2b-8b-video-multimodal-models-adding-small-fraction-time-reversed-video CLM-0341For 2B-8B video multimodal models, adding a small fraction of time-reversed video rollouts into each GRPO group (e.g. 1 reversed out of 7) raises accuracy on action and temporal-reasoning video QA over standard GRPO, while raising the reversed share further (2 or 3 of 7) gives back the gain.
Capability: Reasoning about time in video
Observed on
Qwen3-VL 2B/4B/8B and InternVL2.5-8B, trained on Video-R1 splits, evaluated on 1,000-example samples of STAR, SSv2, NExT-QA, CLEVRER, LLaVA-178K, PerceptionTest; the reversed-ratio.
Sources
- supportsCOMET: Contrastive Motion-Enhanced Temporal Reasoning for Video Multimodal Large Language ModelsValidated by: Same Stage-I checkpoint, Qwen3-VL-8B, group size 7: average accuracy 78.0 for 7 forward + 0 reversed, 78.8 for 6+1, 78.4 for 5+2, 77.6 for 4+3 across six benchmarks.. Measured on one backbone (Qwen3-VL-8B) for the ratio sweep, with 1,000 sampled examples per benchmark and no reported variance or seeds; gains over plain GRPO are under one point on average. Main-table gains confound the motion branch with the reversed-rollout objective; only Table 4 isolates the optimizer.
Disagreeing is the most useful thing you can do here. Nothing on this page waits for a reviewer: the counts above move when evidence lands, and every source stays attached with who brought it.
Related claims
- For 2B-8B video multimodal models fine-tuned for temporal reasoning, placing a small minority of reversed-video rollouts in each GRPO group (1 reversed out of 7) beats both all-forward GRPO and larger reversed shares (2 or 3 of 7), because more reversed samples dilute the main forward-video answering objective.Reasoning about time in video · reviewed by AI
- For 8B multimodal models, preference tuning and plain outcome-reward GRPO lower hallucination rates while making stated confidence worse calibrated than the untuned backbone, because rank- or correctness-based rewards never score the numerical accuracy of the model's own confidence; adding a Brier-plus-group-ECE calibration term to the GRPO reward cuts calibration error sharply and leaves accuracy metrics largely unchanged.Grounding answers in the image · reviewed by AI
- In GRPO-style RL post-training of 7B multimodal reasoning models, composing each optimization group from rollouts stratified by length-normalized sequence surprisal and outcome (confident-correct anchors, confident-wrong negatives, low-confidence successes) within the same step keeps path entropy from collapsing and avoids the late-training performance decline GRPO and DAPO show after a few hundred steps.Retaining skills and safety after fine-tuning · reviewed by AI
- When post-training a small LLM (Qwen2.5 0.5B-3B) to produce weekly asset-return forecasts from news, rewarding rank correlation of a full listwise asset ordering gives a dense learning signal where exact-match GRPO gives almost none (all sampled actions scoring zero on ~35% of steps), and yields higher Spearman correlation and backtest Sharpe than the SFT checkpoint, standard GRPO, and prompted commercial models.Predicting future events · reviewed by AI
- In GRPO-style RLVR on math for 7B models, injecting partial reasoning prefixes from a small out-of-family model (Gemma-2-2B, LLaMA-3.2-1B) into a fraction of rollouts keeps policy entropy higher early in training and raises pass@k, with gains growing as k rises; the effect comes from distributional unfamiliarity rather than prefix quality, since prefixes from same-family models of equal or higher quality give no gain over vanilla GRPO.Retaining skills and safety after fine-tuning · reviewed by AI
Notes
Drafted from the paper by a model and filed unreviewed. Visible here so it can be read, not because anyone has vouched for it: it does not move any technique's standing and does not count toward the internal scorecard. Drafted confidence: low. Falsifier as drafted: Running the same reversed-ratio sweep on other backbones or without the Taylor motion branch and finding no advantage of 6+1 over plain GRPO, or finding gains increase monotonically with the reversed share.