{"id":"dbfecd80-f0dc-46ae-aa35-17ce7b5f4631","arxiv_id":"2412.03293","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A robot policy generates its own language reasoning before acting and injects it into a diffusion action decoder, outperforming several VLA baselines on real-robot manipulation.","lead":"DiffusionVLA couples a large vision-language model's reasoning with a diffusion model for robot actions, using the model's own text rationales to condition the action decoder. The authors report strong real-robot results on sorting, bin-picking, and bimanual tasks, including generalization to unseen objects.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central ablation does not isolate the FiLM reasoning-injection mechanism; the 33.3-point gap in Table 8 could come from added language supervision or pretraining confounds, so the paper's core causal claim is underdetermined.","rationale":"The reader's weakest_assumption (pretraining fairness) is real and load-bearing, but I see an even more direct threat to the paper's central claim. The named contribution is not merely 'a diffusion VLA with a large VLM'; it is the reasoning-injection module. The only experiment that targets this module is Table 8, and that ablation is under-specified: it removes the module rather than the FiLM pathway alone, and it does not say whether reasoning tokens and the NTP loss remain. Without that control, the observed drop could be caused by removing the language-supervision objective or by the general presence of reasoning text, neither of which is the proposed mechanism. This matters because the interpretability and self-correction claims in Section 4.4 are also attributed to this module. The comparison-level confounds the reader identified reinforce the concern: the DiVLA-versus-baseline gaps vary pretraining data, backbone, and action decoder simultaneously, so they cannot substitute for a clean ablation. I therefore keep the CONDITIONAL verdict: the architecture is plausible and the reported gains are large, but the causal role of the central innovation needs a controlled experiment and ideally matched-pretraining comparisons before the claim is accepted.","tokens_in":19212,"tokens_out":9191,"duration_ms":95818,"concrete_test":"Run a three-arm ablation on the same 77-trial multi-task benchmark: (A) full DiVLA; (B) identical model and training objective, but with the FiLM parameters for the reasoning embedding removed or zeroed so reasoning tokens are still autoregressed and Lntp is still optimized, while the diffusion policy never receives them; (C) no reasoning tokens and no Lntp term. If A and B are statistically indistinguishable, the 33.3-point gap in Table 8 is not caused by the injection module; if A clearly beats B, the mechanism is supported. Report per-arm trial counts and binomial confidence intervals.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's core novelty is the reasoning-injection module (Section 3.1), and the main evidence for it is Table 8: removing it drops multi-task success from 83.6% to 50.3%. But the text does not state what 'w/o reasoning injection' retains. If the ablated model also drops the autoregressive reasoning tokens or the Lntp term, the comparison conflates three separate changes: (i) the FiLM conditioning of the diffusion policy, (ii) the extra language-supervision objective, and (iii) the presence of reasoning text as a target. The central claim that injecting self-generated reasoning into the policy is what drives generalization is therefore not isolated by the reported experiment. The other comparisons do not repair this: Table 1 and Figure 4 vary pretraining data (Droid 39K vs OXE 970K), backbone, model size, and action decoder simultaneously, so the gap versus OpenVLA and Octo cannot be attributed to the proposed mechanism. Interpretability claims in Section 4.4 also rest on the same injection mechanism without a control showing that the reasoning text is causally used rather than post-hoc correlated. This is addressable, but as reported the causal contribution of the named innovation is underdetermined.