{"id":"d505162d-5779-4478-b65e-9f85eac0b97b","arxiv_id":"2607.08193","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A lightweight VLM inspecting SMAC policy videos produces open-ended multi-agent curricula that outperform text-only ablations and PLR scalar-score methods on held-out maps.","lead":"VIP feeds episode videos of multi-agent RL policies to a VLM so the model can recommend the next curriculum task. On SMAC it beats text-only and scalar-score baselines, showing visual inspection can capture learning progress that win-rate alone misses.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"The visual-advantage claim rests on a single short, low-fidelity SMAC video per step; if the VLM is mainly using the win-rate text, the multi-modal superiority is overstated.","rationale":"The reader correctly isolates the weakest link: whether the impoverished SMAC rendering + single short video actually gives the 7B VLM a reliable visual signal of learning progress. The paper’s strongest claim is precisely that this visual channel produces better curricula than the text-only ablation (and than scalar PLR). The ablation already removes the video while keeping the same win-rate and history text, and VIP wins on two of three held-out maps; that is the cleanest evidence the paper supplies. My concern is the natural next stress test of that same assumption: if even a blank-frame control still yields VIP-like curricula, then the VLM is not using the pixels and the multi-modal story is overstated. Because that control is absent, the claim remains conditional on an untested visual-utilization hypothesis. No stronger internal inconsistency appears; the evaluation protocol, hyper-parameter sharing, and PLR baselines are transparent. Hence the verdict stays CONDITIONAL and I agree with the reader’s identification of the load-bearing assumption.","tokens_in":21868,"tokens_out":671,"duration_ms":6507,"concrete_test":"Re-run the exact VIP curriculum loop on the same seeds, but replace every video with a static blank (or random-noise) frame while keeping the identical win-rate string and task-history text; if the resulting curricula and held-out fine-tuning win rates (Table 1 maps) remain statistically indistinguishable from full VIP, the visual channel is not load-bearing.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim (VIP > text-only ablation and PLR on held-out fine-tuning win rates, Table 1 / Fig. 3) requires that the single short episode video (1–10 s, 1024×1024, 25 fps, minimal SMAC rendering of colored circles + abbreviated labels + same-color attack lines; §5.1, App. A.3, Fig. 5) supplies the VLM with usable learning-progress signal beyond the win-rate number and task history already present in the text-only ablation. The paper’s own qualitative examples (Fig. 1, App. D) and the large gap on 3s vs 4z / 3s5z are offered as evidence that the video is decisive. Yet the rendering is acknowledged to be information-poor (same-color fire lines, no clear shooter/target distinction), only one video is supplied per curriculum step (no batching), and the 7B VLM has limited context. If the VLM is largely ignoring the pixels and reacting to the scalar win-rate thresholds already written into the system/task prompts (App. F.3: “if win rate >70% increase difficulty … if <30% decrease”), the multi-modal superiority collapses and the result reduces to a prompt-engineered text curriculum that happens to be paired with a video. That is the single most load-bearing empirical assumption.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper proposes Visual Inspection of Policies (VIP), an open-ended autocurriculum method that feeds episode videos of the current multi-agent policy (plus a short textual summary such as win rate and task history) to a Video Language Model, which then recommends the next interesting task. After sanitization via sentence similarity, the recommended task is used for further MAPPO training. Empirically, VIP is evaluated on SMAC with a lightweight open VLM (VideoLLaMa2-7B). On held-out maps and on frequently seen maps, fine-tuning win rates after curriculum training are higher for VIP than for a text-only ablation, two PLR score variants (PVL, MaxMC), random curricula, and training from scratch (Tables 1–2, Figures 3–4). A secondary comparison to supervised grid-search MAPPO* contextualizes sample efficiency. Qualitative frame analyses (Figure 1, Appendix D) illustrate cases where videos appear to capture near-miss strategies invisible to scalar scores.","tokens_in":22267,"tokens_out":1415,"duration_ms":22256,"significance":"If the result holds, VIP offers a simple, agent-count-agnostic route to open-ended multi-agent curricula that exploits visual behavior rather than only scalar learning signals or text summaries. The clean text-only ablation, use of an openly accessible 7B VLM whose inference cost is ~1% of training time, and first unsupervised-autocurriculum results on SMAC are concrete strengths. The work is timely given the rise of multi-modal LLMs and the known limitations of regret/score-based UED in sparse or deceptive multi-agent settings. Credit is due for reporting 95% CIs over five seeds, both held-out and in-distribution fine-tuning, and explicit qualitative case studies. The contribution is primarily empirical and methodological rather than theoretical.","major_comments":[{"comment":"The central