{"id":"2ff7a31d-ee1d-4b02-9b2f-6ab4ecb8f28b","arxiv_id":"2608.05738","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"VLA-Talker injects externally grounded spatial evidence into a VLA prompt, supervises only actions, and beats CoT baselines on manipulation success, latency, and data efficiency.","lead":"This paper argues that robot policies should read grounded spatial language produced by external perception tools instead of generating their own reasoning text. The proposed system, VLA-Talker, injects detector and depth evidence into a vision-language-action model's prompt, supervises only action tokens, and reports better success rates, lower latency, and better data efficiency than chain-of-thought baselines on simulation benchmarks and real robot tasks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline comparison confounds 'generating language' with 'lacking external perception': Gen-CoT must infer object locations from pixels at test time while VLA-Talker receives tool-derived evidence, so the success gap may reflect information asymmetry rather than the consume-vs-generate principle.","rationale":"The paper's central claim is that a VLA needs to consume grounded language rather than generate it. The strongest empirical support for this is the matched comparison against Gen-CoT and Table 3. But the match is only in training labels: at inference, VLA-Talker is given object centroids, depths, and gripper projection from an external tool cascade, while Gen-CoT and BC must recover those facts from pixels. This is a textbook confound: the success gap could come entirely from test-time access to near-ground-truth perception. The reader's weakest_assumption focused on tool-loop accuracy, but that is a robustness concern the paper partially addresses with fallback cascades and noise studies. The confound is more load-bearing because it questions whether the reported experiments can distinguish 'consuming vs generating' at all. I still credit the paper's strong points: the in-context supervision mask is cleanly ablated in (b) vs (c) holding injected evidence fixed, the paraphrase-rendering design is thoughtful, and the real-robot multi-task numbers are suggestive. However, the central design principle is not established without a tool-augmented generation baseline. A conditional acceptance requiring that additional comparison is the appropriate disposition.","tokens_in":24311,"tokens_out":10887,"duration_ms":113525,"concrete_test":"On LIBERO, add a tool-augmented CoT baseline that at inference receives the same evidence tuple c_t (gripper projection, object centroids, depth) as an input prefix before generating its rationale and actions, trained with the same two-stage recipe; compare success and control frequency to VLA-Talker. If this baseline's success rises to within error bars of VLA-Talker, the reported CoT deficit is explained by test-time perception access rather than by consumption-vs-generation. If it still trails substantially, the generation objective itself is implicated.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Gen-CoT is called 'matched evidence' because its training targets are the tool loop's tuples, but at inference it must generate those tuples from the RGB image; VLA-Talker receives the tuples as externally computed context. The main comparison therefore varies two factors simultaneously: (i) test-time access to near-ground-truth spatial evidence, and (ii) whether language is generated or injected. Since BC and Gen-CoT must infer object centroids and depth from pixels, while VLA-Talker is handed them by GroundingDino/DepthAnything/gripper projection, the observed success gaps (e.g., 97.4 vs 96.2 on LIBERO, 72.4 vs 54.7 on SimplerEnv) are as consistent with 'better perception at inference' as with 'consuming beats generating.' Table 3(a) vs 3(c) has the same confound, plus a 4.6x control-rate difference. Table 12 confirms that removing the tool loop drops VLA-Talker to 84.3%, below plain BC (90.4%), so the entire gain is attributable to the external evidence, not to language-consumption behavior per se. A baseline that is given the same tool outputs at inference and still generates a rationale is needed to attribute the gains.