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Diffusion-VLA (DiVLA), a vision-language-action model that combines a pretrained autoregressive VLM (Qwen2-VL) with a diffusion policy head for action generation. The model is trained with a diffusion loss plus a next-token prediction loss so that, given an instruction and observation, it first emits free-form reasoning tokens and then generates actions. A reasoning injection module uses FiLM to condition the diffusion policy on the embedding of the generated reasoning phrases. The authors evaluate DiVLA on real robots across multi-task learning, factory sorting, zero-shot bin picking on 102 unseen objects, bimanual table bussing, novel instruction following, and model scaling from 2B to 72B parameters. The central claims are that self-generated reasoning improves generalization and interpretability, that the reasoning injection module is responsible for a 33.3-point performance gap, and that DiVLA outperforms OpenVLA, Octo, TinyVLA, and Diffusion Policy on the reported tasks while running at 82 Hz on an A6000 GPU.","tokens_in":19484,"tokens_out":3027,"duration_ms":31308,"significance":"If the empirical claims hold, the paper makes a useful architectural contribution: it shows a concrete way to give diffusion policies a chain-of-thought-like reasoning channel without requiring iterative re-prompting, and it provides real-robot evidence for data-efficient training and fast inference. The strengths include genuine real-robot evaluations across multiple embodiments, a large zero-shot bin-picking test set, explicit scaling experiments from 2B to 72B, and an honest discussion of limitations such as sensitivity to low-bit quantization and imperfect VQA behavior. The interpretability examples in Section 4.4 are visually compelling. However, the causal contribution of the named reasoning-injection mechanism is not isolated by the reported ablation, and the headline comparisons with OpenVLA and Octo are confounded by differences in pretraining data, model size, and backbone. The very small trial counts in several evaluations further weaken the quantitative claims. The paper is promising, but the central causal claim and the comparative superiority claims need additional experiments or re-analysis before they can be accepted as stated.","major_comments":[{"comment":"The ablation labeled 'w/o reasoning injection' does not specify which components are removed. If removing the injection module also removes the autoregressive reasoning tokens and the alpha*L_ntp term, then the reported 33.3-point drop conflates (i) the FiLM-based conditioning, (ii) the extra language-supervision objective, and (iii) the presence of reasoning text as a training target. To support the paper's core claim that injecting self-generated reasoning into the policy is what drives the improvement, the authors should report at least two additional conditions: one that keeps the reasoning tokens and L_ntp loss but removes only the FiLM conditioning, and one that keeps FiLM conditioning but removes the reasoning content (e.g., injecting a generic text embedding). As written, the central causal claim is underdetermined by the reported experiment.","section":"Section 8.4, Table 8"},{"comment":"The main comparisons against OpenVLA and Octo do not control for pretraining data scale or model size. DiVLA-2B is pretrained on Droid (39K trajectories), while OpenVLA and Octo are pretrained on OXE (970K trajectories), and DiVLA-2B is 2B parameters while OpenVLA is 7B. The paper acknowledges this asymmetry (Section 4.2: 'OXE ... is 25 times larger than our pre-trained datasets') but still interprets the performance gaps, including the 63.7% versus 28.4% bin-picking result, as evidence of architectural superiority. The gaps could be due to the different pretraining data, the different backbones, or the fine-tuning protocol rather than the proposed reasoning mechanism. At minimum, the authors should provide an ablation in which DiVLA without the reasoning injection is trained and evaluated under the same protocol, and ideally a comparison in which OpenVLA or Octo is fine-tuned on the same data budget or DiVLA is pretrained on the same OXE data.","section":"Section 4.2, Table 1, and Section 4.5, Figure 4"},{"comment":"Several critical evaluations use very small trial counts. For the visual generalization setting, each task has only 9 trials, so a difference of one success changes the reported rate by 11.1 percentage points. For instance, in Task 1 of the visual generalization rows, DiVLA-2B succeeds 4/9 and OpenVLA succeeds 5/9, yet the text in Section 4.2 states that 'our method consistently maintains the highest average success rate' across tasks. No confidence intervals or significance tests are reported anywhere. The main multi-task per-task counts (11 trials for most tasks) also make the differences between DiVLA and baselines statistically fragile. The authors should provide confidence intervals or exact binomial tests, and ideally increase the trial counts for the claims that are central to the paper.","section":"Section 6.1, Table 4"},{"comment":"The interpretability and self-correction claims rely on a few selected examples rather than a systematic evaluation. The claim that 'the integrated reasoning module allows for a form of self-correction' is based on a single illustrative intervention (replacing a toy car with a hex key) with