{"id":"89d41a61-388d-4048-be15-22e299bde5ad","arxiv_id":"2607.04988","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.5,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A VLM-backed robot policy with latent foresight tokens supervised by a frozen video model achieves best overall results on six sim benchmarks and strong real-world compositional generalization.","lead":"InternVLA-A1.5 is a robot policy that keeps a pretrained vision-language model for language understanding while learning continuous actions and short-horizon future dynamics via latent foresight tokens supervised by a frozen video generator. It reports top scores on six simulation benchmarks and stronger held-out instruction following and long-horizon lab tasks in the real world.","discovery_kind":"extension","skeptic_critique":{"model":"grok-4.5","headline":"Ablations do not isolate foresight from continuous expert and co-training; short-horizon frozen WAN may not be the causal driver of claimed gains.","rationale":"The paper’s design thesis is clear and the multi-benchmark + real held-out results are competitive for a VLA preprint. The single most load-bearing soft spot is exactly the one the reader named: necessity of short-horizon frozen-video supervision is not isolated from the continuous expert and co-training scale. Table 8 shows foresight helps but does not rule out that a pure continuous expert under the same data recipe would close most of the gap. That keeps the verdict CONDITIONAL (accept-shaped pending stronger isolation ablations, multi-seed stats, and public artifacts) rather than moving it to ACCEPT or REJECT. No internal inconsistency or fabrication concern; the issue is causal attribution of the headline gains.","tokens_in":22945,"tokens_out":617,"duration_ms":5264,"concrete_test":"Train and evaluate three matched Stage-2 variants on the same 1.2M+3M mixture and identical post-train protocol: (A) full InternVLA-A1.5, (B) unified expert with foresight tokens but L_video=0 (no WAN gradients), (C) continuous flow-matching expert with foresight tokens and WAN path entirely removed (action queries attend only to VLM context). Report LIBERO-Plus total, DOMINO zero-shot SR/MS, and real MOF + held-out Insert/Move success with ≥3 seeds. If (B) and (C) stay within ~2–3 points of (A) on DOMINO/MOF, the foresight-prior claim is not load-bearing.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim attributes best-on-six-benchmarks results and real-world compositional/long-horizon gains to (i) preserved VLM semantics via continued VQA/subtask training plus (ii) latent foresight tokens supervised by frozen WAN2.2 (N=4 frames, same horizon as one action chunk; §3.2 Eqs. 3–4, L_stage2 with α=1, β=10). Table 8 only removes video loss or foresight tokens from the full two-stage model; it never compares against an otherwise identical Stage-2 continuous flow-matching expert without any foresight pathway, nor against a matched-scale continuous expert trained without the 0.15:0.85 multimodal co-training mixture (§4.3). Limitations already note that foresight is short-horizon and the generator is frozen/generic. Thus the reported zero-shot DOMINO (27.7%) and MOF (76.4% vs π0.5 29.3%) lifts could be driven primarily by the continuous expert + data scale rather than by dynamics priors distilled from WAN. The reader’s weakest_assumption correctly flags this causal gap.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"InternVLA-A1.5 proposes a unified VLA that keeps a native VLM backbone under continued VQA/subtask/discrete-action training and attaches a lightweight Mixture-of-Transformers expert for continuous flow-matching actions. Future prediction is cast as latent querying: learnable foresight tokens produce a compact condition for a frozen pretrained video generator (WAN2.2), so the policy inherits dynamics priors without pixel-level generation; the video branch is discarded at inference. Pretrained on 1.2M robot episodes and 3M multimodal samples, the model reports best overall scores on six simulation benchmarks (LIBERO, LIBERO-Plus, RoboTwin, DOMINO, EBench, SimplerEnv) and, in real-world held-out instruction-binding and long-horizon MOF tasks, stronger compositional generalization and sequential execution than π0.5 and Motus.","tokens_in":23285,"tokens_out":894,"duration_ms":11998,"significance":"If the results hold under tighter causal controls, the work is a useful systems contribution to unified robot policies: it shows a practical way to keep VLM semantics while injecting world-model dynamics without paying generation cost at deployment, and it demonstrates strong multi-benchmark and real-world compositional performance. Strengths include broad evaluation (six simulators plus real held-out bindings and MOF), explicit ablations of video loss and foresight tokens (Table 8), qualitative foresight rollouts (Fig. 11), and a clear training/inference split that preserves real-time control. The latent-query