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper argues that a vision-language-action model needs the ability to consume grounded language rather than to generate rationales. VLA-Talker injects tool-derived spatial evidence (open-vocabulary detection, monocular depth, analytical gripper projection, VLM fallback) as read-only <spatial> context at keyframes, supervises only action tokens during an in-context post-training stage, renders evidence in diverse paraphrases through a round-trip-filtered data engine, and then applies GRPO with sparse success rewards. The paper reports success rates on LIBERO, RoboCasa-GR1, SimplerEnv, and eight real-robot AgiBot tasks, comparing against a matched-evidence generative CoT baseline (Gen-CoT) and published VLA baselines, and claims that free-form CoT degrades low-level control through a grounding gap, objective interference, and inference latency.","tokens_in":24622,"tokens_out":26742,"duration_ms":248826,"significance":"If established cleanly, the central claim would be a valuable contribution: it decouples evidence acquisition from evidence consumption, makes the falsifiable prediction that action-only supervision on injected evidence beats rationale generation under equal evidence, and demonstrates a large inference-cost advantage (12.8 Hz vs 2.8 Hz, Table 10). The paper has genuine strengths: Table 3(b) versus (c) is a confound-free ablation of the supervision mask under identical test-time injection; the round-trip consistency filter with explicit pixel/depth tolerances (Appendix A) is a machine-checkable safeguard; Appendix E provides three-seed statistics with a Welch test; Appendices G and H give hyperparameter sensitivity and synthetic perception-noise stress tests; and Appendix N states limitations honestly, including the premise that tool outputs must be trustworthy, which the cascade robustness studies (Table 14, Fig. 13) partially mitigate.","major_comments":[{"comment":"Table 1 reports Gen-CoT at 96.2% LIBERO average, while Table 3(a) and Fig. 4 report the same 'matched-evidence generate-and-supervise-text' scheme at 81.5%, and Fig. 1's motivational bars (CoT ≈ 0.88) match neither value. Table 6 adds a third number for the same method at the full data budget (87.6% at 50 demonstrations), placing Gen-CoT below BC's 90.4%, whereas Table 1 places Gen-CoT (96.2%) well above BC (90.4%, Table 4). If these are different configurations (e.g., single-stage SFT versus the full two-stage recipe, or different per-task data budgets), the text must say so explicitly; as printed, the central claim's strength swings from a 1.2-point margin (Table 1) to a 15.9-point margin (Table 3), and no sentence in the paper allows a reader to reconcile the values. Please unify the definition of the matched-evidence baseline, state its configuration in every table where it appears, and re-derive the affected claims.","section":"Tables 1 and 3; Fig. 4"},{"comment":"The headline comparisons confound test-time evidence access with the generate-versus-inject contrast. Gen-CoT must produce the evidence tuples from the RGB image at inference (the 'grounding gap' defined in the method section), whereas VLA-Talker receives the same tuples from the tool loop on keyframes at inference ('Keyframe gating'). The comparison therefore varies two factors at once: whether the policy is handed near-ground-truth spatial evidence, and whether language is generated or injected. Table 3(a) versus (c) cannot attribute the 15.9-point gap to the consume-versus-generate principle, and Table 12's 'w/o tool loop (self-guess)' result (84.3%, below plain BC at 90.4%) shows the gains track the external evidence itself. A control that receives the identical tool-loop output at inference while still generating (or supervising) the rationale is needed to support the stated attribution; only Table 3(b) versus (c), which shares test-time injection and varies only the supervision mask, is free of this confound.","section":"Tables 1, 2, 3, 5; Table 12"},{"comment":"The method section specifies tool invocation in two incompatible ways. 'Keyframe gating' states that evidence is injected only on a fixed initial/gripper-change/periodic schedule, with 'the same schedule ... applied' at inference, so the tool loop runs independently of any policy decision. The RL section, however, defines rollouts containing policy-emitted tool calls (Eq. (4): C_k), rewards tool-call syntax through I_format in Eq. (5), asserts that 'the policy is free to discover when to call tools rather than being told,' and Fig. 10(b) reports that GRPO reduces average tool calls per episode from ~3.4 to ~1.8. If the keyframe schedule is fixed, the policy cannot choose to skip or add tool calls and the Fig. 10(b) result is unexplained; if the policy can emit tool requests, the interaction between those requests and the gating schedule must be specified, including how the 12.8 Hz inference cost (Table 10) accounts for tool-loop overhead. Please clarify the mechanism and make the claims consistent.","section":"Method: Keyframe gating vs. RL stage"},{"comment":"The abstract claims the paper shows 'both