no quantification of how often the reasoning changes correctly, how often the action follows the reasoning, or whether a model without the reasoning injection would exhibit different behavior under the same intervention. To substantiate the interpretability claim, the authors should provide a protocol that measures the agreement between the generated reasoning and the executed action over multiple trials, and compare against a baseline that produces plausible-looking rationales after the fact.","section":"Section 4.4 and Figure 6"}],"minor_comments":[{"comment":"The paper uses both 'DiffusionVLA' and 'DiVLA' for the same method; please standardize the name.","section":"Throughout"},{"comment":"The sentence 'The original Droid data contains only robotic actions, paired partially with observations and language instructions' is repeated verbatim in consecutive sentences; one copy should be removed.","section":"Section 3.2, Pretraining Data"},{"comment":"The heading contains a typo: 'Implementation details and pretraiend data' should be 'pretrained data'.","section":"Section 4.1"},{"comment":"The column header 'Trails' should be 'Trials'.","section":"Table 4"},{"comment":"The sentence 'Notice that both Octo and OpenVLA is pre-trained on OXE' has a subject-verb agreement error; it should be 'are pre-trained'.","section":"Section 4.2"},{"comment":"The inference-speed comparison omits hardware details for OpenVLA beyond 'A6000 GPU'; please specify whether the same batching, sequence length, and precision settings were used for all methods.","section":"Section 8.2, Table 5"}],"recommendation":"major_revision","confidential_remarks":"The paper is conditionally plausible and the real-robot evaluation is a strength, but the central causal claim about the reasoning injection module is not isolated, and the headline comparisons against OpenVLA/Octo are confounded. The authors should be asked to provide a controlled ablation (reasoning tokens with and without FiLM injection, and injection without reasoning content) and to report confidence intervals or more trials for the per-task results. If the authors can supply these, the paper could become acceptable; without them, the claims outrun the evidence. The novelty disclosure appears adequate, and the fit with the journal's scope is good."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the paper has a plausible and reasonably novel combination—an autoregressive VLM that emits reasoning tokens, a diffusion action head, and a FiLM layer that injects the reasoning embedding into the policy. Real-robot results are impressive on their face: 83.6% multi-task vs 45.5% best baseline, 63.7% zero-shot bin-picking on 102 unseen objects, and scaling from 2B to 72B works. The writing is clear and the related work is handled honestly, including the unified AR+diffusion models and pi0.\n\nWhat's actually new here is narrow but real: nobody else, as far as I know, has coupled self-generated reasoning text into a diffusion action decoder via FiLM and shown gains on real robots. That is a legitimate contribution.\n\nThe soft spots are in the evidence, not the idea. The main ablation in Table 8 removes 'reasoning injection' but doesn't say whether the model still gets the NTP loss and the reasoning tokens as targets. If those go too, the 33-point gap could come from losing language supervision rather than from the FiLM conditioning per se. The stress-test is right that the central causal claim is underdetermined as reported. Also, the zero-shot bin-picking comparison varies pretraining data (Droid 39K vs OXE 970K), backbone, model size, and action decoder all at once. The baselines had more pretraining data, which should favor them, but the point stands that you can't attribute the win to the injection mechanism from that comparison alone. Trial counts are small—9 trials per visual generalization task—and there are no error bars or significance tests. The abstract's claim about training from scratch on fewer than 50 demonstrations is not backed by a dedicated experiment; it looks like Task 3 has 50 demos in the multi-task set, which is not the same thing.