{"id":"a92a1355-c7cf-4e39-ac29-051b09a54094","arxiv_id":"2506.01953","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"FiS-VLA embeds a diffusion-based action module into the final transformer blocks of a vision-language model, achieving 69% mean success on RLBench and a claimed 117.7 Hz control frequency.","lead":"The paper proposes a robot-control model that packs a fast action generator inside a slow vision-language reasoning model by sharing some of the same neural-network blocks. It reports faster and more accurate manipulation than prior methods in simulation and on two dual-arm robots.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's central causal claim—that embedding the fast action module inside the VLM's final blocks is what improves manipulation—is never tested against a matched separate-head baseline; the reported gains could come from extra pretraining, added point-cloud/state inputs, or capacity.","rationale":"The reader flagged the preservation of System 2's reasoning as the weakest assumption, and that is indeed untested: the paper provides no language or reasoning benchmark before/after co-training, and the Lslow-removal ablation (69% to 62%) could reflect regularization rather than preserved reasoning. That concern is real, but I see a more fundamental gap in the causal claim. The paper's contribution is the 'Fast-in-Slow' architecture itself: partially sharing parameters between a VLM-based System 2 and an embedded System 1. The strongest evidence for this design would be a matched control where System 1 is a separate policy head with the same capacity, same conditioning features, and same training data. Without that control, the reported gains over CogACT and pi0 are confounded by multiple simultaneous differences: pretraining data, point clouds, robot-state inputs, and the co-training objective. The existing ablations vary the number of shared blocks and input modalities, but never remove the sharing itself. Therefore, while the empirical system is plausible and the paper is clearly written, the central conceptual claim is not yet isolated. This does not change the reader's CONDITIONAL verdict; it strengthens the reason for requesting a controlled comparison before the architectural paradigm is accepted.","tokens_in":24903,"tokens_out":5413,"duration_ms":57634,"concrete_test":"Implement a matched 'external-head' variant of FiS-VLA: cut at LLM block 30, and mount a randomly initialized 2-block transformer policy head (same hidden size, same diffusion conditioning) outside the VLM, receiving the same block-30 latents and the same high-frequency inputs (images, point cloud, robot state, noised action). Train it with the identical Lfast + Lslow recipe, data mixture, and RLBench fine-tuning budget (including the 860K-trajectory pretraining) as FiS-VLA. If this variant matches FiS-VLA's 0.69 ± 0.03 mean success rate within variance, the embedded-System-1 architecture is not the cause of the reported gains; if it is clearly worse (e.g., more than a 5% gap), the central architectural claim is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The contribution is a specific architectural hypothesis: System 1 should partially share parameters with System 2, because a separate policy head fails to fully leverage the VLM's pretrained knowledge. To make that hypothesis load-bearing, FiS-VLA must be compared with a policy head of matched capacity and inputs that is external to the VLM but conditioned on the same intermediate features (block 30 latents, 2D image, point cloud, robot state, noised action). The paper never runs this control. Table 1 compares against CogACT, pi0, OpenVLA, and ManipLLM, but those differ simultaneously in base model, pretraining corpus, observation modalities, action parameterization, and training recipe—not only in whether System 1 is embedded. The ablation in Fig. 3 varies the number of shared blocks (1/2/4/8), input modalities, and slow/fast frequency ratio; none of these conditions removes the parameter sharing while holding all else fixed. The pretraining on 860K trajectories and the 3D point-cloud/state inputs are themselves candidate explanations for the +8% simulation gain. Thus the central 'Fast-in-Slow' mechanism is confounded. This does not invalidate the empirical system, but it means the paper's headline conceptual claim is not yet established by the reported experiments.