Dynamic Token Compression for Efficient Video Understanding through Reinforcement Learning

Adapting Multimodal Large Language Models (MLLMs) for hour-long videos is bottlenecked by context limits. Dense visual streams saturate token budgets and exacerbate the lost-in-the-middle phenomenon. Existing heuristics, like sparse sampling or uniform pooling, blindly sacrifice fidelity by discarding decisive moments and wasting bandwidth on irrelevant backgrounds. We propose Tempo, an efficient query-aware framework compressing long videos for downstream understanding. Tempo leverages a Small Vision-Language Model (SVLM) as a local temporal compressor, casting token reduction as an early cross-modal distillation process to generate compact, intent-aligned representations in a single forward pass. To enforce strict budgets without breaking causality, we introduce Adaptive Token Allocation (ATA). Exploiting the SVLM's zero-shot relevance prior and semantic front-loading, ATA acts as a training-free $O(1)$ dynamic router. It allocates dense bandwidth to query-critical segments while compressing redundancies into minimal temporal anchors to maintain the global storyline. Extensive experiments show our 6B architecture achieves state-of-the-art performance with aggressive dynamic compression (0.5-16 tokens/frame). On the extreme-long LVBench (4101s), Tempo scores 52.3 under a strict 8K visual budget, outperforming GPT-4o and Gemini 1.5 Pro. Scaling to 2048 frames reaches 53.7. Crucially, Tempo compresses hour-long videos substantially below theoretical limits, proving true long-form video understanding relies on intent-driven efficiency rather than greedily padded context windows.

Video Large Language Models (Video LLMs) achieve strong performance on video understanding tasks but suffer from high inference costs due to the large number of visual tokens. We study training-free, query-agnostic video token compression, where heuristic importance rules can over-represent large redundant visual groups and miss less frequent visual patterns, particularly under aggressive token budgets. We propose KiToke, a distribution-aware compression framework that treats visual tokens as samples from the video's spatiotemporal token distribution. KiToke uses global embedding density as a proxy for representational redundancy and inverse-density sampling to balance token allocation across visual patterns of different frequencies. It further aggregates unselected tokens into retained anchors within detected temporal intervals to reduce cross-transition mixing. Across evaluated backbones, KiToke achieves the highest average accuracy among compared training-free methods and remains robust down to 1% retention, where most evaluated baselines underperform uniform random token sampling. On LLaVA-OneVision, it preserves 98.5% and 89.2% of the uncompressed model's four-benchmark mean accuracy at 10% and 1% token retention, with 6.8x and 12.5x prefill speedups, respectively, including compression overhead.

Long video understanding is inherently challenging for vision-language models (VLMs) because of the extensive number of frames. With each video frame typically expanding into tens or hundreds of tokens, the limited context length of large language models (LLMs) forces the VLMs to perceive the frames sparsely and lose temporal information. To address this, we explore extreme video token compression towards one token per frame at the final LLM layer. Our key insight is that heuristic-based compression, widely adopted by previous methods, is prone to information loss, and this necessitates supervising LLM layers into learnable and progressive modules for token-level compression (LP-Comp). Such compression enables our VLM to digest 2x-4x more frames with improved performance. To further increase the token efficiency, we investigate frame-level compression, which selects the frames most relevant to the queries via the internal attention scores of the LLM layers, named question-conditioned compression (QC-Comp). As a notable distinction from previous studies, we mitigate the position bias of LLM attention in long contexts, i.e., the over-concentration on the beginning and end of a sequence, by splitting long videos into short segments and employing local attention. Collectively, our combined token-level and frame-level leads to an extreme compression model for long video understanding, named XComp, achieving a significantly larger compression ratio and enabling denser frame sampling. Our XComp is finetuned from VideoChat-Flash with a data-efficient supervised compression tuning stage that only requires 2.5% of the supervised fine-tuning data, yet boosts the accuracy from 42.9% to 46.2% on LVBench and enhances multiple other long video benchmarks.