{"id":"40b704fe-cbd3-483b-a59d-2564cc109720","arxiv_id":"2608.07693","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"CosmosAlign adapts Cosmos3-Nano with two-stage LoRA, medoid sample selection, and motion-adaptive blending, achieving first place (76.49) on the AI City Challenge 2026 Track 5 traffic video forecasting benchmark.","lead":"CosmosAlign is a recipe for adapting a large pretrained world model to traffic video forecasting: two stages of lightweight LoRA tuning plus training-free test-time selection and blending. It reports first place on the AI City Challenge 2026 Track 5 with a composite score of 76.49.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Validation split may not be representative of the official test set, so the ablations that select every recipe hyperparameter do not establish the claimed alignment effect.","rationale":"The reader's weakest assumption is exactly the representativeness of the 36-clip validation split. This is the paper's single most load-bearing concern because the recipe selection, the ablations in Tables 1, 3, 4, and the paper's central claim that alignment (rather than capacity) drives the result all rest on it. I agree with the reader that this warrants a CONDITIONAL verdict. The paper is an engineering-competition write-up with plausible methodology and a credible first-place claim; the benchmark result itself, if the leaderboard is accurate, is a factual claim that is unlikely to be wrong. The fragile part is the scientific interpretation: which components actually contributed to the score gain. The concern is not a formal inconsistency; it is an empirical representativeness risk, and the test I propose directly settles it by re-running the official test set.","tokens_in":15747,"tokens_out":1520,"duration_ms":13204,"concrete_test":"Compute the validation-to-test transfer for each ablation component in Table 1. Concretely: rerun the official 71-clip test set for the four extreme configurations that bracket the recipe—plain Stage-1 at g=3, Stage-1+g=3+structured, Stage-1+Stage-2+g=3, and the full recipe with best-of-4+MA blending—using the same code path as the reported Table 2 progression. If the rank order and per-component deltas on the official test set match the validation-based selections, the 36-clip split is representative and the alignment claim survives. If, instead, any component moves in the opposite direction (e.g., MA blend degrades FVD/PSNR on the official test), the validation split is not tracking the test metric and the central claim needs revision.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that two-stage LoRA alignment, not capacity, drives the 76.49 first-place result. But every recipe decision—guidance g=3, LoRA rank 32, Stage 2 length 750, blending thresholds m_lo=3, m_hi=20, w=0.9, and the medoid over seeds 0..3—is selected on a 36-clip validation set that the authors 'built to mirror the test protocol exactly' (Sec. 4.1). The load-bearing premise is that this 36-clip split is representative of the 71-clip official test set. The paper provides no evidence for representativeness beyond the view ratio, and the validation and test metric scales differ materially: validation PSNR is around 23.9 and SSIM around 0.79, while the server-reported test PSNR is around 20.1 and SSIM around 0.65. Test-set results are single-seed and, for the progression in Table 2, based on official 71-clip runs with no error bars. The concern is not that the benchmark result is wrong; it is that the causal claim—that alignment choices yield the observed gains—requires the 36-clip split to track the official test metric. If the split is unrepresentative, the selected hyperparameters could be tuned to the 36-clip noise, and the +1.17 gain from blending, which is also the largest component in the final submission, may not transfer to the official test set as claimed.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript presents CosmosAlign, a system for text-conditioned traffic video forecasting built on the frozen Cosmos3-Nano world model. It proposes a two-stage LoRA adaptation (Stage 1 reweights the conditioning-mode distribution on WTS/BDD windows; Stage 2 continues on re-captioned WTS windows), followed by a training-free test-time procedure that selects the medoid of four samples and blends static regions toward the last observed frame. The authors report validation ablations over guidance, adapter rank, prompt format, post-processing, history length, and training length, a test-set progression of submissions, and a final official score of 76.49 ranking first on