ChatPaper
The papers AI researchers worldwide upvoted this week, explained in plain language
Think of video understanding as watching a movie—most AI models need to see every frame in high resolution, like reading every line of a script. VideoChat3 instead uses a novel 3D vision transformer (I3D-ViT) that adapts frame resolution to the action, skipping unnecessary details. With only 4B parameters, it outperforms larger open-source models on diverse video tasks, thanks to its efficient design and tailored training datasets. This makes advanced video AI more accessible, addressing scalability and reproducibility issues in current open-source models.
Think of it as a small model learning a clever trick, then showing the big model the 'aha' moment directly. Instead of re-running costly RL training, Direct-OPD transfers the small model's improved reasoning pattern as a dense reward signal. On AIME 2024, this boosted Qwen3-1.7B from 48.3% to 58.3% in just 4 hours on 8 A100 GPUs. The takeaway: large models can get smarter without expensive RL from scratch.
Imagine a GPS that sees the road and lets you steer. That's the idea behind ABot-N1, a navigation AI that separates high-level reasoning from low-level control. In urban tests, it boosted the rate of reaching specific places by 35.0% (to 77.3%) and achieved 95.4%/92.9% success in complex indoor/outdoor scenes. By grounding decisions in pixel-level goals, it becomes more robust and interpretable — crucial for real-world deployment.
Think of a student who, after each exam, writes down what worked and what didn’t, then uses those notes to study for the next test. SEED does something similar for AI: it has the policy itself analyze past complete episodes to generate concise natural-language "skills" (like "grab the wrench before opening the valve"). These skills are re-scored against current actions, creating a dense signal that helps the agent learn from sparse rewards. Because the skill analyzer evolves with the policy, the feedback stays relevant. The result: better performance on long, multi-step tasks like tool use or conversation.
Imagine solving a puzzle while forgetting each piece you've placed. That's how most AI search agents work—their progress stays hidden, causing loops. SearchOS changes this by making search state explicit, persistent, and shared across agents. It frames the task as relational schema completion with grounded citations. On benchmarks, SearchOS leads all metrics among evaluated baselines. This matters because it makes information-seeking agents more reliable and efficient, reducing wasted search and improving output quality.
Imagine reading a long book but only highlighting a few sentences at a time—that’s LongStraw. It cuts the training graph by evaluating the shared prompt without autograd and replaying short response branches one by one. This enables million-token RL post-training under a fixed GPU budget. For example, grouped Qwen scoring and response backward at 2.1M positions on eight H20 GPUs, with only 0.21 GB extra memory per group. Why it matters: AI agents need long context for decisions; LongStraw bridges that gap.
Like a chess player who visualizes a winning move but makes a blunder, world-action models (WAMs) can seem to imagine a correct future while executing a wrong action. BadWAM uses small visual perturbations to break this alignment. Under an action-only attack, task success drops from 96.5% to 43.1%. The model's imagined future appears normal, but its actual actions are desynchronized. Why it matters: This vulnerability undermines the promised safety and interpretability of WAMs for embodied control.
5 minutes a day to keep up with AI
5 trending papers daily, explained in plain words, plus one quick puzzle.
Read today's issue →