\n\nNone of this is fatal. The architecture is well motivated, the results are promising, and the issues are addressable with a cleaner ablation (keep the NTP loss, remove only the FiLM injection), more trials, and a matched-pretraining comparison or at least a discussion of the confound. I'd send this to a serious referee. The paper deserves a chance to tighten the evidence, and the community would benefit from either a confirmed or a refuted causal claim here.","headline":"Plausible architecture, promising real-robot results, but the central ablation doesn't isolate the reasoning-injection mechanism; deserves a referee who will demand a cleaner experiment.","tokens_in":20049,"tokens_out":3616,"would_cite":false,"duration_ms":33311,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A vision-language-action model that generates its own reasoning and injects it into a diffusion action decoder outperforms prior robot policies and becomes interpretable.","keywords":["vision-language-action model","robot foundation model","diffusion policy","autoregressive reasoning","reasoning injection","real-robot manipulation","generalization","interpretability"],"falsifier":"Run the same real-robot tasks with all policies pretrained on the same data and matched for camera views, training epochs, and fine-tuning budget, and run an ablation that varies only the reasoning text while keeping the architecture fixed; if a diffusion policy without injected reasoning matches DiVLA on unseen-object bin picking, or if random reasoning phrases produce the same success as task-relevant ones, the paper's central claim is refuted.","tokens_in":19022,"feed_emoji":"🤖","tokens_out":11538,"duration_ms":103491,"temperature":0.7,"pith_summary":"Diffusion-VLA (DiVLA) is a robot policy that lets a vision-language model first write a short reasoning phrase about the scene and task, then feeds that phrase into a diffusion-based action generator that outputs continuous robot commands. The paper's central claim is that this two-part design, with the reasoning embedded directly into the policy, gives better real-world manipulation performance than both pure next-token vision-language-action models and pure diffusion policies, while making the robot's decisions readable. On a five-task multi-task benchmark on a single-arm robot, the 2B model averages 83.6% success versus 45.5% for the best baseline, and it picks 63.7% of 102 never-seen objects in a zero-shot bin-picking test versus 28.4% for the strongest baseline. The authors also report robustness to distractors, backgrounds, and camera shifts, fast adaptation to a bimanual robot, and 82 Hz inference on a single GPU. If these results hold under controlled comparison, they point to a practical recipe for combining language reasoning with high-frequency continuous action generation in robot foundation models.","feed_headline":"Reasoning module lifts robot task success to 83.6%","feed_subtitle":"The same model explains its decisions, runs at 82 Hz, and beats larger baselines on real robots.","key_machinery":"The load-bearing mechanism is the reasoning injection module. In a forward pass, the vision-language model emits a short sequence of reasoning tokens; the final embedding from that reasoning output is used as a Feature-wise Linear Modulation (FiLM) conditioning vector, scaling and shifting the activations of the diffusion policy's layers so that the explicit rationale directly modulates action generation. FiLM is a conditioning layer that applies an affine transformation to features based on a conditioning vector. Because the reasoning is generated inside the same pass, DiVLA avoids the recursive generate-then-feed-back loop used by chain-of-thought policies. The diffusion head follows the standard Diffusion Policy architecture, with a two-layer MLP projection aligning vision-language embeddings to the action decoder, and training combines a diffusion loss with a next-token-prediction loss at weight ratio alpha=10; the vision-language backbone is fine-tuned with LoRA while the vision encoder and base weights are frozen. For pretraining, an external language model rewrites raw robot trajectories into reasoning-augmented form.","core_discovery":"The authors' claim is that autoregressive reasoning and diffusion action generation solve each other's weaknesses: the vision-language head decomposes the user's query into a short rationale, and the diffusion head generates continuous joint-space commands, while a FiLM-based reasoning injection module makes the rationale directly modulate the policy. They report that this design reaches 83.6% average success on five real-robot multi-task skills, 66.2% average on a four-category factory-sorting task with seen and unseen objects, 63.7% zero-shot success on 102 unseen bin-picking objects, and 72.9% and 70.8% on bimanual table bussing for seen and mixed objects. Removing the reasoning injection drops the multi-task average from 83.6% to 50.3%. The same model retains visual-question-answering ability, follows multi-step instructions, and runs at 82 Hz for the smallest variant.","pith_inferences":["The FiLM injection path is not tied to diffusion: the paper does not test it on flow-matching or one-step distillation action heads, which could trade some reported robustness for higher control frequency.","A causal test the paper does not run is to sample