design is a concrete alternative to learning pixel futures from scratch and is of interest to the VLA and world-model communities even if some gains partly come from scale and the continuous expert.","major_comments":[{"comment":"§5.3 Table 8 and §3.2: The central attribution of zero-shot DOMINO (27.7%) and long-horizon MOF gains to dynamics priors from foresight is under-isolated. Table 8 only removes L_video or the foresight tokens from the full two-stage model; it does not compare against an otherwise identical Stage-2 continuous flow-matching expert with no foresight pathway, nor against a matched continuous expert trained without the 0.15:0.85 multimodal co-training mixture (§4.3). Without those controls, the reported lifts could be driven mainly by the continuous expert, data scale, or co-training rather than WAN-distilled dynamics. Please add at least the no-foresight continuous-expert baseline (same architecture, data, and Stage-1 init) on LIBERO-Plus, RoboTwin, DOMINO, and one real task, and discuss effect sizes relative to removing video loss alone.","section":null},{"comment":"§3.2 Eqs. (3)–(4), N=4, and Limitations: Foresight supervision spans only one action-chunk horizon (N=4 frames) through a frozen generic WAN2.2. The narrative links this short-horizon signal to long-horizon MOF success (76.4% vs π0.5 29.3%, Fig. 8) and dynamic DOMINO. That causal chain is not yet demonstrated: either provide intermediate analyses (e.g., foresight-token ablation on MOF subtask stages, or correlation between foresight quality and failure modes) or soften claims that latent foresight is what “sustains long-horizon execution,” and state more clearly that long-horizon gains may also come from explicit subtask prediction retained from Stage 1.","section":null},{"comment":"§5.1 real-world protocol: Trial counts are modest (e.g., 15 trials per binding for Sort/Insert, 16 for Move, 20 for MOF) with no confidence intervals or statistical tests. Given that Sort Tubes overall is slightly behind π0.5 (75.9 vs 77.8) while OOD bindings favor InternVLA-A1.5, please report per-condition n, success counts or CIs, and whether differences on held-out bindings remain significant under a simple binomial or bootstrap test. This is load-bearing for the “strongest compositional generalization” claim.","section":null}],"minor_comments":[],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The useful takeaway is simple: keep the native VLM on VQA and subtask prediction so semantics do not drift, attach a light MoT expert for continuous flow-matching actions, and train a small set of foresight tokens so a frozen video model (WAN2.2) can supervise them—then throw the video branch away at inference. That combination is the real design move, and it is cleaner than most “unify understanding + generation + action” stacks that either retrain pixel generation or let the backbone rot.\n\nWhat they do well is execution and reporting. Six sim benchmarks (LIBERO, LIBERO-Plus, RoboTwin, DOMINO, EBench, SimplerEnv) plus real held-out instruction bindings and a long-horizon MOF chemistry procedure, all against π0.5 and Motus. Best or near-best averages across the board; real-world OOD bindings and MOF (76% vs 29% for π0.5) are the numbers that matter for the compositional claim. Attention masks, chat-template action/state encoding, two-stage recipe, and the “discard generator at test time” choice are all carefully specified. Ablations that drop video loss or foresight tokens move the needle in the right direction, especially on zero-shot LIBERO-Plus and DOMINO. Data scale (1.2M episodes + 3M multimodal) and sampling weights are documented.\n\nSoft spots, in proportion: Table 8 does not fully isolate foresight from the continuous expert and the heavy multimodal co-training mixture. You cannot yet rule out that a matched Stage-2 flow expert without any foresight path would close a lot of the gap. Foresight is short-horizon (N=4, same as one action chunk) and the generator stays frozen and generic—the authors already say this in Limitations. Real-world trial counts are modest. Those are real but ordinary for this literature; they do not collapse the main empirical claim.\n\nMath and citation pattern look fine: flow-matching and next-token losses are standard, self-cites to InternVLA-A1/M1 are expected lineage, not circular evaluation. Free parameters (α, β, M, sampling ratios) are listed; nothing is hidden as a free lunch.