{"id":"f38e99e8-8ade-4e6d-886d-e20d9aebfe01","arxiv_id":"2607.02865","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A frozen VLA plus latent world-model rollouts and a value model can raise real-robot OOD manipulation success from 23.75% to 66.25% without any target-environment finetuning.","lead":"DREAMSTEER steers frozen vision-language-action robot policies at deployment by sampling action chunks, imagining their futures with a latent world model, and ranking them with a language-conditioned value model—no finetuning. On real-robot pick-and-place with unseen objects it lifts success from 24% to 66% and instruction following from 39% to 56%.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified beyond the reader's already-flagged ranking-correlation weakness.","rationale":"The paper's strongest claim is empirical and well-supported by the real-robot tables and ablations under a clear zero-shot protocol (different lab, cameras, objects; no finetuning). Complementary generalization across policy/WM/value (§5.1) is a coherent explanation for why steering works when single-sample π0 fails. The ranking-correlation evidence is only moderate, and latency/candidate-coverage limits are acknowledged (§5.2), but these are already the reader's stated reasons for CONDITIONAL rather than full ACCEPT. No independent contradiction or unstated assumption that would overturn the reported gains was found; therefore the reader's CONDITIONAL verdict and medium correctness_risk stand without adjustment.","tokens_in":14522,"tokens_out":480,"duration_ms":5517,"concrete_test":"On the same 80 OOD trials, recompute selection using ground-truth future frames (oracle ranking) versus WM-decoded frames under identical VLAC scoring; if oracle success exceeds full DREAMSTEER by >15 absolute points while WM ranking still beats random, the moderate r=0.66 is confirmed as the binding constraint rather than a fatal flaw.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (frozen π0 + multi-embodiment latent WM + VLAC ranking yields 23.75%→66.25% OOD success and 38.75%→56.25% IF accuracy with no target data) is internally consistent with the reported ablations. Table 2 shows both candidate diversity and value ranking are necessary (π0 alone 23.75%, π0+DREAMSTEER 42.5%, full method 66.25%, random/primitives-only 0%). The reader's weakest assumption—that H=10 DINOv2 rollouts preserve relative VLAC order (Pearson r=0.66 / Spearman ρ=0.69 on RoboArena clips, §4.2)—is the softest link, but it is already quantified, does not contradict the real-robot gains, and is not a hidden inconsistency. No stronger load-bearing flaw (e.g., target-data leakage, non-comparable baselines, or inverted ranking) appears in the manuscript.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper proposes DREAMSTEER, a deployment-time steering method that improves a frozen pretrained VLA (π0) without any target-environment finetuning or parameter updates. At each steering step it forms a candidate set of action chunks from stochastic VLA samples plus a small library of Cartesian primitives, rolls each candidate out with a multi-embodiment action-conditioned latent world model (DINOv2 latents, spatio-temporal transformer), decodes the imagined observations, and ranks them with a frozen language-conditioned value model (VLAC). The highest-scoring chunk is executed. Real-robot experiments on a Franka/DROID-style setup under substantial distribution shift report OOD pick-and-place success rising from 23.75% to 66.25% and instruction-following accuracy from 38.75% to 56.25% over single-sample π0, with ablations showing that both candidate diversity and value ranking are necessary.","tokens_in":14787,"tokens_out":1115,"duration_ms":11090,"significance":"If the reported gains hold under broader evaluation, the work offers a practical, training-free route to improve generalist VLAs at deployment—an important operational need given the cost and risk of collecting target-domain demonstrations. Strengths include a clean training-free composition claim, multi-embodiment world-model training, explicit ablations (Table 2) that isolate diversity versus ranking, Wilson confidence intervals on real-robot trials, and a quantified check that imagined rollouts preserve relative value order (Pearson r=0.66, Spearman ρ=0.69). The approach is complementary to finetuning and is immediately usable with existing frozen VLAs, latent world models, and progress estimators.","major_comments":[{"comment":"§4.2 / Fig. 6: The ranking assumption that H=10 latent rollouts preserve VLAC order is supported only by moderate correlation (Pearson r=0.66, Spearman ρ=0.69) on held-out RoboArena clips. This is the softest load-bearing link for the central claim. The paper should either (i) report ranking accuracy / top-1 selection agreement on the same clips (not only correlation of scalar scores), or (ii) show that the real-robot gains remain when the value model is replaced by an oracle ranking of ground-truth futures, so readers can bound how much of the 23.75%→66.25% lift is limited by ranking error.","section":null},{"comment":"§4.3 / Tables 2–3: Evaluation covers only four OOD objects and four IF targets (20 trials each). Aggregate Wilson CIs are reported, but object-level variance is large (e.g., Tape 6/20→16/20 vs. Mustard 3/20→11/20). For a claim of broad OOD robustness under deployment shift, the manuscript needs either