AdaCodec: A Predictive Visual Code for Video MLLMs
arXiv:2606. 02569v1 Announce Type: cross Abstract: Video is temporally redundant: adjacent frames usually share most objects, background, and layout.
arXiv:2606. 02569v1 Announce Type: cross Abstract: Video is temporally redundant: adjacent frames usually share most objects, background, and layout.
arXiv:2605.08158v2 Announce Type: replace-cross Abstract: Long video understanding with multimodal language models suffers from three compounding bottlenecks: heavy decode cost to obtain dense RGB fr...
arXiv:2604.06036v4 Announce Type: replace-cross Abstract: Continuous inference over concurrent video streams imposes substantial compute and memory demands on vision-language model (VLM) serving. Str...
Video-Mirai is a training-only method that improves autoregressive video diffusion models by adding foresight. It uses a frozen foresight encoder to read completed rollouts and a lightweight predictor to distill future-informed targets into causal states, without altering the generator’s inference process. The approach raises the 5‑second VBench Total Score from 83.8 to 84.6 and improves subject and background consistency on 30‑second rollouts.
arXiv:2609.39924v1 Announce Type: cross Abstract: Vision-language models face a fundamental scaling bottleneck: the number of visual tokens grows with both temporal duration and spatial resolution, m...
LayerRecall is a memory router for autoregressive video diffusion that selectively retrieves and injects historical key/value states into specific layers of the model, based on the current context. It addresses the problem that existing memory mechanisms expose nonlocal history but do not guarantee effective use, by recognizing that different layers prefer current, recent, or distant context. The method, combined with Cross‑Horizon Prediction Matching, achieves state‑of‑the‑art long‑range consistency on MemoBench and MovieBench while maintaining local continuity and incurring negligible inference overhead.
The paper introduces PreSTP, a training‑free method for speeding up few‑step autoregressive video diffusion models by pruning redundant tokens. It identifies two redundancy patterns—cross‑frame similarity from static content and cross‑step convergence during denoising—and uses early clean estimates from the model’s own trajectory to build a token‑wise redundancy mask. PreSTP achieves significant speedups (up to 3.20× for 5‑second videos and 1.96× for 30‑second videos) while maintaining or improving fidelity compared to existing caching baselines.
arXiv:2606. 06991v1 Announce Type: cross Abstract: Online Video Large Language Models (Video-LLMs) have advanced toward seamless human-AI interaction through frame-by-frame processing and proactive responding.
MoCAR (Motion-code Coordinate-aware AutoRegression) is a decoder‑only framework that treats continuous trajectory forecasting as next‑code prediction in a coordinate‑aware latent space. It learns a continuous motion‑code space from endpoint‑normalized trajectory segments, using historical codes as a teacher‑forced prefix and generating future codes autoregressively while updating local scene context. The method achieves top‑tier performance on Argoverse benchmarks, transfers well from AV2 to AV1 in zero‑shot evaluation, and improves on turn‑heavy scenarios without requiring trajectory‑space re‑tokenization or proposal‑and‑refinement pipelines.
arXiv:2609.23601v1 Announce Type: new Abstract: Long-video understanding must capture transient visual evidence under strict token budgets, yet existing methods compress frames, append memory tokens,...
The paper introduces Recency Forcing, a technique that addresses the long‑horizon degradation in autoregressive video generation caused by KV eviction mismatch. By applying a timestep‑dependent bias—Temporal Response Bias—derived from a positional response measure, the method reduces the influence of distant frames during inference without altering context length or training objectives. An exact reformulation, Biased Attention Reparameterization, enables this bias to be applied as a standard FlashAttention call with zero overhead, achieving state‑of‑the‑art long‑horizon generation quality on VBench datasets.
LongVU‑TTT is a causal test‑time training method for long‑video multimodal large language models that inserts a convolutional resampler with fast‑weight updates between the vision encoder and the LLM. The fast weights adapt per video and contextualize frame features before compression, while a hybrid selector keeps explicit visual evidence for downstream reasoning. Experiments show that TTT‑Conv outperforms TTT‑MLP and bidirectional Mamba2 on MLVU, and beats attention‑ and fixed‑state recurrent resamplers on three benchmarks, achieving competitive results on five video‑understanding tasks after reducing 512 frames to 128 LLM frames.