{"id":"b4b498bb-7558-4969-a07e-0d6097778f29","arxiv_id":"2606.08962","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"C³ache caches and reuses cross-chunk residuals at matching denoising steps in WAMs, delivering up to 2.5× inference speedup with negligible drop in task success rate.","lead":"The paper presents C³ache, a training-free technique that reuses denoising residuals across consecutive inference chunks in World Action Models. This targets a previously overlooked redundancy to cut wall-clock inference time by up to 2.5× on robot benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"The assumption that residuals at a fixed denoising step remain strongly correlated across chunks may fail to hold with sufficient precision for long-horizon or non-smooth trajectories.","rationale":"The reader's weakest_assumption directly identifies the same unverified correlation premise that the speedup claim rests on. Because the original verdict was formed from the abstract alone, access to the full manuscript does not remove the need for the quantitative check above; it merely makes the check feasible. The verdict therefore moves from UNVERDICTED to CONDITIONAL pending that measurement.","tokens_in":1656,"tokens_out":393,"duration_ms":10861,"concrete_test":"On the same Fast-WAM backbone and benchmark tasks, instrument the inference loop to record the L2 norm ||r_t^{i} - r_t^{i+1}|| at every denoising step t and chunk boundary i; also run an ablation that recomputes every residual from scratch versus the cached version and report both wall-clock time and task success rate. If the median per-step L2 difference exceeds 5% of the residual magnitude or if success rate drops >3% relative to the recompute baseline, the negligible-degradation claim does not hold.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"C³ache reuses the residual computed at denoising step t of chunk i directly for chunk i+1 at the same t. This is justified by an empirical observation of correlation under smooth robot behavior. However, the central claim of 2.5× wall-clock speedup with negligible success-rate drop requires that any approximation error introduced by reuse does not accumulate or amplify over the full sequence of chunks and denoising steps. No section quantifies the per-step residual difference (e.g., via cosine similarity or L2 norm) as a function of chunk index, behavior smoothness, or task horizon, nor does it bound the downstream effect on the final action or video prediction.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper proposes C³ache, a training-free method for accelerating inference in World Action Models (WAMs) by caching and reusing residuals across consecutive inference chunks at the same denoising step. This exploits an observed correlation in residuals for smooth robot behaviors. On benchmarks using a Fast-WAM backbone, it reports up to 2.5× wall-clock speedup with negligible degradation in task success rate.","tokens_in":1796,"tokens_out":393,"duration_ms":12285,"significance":"If the speedup holds with the claimed negligible impact on success rate, the result would be significant for practical deployment of WAMs, which currently suffer from expensive multi-chunk denoising. The training-free nature and reliance on an empirical correlation (rather than fitted parameters) are strengths that could enable immediate adoption without retraining costs.","major_comments":[{"comment":"The central claim of 2.5× speedup with negligible success-rate drop depends on the assumption that residual reuse introduces no accumulating error. However, no section quantifies the per-step residual difference (e.g., via cosine similarity or L2 norm) as a function of chunk index, behavior smoothness, or task horizon, nor bounds the downstream effect on final actions or predictions.","section":"Method and Experiments sections"},{"comment":"The empirical results report concrete speedup numbers but provide no details on experimental controls, variance across runs, or exact baseline implementations. This leaves the 'negligible degradation' claim only moderately supported, as factors like task horizon or non-smooth trajectories could affect outcomes.","section":"Experiments section"}],"minor_comments":[{"comment":"Notation for residuals and denoising steps could be clarified with an explicit equation or diagram showing the cross-chunk reuse operation.","section":"Method section"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the constructive comments. We address each major comment below.","responses":[{"response":"We agree that explicit quantification of residual differences would strengthen the support for the central claim. While the manuscript presents empirical evidence via observed speedups and success rates under the correlation for smooth behaviors, we will add a new analysis subsection (in Methods) reporting cosine similarity and L2 norms of residuals across chunk indices and denoising steps, plus discussion of downstream effects for varying horizons and smoothness levels.","revision_made":"yes","referee_comment":"[Method and Experiments sections] The central claim of 2.5× speedup with negligible success-rate drop depends on the assumption that residual reuse introduces no accumulating error. However, no section quantifies the per-step residual difference (e.g., via cosine similarity or L2 norm) as a function of chunk index, behavior smoothness, or task horizon, nor bounds the downstream effect on final actions or predictions."},{"response":"We acknowledge the need for fuller experimental reporting. In the revised Experiments section we will add: exact baseline implementation details, number of runs with standard deviations for all metrics, explicit controls for task horizon, and new results on non-smooth trajectories to better substantiate robustness of the negligible-degradation claim.","revision_made":"yes","referee_comment":"[Experiments section] The empirical results report concrete speedup numbers but provide no details on experimental controls, variance across runs, or exact baseline implementations. This leaves the 'negligible degradation' claim only moderately supported, as factors like task horizon or non-smooth trajectories could affect outcomes."