AI City Challenge 2026 Track 5. They conclude that adapting world models to this task depends primarily on distribution alignment rather than model capacity.","tokens_in":15999,"tokens_out":8751,"duration_ms":75002,"significance":"If the reported result is taken at face value, the paper provides a strong empirical data point that parameter-efficient alignment of a large pretrained world model can win a competitive forecasting benchmark, and the public code and detailed leaderboard progression are useful to the community. The training-free medoid selection and motion-adaptive blending are simple, clearly described, and additive in the reported numbers. I found no circular derivation: the medoid and blending are evaluated on held-out test clips, and the validation set is separate from the test set. However, the broader conceptual claim about distribution alignment versus capacity is not established at the same level of confidence as the benchmark result, because the component attributions rest on a 36-clip validation set and single test runs.","major_comments":[{"comment":"Every recipe hyperparameter (guidance g=3, LoRA rank/alpha 32/64, Stage 2 length 750, blending thresholds m_lo=3 and m_hi=20, blending weight w=0.9, and the medoid over seeds 0-3) is selected on the 36-clip validation set using seed 0, and each row of the Table 2 progression is a single official test run. The validation set is described as mirroring the test protocol only in view ratio and horizon distribution, and the validation and test metric scales differ materially (validation PSNR around 23.9 versus test PSNR around 20.1). The paper therefore does not currently establish that the validation-selected choices, or the component gains in Table 2 (notably guidance +1.45 and blending +1.17), transfer to the official test set beyond the single final score. Please add evidence of validation-to-test consistency (e.g., rank correlation across configurations, multiple seeds, or a bootstrap over the 36 clips) and, where possible, repeated test submissions for the key incremental rows.","section":"Sec. 4.1 and Table 2"},{"comment":"The 'History 5 frames' row, as displayed, reports PSNR 24.40, LPIPS 0.163, and CLIP-S 28.12, which are better than every fine-tuned configuration in the same table, including the Stage 2 +750 checkpoint (PSNR 24.25, LPIPS 0.166, CLIP-S 27.92). The 'Longer Conditioning History' paragraph interprets the history-length trend as a property of the adapted model, but if '5 frames' is actually the zero-shot Cosmos3-Nano baseline, the table neither labels it nor reports a zero-shot score on the official test set; if it is the adapted model, it contradicts the g=3 and Stage 2 rows. Please clarify this configuration and, if it is the zero-shot baseline, use it to quantify the fine-tuning gain on both validation and test.","section":"Table 1"},{"comment":"The claim that adaptation 'depends primarily on distribution alignment rather than increased model capacity' is supported only by a rank-32 versus rank-128 comparison within the same 16B backbone, while the authors state in the conclusion that validation on larger variants such as Cosmos3-Super is future work. The reported evidence shows that a larger adapter does not help in this data-limited setting, but it does not support the general comparative claim about capacity. Please either restrict the claim to the data-limited LoRA setting or add a positive control against a larger backbone or a full fine-tune.","section":"Abstract, Sec. 4.2, Sec. 5"},{"comment":"The final incremental rows are close to plausible run-to-run noise. The deflickering stack adds +0.035 points (75.35 versus 75.32), and the text attributes +0.32 to best-of-3 despite an intervening deflickering row (75.03 to 75.32); with single submissions per configuration, these margins are not demonstrated to be significant. Please provide repeated official runs or an estimate of test-set noise so the component attributions are credible.","section":"Table 2"}],"minor_comments":[{"comment":"In the 'Stage 1: Forecasting Adaptation' paragraph, 'The conditioning-mode distribution ... are reweighted' should be 'is reweighted'.","section":"Sec. 3.2"},{"comment":"In the 'Stage 2 Fine-Tuning' paragraph, 'best-of-3 selection adds +0.32 on the Stage 2 recipe (75.00 to 75.32)' skips the intermediate deflickering row; please state the exact comparison base.","section":"Sec. 4.2"},{"comment":"The phrase 'stack separatedly' appears to be a typo for 'stack separately'.","section":"Table 2 caption"},{"comment":"Reference [24] contains the placeholder 'fill