more objects/scenes or a clearer statement that the result is a controlled demonstration on this object set rather than a general guarantee. Expanding the object suite or adding a leave-one-object-out style check would strengthen the central numbers.","section":null},{"comment":"§3.2 / Eq. (1) and candidate construction: The full method’s gain over π0+DREAMSTEER (42.5%→66.25%) comes largely from the hand-designed Cartesian primitive library. The paper should quantify how sensitive results are to the primitive set (e.g., remove gripper open/close, change step sizes, or replace primitives with pure noise around the current pose). Without that, it remains unclear how much of the reported improvement is portable versus dependent on a carefully chosen C_prim for this workbench.","section":null}],"minor_comments":[{"comment":"§4.3: Steering latency (~13 s for 13 candidates) is acknowledged but not placed in context of closed-loop control rates; a short note on how often steering is applied relative to the low-level controller would help practitioners.","section":null},{"comment":"Fig. 5 / Fig. 10: Rollout visualizations are useful; adding the corresponding VLAC scores next to selected frames would make the ranking story more self-contained.","section":null},{"comment":"Table 1: “Training-free composition” is a useful axis; a one-sentence clarification that VLAC and the world model are themselves pretrained (just not on the target lab) would avoid any ambiguity about “zero training.”","section":null},{"comment":"Notation: C_t = C_VLA ∪ C_prim is clear; stating the exact |C_prim| and the numerical K used in the main tables once in the main text (not only in the supplement) would improve reproducibility.","section":null},{"comment":"Related work: V-GPS, FOREWARN, GPC, VLA-Reasoner, and LaDi-WM are fairly compared; a brief note on whether any of those methods could be run zero-shot with the same frozen π0 would make the novelty claim sharper.","section":null}],"recommendation":"major_revision","confidential_remarks":"The core idea is solid and the real-robot ablations are more careful than many concurrent VLA-steering papers. The ranking-correlation and limited object suite are genuine but fixable; I would not reject on novelty or circularity grounds. Fit for a robotics venue is good if the authors tighten the evaluation scope statement and add the ranking/primitive sensitivity checks."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The punchline is simple: they take a frozen π0, sample action chunks, mix in a few Cartesian primitives, roll them out in a multi-embodiment DINOv2 latent world model, and rank with an off-the-shelf VLAC value model. No target-environment data, no parameter updates. On their Franka setup they move OOD pick-and-place from 23.75% to 66.25% and instruction-following contact accuracy from 38.75% to 56.25% across 80 trials each, with Wilson CIs and clean ablations.\n\nWhat is actually new is the fully plug-and-play composition. Prior steering work (V-GPS, FOREWARN, GPC, LaDi-WM, etc.) either finetunes something on the target task or lacks the generalized latent WM + language value model pairing. Their Table 1 is fair about that. The ablations in Table 2 are the strongest part of the paper: policy alone, policy+DreamSteer, primitives alone, and random selection over the same candidate set. Random and primitives-only go to zero; adding ranking and diversity both matter. That is honest experimental design.\n\nThe soft spot is exactly the one the reader flagged: imagined vs ground-truth value scores only correlate at Pearson r=0.66 / Spearman ρ=0.69 on held-out RoboArena clips. That is moderate, not terrible, and they report it rather than hide it. Candidate coverage is the other real limit—if neither π0 samples nor the hand primitives make progress, the system cannot invent a solution. Latency (~13 s per steering step with 13 candidates) is a practical issue they own. No code or weights yet, which keeps reproducibility moderate. None of these invert the central claim.\n\nMath and citation pattern look solid; free parameters (K, H, steering frequency, primitive library) are stated and ablated where it matters. This is for people who deploy generalist VLAs and care about test-time compute without collecting new demos. It deserves a serious referee. I would engage with it and expect it to survive peer review with requests for multi-view ranking evidence or code release.","headline":"Training-free VLA steering with real-robot gains that hold up under the ablations; ranking correlation is the softest link, not a contradiction.","tokens_in":15391,"tokens_out":544,"would_cite":true,"duration_ms":5288,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"A frozen robot policy can be steered at deployment by imagining candidate futures with a latent world model and ranking them with a language-conditioned value model—no finetuning required.","keywords":["latent world models","vision-language-action policies","deployment-time steering","robot manipulation","zero-shot generalization","value-guided ranking","action chunks"],"falsifier":"On the same OOD pick-and-place trials, if ranking candidates by value scores on world-model rollouts no longer beats random selection among the same candidate set—or if value scores on imagined clips systematically reverse the order of matched ground-truth clips—the claimed steering gain