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FiS-VLA, a 7B vision-language-action model that embeds a fast diffusion-based action module (System 1) inside the final transformer blocks of a pretrained VLM-based System 2, rather than using a separate policy head. System 2 runs at low frequency on language and 2D images, producing intermediate latent features; System 1 runs at high frequency on 2D images, 3D point clouds, and robot state, producing end-effector or joint actions. The two systems are co-trained with a diffusion denoising loss for System 1 and an autoregressive next-token loss for System 2. The model is pretrained on 860K trajectories and fine-tuned on RLBench and two self-collected dual-arm real-world datasets. Reported results include 69% average success on 10 RLBench tasks versus 61% for CogACT, real-world success rates of 68% and 74% on Agilex and AlphaBot versus 59% and 61% for pi0, and control frequencies of 21.9 Hz at action chunk 1 and a claimed 117.7 Hz at action chunk 8.","tokens_in":25120,"tokens_out":3724,"duration_ms":35142,"significance":"If the central architectural hypothesis is substantiated, the paper makes a useful conceptual and engineering contribution: it challenges the common design of attaching a separate fast policy head to a VLM, and it demonstrates a concrete way to share parameters between reasoning and execution within a single foundation model. The paper's strengths include a large-scale pretraining recipe, systematic ablations over shared block count, input modalities, and frequency ratio, evaluation across two real dual-arm platforms with different control modes, and a failure-case appendix that adds practical value. However, the paper's headline conceptual claim is currently confounded by the absence of a matched separate-head control, and the speed claim is presented inconsistently as both measured and theoretical.","major_comments":[{"comment":"The paper's central claim is that embedding System 1 inside the VLM's final blocks is what improves manipulation performance. This claim is not tested against a matched separate-head baseline. The comparisons in Table 1 (ManipLLM, OpenVLA, pi0, CogACT) differ simultaneously in base model, pretraining corpus, observation modalities, action parameterization, and training recipe, while the ablations in Figure 3 vary the number of shared blocks but never remove parameter sharing while holding all other factors fixed. The observed +8% simulation gain over CogACT could therefore be explained by the added point-cloud/state inputs, the 860K-trajectory pretraining, or the increased effective capacity, rather than by the Fast-in-Slow architecture. Please add a control policy head of matched capacity and inputs that is external to the VLM but conditioned on the same block-30 intermediate features, 2D image, point cloud, robot state, and noised action, trained with the same pretraining and fine-tuning protocol. This control is load-bearing for the paper's conceptual contribution.","section":"§3.2, §4.2, Table 1"},{"comment":"The 117.7 Hz figure is presented as an achieved control frequency in the abstract and introduction, but Appendix B.1 states that 'the theoretical control frequency reaches up to 117.7 Hz' when predicting eight future actions. Table 1 reports 21.9 Hz at action chunk one, and no measured end-to-end frequency at chunk eight is reported anywhere. Please clarify whether 117.7 Hz includes the full perception stack (image encoding, point-cloud tokenization, and denoising) or whether it is an idealized arithmetic upper bound derived from action chunking. If it is theoretical, state this clearly in all claims and report a measured closed-loop frequency under the actual deployment conditions.","section":"Abstract, §1, Appendix B.1"},{"comment":"The real-world evaluation is weaker than the simulation evaluation in several ways that affect the headline '11% real-world improvement' claim. Success is determined by human evaluation, results are reported as single values with no variance or number of repeated rollouts, and the only baseline is pi0; CogACT, which is the stronger simulation baseline, is not evaluated in the real world. Given the small number of tasks and the absence of variance, the cross-method differences in Tables 2 and 3 should be interpreted cautiously. Please report the number of rollouts per task, per-task variance or confidence intervals, and, where feasible, include at least one additional strong baseline or temper the real-world superiority claim accordingly.","section":"§4.3, Table 2, §4.4, Table 3"},{"comment":"The paper claims that the dual-aware co-training strategy preserves System 2's reasoning capability while adding action generation, but this is only supported indirectly by the downstream manipulation ablation in which removing Lslow drops success from 69% to 