{"id":"b1b44895-d358-40fc-90e4-1fe5b36e83f5","arxiv_id":"2607.19919","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"Selective re-noising (ReRoll) during diffusion decoding improves long-horizon robot planning and policy success on maze, LIBERO-10, and unified video-action benchmarks.","lead":"A new diffusion decoding scheme, Diffusion ReRoll, selectively re-noises parts of a predicted robot trajectory while other parts continue denoising, so earlier and later segments can revise each other. Across maze planning, LIBERO-10 policy learning, and video-action modeling, it reports sizable success-rate gains over Diffusion Forcing and Diffusion Policy.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"ReRoll-event ablation conflates re-noising with extra DDIM updates; no same-step no-reset control is shown.","rationale":"The reader's CONDITIONAL verdict is appropriate, and this concern does not change it. The main empirical comparisons are not simply a compute artifact: ReRoll is compared against DF at matched DDIM rows in planning and against DP with a larger sampling budget in policy. However, the internal evidence that selective re-noising is the active mechanism is not fully controlled. The Nroll ablation varies both the presence of resets and the total number of DDIM updates, so the conclusion that 'gains are largely driven by selective re-noising' is not established by that figure alone. A single additional control—matching DDIM rows while removing resets—would settle whether the mechanism is re-noising or simply more denoising steps. This is an addressable experimental gap rather than a fundamental flaw, so the verdict remains CONDITIONAL.","tokens_in":27451,"tokens_out":8969,"duration_ms":99985,"concrete_test":"On AntMaze Giant guidance-based planning, use the DR-trained checkpoint and deploy a monotone forward schedule with no resets (Nroll=0) but with the denoising slope lowered so that the number of DDIM rows equals the six-event ReRoll matrix (202 rows). Run three seeds. If success is near 86% (the reported six-event ReRoll value), the re-noising mechanism is not necessary to explain the gain; if success stays near the 124-row zero-event baseline, re-noising is the active ingredient.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central mechanistic claim—that structured re-noising, not extra denoising compute, drives gains—rests heavily on the ReRoll-event ablation in Sec. 4.1. Table 7 shows that increasing Nroll from 0 to 6 changes the schedule from 124 to 202 DDIM rows in planning, and from 11/23/31 to 29/53/85 updates in the policy settings. Fig. 7 therefore compares a re-noised schedule against a causal no-reset schedule with far fewer denoising updates. A monotonic no-reset schedule with the same 202-row budget might match the six-event ReRoll result, which would make 'revisable denoising' a label for extra DDIM budget rather than a distinct mechanism. The main comparisons are roughly compute-matched (ReRoll 202 rows vs DF 199; ReRoll policy uses fewer calls than the 100-step DP baseline), so the headline empirical gains are not invalidated. The weak spot is specifically the attribution of those gains to selective re-noising as opposed to additional denoising iterations.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"Diffusion ReRoll (DR) is a scheme for robotic sequential prediction that generalizes per-token noise conditioning à la Diffusion Forcing. The schedule matrix is organized into virtual piecewise-linear 'chunks'; when a chunk's leading edge reaches a reset noise level, a ReRoll event re-noises that region to full noise and attaches it to an adjacent chunk, while other regions continue denoising. Training uses randomized linear-chunk noise profiles, and the paper introduces forward and bidirectional schedule variants. The method is evaluated on OGBench maze planning, LIBERO-10 and RoboCasa policy learning, and UWM unified video-action prediction, reporting large relative gains over Diffuser, Diffusion Forcing, and Diffusion Policy, plus schedule ablations and an analysis of two LIBERO-10 reset protocols.","tokens_in":27714,"tokens_out":10531,"duration_ms":110852,"significance":"If the reported results hold, DR identifies a new and potentially useful design axis for diffusion sequence predictors—non-monotonic, structured re-noising—and demonstrates gains across three diverse robotics benchmarks. The paper's transparency is a notable strength: matched DF/DR architectures, the zero-event ablation, the action-representation control, the causal-masking control, and the explicit reporting of two LIBERO-10 reset protocols. However, the central mechanistic attribution is currently weakened by a compute confound: the ReRoll-event