length. If the perturbed-split success rates stay near 70-76%, the semantic content of reasoning is not load-bearing and the effect is generic regularization; if performance drops back toward the standard-VLA baseline, the reasoning content itself is doing the work.","supporting_citations":[{"cited_title":"Zawalski, W","cited_arxiv_id":null,"evidence_quote":"Supplies the embodied chain-of-thought formulation, the Bridge reasoning annotation dataset, and the training pipeline that the paper's recipes modify."},{"cited_title":"Belkhale and D","cited_arxiv_id":null,"evidence_quote":"Supplies the MiniVLA backbone and VQ-VAE action-chunking scheme used for all LIBERO-90 experiments."},{"cited_title":"Co-training Improves Prompt-based Learning for Large Language Models","cited_arxiv_id":"2202.00828","evidence_quote":"Sets the previous state-of-the-art LIBERO-90 success rate that the paper's best policies are compared against."}],"review_version":1}