62%. There is no direct evaluation of System 2's language understanding, instruction following, or multimodal reasoning quality before and after co-training. Since the paper's motivation depends on keeping the VLM's pretrained reasoning intact, please add a direct measure of System 2 capability (e.g., performance on a VLM benchmark or instruction-following accuracy) or explicitly limit the claim to the observed manipulation outcomes.","section":"§3.4, §4.2"}],"minor_comments":[{"comment":"The first sentence of the introduction contains a typo: 'undamental' should be 'fundamental'.","section":"§1"},{"comment":"The wording of the 117.7 Hz claim should be unified across the abstract, introduction, and appendix so that the measured and theoretical nature of the number is unambiguous.","section":"Abstract and Appendix B.1"},{"comment":"The symbol D_t is used in Eq. (2) before it is defined; define it before or immediately after the equation.","section":"Eq. (2)"},{"comment":"The caption says 'Left images show the three generalization test scenarios,' but the table as typeset contains no images; either include the images or revise the caption.","section":"Table 3"},{"comment":"There are minor grammar and consistency issues, for example 'Table 2 showcase' should be 'Table 2 showcases', and the baseline name is written both as 'π0' and 'pi0'.","section":"Throughout"},{"comment":"The phrases '8%' and '11%' in the abstract refer to percentage-point differences in success rates; consider using 'percentage points' to avoid ambiguity.","section":"Abstract and §4"},{"comment":"The notation 'θf ⊆ θ' is informal; clarify how the shared-parameter subset relationship is defined.","section":"§3.2"}],"recommendation":"major_revision","confidential_remarks":"This is a strong systems paper with a clear and interesting architectural idea, and the empirical coverage is substantial. The main obstacle is not the validity of the system but the lack of a matched control for the central 'Fast-in-Slow' claim; without it, the paper overstates what the experiments establish. I believe this is fixable within the manuscript's scope by adding the control experiment or by substantially reframing the claims, so I recommend major revision rather than rejection. I would also ask the editor to ensure the speed claim is reported honestly as measured versus theoretical."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read it this way: FiS-VLA is a solid, readable systems paper, and the core idea—embedding a diffusion-based System 1 inside the last two blocks of a 7B VLM instead of attaching a separate policy head—is genuinely new relative to CogACT and pi0. The ablations are more thorough than most: block count, modality mix, slow/fast ratio, action chunk, input variants. The real dual-arm results, including failure cases, add credibility. I rate this conditional rather than skeptical.\n\nThe main soft spot is exactly what the stress-test note says: the central causal claim, that parameter sharing itself is what buys the +8% over CogACT, is never isolated. The full model differs from baselines in pretraining data, point-cloud and state inputs, and training recipe simultaneously. A matched separate-head control—same 860K pretrain, same modalities, same diffusion objective, taking features from block 30—would settle it. Without that, the 'Fast-in-Slow' mechanism remains plausible but unproven as a mechanism. That does not invalidate the system; it just means the paper's conceptual headline outruns its experiments.\n\nTwo smaller issues. First, Appendix B.1 calls 117.7 Hz 'theoretical control frequency', but the abstract states it as achieved; the abstract should carry the same qualifier. Second, the real-world table compares only with pi0, not CogACT, and has no variance; with single evaluations, the 68 vs 59 and 74 vs 61 gaps are believable but not quantified. Also, the overlap with HybridVLA from the same group is not mentioned; readers should be told exactly how FiS-VLA differs.\n\nAlso note the paper's own limitation section admits static configuration of shared blocks and frequency ratio; that is honest and should stay.