\n\nThis is for people building or comparing large VLAs who care about keeping VLM semantics while stealing dynamics priors without test-time generation cost. It deserves a serious referee. I would engage: read the architecture and the real-world binding design carefully, and ask for one cleaner ablation if you re-implement.","headline":"Solid engineering VLA paper: latent foresight tokens + frozen WAN supervision + kept VQA/subtask co-training is a clean practical pattern with strong multi-benchmark numbers; causal isolation of foresight is incomplete but not fatal.","tokens_in":24031,"tokens_out":628,"would_cite":true,"duration_ms":7430,"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 robot policy keeps VLM semantics and inherits video-model dynamics via short latent foresight tokens, without pixel generation at test time.","keywords":["vision-language-action","latent foresight","robot manipulation","compositional generalization","world-model priors","flow matching","mixture-of-transformers"],"falsifier":"An ablation that matches data, co-training, and the continuous expert but removes video supervision and foresight tokens (or replaces the frozen generator with a weak random conditioner) and still matches InternVLA-A1.5 on DOMINO zero-shot, LIBERO-Plus, and real long-horizon MOF success.","tokens_in":23834,"feed_emoji":"🤖","tokens_out":631,"duration_ms":5093,"temperature":0.7,"pith_summary":"Unified robot policies try to combine language understanding from vision-language models with a sense of physical dynamics from future prediction, but typical designs blur the pretrained backbone's semantics, pit mismatched losses against each other, and retrain pixel generation from scratch. InternVLA-A1.5 keeps the native VLM training on VQA and subtask prediction, attaches a light expert for continuous actions, and treats foresight as latent querying: a few learnable tokens compress the task-relevant future and are supervised by a frozen pretrained video generator. That generator is dropped at deployment, so control stays real-time. After pretraining on large robot and multimodal corpora, the paper reports the best overall scores on six simulation suites and, in real settings, stronger compositional success on held-out instruction bindings plus better long-horizon sequential execution.","feed_headline":"Robot policy gets video dynamics without generating pixels at test","feed_subtitle":"Latent foresight tokens and kept VQA training yield top sim scores and real compositional success","key_machinery":"Latent foresight tokens: a fixed set of learnable queries that attend to the shared multimodal context, produce a compact conditioning code for a frozen video generator over a short future horizon, and feed that future-aware signal into flow-matching action prediction; the video branch exists only in training.","core_discovery":"The paper claims that preserving ongoing VQA and subtask training on a native VLM backbone, while supervising a small set of learnable foresight tokens through a frozen video generator, unifies understanding, dynamics priors, and continuous action without eroding semantics or paying pixel-generation cost at inference, and that this yields the strongest overall results on six simulation benchmarks plus the best real-world compositional generalization and long-horizon success among the compared methods.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Latent foresight tokens inject video dynamics without pixel gen at test","Native VLM keeps VQA while foresight tokens borrow frozen video priors","Unified expert plus latent queries unifies semantics dynamics and action","Foresight tokens condense future under video model for compositional gains","Preserved VQA and latent foresight yield top sim scores and real generalization"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"That short-horizon supervision through a frozen, generic video generator is what mainly supplies the dynamics needed for the reported zero-shot dynamic and long-horizon gains, rather than data scale, co-training, or continuous-action training alone.","fun_headline_variants_meta":{"raw":{"variants":["Latent foresight tokens inject video dynamics without pixel gen at test","Native VLM keeps VQA while foresight tokens borrow frozen video priors","Unified expert plus latent queries unifies semantics dynamics and action","Foresight tokens condense future under video model for compositional gains","Preserved VQA and latent foresight yield top sim scores and real generalization"]},"model":"grok-4.5","effort":"low","cost_usd":0.005198,"raw_usage":{"total_tokens":1403,"prompt_tokens":805,"num_sources_used":0,"completion_tokens":94,"cost_in_usd_ticks":51980000,"prompt_tokens_details":{"text_tokens":805,"audio_tokens":0,"image_tokens":0,"cached_tokens":0},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":504,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":805,"tokens_out":94,"duration_ms":4069,"temperature":1.0,"reasoning_tokens":504,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-11T10:27:59.308291+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"An ablation that matches data, co-training, and the continuous expert but removes video supervision and foresight tokens (or replaces the frozen generator with a weak random conditioner) and still matches InternVLA-A1.5 on DOMINO zero-shot, LIBERO-Plus, and real long-horizon MOF success.","supporting_citations":[],"review_version":1}