{"id":"a80787f6-229e-4340-a420-84edbf09de17","arxiv_id":"2606.16447","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Naive long-context Diffusion Policies succeed with UNet+Cross-Attention and sufficient data; variable-history training cuts sample complexity in the low-data regime.","lead":"Longer observation histories for diffusion robot policies work better than prior work claimed, if you use UNet with cross-attention and enough data. A multi-length training curriculum further helps when data is scarce, challenging the idea that naive long context is brittle.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"Single-run best-checkpoint selection can systematically favor long-context policies that overfit earlier, undermining architecture and curriculum rankings.","rationale":"The reader already flags single-seed training and best-of-checkpoints as the weakest assumption and correctly assigns CONDITIONAL. My stress-test confirms that this is the single most load-bearing concern for the strongest claim: without multi-seed evidence, the large multi-factor sweep cannot reliably establish that naive scaling “is not as brittle as advertised” or that Cross-Attention / variable history are robust remedies. No deeper internal inconsistency or experimental design flaw was found; the paper is transparent about the limitation (B.3, D.3). Therefore the verdict stays CONDITIONAL and agreement is full. The concrete multi-seed re-run is the minimal check that would either solidify or demote the claim.","tokens_in":20181,"tokens_out":538,"duration_ms":4902,"concrete_test":"Re-train the three architectures (UNet+FiLM, UNet+Cross-Attention, DiT) and the progressive+short variable-history method for push-and-return and square at N/2 and N, To=20 (or 80), with three independent random seeds each; report mean±std of best-checkpoint success and of a fixed late-checkpoint success. If any architecture or curriculum ranking reverses or the mean gap shrinks below the original Wilson intervals, the headline claim weakens.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim that UNet+Cross-Attention + naive scaling (and variable-history training) yields high success even at To=80–92 rests on one training run per configuration, with the reported success taken as the max over a set of checkpoints that includes early/late snapshots chosen partly by closed-loop success (Appendix B.3). The paper itself notes that longer-context policies in the N/2 regime often converge earlier and that training/validation loss can improve while closed-loop success falls (Appendix D.3, Fig. 12). Because checkpoint selection is success-aware and multi-seed variance is unmeasured, the apparent ranking of architectures (Fig. 7) and the low-data gains of progressive+short variable history (Fig. 8) could be inflated by lucky early checkpoints rather than by a stable inductive bias. Wilson intervals on 200 rollouts capture only evaluation noise, not training stochasticity. If seed-to-seed variance is comparable to the reported gaps, the corrective claim against prior “naive scaling is brittle” literature is not yet secure.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper systematically studies how Diffusion Policy performance changes as observation context length To is scaled from short to long (up to 80–92). Across five tasks that vary in local manipulation stability and memory demand, three data regimes (N/2, N, 2N), and three architectures (UNet+FiLM, UNet+Cross-Attention, DiT), the authors report that naive long-context scaling is not catastrophically brittle when UNet+Cross-Attention is used and data is at usual single-task scale. They introduce Variable History Training (Algorithm 1), a curriculum over multiple context lengths that improves low-data long-context success while preserving high-data performance, and re-examine past-token prediction with frozen encoders. Supporting metrics (task success, manipulation completion, contextual success) and limited hardware results on marshmallows (To=92) are provided.","tokens_in":20520,"tokens_out":1258,"duration_ms":10625,"significance":"If the empirical picture holds, the work revises a widely cited premise in long-context imitation learning: that naive history scaling fails and therefore requires heuristic compression, VLM filtering, or auxiliary losses. The breadth of the sweep (~200 policies), the introduction of memory-demanding tasks with interpretable success decompositions, the architecture ablation that isolates UNet+Cross-Attention as a useful inductive bias, and the simple curriculum (Algorithm 1) are concrete contributions that practitioners can adopt. The re-evaluation of past-token prediction and encoder freezing is also useful. Strengths include clear success metrics, public website/code intent, and explicit discussion of failure modes (overfitting vs. memory tracking).","major_comments":[{"comment":"Appendix B.3 and D.3: each reported success rate is from a single training run, with the best of many checkpoints chosen by closed-loop success (including early/late snapshots). The paper itself notes that longer-context policies in the N/2 regime often converge earlier and that training/validation loss can fall while closed-loop success falls (Fig. 12). Wilson intervals on 200 rollouts capture only evaluation noise. Without multi-seed variance, the architecture ranking in Fig. 7 and the low-data gains of progressive+short variable history in Fig. 8 could be inflated by lucky early checkpoints. At least 2–3 seeds on the key comparisons (UNet+xAttn vs FiLM at long To; variable history vs naive at N/2) are needed for the central claim against prior “naive scaling is