\n\nOverall: this is a serious paper, written by people who know the area, with a reproducible-ish recipe but no code or weights yet. It deserves a real referee slot. I would ask the authors to run the matched separate-head control, report real-world variance, and fix the Hz wording; then the empirical claims would stand.","headline":"Solid dual-system VLA with a genuinely new shared-parameter twist; the mechanism itself is confounded by pretraining and modality changes, but the system and ablations still deserve review.","tokens_in":25740,"tokens_out":1983,"would_cite":true,"duration_ms":18449,"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 single 7B VLM can serve as both slow reasoner and fast robot controller.","keywords":["vision-language-action model","dual-system architecture","fast and slow systems","diffusion policy","autoregressive reasoning","robotic manipulation","high-frequency control","point cloud conditioning"],"falsifier":"Run the co-trained model on a standard vision-language reasoning benchmark that the untouched base VLM passes; a large drop relative to the base would contradict the claim that System 2's reasoning is preserved. Alternatively, replace the block-30 latent features given to System 1 with random or frozen features in the same training setup: if success stays near 69 percent, the latent-conditioning mechanism is not the source of the reported gain.","tokens_in":24660,"feed_emoji":"🤖","tokens_out":4680,"duration_ms":50740,"temperature":0.7,"pith_summary":"The paper tries to show that a robot's fast, reactive action loop does not need a separate lightweight policy model; instead, the final transformer blocks of an intact vision-language model can themselves become the fast executor. If true, this would mean internet-scale pretrained knowledge, which normally lives in the slow reasoning layers, is directly available to the high-frequency control loop, closing the usual gap between smart planning and quick acting. The authors embed System 1 inside System 2 by sharing parameters, give the two systems different input modalities and asynchronous operating frequencies, and train them jointly with both a diffusion action loss and an autoregressive next-token loss. On their reported evaluations, the approach lifts average success by 8 percent in simulation and 11 percent on two real-world dual-arm platforms relative to prior state-of-the-art dual-system methods, while pushing control frequency above 20 Hz at action chunk one and over 117 Hz at chunk eight.","feed_headline":"Robot actions at 117 Hz from one VLM brain","feed_subtitle":"Putting fast execution inside the slow reasoner beats separate dual-system models by 8-11%.","key_machinery":"The load-bearing object is the shared-block split: the final $N$ transformer blocks of a pretrained 7B LLM are detached from the autoregressive path and turned into the fast System 1, while the earlier blocks plus the full LLM remain the slow System 2. System 2's intermediate latent features, taken after block 30, periodically condition System 1, and System 1 additionally reads high-frequency modalities, including a lightweight 3D tokenizer whose point-cloud tokens pass through the shared vision encoder. Training uses $\\mathcal{L}_{\\mathrm{FiS\\text{-}VLA}} = \\mathcal{L}_{\\mathrm{fast}} + \\mathcal{L}_{\\mathrm{slow}}$, where $\\mathcal{L}_{\\mathrm{fast}}$ is a diffusion noise-prediction objective over action chunks and $\\mathcal{L}_{\\mathrm{slow}}$ is a cross-entropy next-token objective over discrete actions or language, so that the two objectives share one parameter set and each system's representation is preserved.","core_discovery":"The central claim is that a unified vision-language-action model can house both Kahneman-style systems in one network: the slow, deliberate reasoning path keeps the whole transformer, while the final two transformer blocks are repurposed as the fast System 1 execution module that generates continuous action chunks through diffusion. System 2 runs at low frequency on language and 2D images, emitting latent features from block 30 that condition System 1, which runs at high frequency on 2D images, 3D point clouds, and robot state. A dual-aware co-training objective, the sum of a diffusion denoising loss and an autoregressive next-token prediction loss, teaches System 1 to act without destroying System 2's reasoning. The paper reports 69 percent mean success on a 10-task simulated manipulation benchmark versus 61 percent for the previous best method, 68 and 74 percent on two real-world dual-arm platforms versus 59 and 61 percent, and control frequencies of 21.9 Hz at action chunk one and 117.7 Hz at chunk eight.","pith_inferences":["The paper does not directly evaluate System 2's language or reasoning quality after co-training; a standard VLM benchmark run on the final checkpoint would test the preservation claim outside the action domain.","The fixed two-block boundary and fixed 1:4 frequency suggest a natural extension to adaptive gating, where the model decides how many blocks and what update rate a task needs; the authors