# Video Understanding

On CARLA third-person driving videos, six open-source video-language models judging forward/backward, left/right, and accelerate/decelerate stayed near chance under standard, visual, few-shot, and chain-of-thought prompting, with systematic class biases such as every model scoring 100% on forward and 0% on backward. LoRA finetuning then reached near-perfect held-out accuracy for all models except VideoLLaMA2 on accelerate/decelerate, and simple classifiers on pre- and post-compression visual features decoded motion cues at high accuracy. The motion information is present inside the models, but the prompted prediction pathway does not reliably use it.

On CARLA third-person driving videos, six open-source video-language models judging forward/backward, left/right, and accelerate/decelerate stayed near chance under standard, visual, few-shot, and chain-of-thought prompting, with systematic class biases such as every model scoring 100% on forward and 0% on backward. LoRA finetuning then reached near-perfect held-out accuracy for all models except VideoLLaMA2 on accelerate/decelerate, and simple classifiers on pre- and post-compression visual features decoded motion cues at high accuracy. The motion information is present inside the models, but the prompted prediction pathway does not reliably use it.

Compressing visual tokens for video large language models costs feature comparisons in the compression decisions themselves. This method aligns reference-frame dependencies, motion vectors and residuals already present in compressed video to the visual token grid, uses them to propose small candidate sets, then verifies with local visual similarity. Across three video LLMs and four video-understanding benchmarks, performance matches the training-free baseline FlashVID and sometimes exceeds it, while comparisons for token protection and temporal matching drop substantially.

Compressing visual tokens for video large language models costs feature comparisons in the compression decisions themselves. This method aligns reference-frame dependencies, motion vectors and residuals already present in compressed video to the visual token grid, uses them to propose small candidate sets, then verifies with local visual similarity. Across three video LLMs and four video-understanding benchmarks, performance matches the training-free baseline FlashVID and sometimes exceeds it, while comparisons for token protection and temporal matching drop substantially.

The work presents Omni-Decision, an omni-modal agent that replaces the growing dialogue history with an explicit evidence ledger, where a critic reads each noisy observation and passes only usable content to the ledger so the planner works from a compact context throughout the task; each run records state, action, and verdict at every step, and supervised fine-tuning plus decision-level reinforcement learning on these trajectories further improve the planner, yielding state-of-the-art 81.4% accuracy on OmniGAIA at roughly 43% of Gemini-3.1-Pro's cost per question and 65.0% on WorldSense long-video understanding, level with the strongest end-to-end model.

The work presents Omni-Decision, an omni-modal agent that replaces the growing dialogue history with an explicit evidence ledger, where a critic reads each noisy observation and passes only usable content to the ledger so the planner works from a compact context throughout the task; each run records state, action, and verdict at every step, and supervised fine-tuning plus decision-level reinforcement learning on these trajectories further improve the planner, yielding state-of-the-art 81.4% accuracy on OmniGAIA at roughly 43% of Gemini-3.1-Pro's cost per question and 65.0% on WorldSense long-video understanding, level with the strongest end-to-end model.