brittle” literature to be secure.","section":null},{"comment":"Section 4.1 / Fig. 7 and Appendix E.1: DiT is shown to fail, but the authors substantially alter the original DiT design (more layers, non-causal attention within the action chunk, extra sampling steps) while still reporting catastrophic failure on Drake tasks. The claim that DiT “should clearly not be chosen as a baseline” is therefore only partially supported; either a closer reproduction of Torne et al.’s hyperparameters or an explicit statement that even a capacity-matched, non-causal DiT fails is required so that the architecture comparison fairly represents prior long-context work.","section":null},{"comment":"Section 3.2 and Appendix D.2: the claim that sample complexity of long-context learning is primarily dictated by local stability of the manipulation primitive rests on a small set of tasks (lift/grasp-and-return vs push-T/square/push-and-return) and a single contact-offset diagnostic (Table 4). The axiom is plausible but not yet isolated from other confounds (action dimensionality, data-generation method, observability of phase). A controlled ablation that holds memory structure fixed while varying only local stability would strengthen this load-bearing interpretation.","section":null}],"minor_comments":[{"comment":"Algorithm 1: the notation Tpast_p(m)=min{m,Tpast_p} and the role of ρ_i are clear, but the recommended default (progressive+short vs random sprinkle+full) is only stated in prose in Appendix E.2; a short decision rule in the main text would help practitioners.","section":null},{"comment":"Figure 1 caption and abstract claim “first study to investigate context length … at this level of detail”; Mark et al. [9] already provide limited data-scaling ablations. Soften the priority claim or cite their agreement more precisely.","section":null},{"comment":"Table 1: parameter counts for FiLM grow with To while xAttn stays fixed; the main comparison keeps UNet channel widths fixed rather than total parameters. Appendix E.1 partially addresses this, but a one-sentence pointer in Section 4.1 would avoid confusion.","section":null},{"comment":"Typos / wording: “auxilliary” (Section 1.1), “datagrams” (Fig. 1 / Algorithm 1) is nonstandard for trajectory segments; “insentive” (Appendix B.3).","section":null},{"comment":"Hardware: marshmallows uses only 20 trials and the latest checkpoint; state this limitation more prominently when claiming high success at To=92.","section":null}],"recommendation":"major_revision","confidential_remarks":"The single-seed, success-aware checkpoint selection is the main risk to the corrective claim against prior work. If the authors can add a modest multi-seed study on the critical cells, the paper becomes a strong empirical contribution for a robotics venue. Scope is appropriate for a methods/empirical robotics journal; novelty is incremental but the breadth of the sweep is valuable."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The punchline is simple: this paper shows that “naive long context is brittle” was overstated for single-task Diffusion Policy. With UNet+Cross-Attention and ordinary data scales, success holds or improves out to To=80–92 on several tasks; DiT is the architecture that actually collapses. Variable-history training (Algorithm 1) then closes the remaining low-data gap without hurting the high-data regime.\n\nWhat is new is the systematic sweep—task local stability × data scale × architecture × context length—plus the multi-context curriculum and a careful re-check of Torne et al.’s past-token + frozen-encoder recipe. The invented push/grasp-and-return tasks cleanly separate manipulation completion from contextual success, which is useful. The main claim is measured on held-out rollouts, not circular. Citations are standard baselines, not definitional loops. Hardware on marshmallows (To=92, 100 demos) is limited but consistent with the sim story.\n\nSoft spots, in proportion: one training seed per config and success-aware best-checkpoint selection (Appendix B.3) can favor early-converging long-context runs, exactly as the paper notes in D.3 and Fig. 12. Wilson intervals only cover evaluation noise. That weakens the precision of architecture and curriculum rankings more than it kills the directional claim—the data-scale and architecture trends are large and repeated across tasks. Free parameters (N, p=0.8, short vs full past horizon) are ordinary hyperparams, not hidden fitting. Inference cost of long image histories is acknowledged but unsolved.\n\nThis is for people who train single-task visuomotor diffusion policies and for anyone citing “naive history fails.” It will not rewrite multi-task VLA memory, but it should change how we set baselines. I would send it to peer review; the evidence is sharp enough to deserve referee time even if multi-seed ablations are requested. Engage with it.","headline":"Solid multi-factor empirical corrective: naive long-context Diffusion Policy works with UNet+Cross-Attention and enough data; variable-history training is a practical low-data fix. Single-seed best-checkpoint selection is the real soft spot, not a load-bearing collapse.","tokens_in":21072,"tokens_out":514,"would_cite":true,"duration_ms":5450,"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":"Naively scaling Diffusion Policy context length is not as brittle as prior work claimed.","keywords":["diffusion policy","imitation learning","long context","robotic manipulation","cross-attention","variable history training","sample complexity"],"falsifier":"Retrain the same UNet+Cross-Attention, UNet+FiLM, and DiT configurations on push-and-return and square at N/2 