random reasoning phrases at inference and check whether actions change; the mid-task hex-key swap suggests, but does not prove, that the rationale drives the action.","If reasoning quality is the active ingredient, then longer or multi-step rationales generated at inference could further improve novel-object and multi-step instruction following, since the architecture already supports autoregressive generation of arbitrary text.","The pretraining-data mismatch is the main threat to interpreting the reported gains as architectural; a matched-data replication might shrink the raw gaps while still leaving interpretability and speed as separable advantages."],"forward_implications":["A policy can be both conversational and control a robot: DiVLA retains visual-question-answering ability even without co-training on vision-language chat data.","Generalization to new objects can arise by analogy: DiVLA labels a screwdriver as a hex key and a Sprite can as a green can, then sorts them correctly.","Adding a new robot embodiment reduces to training a new MLP action head, so pretrained knowledge is reused rather than rebuilt.","Scaling the same recipe from 2B to 72B improves sorting from 66.2% to 82.4% and bin picking from 63.7% to 75.9%.","The reasoning injection itself is essential: removing it lowers the five-task average from 83.6% to 50.3%."],"supporting_citations":[{"why":"Supplies the diffusion policy action decoder that DiVLA reuses and the Diffusion Policy baseline it must beat.","marker":"Chi et al., 2023"},{"why":"The main autoregressive vision-language-action baseline and comparison point for speed, instruction following, and bin picking.","marker":"Kim et al."},{"why":"Provides the FiLM conditioning mechanism used in the reasoning injection module.","marker":"Perez et al., 2018"},{"why":"Droid is the pretraining dataset for DiVLA-2B and DiVLA-7B, the source of the reasoning-augmented action data.","marker":"Khazatsky et al., 2024"},{"why":"OXE is the much larger pretraining dataset used by Octo and OpenVLA baselines and, combined with Droid, for DiVLA-72B; it defines the fairness baseline.","marker":"O'Neill et al., 2023"},{"why":"TinyVLA is the diffusion-based vision-language-action baseline closest in spirit to DiVLA, providing a direct comparison.","marker":"Wen et al., 2024"},{"why":"Qwen2-VL is the pretrained vision-language backbone that supplies autoregressive reasoning and visual understanding.","marker":"Wang et al., 2024b"},{"why":"LoRA is the fine-tuning method used to update the vision-language backbone during robot-data training.","marker":"Hu et al., 2021"},{"why":"Octo is a diffusion-based generalist baseline whose separate action-decoder design is contrasted with DiVLA's embodiment adaptation.","marker":"Octo Model Team et al., 2024"}],"fun_headline_variants":["Self-reasoning robot policy: 83.6% success, 82 Hz, explainable","Robots that reason while acting hit 83.6% on real tasks","DiffusionVLA: reasoning + diffusion reaches 83.6% task success","Reasoning injection boosts robot success to 83.6% from 50.3%","83.6% robot success with a model that explains its decisions"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the comparisons are fair: in the paper's Section 4.1, DiVLA is pretrained on a 39K-trajectory dataset while Octo and OpenVLA are pretrained on a 970K-trajectory dataset, and the zero-shot bin-picking evaluation in Figure 4 does not control for that difference.","fun_headline_variants_meta":{"raw":{"variants":["Self-reasoning robot policy: 83.6% success, 82 Hz, explainable","Robots that reason while acting hit 83.6% on real tasks","DiffusionVLA: reasoning + diffusion reaches 83.6% task success","Reasoning injection boosts robot success to 83.6% from 50.3%","83.6% robot success with a model that explains its decisions"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000251,"raw_usage":{"total_tokens":1606,"prompt_tokens":1040,"completion_tokens":566,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":656,"completion_tokens_details":{"reasoning_tokens":458}},"tokens_in":656,"tokens_out":566,"duration_ms":5214,"temperature":1.0,"reasoning_tokens":458,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T22:32:56.103386+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same real-robot tasks with all policies pretrained on the same data and matched for camera views, training epochs, and fine-tuning budget, and run an ablation that varies only the reasoning text while keeping the architecture fixed; if a diffusion policy without injected reasoning matches DiVLA on unseen-object bin picking, or if random reasoning phrases produce the same success as task-relevant ones, the paper's central claim is refuted.","supporting_citations":[],"review_version":1}