ablation varies both the number of re-noising events and the number of DDIM updates. Thus, the empirical contribution is solid as a system-level result, but the claim that 'revisable denoising' is the active ingredient needs an additional control experiment before it can be regarded as established.","major_comments":[{"comment":"The central attribution claim—'DR's gains are largely driven by selective re-noising'—is not yet isolated from inference compute. In the ReRoll-event ablation, increasing Nroll from 0 to 6 changes the schedule matrix from 124 to 202 DDIM rows in planning and from 11/23/31 to 29/53/85 updates in policy settings (Table 7). The zero-event condition is therefore not a same-budget no-reset control; a monotonic causal schedule with the same number of rows might recapture most of the gain. Please add a control that matches the total DDIM update count while keeping the schedule monotonic (e.g., a slower causal ramp with the same number of rows, or additional intermediate rows), and re-evaluate the conclusion in Sec. 4.1. Without this control, the reported gains remain consistent with 'more denoising steps' as an alternative explanation.","section":"Sec. 4.1, Fig. 7, Table 7"},{"comment":"The abstract's headline '56.5% relative gain' over Diffusion Policy on LIBERO-10 is based on the official gripper-reset protocol. Under the open-gripper protocol reported in Table 2, the same 2/16/8 comparison is 82.5% vs. 76.2% (about 8% relative), and the average improvement over the five settings drops to roughly 18% relative. The official 2/16/8 ReRoll result also has a very large standard deviation (51.0 ±24.1). The paper does transparently report both protocols and discusses the reset discrepancy, but the summary and conclusion select the most favorable number. Please add a protocol caveat in the abstract/conclusion, or report the average across both protocols, so the reader is not left with an overstatement of the policy-learning gain.","section":"Abstract and Sec. 4.3, Table 2, Appendix C.3.2"}],"minor_comments":[{"comment":"'Relative gains' are defined only in Appendix C.1; consider a parenthetical in the abstract (e.g., 'relative to the baseline average') to avoid misreading as absolute percentage-point gains.","section":"Abstract"},{"comment":"Please add a footnote to Table 2 pointing to Appendix C.3.2 and D.3 for the reset-protocol sensitivity and the per-seed results behind the large standard deviation.","section":"Table 2"},{"comment":"Ensure the y-axis label ('Success rate') is visible in the final figure; the text references fractions but the figure axes are not labeled in the submitted version.","section":"Fig. 7"},{"comment":"M is undefined when introduced. Define M as the number of schedule rows (denoising steps).","section":"Sec. 2.2, Eq. (2)"},{"comment":"The forward noising transition in Eq. (6) uses the same alpha-bar scaling; a short derivation or reference to standard forward diffusion would help readers verify the variance.","section":"Appendix A.1"},{"comment":"The Diffuser guidance results in Table 1 are very low (e.g., 16% on AntMaze Giant). Please state whether this reflects the re-enabled OGBench start-goal randomization and whether the Diffuser model was trained with the same guidance objective, so readers can compare with other Diffuser evaluations.","section":"Sec. 4.2"}],"recommendation":"major_revision","confidential_remarks":"The paper makes a strong empirical case for a new schedule-based mechanism, and its transparency (matched baselines, protocol sensitivity, per-seed results) is a credit. My main concern is the compute confound in the event ablation; I believe a single additional control would resolve it. I recommend major revision, with the expectation that the paper can be accepted after that experiment and a more balanced headline."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Diffusion ReRoll is a genuine new mechanism for diffusion-based sequence generation, and the empirical work is mostly careful. The schedule matrix with linear chunks and explicit ReRoll events is a real extension of Diffusion Forcing's per-token noise conditioning, and the co-designed training on randomized linear chunks is sensible and shown to matter. The evaluation spans planning, policy learning, and video-action prediction, which is broader than most papers in this area.\n\nWhat the paper does well: comparisons are mostly matched — same architecture for DF versus ReRoll, same action representation — and the ablations for the action representation and for linear-chunk training are clean. The authors also deserve credit for disclosing the gripper-reset protocol sensitivity in LIBERO and reporting per-seed numbers for the high-variance setting. That is honest reporting.