claim that policy videos supply usable learning-progress signal beyond win rate and task history (Q1; Table 1 / Figure 3 gaps on 3s vs 4z and 3s5z) rests on a single short (1–10 s), low-fidelity SMAC rendering (colored circles, abbreviated labels, same-color attack lines; §5.1, Appendix A.3, Figure 5) fed once per curriculum step to a 7B VLM. The system/task prompts already encode explicit win-rate thresholds for difficulty/map changes (Appendix F.3: “if win rate >70% … if <30% …”). The text-only ablation is necessary but not fully sufficient to isolate visual contribution: a control with scrambled/blank/random videos (or attention/saliency analysis) is needed to rule out that the VLM is largely ignoring pixels and reacting to the scalar thresholds and history already present in text. Without such a control the multi-modal superiority remains under-supported.","section":"§5.1, §5.2, Appendix A.3, F.3"},{"comment":"PLR-PVL/MaxMC curricula exhibit extremely low map diversity and erratic non-adjacent repeats (Appendix G, Table 6), while VIP produces more progressive sequences. The paper attributes PLR’s poor fine-tuning results to scalar-score limitations, which is plausible, yet the comparison would be stronger if PLR were given the same task-space enumeration and a curriculum-length / replay-buffer budget matched to VIP’s 16 tasks, or if an additional UED baseline better suited to cooperative multi-agent settings were included. As written, it is hard to separate “scalar scores are insufficient” from “these particular PLR instantiations were poorly tuned for SMAC’s discrete map space.”","section":"§5.2–5.3, Appendix G"},{"comment":"Generalization claims rely on fine-tuning for another 10^7 steps with shared default MAPPO hyperparameters (Table 5) on only three held-out maps (Table 4). While the unsupervised protocol is correctly motivated, three maps (two of which share stalker/zealot composition) and a fixed 80/20 split leave open whether the advantage transfers to other SMAC compositions, difficulties, or non-SMAC multi-agent domains. A sensitivity analysis on the train/test split or at least one additional domain would substantially strengthen the “open-ended multi-agent” claim.","section":"§5.1–5.2, Table 4"}],"minor_comments":[{"comment":"Figure 5 and the accompanying text note that same-color fire lines make shooter/target identity ambiguous; this limitation should be flagged earlier (e.g., in §5.1) when the reader first encounters the rendering assumption.","section":"§5.1 / Appendix A.3"},{"comment":"Algorithm 1 uses non-standard notation (A^i_t, ϕ^video_t, ϑ_t). A short glossary or alignment with standard multi-agent POMDP notation would improve readability.","section":"Algorithm 1"},{"comment":"The sentence-similarity sanitization is reported to recover all hallucinations (5.6% rate; Appendix F.2), yet the similarity threshold and embedding model are only lightly described. Reporting the exact threshold and a few recovered vs. unrecovered examples would aid reproducibility.","section":"§4.1, Appendix F.2"},{"comment":"Typos / formatting: “OPEN-ENDEDMULTI-AGENT…” missing spaces in the title block; “Franc ¸ois-Lavet” and similar accent artifacts; “PLR ⊥” rendering inconsistency.","section":"Title / References"},{"comment":"Table 3 correctly notes that MAPPO* is supervised; the main text could more explicitly state that the 100× interaction gap is the intended takeaway rather than raw win-rate superiority.","section":"§5.4, Table 3"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a solid empirical contribution but currently over-claims the decisiveness of visual inspection relative to the strength of the controls. A scrambled-video ablation and tighter PLR matching would convert a borderline major-revision into a clear minor-revision or accept. Scope is appropriate for a methods-oriented ML venue; the SMAC-only evaluation is the main external-validity risk."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The useful takeaway is simple: feeding one short SMAC episode video plus win rate into VideoLLaMa2-7B produces better unsupervised multi-agent curricula than the same prompt without the video, and better than PLR-PVL/MaxMC or random. That is a concrete, ablated result on a domain that had not seen unsupervised autocurricula before.\n\nWhat is new is the systematic use of raw policy videos (not success detectors or text summaries) for task selection in an open-ended multi-agent setting. The method itself is deliberately thin—train MAPPO, record one video, prompt the VLM, sanitize with sentence similarity—and that thinness is a feature. The text-only ablation is cleanly done, 95% CIs over five seeds are reported, both held-out and frequently-seen maps are evaluated, and the qualitative frames (Figure 1, Appendix D) actually illustrate the near-miss intuition. Compute is honest: VLM inference is ~1% of wall time. Prompts and failure-mode rates are in the appendix. Citation pattern is fair; they position against OMNI, UED/PLR, and MAESTRO without overclaiming.\n\nSoft spots are real but proportionate. The SMAC rendering is minimal (same-color attack lines, colored circles with abbreviations), only one short video is used per step, and the system prompt already contains win-rate thresholds. That is the load-bearing assumption the stress-test flags. I do not think it collapses the result: the text-only ablation keeps the identical thresholds and history and still loses badly on 3s vs 4z and 3s5z, so the pixels are doing work. Still, it is a single domain, single 7B VLM, fixed MAPPO defaults, no public code, and the train/test map split is small. The comparison to supervised grid-search MAPPO* is appropriately framed as context, not as a claim of superiority.