empirically and analytically' that free-form CoT degrades low-level control, but the analytical component is the token-count argument around Eq. (2): because language tokens outnumber action tokens, 'gradient mass is dominated by sounding reasonable.' Token-count imbalance alone does not establish gradient domination, since per-token cross-entropy terms are typically normalized, action-token losses can be large early in training, and either term can be masked or reweighted (as the paper itself does in Table 3). Table 3(b) versus (c) empirically supports the value of action-only supervision, so the paper should either provide measured per-stage gradient norms or withdraw the 'analytically' claim.","section":"Why Generative CoT Hurts Low-Level Control; Eq. (2)"},{"comment":"Appendix G states that all hyperparameters were selected by a coarse grid search on the LIBERO validation split and then frozen across benchmarks, and Fig. 12 selects the GRPO defaults (β = 0.01, M = 16) on LIBERO as well; the headline LIBERO average of 97.4% is therefore tuned on the same suite on which it is presented as a new SOTA (the margin over VLA-Thinker in Table 1 is 0.4 points). Please either report LIBERO with a configuration chosen without access to LIBERO, or clearly separate development-split from held-out results. The main tables (1, 2, 5) also carry no seed variance; the three-seed analysis in Appendix E covers only BC, Gen-CoT, and VLA-Talker, not the published baselines in those tables.","section":"Appendix G; Table 1"}],"minor_comments":[{"comment":"The introductory bar chart gives no benchmark or configuration for its values (CoT ≈ 0.88, ICL ≈ 0.97), and neither number matches Tables 1 or 3; please label the source or remove the chart.","section":"Fig. 1"},{"comment":"Setting (b) is annotated as 0.9x latency (73 ms) and setting (c) as 1.0x (78 ms), implying that supervising the injected text yields a faster policy than action-only supervision; please explain this or correct the annotation.","section":"Fig. 4"},{"comment":"The paragraph contains a dangling cross-reference, 'the two-stage recipe of Section :', with the section number missing.","section":"Implementation paragraph"},{"comment":"The x-axis label renders as 'Injected pixel-localization noise ( , px)' with the σ symbol missing.","section":"Fig. 13"},{"comment":"The caption contains an editorial annotation ('[highlighting applied to family- and overall-average rows only]') that should be removed from the final caption.","section":"Table 13"},{"comment":"The real-robot table reports 20 trials per subtask with no variance or repeated-initialization information, and several margins (e.g., Pen single-task, 35 vs 15) are within plausible binomial sampling noise at this sample size; please add confidence intervals or more trials, and note that the single-task +CoT average (41.9) coincides exactly with the Baseline average (41.9).","section":"Table 8"},{"comment":"The phrase 'agentic tool-use interface, in which the policy queries ...' is not reflected in the described training and inference procedure, where the tool loop runs on a fixed schedule; please align the wording with the mechanism clarified in response to the major comments.","section":"Introduction"},{"comment":"The table reports 1,640 training episodes for LIBERO, which implies roughly 41 demonstrations per task under the standard 40-task split, a number that does not match the 50-demonstration budget used in Table 6; please state the exact per-task demonstration budget used for the main results.","section":"Table 16"},{"comment":"The abstract's claim of 'nearly 10%' improvement over standard behavior-cloning baselines is not tied to a specific table: on LIBERO the gain over the BC backbone is 7.0 points (Table 4), and no BC baseline is reported for RoboCasa-GR1 or SimplerEnv; please make the claim quantitative and traceable.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":"For the editor: I find no indication of misconduct; the numerical inconsistencies across tables most plausibly stem from unstated configuration differences (training stage, data budget) and should be resolved by requiring the authors to report the exact configuration of Gen-CoT in every table. The selection of hyperparameters on the LIBERO validation split is an additional reason to request a held-out evaluation before any acceptance. The citation style is heavy on same-year arXiv preprints, which is normal for this subfield, and the novelty relative to CoT-VLA, VLA-Thinker, and SpaceTools is discussed adequately."