author list from the ICCW 2025 paper, T.' and needs to be completed before publication.","section":"References"},{"comment":"The statement that 'all configuration sweeps are validated at seed 0' should also clarify whether validation metrics are averaged over the 36 clips or pooled over frames, since this affects the interpretation of the reported FVD.","section":"Sec. 4.1"}],"recommendation":"major_revision","confidential_remarks":"This is an empirical challenge report whose main value is the verified leaderboard result and the public code. The revisions requested in the major comments are mostly about calibration of the conceptual claims; if the authors can provide a zero-shot baseline and validation-to-test consistency evidence, I would support acceptance. The citation of the authors' own TrafficVLM is appropriate and is not load-bearing for the technical claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a solid first-place competition writeup, more useful as a recipe than as a scientific claim. The authors are honest about what didn't work, and the ablation coverage is unusually thorough for a challenge paper. But the central claim—that distribution alignment, not capacity, is what matters—is not actually tested against the pretrained model on the official test set, and the entire hyperparameter stack is selected on a 36-clip validation split that we have no reason to believe tracks the test metric.\n\nWhat's new: the two-stage LoRA recipe (reweighting the conditioning-mode distribution, then re-captioning to the model's native structured prompt format) paired with medoid selection and a self-regulating motion-adaptive blend is a sensible, previously unpublished combination. The negative results are refreshing: chunked autoregressive rollout hurts, rank 128 hurts, longer conditioning history hurts, data rebalancing hurts. Those all support the alignment-over-capacity narrative, but they are validation-set results.\n\nSoft spots, in order of importance. First, the validation set is 36 clips, and every hyperparameter—guidance, LoRA rank, Stage 2 length, blend thresholds, medoid seeds—is chosen there. The paper says the split 'mirrors the test protocol exactly,' but the metric scales differ by a lot (validation PSNR ~24 vs test PSNR ~20), and there is no evidence that rank-ordering of configurations on validation matches rank-ordering on the test set. The test progression in Table 2 confirms the chosen values work, but it doesn't test alternatives. Second, Table 1's 'History 5 frames' row shows PSNR 24.40, LPIPS 0.163, and CLIP-S 28.12, all better than the LoRA-tuned rows below it, and the caption doesn't explain what that row is. If it's zero-shot Cosmos3-Nano, that's a huge result and kills the adaptation narrative; if it's something else, the table needs a footnote. Third, there is no zero-shot baseline on the official test set, so we can't see what the untuned model would score. Fourth, minor but real: reference [24] is an unfinished placeholder ('fill author list...'), and the code link points to a project page, not a repository.\n\nNone of this makes me think the first-place result is fake—leaderboard results are what they are. But the paper's scientific claim about alignment versus capacity is under-supported as written. The recipe itself is valuable; I'd be happy to cite it as a challenge solution.\n\nWho's this for? Someone adapting a large world model to a specific forecasting benchmark wants this paper. It deserves a serious referee, but only after the authors fix Table 1, add a zero-shot test baseline, discuss the validation-set representativeness, and clean up the references. If those revisions land, it's a decent workshop paper; as it stands, it's a useful tech report.","headline":"A genuinely useful first-place competition recipe with unusually honest ablations, but the alignment-over-capacity claim rests on a 36-clip validation split and a zero-shot baseline is missing.","tokens_in":16594,"tokens_out":4693,"would_cite":true,"duration_ms":42351,"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":"The paper argues that distribution alignment, not model capacity, is the key to adapting pretrained world models to traffic forecasting, and demonstrates it by winning Track 5 with 76.49.","keywords":["traffic video forecasting","world models","autonomous driving","LoRA fine-tuning","distribution alignment","text-conditioned video generation","medoid selection","AI City Challenge 2026"],"falsifier":"Re-run the hyperparameter sweeps on a different 36-clip validation split drawn from the same WTS training pool; if