\n\nThis is for people who build multi-agent curricula or who want a practical way to inject visual priors into open-ended RL without inventing new score functions. It is not a theory paper. I would send it to peer review; the central comparison is solid enough to deserve referee time even if reviewers demand more domains or a stronger VLM. Worth reading and, for anyone working on LLM-guided curricula, worth citing.","headline":"Clean first empirical win for raw policy videos in multi-agent open-ended curricula on SMAC; the text-only ablation is the real control and it holds.","tokens_in":22864,"tokens_out":589,"would_cite":true,"duration_ms":7197,"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":"Watching policy videos lets a lightweight VLM build better multi-agent training curricula than win rates or text summaries alone.","keywords":["open-ended curricula","multi-agent RL","video language models","policy inspection","SMAC","unsupervised environment design","curriculum learning"],"falsifier":"Replace the video input with pure noise or with a non-informative static frame while keeping every other component identical; if fine-tuning win rates on the held-out maps then fall to the level of the text-only ablation, the claimed advantage of visual inspection disappears.","tokens_in":22767,"feed_emoji":"🎬","tokens_out":597,"duration_ms":5920,"temperature":0.7,"pith_summary":"Open-ended curricula try to keep an RL agent on the edge of its abilities so that it gradually acquires harder skills without a fixed final task. The hard part is knowing, for the agent's current policy, which next task is interesting yet still solvable. Prior methods reduce that judgment to a scalar learning score or a short text summary; both can miss near-miss strategies that a human coach would spot by watching the match. This paper shows that feeding short episode videos of the current multi-agent policy to a modest Video Language Model is enough for the model to recommend more useful next tasks. On the StarCraft Multi-Agent Challenge the resulting VIP curricula produce agents that fine-tune to substantially higher win rates on held-out maps than either a text-only ablation or scalar-score baselines, even though the VLM is only a 7-billion-parameter open model and the videos use the environment's crude built-in rendering.","feed_headline":"Policy videos beat win rates for multi-agent curricula","feed_subtitle":"A 7B VLM watching SMAC episodes builds better training sequences than scalar scores or text alone","key_machinery":"Visual Inspection of Policies (VIP): after each training segment, one or more episode videos plus a win-rate summary are given to a VLM; the model proposes the next interesting task; a sentence-similarity sanitizer maps the free-form answer onto a valid map-difficulty pair.","core_discovery":"Direct visual inspection of policy episode videos by a Video Language Model yields open-ended multi-agent curricula that are more effective, measured by fine-tuning win rates on held-out SMAC maps, than both the identical pipeline without video and established scalar-score curriculum methods.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["VLM policy videos beat scores for multi-agent curricula","Watching SMAC episodes with VLMs builds better autocurricula","Video inspection outperforms text and win rates in multi-agent RL","Policy videos via 7B VLM yield stronger multi-agent training curricula","Direct visual policy checks create superior open-ended multi-agent curricula"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"That a single short video of SMAC's minimal rendering (colored circles with abbreviated labels and monochrome attack lines) already supplies the VLM with enough behavioral signal to judge progress better than the win rate alone.","fun_headline_variants_meta":{"raw":{"variants":["VLM policy videos beat scores for multi-agent curricula","Watching SMAC episodes with VLMs builds better autocurricula","Video inspection outperforms text and win rates in multi-agent RL","Policy videos via 7B VLM yield stronger multi-agent training curricula","Direct visual policy checks create superior open-ended multi-agent curricula"]},"model":"grok-4.5","effort":"low","cost_usd":0.002936,"raw_usage":{"total_tokens":1022,"prompt_tokens":703,"num_sources_used":0,"completion_tokens":73,"cost_in_usd_ticks":29360000,"prompt_tokens_details":{"text_tokens":703,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":246,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":703,"tokens_out":73,"duration_ms":3554,"temperature":1.0,"reasoning_tokens":246,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-10T11:33:50.539305+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Replace the video input with pure noise or with a non-informative static frame while keeping every other component identical; if fine-tuning win rates on the held-out maps then fall to the level of the text-only ablation, the claimed advantage of visual inspection disappears.","supporting_citations":[],"review_version":1}