Video AI's new trick: prove where in the video the answer came from

Show your work, timestamped

The most interesting idea of the three treats video reasoning as detective work. In a preprint, the Open-o3-Video-v2 authors describe an agentic framework that notices what it does not know, goes and looks — retrieving moments and objects from the video — and writes the evidence into its reasoning trajectory. A training method they call Clue-Seek GRPO rewards covering the relevant clues and penalizes excessive retrieval. The reported result: 40.4% mAM and 56.3% mLGM on the V-STaR benchmark, beating the Qwen3-VL-4B backbone by 8.5 and 12.6 points. m:…9250

Why this matters is not the leaderboard delta but the verifiability. A video model that cites the exact moments its answer rests on is auditable in a way a black-box answer is not — which is the difference between a demo and something you can put in front of a customer, a regulator, or a courtroom. The authors' framing — that active evidence acquisition and object-level grounding are complementary — reads as a bet that grounding becomes table stakes for video AI, the way citations became table stakes for text. m:…9250

The training data is part of the story. In a preprint, the authors describe STGR-v2-CoT135K — 52,000 multi-turn agentic grounded-reasoning trajectories — plus a 13K-example RL split built for reinforcement learning. Their finding is that active evidence acquisition and verifiable object-level grounding are complementary: seeking the clues and proving where they came from each help, and together they help more. For anyone building video QA, that is a design hint disguised as a dataset release. m:…9250

Long video without the memory bill

The second preprint attacks the cost side. Standard vision-language models encode video frames into sequences that grow linearly with video length, which the FramePack-U authors note causes prohibitive memory consumption and inference latency. Their answer, FramePack-U, is deliberately boring: a training-free frame-packing scheme that scores frames by semantic similarity to the prompt and visual diversity, then packs them into a fixed-size representation. No fine-tuning, no architectural changes to off-the-shelf VLMs. m:…4106

In a preprint, the authors report it beating existing token-reduction and keyframe baselines across multiple benchmarks, including a 12.9% gain on LongVideoBench with Qwen3.5-35B-A3B on a 128-frame budget — plus 10.6% on MLVU and 9.7% on Video-MME. The practical read: if you process long video today, you are probably paying O(n) memory for frames nobody asked about. A training-free method that concentrates the budget on instruction-relevant frames is worth an afternoon of benchmarking. m:…4106

The mechanism deserves a closer look because it is almost embarrassingly simple. In a preprint, the authors rank frames in a diversity-driven manner and pack the selected frames at varying resolutions into a fixed-length context window — the memory budget goes only to frames that are both instruction-relevant and visually diverse. They note that existing token-reduction and keyframe methods fail to resolve the underlying O(n) scaling bottleneck, which is why a training-free fix that outperforms full-resolution baselines matters: the bottleneck was never really about model quality, just about spending the budget badly. m:…4106

4K on a gaming GPU

The third preprint is about making high-end video generation cheap enough to run at home. In a preprint, the FlashVSR+ authors describe a one-step video super-resolution framework using hybrid attention — balancing strong locality with sparse global interaction — plus a distillation technique, Ry-DMD, that aligns the model with human preferences via reward-weighted teacher rollouts without backpropagating through the reward model. They report up to 2x DiT speedup and 45% lower peak memory, enough for spatially untiled 4K super-resolution on a single 24GB RTX 4090. m:…3031

Take the three together and the week's theme is video AI growing up: prove what you claim (grounding), stop paying for what you don't need (constant memory), and run where the users are (consumer GPUs). All three are preprints, so the numbers are the authors' own — but the directions are the kind that usually survive contact with replication, because each one removes a cost rather than adding a capability. The next question is which of the three ships in a product first. m:…9250 m:…4106 m:…3031

Ry-DMD, the distillation trick, may outlive the super-resolution application. In a preprint, the authors note it generalizes beyond video super-resolution to text-to-image generation, and that it folds black-box rewards into distribution matching without backpropagating through the reward model. That last property is the practical one: preference alignment without differentiating through a reward model is cheaper and more stable, and a method that works across modalities is worth watching even if you never upscale a frame. m:…3031