the optimal settings (guidance 3, 750 Stage-2 iterations, blending thresholds 3/20/0.9) shift materially or the official test score drops well below 76.49, the small-split representativeness assumption is refuted.","tokens_in":15466,"feed_emoji":"🚗","tokens_out":10492,"duration_ms":85567,"temperature":0.7,"pith_summary":"The paper sets out to prove that adapting a large pretrained world model to a specific traffic-forecasting benchmark is primarily a distribution-alignment problem, not a model-capacity problem. It builds CosmosAlign on the 16B Cosmos3-Nano model with a two-stage LoRA recipe: first rewiring the conditioning-mode distribution so that 70% of gradient steps supervise the five-frame video-conditioned forecasting mode, then re-captioning the training data into the model's native structured prompt format. Paired with a training-free inference step of consensus-based medoid selection and motion-adaptive blending of static regions, the adapted model scores 76.49 on the AI City Challenge 2026 Track 5 leaderboard and ranks first. A sympathetic reader would care because it suggests that strong downstream forecasting can be achieved cheaply by aligning the interface, rather than by scaling the model.","feed_headline":"Alignment, not scale, wins traffic video forecasting","feed_subtitle":"Two-stage LoRA tuning of Cosmos3-Nano plus medoid selection reaches 76.49, first on AI City Track 5.","key_machinery":"The carrying mechanism is the two-stage LoRA alignment recipe applied to Cosmos3-Nano, a 16B world foundation model with a Reasoner–Generator architecture. Stage 1 trains low-rank adapters (rank 32, $\\alpha=64$) on the Generator attention projections while freezing the base model, reweighting the conditioning-mode distribution $\\{k=0,k=1,k=2\\}$ from the default $\\{0.7,0.2,0.1\\}$ to $\\{0.1,0.2,0.7\\}$ so that 70% of gradient steps supervise the five-frame video-conditioned forecasting mode. Stage 2 continues training on 3,107 WTS windows re-captioned into the model's native JSON temporal-caption format, aligning the prompt distribution. At inference, the payload holds five history frames, the structured prompt, and an extended negative prompt; four samples are drawn with a 35-step UniPC solver at guidance 3, the medoid (minimum mean pairwise distance on a downsampled grid) is selected, and a motion-adaptive blend weights the last observed frame into static regions only.","core_discovery":"The paper's central claim is that on the WTS-based Track 5 task, the dominant bottleneck in adapting Cosmos3-Nano is distribution mismatch rather than model capacity, and that a two-stage LoRA adaptation—Stage 1 reweights the conditioning-mode distribution from the pretraining default toward the evaluated forecasting mode, and Stage 2 aligns the training captions to the model's native structured prompting format via LLM re-captioning—plus a training-free inference of medoid sample selection and motion-adaptive blending, yields the top official score of 76.49. The paper supports this with ablations: raising the LoRA rank from 32 to 128 degrades four of five validation metrics, longer Stage 1 training shows monotone CLIP-S decay, and longer conditioning histories hurt quality; conversely, guidance tuning to $g=3$, structured prompts, the Stage 2 checkpoint, best-of-$N$ medoid selection, and the motion-adaptive blend each improve the final score on the official test set.","pith_inferences":["If this alignment-over-capacity claim holds beyond this benchmark, the field's default reflex of scaling up may be misplaced; the first diagnostic when adapting a foundation model should be an analysis of the conditioning-mode and prompt-format gap between pretraining and target.","The motion-adaptive blending rule is a test-time constraint that anchors static regions to the last observed frame; this suggests a broader class of physics-aware post-processing that enforces scene-structure priors without retraining.","Because the medoid selection is ground-truth-free, it could be applied to online adaptation or self-training loops where no labels exist but consensus across samples is a proxy for reliability.","A direct extension: run the same recipe on a non-traffic video-forecasting benchmark (for example, general action prediction) to see whether conditioning-mode reweighting alone reproduces the gains; if it does not, the alignment claim is specific to safety-critical traffic scenes rather than general video generation."],"forward_implications":["If the alignment thesis is right, strong traffic-video forecasting can be obtained from a 16B model