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague—\n\nThe paper has a genuinely useful idea and a real empirical confound. The useful idea: a VLA should be trained to consume externally grounded spatial evidence injected as read-only context, with supervision only on action tokens, rather than generate free-form rationales before acting. The diverse paraphrase rendering is a nice touch, and the keyframe-gating and tool-cascade robustness studies are done carefully. The real-robot results and data-efficiency curves give the framework practical weight.\n\nBut the headline comparison does not support the central claim. Gen-CoT is called 'matched evidence' because it is trained on the same tool-loop tuples, but at inference it must generate those tuples from the RGB image, while VLA-Talker is handed them by GroundingDino/DepthAnything/gripper projection. That varies two things at once: whether language is generated or injected, and whether the policy has test-time access to near-ground-truth spatial evidence. The success gaps (97.4 vs 96.2 on LIBERO, 72.4 vs 54.7 on SimplerEnv) are as consistent with 'better perception at inference' as with 'consuming beats generating.' Table 12 makes this concrete: removing the tool loop drops VLA-Talker to 84.3%, below plain BC at 90.4%. So the entire gain is attributable to the external evidence, not to the consumption behavior per se.\n\nThere is a clean comparison hiding in Table 3: (b) inject+supervise text vs (c) inject+action-only. There the evidence is matched at inference, and the 89.7 to 97.4 gap isolates the benefit of action-only supervision. That is the comparison to build the paper around. The paper instead leads with the confounded (a) vs (c).\n\nOther soft spots, in decreasing order: the analytical claim about objective interference is asserted (token-count imbalance is plausible but not derived); several GRPO hyperparameters were tuned on the LIBERO validation split before reporting the LIBERO headline; no code or data is released; error bars appear only for the 3-seed LIBERO analysis in the appendix. These are addressable.\n\nOn balance: the framework is sensible, the experiments are extensive, and the limitation section is honest. But the central attribution is not yet established. A referee should ask for a baseline given the same tool outputs at inference that still generates a rationale. That would settle it.\n\nI'd send this to peer review rather than desk-reject; the idea deserves referee time, and the confound is fixable.","headline":"Strong recipe and honest experiments, but the 'consume over generate' claim is confounded by test-time information asymmetry.","tokens_in":25119,"tokens_out":3858,"would_cite":true,"duration_ms":37048,"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":"Vision-language-action robots should consume grounded language rather than generate it: injecting tool-measured spatial facts as context, with supervision only on actions, beats chain-of-thought on success, speed, and data efficiency.","keywords":["vision-language-action models","in-context post-training","agentic tool use","chain-of-thought","grounded language","robot manipulation","behavior cloning"],"falsifier":"Corrupt the injected evidence deliberately on a held-out scene set, shifting every object centroid by about 30 pixels or making the VLM fallback mislabel target categories, and measure success; the paper's account predicts success drops sharply and monotonically with evidence error because the policy acts on what it reads, whereas flat success near behavior-cloning level would show the gain comes from something other than consuming grounded evidence. A finer variant the paper's ablations make available: with evidence fixed, sweep the supervision weight on the injected text tokens from zero (the paper's setting) up to full language-modeling loss, and check whether accuracy declines smoothly toward the Gen-CoT level as predicted.","tokens_in":24110,"feed_emoji":"🤖","tokens_out":15071,"duration_ms":121136,"temperature":0.7,"pith_summary":"Vision-language-action (VLA) models for robot manipulation are trained by behavior cloning: imitate expert actions from an image and a fixed instruction. The field's natural fix is chain-of-thought (CoT), letting the model write a textual rationale before acting, but this paper argues CoT is the wrong tool for low-level control, for three reasons: the rationale is generated from the same image the action head already sees (so it adds no information and can mislead), language tokens dominate the gradient so the policy learns to narrate rather than act, and autoregressively sampling hundreds of tokens per decision breaks closed-loop timing. The paper's alternative, VLA-Talker, gives the model the ability to consume grounded language: an external