\n\nThe soft spots are real but not fatal. First, the ReRoll-event ablation in Fig. 7/Table 7 confounds re-noising with extra DDIM rows: more events also mean more denoising updates. There is no no-reset schedule with the same DDIM budget, so the mechanistic claim that selective re-noising, rather than additional compute, drives the gains is not fully isolated. This matters for the framing, but the main comparisons against DF, Diffuser, and Diffusion Policy are roughly compute-matched, so the headline results are not invalidated. Second, no code or data is released, so the complex training/sampling pipeline cannot be independently checked. Third, the unified video-action table has no error bars, and the bidirectional schedule comparison is single-seed. Fourth, the LIBERO headline gain is protocol-sensitive: under the open-gripper reset the advantage over Diffusion Policy shrinks substantially. The authors transparently note this, but it should temper how the 56.5% number is quoted.\n\nThis paper deserves a serious referee. The core idea is plausible, the experiments are broad, and the flaws are addressable. I would ask for a same-budget no-reset control, code release, and error bars on the unified table before accepting. If the control shows re-noising is mostly extra compute, the framing would need to change, but the empirical contribution would still stand.","headline":"Solid empirical paper with a real new mechanism; the core ablation doesn't fully separate re-noising from extra DDIM budget, but the compute-matched baseline gains hold up.","tokens_in":28221,"tokens_out":2160,"would_cite":true,"duration_ms":24922,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Diffusion ReRoll makes robotic sequence generation revisable by selectively re-noising locally stable regions, so earlier and later segments can refine one another during denoising, and reports consistent success gains over monotonic and ca","keywords":["diffusion models","robotic sequential prediction","revisable denoising","schedule matrix","linear chunks","cross-horizon revision","policy learning","long-horizon planning"],"falsifier":"Take a denoiser trained with independently sampled per-token noise and deploy the exact ReRoll schedule matrix on AntMaze Giant guidance planning and LIBERO-10; if its success matches the linear-chunk-trained ReRoll model, the paper's claim that structured chunk training is necessary would be refuted. A second decisive check: on the same trained model, compare zero ReRoll events against six; if success does not rise with event count, the gains would come from something other than re-noising.","tokens_in":27265,"feed_emoji":"🤖","tokens_out":8026,"duration_ms":86920,"temperature":0.7,"pith_summary":"Diffusion ReRoll argues that a diffusion sequence predictor need not commit monotonically to an early plan. It trains a per-token-noise denoiser on randomized piecewise-linear 'linear chunks' and then deploys a schedule matrix that periodically returns selected horizon regions to high noise—ReRoll events—while the rest of the sequence keeps denoising. This lets early and late segments revise each other before a final plan is produced. The paper's evidence is a set of matched comparisons: a 21% relative gain over Diffusion Forcing in guidance-based maze replanning, a 23% relative gain over Diffuser in goal inpainting, a 56.5% average relative gain over Diffusion Policy on official LIBERO-10, and improved out-of-distribution joint-policy success (57.2% to 71.0%) in unified video-action prediction. A reader should care because closed-loop robots are hurt by early mistakes that later context could correct, and this mechanism makes the correction built into denoising rather than bolted on as a separate repair module.","feed_headline":"Structured re-noising lifts robot planning success up to 56 percent","feed_subtitle":"ReRoll lets early and late segments revise each other, lifting maze, policy, and video-action results.","key_machinery":"The load-bearing object is the schedule matrix with virtual linear chunks. Each entry is the noise level assigned to a token at a given denoising row; linear chunks are piecewise-linear noise profiles across the horizon, so denoising proceeds as waves rather than at a single global level. A ReRoll event is a scheduled upward transition that re-injects full noise into a chunk of tokens, attaching it to the next wave. The matrix alone is not enough: the paper also changes the training noise distribution to randomized linear chunks, and the deployment matrix controls whether information flows forward, from both ends, or mostly from the terminal side.","core_discovery":"The central claim is that revisability can be