with low-rank adapters and no capacity increase, making the whole pipeline runnable on a single A100 GPU.","The two-stage recipe of conditioning-mode reweighting plus caption-format alignment should transfer to other world models that share a structured prompting interface, likely saving compute on future benchmarks.","The training-free medoid selection and motion-adaptive blending improve fidelity without ground truth or auxiliary models, so they can be dropped into other sampling-based generative forecasting pipelines.","The ablations imply that in the data-limited regime, memorization and train-test distribution mismatch, not model size, limit forecast quality; the paper's stated next step is to test the recipe on larger variants such as Cosmos3-Super and on other world model families."],"supporting_citations":[{"why":"Supplies the Cosmos3-Nano base world model, its Reasoner–Generator architecture, and the native structured-prompt interface the paper aligns to.","marker":"[1]"},{"why":"Supplies the LoRA method used for the two-stage parameter-efficient fine-tuning.","marker":"[22]"},{"why":"Supplies the WTS dataset that defines the traffic-safety forecasting task and the training/validation windows.","marker":"[25]"},{"why":"Defines the AI City Challenge 2026 Track 5 benchmark and the official composite score that the paper's 76.49 result comes from.","marker":"[48]"},{"why":"Supplies the BDD100K external dashcam videos used in Stage 1 training to expand the forecasting corpus.","marker":"[60]"},{"why":"Supplies the UniPC 35-step solver used for sampling at inference.","marker":"[64]"},{"why":"Provides the Minimum Bayes Risk decoding principle that inspires the consensus-based medoid sample selection.","marker":"[26]"},{"why":"Provides the self-consistency idea that also motivates selecting the sample with minimum mean pairwise distance.","marker":"[54]"}],"fun_headline_variants":["Alignment, not scale, wins traffic video forecasting","Two-stage LoRA tuning tops AI City Track 5","CosmosAlign: alignment over capacity wins track","Medoid selection boosts CosmosAlign to first","Distribution alignment, not size, wins traffic"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The 36-clip validation set the authors built to mirror the test protocol is representative enough of the 71-clip official test set that the hyperparameters chosen on it (guidance 3, Stage 2 length 750 iterations, blending thresholds $m_{\\mathrm{lo}}=3$, $m_{\\mathrm{hi}}=20$, $w=0.9$) transfer to the real test.","fun_headline_variants_meta":{"raw":{"variants":["Alignment, not scale, wins traffic video forecasting","Two-stage LoRA tuning tops AI City Track 5","CosmosAlign: alignment over capacity wins track","Medoid selection boosts CosmosAlign to first","Distribution alignment, not size, wins traffic"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000296,"raw_usage":{"total_tokens":1721,"prompt_tokens":952,"completion_tokens":769,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":568,"completion_tokens_details":{"reasoning_tokens":697}},"tokens_in":568,"tokens_out":769,"duration_ms":6720,"temperature":1.0,"reasoning_tokens":697,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T00:24:25.458924+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the hyperparameter sweeps on a different 36-clip validation split drawn from the same WTS training pool; if the optimal settings (guidance 3, 750 Stage-2 iterations, blending thresholds 3/20/0.9) shift materially or the official test score drops well below 76.49, the small-split representativeness assumption is refuted.","supporting_citations":[{"cited_title":"In: ECCV","cited_arxiv_id":null,"evidence_quote":"Supplies the WTS dataset that defines the traffic-safety forecasting task and the training/validation windows."},{"cited_title":"In: ECCV Workshops","cited_arxiv_id":null,"evidence_quote":"Defines the AI City Challenge 2026 Track 5 benchmark and the official composite score that the paper's 76.49 result comes from."},{"cited_title":"In: Thirty-seventh Conference on Neural Information Processing Systems (2023),https://openreview.net/forum? id=hrkmlPhp1u","cited_arxiv_id":null,"evidence_quote":"Supplies the UniPC 35-step solver used for sampling at inference."},{"cited_title":"In: Proceedings of the Human Language Technology Conference of the North American Chapter of the Association for Computational Linguistics: HLT- NAACL 2004","cited_arxiv_id":null,"evidence_quote":"Provides the Minimum Bayes Risk decoding principle that inspires the consensus-based medoid sample selection."}],"review_version":1}