themselves flag dynamic adaptation as future work.","The same 'repurposed blocks inside a pretrained model' trick may transfer to other continuous high-frequency inputs, such as tactile or audio streams, though only 2D images, point clouds, and robot state are tested here.","The claimed gain rests on intermediate latent features being genuinely informative; a direct comparison between real block-30 features and frozen or randomized conditioning features would isolate whether the latent guidance or the shared parameters carry the improvement."],"forward_implications":["A single network, not two separately trained models, can deliver both deliberate reasoning and real-time control, with a 1:4 System 2 to System 1 frequency ratio sufficient for stable manipulation.","Control-frequency gains scale with action chunking: predicting eight actions in one forward pass reaches 117.7 Hz, meaning the policy network stops being the bottleneck for real-time closed-loop robot control.","Co-training the autoregressive objective is part of action accuracy, not just a side effect: dropping $\\mathcal{L}_{\\mathrm{slow}}$ lowers simulated success from 69 to 62 percent.","Because the fast executor shares the VLM's pretrained layers, scaling up the underlying VLM should improve execution precision without retraining a separate action head from scratch."],"supporting_citations":[{"why":"Supplies the pretrained VLM whose weights initialize the whole FiS-VLA model.","marker":"[16]"},{"why":"Establishes the VLA training recipe and data preprocessing pipeline that the paper builds on.","marker":"[7]"},{"why":"The main simulation baseline and the prior dual-system design with a separate diffusion policy head.","marker":"[22]"},{"why":"The real-world dual-arm baseline that FiS-VLA is compared against.","marker":"[23]"},{"why":"Prior evidence that intermediate LLM representations, rather than final-layer outputs, can drive action generation.","marker":"[24]"},{"why":"Provides the diffusion action-generation objective used for System 1.","marker":"[3]"},{"why":"Large-scale cross-embodiment dataset used in the 860K-trajectory pretraining corpus.","marker":"[19]"},{"why":"Large-scale in-the-wild robot manipulation dataset used in pretraining.","marker":"[20]"},{"why":"Motivates the fast-versus-slow division of labor that the architecture instantiates.","marker":"[21]"},{"why":"The strategy of tokenizing point clouds and feeding them through the shared vision encoder to reuse pretrained alignment.","marker":"[30]"}],"fun_headline_variants":["Fast-in-Slow: unified VLA hits 117.7 Hz, beats separate systems by 8-11%","One model, two speeds: Fast-in-Slow achieves 117.7 Hz control","Robot actions at 117.7 Hz from a single VLM brain, beating dual-system by 8-11%","Fast-in-Slow: a single VLA with slow reasoning and 117.7 Hz action bursts","117.7 Hz robot control from a unified slow-fast policy (Fast-in-Slow)"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The intermediate latent features from block 30 of the VLM carry enough task-relevant semantic and spatial information for the final two blocks to produce precise high-frequency actions, and co-training with the diffusion objective does not silently destroy System 2's reasoning ability.","fun_headline_variants_meta":{"raw":{"variants":["Fast-in-Slow: unified VLA hits 117.7 Hz, beats separate systems by 8-11%","One model, two speeds: Fast-in-Slow achieves 117.7 Hz control","Robot actions at 117.7 Hz from a single VLM brain, beating dual-system by 8-11%","Fast-in-Slow: a single VLA with slow reasoning and 117.7 Hz action bursts","117.7 Hz robot control from a unified slow-fast policy (Fast-in-Slow)"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000707,"raw_usage":{"total_tokens":3249,"prompt_tokens":1071,"completion_tokens":2178,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":687,"completion_tokens_details":{"reasoning_tokens":2050}},"tokens_in":687,"tokens_out":2178,"duration_ms":16071,"temperature":1.0,"reasoning_tokens":2050,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T11:30:02.128502+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the co-trained model on a standard vision-language reasoning benchmark that the untouched base VLM passes; a large drop relative to the base would contradict the claim that System 2's reasoning is preserved. Alternatively, replace the block-30 latent features given to System 1 with random or frozen features in the same training setup: if success stays near 69 percent, the latent-conditioning mechanism is not the source of the reported gain.","supporting_citations":[],"review_version":1}