engineered directly into the diffusion schedule. A schedule matrix assigns each sequence token a noise level at every denoising row; downward rows are denoising updates, while scheduled upward rows are explicit forward noising that resets a token region to full noise. These reset transitions—ReRoll events—let a locally stable segment be re-generated under the updated context of the rest of the horizon. The paper shows that training on randomized linear noise chunks, rather than independent per-token noise, is what makes this work: applying the ReRoll matrix to a Diffusion-Forcing-trained model does not reliably improve success, and adding a causa","pith_inferences":["Left implicit is that the schedule matrix could be adapted per rollout—e.g., increasing ReRoll events in regions flagged as high-risk—without retraining, since the model already saw variable chunk slopes; this would be a cheap test of the revisability interface.","The two-protocol LIBERO results imply that absolute gains depend on evaluation reset details; future comparisons should report the gripper-reset protocol, and the true advantage of re-noising may be smaller under easier resets.","If ReRoll's benefit is an inductive bias from structured noise, then other per-token diffusion backbones, including larger robot models, might inherit it simply by changing their training noise distribution without adding the schedule machinery.","The zero-action diagnostic offers a reusable metric: action-video consistency can serve as an early signal for downstream control robustness under out-of-distribution visual changes, before full policy rollouts are run."],"forward_implications":["Long-horizon planning should improve most where early mistakes are hard to reverse; the largest reported maze gains are in Large and Giant variants, with ReRoll recovering paths that full-sequence diffusion commits to incorrectly.","The same trained model can change its information-flow pattern by swapping the schedule matrix, so forward, backward, and endpoint-conditioned revision become deployment-time choices rather than architecture changes.","Multi-task action-chunk prediction improves across all evaluated horizons and observation histories on LIBERO-10; the paper hypothesizes the gain is largest when multiple plausible action structures compete, which is consistent with the smaller single-task RoboCasa gains.","Unified video-action models gain in out-of-distribution joint-policy success and show the lowest zero-action error, indicating stronger action-video consistency than monotonic or causal variants.","ReRoll's benefit is conditional on training with randomized linear chunks and non-causal attention; the schedule matrix alone, applied to independently noised training, is not enough."],"fun_headline_variants":["Revisable denoising lifts robot success up to 56%","Diffusion ReRoll: re-noise to refine robot plans","Re-noising diffusions let robot segments correct each other","Iterative re-noising boosts robot planning and policy","ReRoll's re-noise step sharpens robot foresight"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"ReRoll's gains rest on the model being trained with randomized linear noise chunks and without causal masking; if that training alignment is absent, the schedule matrix alone does not reliably beat causal denoising, and the headline LIBERO gain also narrows under the open-gripper reset protocol.","fun_headline_variants_meta":{"raw":{"variants":["Revisable denoising lifts robot success up to 56%","Diffusion ReRoll: re-noise to refine robot plans","Re-noising diffusions let robot segments correct each other","Iterative re-noising boosts robot planning and policy","ReRoll's re-noise step sharpens robot foresight"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001004,"raw_usage":{"total_tokens":4104,"prompt_tokens":785,"completion_tokens":3319,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":529,"completion_tokens_details":{"reasoning_tokens":3236}},"tokens_in":529,"tokens_out":3319,"duration_ms":22485,"temperature":1.0,"reasoning_tokens":3236,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T11:17:38.470613+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a denoiser trained with independently sampled per-token noise and deploy the exact ReRoll schedule matrix on AntMaze Giant guidance planning and LIBERO-10; if its success matches the linear-chunk-trained ReRoll model, the paper's claim that structured chunk training is necessary would be refuted. A second decisive check: on the same trained model, compare zero ReRoll events against six; if success does not rise with event count, the gains would come from something other than re-noising.","supporting_citations":[],"review_version":1}