tool loop (open-vocabulary detector, monocular depth, and a vision-language fallback) measures where the gripper and target objects are, renders that evidence into diverse paraphrased sentences, and injects it into the context while the loss touches only action tokens. Under matched evidence, this configuration tops generative CoT on LIBERO (97.4% vs 81.5%), runs about 4.6 times faster per decision, needs fewer demonstrations, and transfers to eight real-world tabletop tasks. If the claim is right, the design principle for VLA language competence shifts from producing rationales to reading externally grounded evidence.","feed_headline":"Reading beats writing: evidence-injected robots top chain-of-thought","feed_subtitle":"Injected spatial facts as read-only context beat generated rationales on success, speed, and data efficiency.","key_machinery":"The mechanism that carries the argument is in-context post-training with an action-only supervision mask. The training sequence is `[instruction, image, <spatial>evidence</spatial>, action chunk]`, and the loss is computed only on action tokens (plus a single separator), so the model never learns to write the evidence and only learns to attend to it, eliminating the objective interference and generation latency of CoT in one design choice. The evidence itself comes from an agentic tool loop that answers a single question, where are the gripper and the task-relevant objects in image space and which is farther, emitting a structured tuple of gripper pixel and depth (from analytic projection via known camera intrinsics and extrinsics), per-object centroids and normalized depths (from an open-vocabulary detector and depth estimator, with a VLM fallback), and gripper-object relations. A rendering engine then paraphrases each tuple along six axes (reference modality, referential frame, lexicon, depth verbalization, verbosity, evidence-conditioned content) so the same geometric fact appears under many surface forms, teaching the policy to interpret rather than memorize spatial language. A final trajectory-level GRPO stage with sparse success rewards aligns when the policy invokes tools with actual task outcome, shrinking average tool calls per episode from about 3.4 to 1.8.","core_discovery":"The paper's central assertion is that \"what a VLA needs is not the ability to generate language, but the ability to consume grounded language.\" VLA-Talker instantiates this by decoupling evidence acquisition from evidence use: an agentic tool-use cascade projects the gripper into image space analytically, localizes task-relevant objects with an open-vocabulary detector (falling back to a VLM locator when the detector is uncertain), and samples relative depth; the resulting evidence tuple is rendered into diverse, paraphrased spatial descriptions wrapped in `<spatial>` tags; and the policy is post-trained with the language-model loss masked out, so it learns only to condition on the injected evidence, never to produce it. The paper reports that this design improves the LIBERO average success rate from 81.5% for the matched generative-CoT variant to 97.4%, reaches 59.5% on RoboCasa-GR1 and 72.4% on SimplerEnv, and achieves 58.1% single-task success on a real AgiBot G1 humanoid. The injection-vs-generation comparison holds all evidence identical, so the only differences are whether evidence is generated or injected and what is supervised.","pith_inferences":["The paper's error decomposition implies a next bottleneck: once grounding is solved, control precision near contact becomes 69.4% of residual failures, a direct argument for pairing evidence injection with a contact-rich refinement controller, which the paper itself names as a promising direction.","A testable extension the paper leaves implicit: if consuming grounded evidence is the true source of the gain, the same action-only in-context recipe should transfer across VLA backbones and action heads (including diffusion or flow-based heads) without re-engineering the rendering pipeline, since the mechanism lives in the supervision mask.","Because the policy trusts injected context by design, a confidence-aware tool loop that abstains or flags low-certainty evidence could push the approach into the out-of-distribution scenes the paper identifies as its main limitation.","The keyframe gating schedule is hand-designed, and the RL stage already learns to call tools less often on its own, so learning when to request fresh evidence end-to-end is a natural next step the paper leaves open."],"forward_implications":["Under matched evidence, supervising only action tokens on injected context beats both generating-and-supervising text (81.5% vs 97.4% LIBERO average) and injecting-but-still-supervising text (89.7%), so the operative design choice is the supervision mask, not the evidence alone.","Language competence for a VLA becomes measurable as robustness to phrasing: with roughly 24 diverse paraphrase realizations per evidence tuple, success under paraphrased instructions stays at 94.6% versus 77.2% for a single template.","Data efficiency improves: 25 demonstrations per task with VLA-Talker outperform behavior cloning trained on 50, and the gap widens as data shrinks.","The tool cascade's fallback routing is what keeps injected evidence reliable: at a simulated 60% detector drop rate the cascade retains 93.8% success where detector-only falls to 70.6%.","Trajectory-level GRPO adds +1.8% over the in-context cold-start alone and teaches economical tool use, but applied to the raw backbone it degrades to 87.8%, so RL is a complement to, not a replacement for, the grounded post-training."],"supporting_citations":[{"why":"Supplies the OpenVLA backbone the method post-trains from and the behavior-cloning baseline the whole comparison is built on.","marker":"Kim et al. 2024"},{"why":"CoT-VLA, the canonical visual chain-of-thought approach the paper argues against and includes as a benchmark.","marker":"Zhao et al. 2025"},{"why":"VLA-Thinker, the strongest thinking-with-image CoT baseline on LIBERO that VLA-Talker must exceed.","marker":"Wang et al. 2026"},{"why":"Source of the dense in-context post-training idea that gives the method its name and training scheme.","marker":"Sirko-Galouchenko et al. 2025"},{"why":"Supplies the GRPO group-relative RL objective used for the trajectory-level alignment stage.","marker":"Shao et al. 2024"},{"why":"Grounding DINO, the open-vocabulary detector that carries object localization in the tool cascade.","marker":"Liu et al. 2024b"},{"why":"Depth Anything, the monocular depth estimator that produces the relative-depth evidence.","marker":"Yang et al. 2024"},{"why":"LIBERO, the primary benchmark where the matched injection-vs-generation comparison is run.","marker":"Liu et al. 2023a"},{"why":"The visual-instruction-tuning line that backs the VLM locator used as fallback in the cascade.","marker":"Liu et al. 2023b"},{"why":"GR00T N1 and the RoboCasa-GR1 humanoid benchmark, both a baseline and a testbed for embodiment transfer.","marker":"Bjorck et al. 2025"}],"fun_headline_variants":["Inject language, don't generate it: VLA robots outperform CoT","Evidence-injected VLA beats chain-of-thought reasoning","Robots read, don't write: injected context tops CoT","VLA with tool use: consuming facts beats generating rationales","Grounded language in, action out: VLA beats CoT"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the tool loop's measurements are accurate enough at every keyframe for the policy to act on them; the detector, depth estimator, and VLM fallback can return wrong or misleading evidence, and the paper's own limitation notes that in a sufficiently out-of-distribution scene the policy would trust that misleading evidence and act on it.","fun_headline_variants_meta":{"raw":{"variants":["Inject language, don't generate it: VLA robots outperform CoT","Evidence-injected VLA beats chain-of-thought reasoning","Robots read, don't write: injected context tops CoT","VLA with tool use: consuming facts beats generating rationales","Grounded language in, action out: VLA beats CoT"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000746,"raw_usage":{"total_tokens":3410,"prompt_tokens":1113,"completion_tokens":2297,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":729,"completion_tokens_details":{"reasoning_tokens":2208}},"tokens_in":729,"tokens_out":2297,"duration_ms":18690,"temperature":1.0,"reasoning_tokens":2208,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T00:20:21.714907+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Corrupt the injected evidence deliberately on a held-out scene set, shifting every object centroid by about 30 pixels or making the VLM fallback mislabel target categories, and measure success; the paper's account predicts success drops sharply and monotonically with evidence error because the policy acts on what it reads, whereas flat success near behavior-cloning level would show the gain comes from something other than consuming grounded evidence. A finer variant the paper's ablations make available: with evidence fixed, sweep the supervision weight on the injected text tokens from zero (the paper's setting) up to full language-modeling loss, and check whether accuracy declines smoothly toward the Gen-CoT level as predicted.","supporting_citations":[],"review_version":1}