and N with several random seeds; if long-context Cross-Attention no longer outperforms FiLM or if naive long context still collapses relative to short context across seeds, the central claim fails.","tokens_in":21116,"feed_emoji":"🤖","tokens_out":912,"duration_ms":11464,"temperature":0.7,"pith_summary":"Robotic imitation policies usually see only a short slice of recent camera frames, so they cannot remember earlier events and often loop on the same failed motion. This paper asks whether simply feeding much longer observation histories into Diffusion Policies really fails as badly as recent papers have said. Across five tasks that differ in how stable the manipulation is and how much memory they need, and across three data budgets per task, the authors find that with a UNet denoiser conditioned by cross-attention, long histories often work well once you have a normal amount of demonstration data. They also introduce a curriculum that trains one policy on many history lengths at once, which closes much of the remaining gap when data is scarce, and they re-examine an earlier past-action prediction trick, finding that freezing the vision encoder is doing more work than previously advertised.","feed_headline":"Long-context Diffusion Policies work without fancy memory tricks","feed_subtitle":"UNet plus cross-attention and multi-length training close the gap prior work called brittle","key_machinery":"UNet+Cross-Attention conditioning, which keeps per-timestep observation tokens separate and cross-attends them into the UNet rather than collapsing them through FiLM; plus Variable History Training, which samples training windows from a curriculum of shorter and longer context lengths so the same policy learns both short-horizon control and useful memory.","core_discovery":"With an appropriate conditioning method and denoising backbone (UNet plus cross-attention), single-task Diffusion Policies achieve high success rates on many robotic manipulation tasks in the usual data regime even when context length is naively scaled to long horizons (up to tens of frames, and in one hardware case 92). The sample complexity of long-context learning is driven largely by how locally stable the manipulation primitive is, not by history length alone.","pith_inferences":["If long-context failure is mainly covariate shift from overfitting, similar multi-length curricula may help other visuomotor policy classes beyond diffusion.","Locally stable prehensile tasks may systematically understate the data cost of long memory; hard contact and non-prehensile tasks remain the stricter testbed.","Inference-time adaptive context length (short when local control is enough, long when memory is needed) is a natural next step for policies already trained on variable histories.","Hardware success with 100 demos and To=92 suggests some memory-heavy kitchen tasks tolerate rough grasps enough that naive scaling can already be practical."],"forward_implications":["Default single-task Diffusion Policy baselines should prefer UNet with cross-attention when context is long and data is limited.","Long-context failure is often a manipulation-skill learning problem, not a pure memory problem: policies that finish the motion usually also track history.","Variable History Training can be used without first measuring whether the dataset is large enough for naive scaling.","Prior criticism of naive history scaling was partly driven by architecture choice (especially DiT) and missing data-scale ablations.","Past-action prediction alone is not a reliable fix; freezing a short-context vision encoder contributes materially to reported gains."],"fun_headline_variants":["UNet+cross-attention lets Diffusion Policies scale context without memory modules","Naive long context works for single-task Diffusion Policies with right backbone","Local stability—not history length—drives sample needs in long-context Diffusion","Multi-length training cuts sample complexity for long-horizon robotic Diffusion","Long-context Diffusion Policies succeed on many tasks without fancy memory"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"That ranking architectures and curricula from single training runs (best closed-loop checkpoint, Wilson intervals on 200 rollouts) is enough, even though multi-seed training variance is not measured.","fun_headline_variants_meta":{"raw":{"variants":["UNet+cross-attention lets Diffusion Policies scale context without memory modules","Naive long context works for single-task Diffusion Policies with right backbone","Local stability—not history length—drives sample needs in long-context Diffusion","Multi-length training cuts sample complexity for long-horizon robotic Diffusion","Long-context Diffusion Policies succeed on many tasks without fancy memory"]},"model":"grok-4.5","effort":"low","cost_usd":0.003058,"raw_usage":{"total_tokens":1075,"prompt_tokens":752,"num_sources_used":0,"completion_tokens":96,"cost_in_usd_ticks":30580000,"prompt_tokens_details":{"text_tokens":752,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":227,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":752,"tokens_out":96,"duration_ms":2943,"temperature":1.0,"reasoning_tokens":227,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-12T13:48:26.958606+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Retrain the same UNet+Cross-Attention, UNet+FiLM, and DiT configurations on push-and-return and square at N/2 and N with several random seeds; if long-context Cross-Attention no longer outperforms FiLM or if naive long context still collapses relative to short context across seeds, the central claim fails.","supporting_citations":[],"review_version":1}