collapses.","tokens_in":15459,"feed_emoji":"🦾","tokens_out":967,"duration_ms":22039,"temperature":0.7,"pith_summary":"Pretrained vision-language-action policies often fail when the robot meets new objects, cameras, or scenes at deployment. Collecting target demonstrations and finetuning is the usual fix, but it is expensive and not always allowed. This paper argues that the policy can stay completely frozen: sample several candidate action chunks (plus a few simple motion primitives), predict what each would look like with an action-conditioned latent world model, and execute the chunk that a language-conditioned value model scores highest. On real pick-and-place tasks with unseen objects, success rose from 23.75% to 66.25% and instruction-following accuracy from 38.75% to 56.25% over single-sample execution. The result separates “having a useful action prior” from “choosing the right sample before acting,” and shows that look-ahead ranking can recover reliability without any target-environment data.","feed_headline":"Robot success jumps from 24% to 66% by ranking imagined futures","feed_subtitle":"A frozen VLA samples actions; a latent world model and value critic pick the best—no target data needed.","key_machinery":"DreamSteer: a training-free selection loop that samples action chunks, predicts their futures with an action-conditioned latent world model (frozen visual latents plus a factorized spatio-temporal transformer), and ranks the decoded rollouts with a language-conditioned value model before execution.","core_discovery":"Deployment-time steering of a frozen VLA is possible by ranking short imagined rollouts: candidate action chunks from the policy and simple Cartesian primitives are rolled out in a multi-embodiment latent world model and scored by a language-conditioned progress value model; the highest-scoring chunk is executed. With no component finetuned on target data, this composition alone lifts real-robot OOD manipulation success from 23.75% to 66.25% and instruction-following accuracy from 38.75% to 56.25% over single-sample policy execution.","pith_inferences":["The same sample–imagine–rank loop could wrap any stochastic generative policy, not only VLAs, wherever a latent dynamics model and a progress critic exist.","Multi-view value scoring would cut ranking errors caused by single-camera visual similarity and tighten imagined versus true preference order.","Candidate coverage is a hard ceiling: when neither policy samples nor fixed primitives approach the goal, steering cannot invent a solution—latent planning or iterative proposal refinement is the natural next lever.","The complementary-data story invites deliberate system-level training splits (policy on demos, world model on mixed play, value on action-free video) rather than only post-hoc assembly of frozen parts."],"forward_implications":["Useful action chunks often already sit inside a pretrained VLA’s sampling distribution; under distribution shift the bottleneck is selection, not generation.","A world model and value model trained on broader interaction data can correct a policy trained only on successful demonstrations.","Zero-shot rollout ranking becomes a practical alternative when target-environment finetuning is unwanted or infeasible.","Short-horizon latent prediction, rather than pixel video generation, is fast enough for online ranking of many candidates on real hardware.","Adding simple fixed Cartesian primitives expands coverage when policy samples alone stall near the object."],"fun_headline_variants":["Frozen VLAs steered at deploy by ranking latent imagined rollouts","World model + value critic lifts robot success 24% to 66% zero-shot","No finetuning: sample actions, dream outcomes, pick best for VLA","Latent dreams alone triple OOD manipulation success on real robots","DREAMSTEER ranks short futures to fix distribution-shift failures"],"cache_read_input_tokens":128,"weakest_assumption_plain":"Short imagined futures from the latent world model must keep the same preference order as real futures under the value model well enough that the top-ranked candidate is usually the right one to run.","fun_headline_variants_meta":{"raw":{"variants":["Frozen VLAs steered at deploy by ranking latent imagined rollouts","World model + value critic lifts robot success 24% to 66% zero-shot","No finetuning: sample actions, dream outcomes, pick best for VLA","Latent dreams alone triple OOD manipulation success on real robots","DREAMSTEER ranks short futures to fix distribution-shift failures"]},"model":"grok-4.5","effort":"low","cost_usd":0.004762,"raw_usage":{"total_tokens":1391,"prompt_tokens":800,"num_sources_used":0,"completion_tokens":80,"cost_in_usd_ticks":47620000,"prompt_tokens_details":{"text_tokens":800,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":511,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":800,"tokens_out":80,"duration_ms":5519,"temperature":1.0,"reasoning_tokens":511,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-12T06:29:55.368799+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"On the same OOD pick-and-place trials, if ranking candidates by value scores on world-model rollouts no longer beats random selection among the same candidate set—or if value scores on imagined clips systematically reverse the order of matched ground-truth clips—the claimed steering gain collapses.","supporting_citations":[],"review_version":1}