}],"tokens_in":1312,"tokens_out":356,"duration_ms":20318,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The main thing to know is that C³ache reuses denoising residuals across inference chunks at the same step to accelerate world action models, claiming up to 2.5 times faster inference with little loss in task success.\n\nWhat is new is the cross-chunk mechanism. Prior acceleration work cached computation inside one chunk's denoising trajectory. This paper observes that residuals correlate between chunks for smooth behaviors and reuses them directly. The method stays training-free and rests on that correlation rather than any learned parameters.\n\nThe paper does well by identifying a redundancy that within-chunk methods miss and by giving concrete speedup numbers on benchmarks with a Fast-WAM backbone.\n\nThe soft spots are around the strength of the evidence. The description provides no details on experimental controls, variance across multiple runs, or how baselines were set up exactly. The stress-test note points out that there's no quantification of residual differences as chunks progress or for different task horizons. If the correlation weakens on non-smooth trajectories, the error could accumulate and affect the action predictions more than claimed. That makes the \"negligible degradation\" part hard to assess without more data.\n\nThis paper is for researchers focused on making video-based robot policies run faster in practice. Someone implementing WAMs would get a useful idea from it, though they would want to verify the correlation assumption in their own setting.\n\nI would send it to peer review. The core claim is specific and the method is easy to test, so referees can evaluate whether the speedup holds under closer scrutiny.","headline":"C³ache reuses residuals across chunks for a claimed 2.5× WAM speedup, but the reported experiments give no controls, variance, or error measurements.","tokens_in":2261,"tokens_out":390,"would_cite":false,"duration_ms":20413,"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":"C³ache reuses residuals across chunks to speed World Action Model inference up to 2.5 times.","keywords":["world action models","inference acceleration","denoising cache","robot policies","diffusion models","cross-chunk reuse","video prediction"],"falsifier":"Measure the actual correlation between residuals of consecutive chunks at each denoising step on a smooth motion sequence; then replace later residuals with the cached values and check whether task success rate remains within a few percent of the uncached baseline.","tokens_in":2565,"feed_emoji":"⚡","tokens_out":585,"duration_ms":10318,"temperature":0.7,"pith_summary":"World Action Models generalize better than standard robot policies because they learn from unlabeled video, yet they run a costly denoising process over multiple inference chunks. The paper shows that during smooth robot motions the residuals at any fixed denoising step remain strongly correlated from one chunk to the next. C³ache stores those residuals once and reuses them for later chunks at the same denoising step. The result is a training-free acceleration that cuts total wall-clock time by up to 2.5 times while task success stays nearly unchanged. Existing single-chunk accelerators can still be applied on top of this cross-chunk reuse.","feed_headline":"Cross-chunk residual cache speeds WAM inference 2.5 times","feed_subtitle":"Reusing correlated denoising residuals between chunks cuts wall-clock time while keeping task success nearly unchanged.","key_machinery":"The cross-inference-chunk residual cache that stores denoising residuals at each step and re-applies them to later chunks instead of recomputing them.","core_discovery":"C³ache caches and reuses the residuals computed at each denoising step across successive inference chunks, exploiting the correlation that appears when a robot performs smooth behavior.","pith_inferences":["The same chunk-to-chunk residual correlation may appear in other diffusion-based video planners outside robotics.","A tunable similarity threshold on cached residuals could let users trade extra speed for higher accuracy when needed.","Monitoring correlation strength in real time could allow dynamic switching between cached and fresh computation for variable-length behaviors."],"forward_implications":["Total wall-clock inference time for a full task drops by up to 2.5 times.","Task success rate shows negligible degradation on standard benchmarks.","World Action Models can run faster without losing their generalization benefit from video pretraining.","Single-chunk acceleration methods remain compatible and can be stacked with the cross-chunk cache."],"fun_headline_variants":["C³ache reuses cross-chunk residuals for 2.5x WAM speedup","Cross-chunk cache speeds WAM inference 2.5 times","Residual reuse across chunks cuts WAM time 2.5x","C³ache exploits chunk correlations to accelerate denoising"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"When a robot executes a smooth behavior, the residuals computed at a given denoising step are strongly correlated from one chunk to the next.","fun_headline_variants_meta":{"raw":{"variants":["C³ache reuses cross-chunk residuals for 2.5x WAM speedup","Cross-chunk cache speeds WAM inference 2.5 times","Residual reuse across chunks cuts WAM time 2.5x","C³ache exploits chunk correlations to accelerate denoising"]},"model":"grok-4.3","cost_usd":0.009994,"raw_usage":{"total_tokens":4399,"prompt_tokens":588,"num_sources_used":0,"completion_tokens":74,"cost_in_usd_ticks":99937000,"prompt_tokens_details":{"text_tokens":588,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":3737,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":588,"tokens_out":74,"duration_ms":19201,"temperature":1.0,"reasoning_tokens":3737,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-27T17:43:35.080062+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"Measure the actual correlation between residuals of consecutive chunks at each denoising step on a smooth motion sequence; then replace later residuals with the cached values and check whether task success rate remains within a few percent of the uncached baseline.","supporting_citations":[],"review_version":1}