AdaptToken: Entropy-based Adaptive Token Selection for MLLM Long Video Understanding

Dimensions overall score 8.0

GitHub Code Pulse

No public code linked for this paper yet.

• AdaptToken improves average accuracy by +6.7 points over the Qwen2.5-VL 7B base model on long-video benchmarks.result 95%
• AdaptToken-Lite reduces inference time by about half while maintaining comparable performance to the full method.result 90%
• AdaptToken scales to process extremely long video inputs of up to 10,000 frames and continues to improve performance at this scale.technical 90%
• AdaptToken is a training-free framework that uses an MLLM's self-uncertainty (response entropy) as a global control signal for token selection.method 85%
• Prior frame/token selection methods lack a principled mechanism to compare relevance across distant video clips and to stop processing once sufficient evidence is gathered.limitation 80%
• AdaptToken delivers consistent performance gains over base MLLMs and prior frame/token selection methods across four challenging long-video benchmarks.result 85%
• The framework's performance may be limited if the base MLLM cannot capture new visual or contextual information, as it relies on the existing architecture.limitation 75%
• Across benchmarks, a higher certainty score (lower response entropy) from the MLLM correlates with a higher probability of answering correctly.result 80%

Competitive landscape

Competitor map is still being generated for this paper. Enable generation or check back soon.

Startup potential card