REVIEW 9 cited by
Unhackable Temporal Rewarding for Scalable Video MLLMs
Not yet reviewed by Pith; the record is open.
This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.
SPECIMEN: schema-true, not a live event
T0 review · schema-true
One-sentence machine reading of the paper's core claim.
pith:XXXXXXXX · record.json · timestamp
read the original abstract
In the pursuit of superior video-processing MLLMs, we have encountered a perplexing paradox: the "anti-scaling law", where more data and larger models lead to worse performance. This study unmasks the culprit: "temporal hacking", a phenomenon where models shortcut by fixating on select frames, missing the full video narrative. In this work, we systematically establish a comprehensive theory of temporal hacking, defining it from a reinforcement learning perspective, introducing the Temporal Perplexity (TPL) score to assess this misalignment, and proposing the Unhackable Temporal Rewarding (UTR) framework to mitigate the temporal hacking. Both theoretically and empirically, TPL proves to be a reliable indicator of temporal modeling quality, correlating strongly with frame activation patterns. Extensive experiments reveal that UTR not only counters temporal hacking but significantly elevates video comprehension capabilities. This work not only advances video-AI systems but also illuminates the critical importance of aligning proxy rewards with true objectives in MLLM development.
Forward citations
Cited by 9 Pith papers
-
MentalThink: Shaping Thoughts in Mental SVG World
MLLMs that generate and render SVG sketches as multi-turn intermediate reasoning steps reach 55.1% on VSIBench and 76.0% on MindCube, far above the Qwen2.5-VL-7B backbone.
-
Latent Visual Cache for Video Reasoning
A recurrent latent visual cache in the decoder, trained with contrastive key-frame alignment and a vision-grounded GRPO reward, improves video reasoning accuracy while shortening answers.
-
Video Reasoning without Training
An entropy-guided, inference-time value-cache controller improves video reasoning accuracy and cuts output tokens versus RL-trained baselines.
-
GLIMPSE: Do Large Vision-Language Models Truly Think With Videos or Just Glimpse at Them?
Introduces GLIMPSE, a video-QA benchmark whose questions cannot be answered from single frames; best model GPT-o3 scores 66.43% vs 94.82% human accuracy.
-
GRPO-CARE: Consistency-Aware Reinforcement Learning for Multimodal Reasoning
GRPO-CARE improves answer accuracy and reasoning coherence over standard GRPO on a new video reasoning benchmark, with a 6.7 point gain on the hardest level and a 24.5 point higher consistency rate.
-
Reinforcement Learning Tuning for VideoLLMs: Reward Design and Data Efficiency
A GRPO-based post-training recipe for video LLMs using discrete QA rewards plus continuous temporal IoU rewards with variance-based data selection outperforms SFT and Video-R1.
-
Vad-R1: Towards Video Anomaly Reasoning via Perception-to-Cognition Chain-of-Thought
A two-stage trained MLLM with a perception-to-cognition chain-of-thought and a self-verification RL reward outperforms prior models on video anomaly detection and reasoning.
-
Video-CoT: A Comprehensive Dataset for Spatiotemporal Understanding of Videos Based on Chain-of-Thought
Video-CoT contributes a new public dataset and benchmark that add fine-grained chain-of-thought annotations to six spatiotemporal video tasks, with fine-tuning experiments showing moderate gains.
-
Reinforcing Video Reasoning with Focused Thinking
A GRPO variant with token-level KL weighting and partial-credit rewards improves video-QA on modified multi-answer benchmarks, but transfer to original single-answer benchmarks is not established.
Discussion (0). Sign in to comment.