{"id":"2164255f-d610-44d5-9405-3db08220d252","arxiv_id":"2606.08684","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"BLUE trains a lightweight gate on frozen VLA hidden states to selectively activate language generation only when beneficial, achieving SOTA results with 2.54x inference speedup on driving benchmarks.","lead":"BLUE adds a 0.11M-parameter gate to vision-language-action models for autonomous driving that decides per frame whether to generate language or predict actions directly from hidden states. A smart generalist might read it to see a practical way to cut computation in language-augmented AI systems while keeping performance gains on critical frames.","discovery_kind":"unclear","skeptic_critique":{"model":"grok-4.3","headline":"No significant objection identified","rationale":"The reader's weakest_assumption directly identifies the load-bearing point. With only the abstract available in the initial verdict and no contradictory detail supplied in the prompt, no stronger or different concern can be substantiated. The reported numbers and efficiency claims stand or fall on that encoding assumption, which the paper asserts holds.","tokens_in":1796,"tokens_out":240,"duration_ms":17583,"concrete_test":"Reproduce the gate training procedure on the released checkpoints and logs; measure gate accuracy and end-to-end success rate on a fresh set of Bench2Drive routes never seen during label generation for the gate.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim depends on the gate learning a reliable decision from frozen hidden states. The abstract presents this as feasible after analysis showing that hidden states encode benefit information better than scene/kinematic features, with no backbone changes or extra labels needed. No internal inconsistency, circularity, or unsupported derivation is visible from the given material that would falsify the reported 76.2% / 36 driving score or the 2.54x speedup.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper introduces BLUE, a minimal method for efficient language use in vision-language-action (VLA) models for autonomous driving. Through analysis, the authors find that language benefits performance on only a small fraction of routes but can greatly improve or degrade results there; pretrained VLA hidden states encode whether language will help on a given frame better than scene or kinematic features. BLUE trains a 0.11M-parameter gate on frozen hidden states to decide per-frame whether to invoke language generation or predict actions directly, without backbone changes or extra labels. This yields SOTA results of 76.2% success rate on Bench2Drive and 36 driving score on Longest6 v2, plus 2.54x speedup and 8.9% success improvement over the backbone.","tokens_in":1876,"tokens_out":541,"duration_ms":18800,"significance":"If the hidden-state analysis and reported gains hold, the work demonstrates a practical, low-overhead way to retain language benefits in VLA models while cutting unnecessary computation, which is relevant for real-time autonomous driving. The explicit release of code, data, logs, and checkpoints supports reproducibility and is a positive aspect of the contribution.","major_comments":[{"comment":"The central claim that hidden states encode language-benefit information better than scene/kinematic features (and therefore justify training the gate on them) is load-bearing; the analysis section must supply concrete quantitative comparisons (e.g., classification accuracy, AUC, or F1 for benefit prediction using each feature type) and statistical tests to establish this superiority.","section":"Analysis section (referenced in abstract as 'extensive analysis')"},{"comment":"The reported SOTA numbers and 8.9% improvement are load-bearing for the efficiency claim; the results section should include ablations isolating the gate's contribution, full baseline tables with recent VLA methods, and error bars or multiple runs to confirm the gains are not due to training variance.","section":"Results / Experiments section"}],"minor_comments":[{"comment":"Provide the exact architecture, input dimension, and training hyperparameters (loss, optimizer, epochs) of the 0.11M gate so that the 'lightweight' claim can be verified.","section":null},{"comment":"Clarify the precise definition of 'language matters' used in the route-level analysis (e.g., success-rate delta threshold) and how frames were labeled without additional human annotation.","section":null}],"recommendation":"minor_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the positive recommendation of minor revision and the constructive comments on the analysis and results sections. We appreciate the recognition of the work's significance for efficient VLA models in autonomous driving and the value placed on reproducibility. We address each major comment below.","responses":[{"response":"We agree that explicit quantitative metrics and statistical validation would make the superiority claim more rigorous. The current analysis demonstrates the advantage of hidden states over scene and kinematic features through comparative experiments on benefit prediction, but we will revise the analysis section to include the requested classification accuracy, AUC, and F1 scores for each feature type, along with statistical tests such as paired t-tests or Wilcoxon tests to establish significance.","revision_made":"yes","referee_comment":"[Analysis section (referenced in abstract as 'extensive analysis')] The central claim that hidden states encode language-benefit information better than scene/kinematic features (and therefore justify training the gate on them) is load-bearing; the analysis section must supply concrete quantitative comparisons (e.g., classification accuracy, AUC, or F1 for benefit prediction using each feature type) and statistical tests to establish this superiority."},{"response":"We thank the referee for highlighting opportunities to strengthen the empirical validation. The manuscript already reports the 8.9% improvement over the backbone along with some ablations on gate design choices, but we will expand the results section with additional ablations that more precisely isolate the gate's contribution, include a fuller baseline table incorporating recent VLA methods, and report performance across multiple independent runs with error bars and standard deviations to address variance concerns.","revision_made":"yes","referee_comment":"[Results / Experiments section] The reported SOTA numbers and 8.9% improvement are load-bearing for the efficiency claim; the results section should include ablations isolating the gate's contribution, full baseline tables with recent VLA methods, and error bars or multiple runs to confirm the gains are not due to training variance."}],"tokens_in":1463,"tokens_out":433,"duration_ms":13750,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The core contribution is a 0.11M-parameter gate trained on pretrained VLA hidden states to decide per frame whether to generate language or skip straight to action prediction. The authors report that language only matters on a small fraction of routes, that those hidden states already carry a usable signal for when it helps, and that this yields 76.2% success on Bench2Drive, 36 driving score on Longest6 v2, an 8.9% lift over the backbone, and the stated speedup, all without touching the backbone or adding labels.\n\nWhat stands out is the open release of code, data, logs, and checkpoints. That makes the efficiency claim and the hidden-state analysis directly testable, which is useful in a field where real-time constraints matter. The approach stays minimal: no new annotations, no architecture changes, and the gate is cheap enough to run alongside the model.\n\nThe main soft spot is the reliance on the claim that hidden states encode language utility better than scene or kinematic features. The abstract presents supporting analysis, but the strength of that analysis, the exact distributions of when language helps or hurts, and any edge cases or failure modes need to be shown in the full paper. If the gate was fit on the same route distributions used for evaluation, the reported gains could shrink on truly held-out data. The benchmarks themselves look standard, but the paper should clarify how much of the SOTA margin comes from the gate versus other training details.\n\nThis is for groups working on deployable VLA systems for autonomous driving who need to keep language benefits without paying full inference cost every frame. A reader focused on practical efficiency would find the numbers and the released artifacts worth examining. The work is coherent on its own terms and the reproducibility steps are in place, so it deserves a serious referee.","headline":"BLUE's small gate on frozen VLA states lets language run only when it helps, with claimed SOTA numbers and 2.54x speedup on driving benchmarks.","tokens_in":2368,"tokens_out":446,"would_cite":false,"duration_ms":19243,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"A 0.11M-parameter gate on pretrained VLA hidden states decides per frame whether language generation will help autonomous driving.","keywords":["vision-language-action models","autonomous driving","language gating","efficient inference","Bench2Drive","Longest6","hidden states"],"falsifier":"A gate trained on random noise or only kinematic features achieving similar success-rate gains and speedup on the same benchmarks would show the VLA hidden states are not the key signal.","tokens_in":2690,"feed_emoji":"🚗","tokens_out":644,"duration_ms":16254,"temperature":0.7,"pith_summary":"The paper shows that language improves VLA performance on driving tasks only on a small fraction of routes and can even degrade results on others. Generating language at every frame wastes computation because most frames do not benefit. Pretrained VLA hidden states already contain signals about whether language will help on a given frame, even when scene complexity or kinematics do not. BLUE therefore trains a tiny gate on these frozen states to activate language only when useful, leaving the backbone unchanged and requiring no new labels. This selective mechanism delivers higher success rates and faster inference on standard driving benchmarks.","feed_headline":"Tiny gate boosts driving model success 8.9% at 2.54x speed","feed_subtitle":"0.11M-parameter gate on VLA hidden states activates language only when useful, improving Bench2Drive and Longest6 results without new labels","key_machinery":"The 0.11M-parameter gate trained on frozen VLA hidden states to predict per-frame language utility.","core_discovery":"BLUE trains a lightweight gate on frozen VLA hidden states to decide per frame whether to activate language generation or predict actions directly. With just 0.11M parameters, it achieves 76.2% success rate on Bench2Drive and 36 driving score on Longest6 v2, delivering 2.54x inference speedup and 8.9% success rate improvement over the backbone.","pith_inferences":["The same hidden-state signal could be used to gate other expensive modules in multimodal driving models.","Gating during pretraining might reduce overall training cost for future VLA systems.","The approach may transfer to non-driving VLA tasks where language is useful only sporadically.","Making the gate output a continuous score instead of a binary decision could allow partial language use."],"forward_implications":["Language generation can be skipped on most frames without hurting driving performance.","Pretrained VLA hidden states contain frame-specific information about language utility.","Selective language activation produces both higher success rates and faster inference.","No backbone modification or extra human annotation is required to gain these benefits.","State-of-the-art results appear on Bench2Drive and Longest6 v2 benchmarks."],"fun_headline_variants":["0.11M gate decides language activation on VLA hidden states","Selective language use with gate at 2.54x VLA speedup","76.2 percent success on Bench2Drive via BLUE gate","Language utility encoded in pretrained VLA hidden states"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"The hidden states of a pretrained VLA model already encode whether language generation will benefit performance on a given frame.","fun_headline_variants_meta":{"raw":{"variants":["0.11M gate decides language activation on VLA hidden states","Selective language use with gate at 2.54x VLA speedup","76.2 percent success on Bench2Drive via BLUE gate","Language utility encoded in pretrained VLA hidden states"]},"model":"grok-4.3","cost_usd":0.009265,"raw_usage":{"total_tokens":4173,"prompt_tokens":719,"num_sources_used":0,"completion_tokens":69,"cost_in_usd_ticks":92649500,"prompt_tokens_details":{"text_tokens":719,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":3385,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":719,"tokens_out":69,"duration_ms":19914,"temperature":1.0,"reasoning_tokens":3385,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-27T18:28:26.675936+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"A gate trained on random noise or only kinematic features achieving similar success-rate gains and speedup on the same benchmarks would show the VLA hidden states are not the key signal.","supporting_citations":[],"review_version":1}