Home
Topics
Search
Library
Get Started
Research papers, read to you in five minutes.
Research papers, read to you in five minutes.
Every new AI paper that matters, distilled into a short audio episode.
Every new AI paper that matters, distilled into a short audio episode. Follow topics, listen on your commute, skim the recap when you're back.
Today's lead
4DAnyone: Create Anyone in 4D from a Casual Monocular Video
Now playing · Today's lead
4DAnyone: Create Anyone in 4D from a Casual Monocular Video
This week
all topics →
All
Video Generation
Computer Vision
Multimodal
Reinforcement Learning
4DAnyone: Create Anyone in 4D from a Casual Monocular Video
Multimodal · Aug 20
Learn What's Left, Not What's Mastered: Saturation Aware Advantage Reweighting for Multi-Reward Policy Optimization
Reasoning · Aug 17
DreamX-Phi 1.0: Action-Conditioned Video World Model for Robotic Manipulation
Mechanistic Interpretability · Aug 13
Beyond Final Scores: A Systematic Evaluation of Agents for Long-Horizon AI Research and Development
Agents · Aug 13
ComBodied Agents: a New Paradigm of Human-Centric Agentic AI
Agents · Aug 11
Never fall behind the literature again.
Sign up free
Multimodal · Aug 20
4DAnyone: Create Anyone in 4D from a Casual Monocular Video
Turning one phone video into a free-viewpoint avatar breaks because diffusion models can't hold 16+ novel views in one attention pass. 4DAnyone compresses accumulated reference views into a fixed token budget and rotates target-view groupings during high-noise steps, letting global structure propagate even when memory can't.
Reasoning · Aug 17
Learn What's Left, Not What's Mastered: Saturation Aware Advantage Reweighting for Multi-Reward Policy Optimization
Multi-reward RLVR wastes gradient reinforcing objectives the model already maxed out. SA-MRPO reweights each objective's advantage by (1 − saturation)^γ before normalization, redirecting pressure to unsolved rewards — +9.2pp on AMC23 when a length reward saturates, with γ≈0.5 as the sweet spot.
Mechanistic Interpretability · Aug 13
DreamX-Phi 1.0: Action-Conditioned Video World Model for Robotic Manipulation
Robot video world models look right but drift from commanded actions, making them useless for scoring candidate rollouts. DreamX-Phi wires SE(3) end-effector transforms directly into attention as relative rigid motions, splitting heads per arm — ranking first of 31 on WorldArena 2.0 Track 1 at 60.65 EWMScore-P.
Agents · Aug 13
Beyond Final Scores: A Systematic Evaluation of Agents for Long-Horizon AI Research and Development
Ranking long-horizon coding agents by final score hides why they fail. Decomposing 36 AutoLab runs across seven frontier models into framing, execution, and feedback-control reveals reliability — not peak skill — separates them: 0.237 spread on avg@3 vs just 0.122 on best@3. Fix the loop bottleneck, not the leaderboard.
Agents · Aug 11
ComBodied Agents: a New Paradigm of Human-Centric Agentic AI
Most agent work optimizes software or physical state; this position paper argues the missing paradigm optimizes the human over months, scoring preserved agency and capability growth instead of task completion. The concrete ask: instrument an intervention-conditioned personal world model where "do nothing" and "escalate" are first-class actions.
Never fall behind the literature again.
Free account. Follow topics, build your queue.
Sign up free
Multimodal · Aug 20
4DAnyone: Create Anyone in 4D from a Casual Monocular Video
Turning one phone video into a free-viewpoint avatar breaks because diffusion models can't hold 16+ novel views in one attention pass. 4DAnyone compresses accumulated reference views into a fixed token budget and rotates target-view groupings during high-noise steps, letting global structure propagate even when memory can't.
Reasoning · Aug 17
Learn What's Left, Not What's Mastered: Saturation Aware Advantage Reweighting for Multi-Reward Policy Optimization
Multi-reward RLVR wastes gradient reinforcing objectives the model already maxed out. SA-MRPO reweights each objective's advantage by (1 − saturation)^γ before normalization, redirecting pressure to unsolved rewards — +9.2pp on AMC23 when a length reward saturates, with γ≈0.5 as the sweet spot.
Mechanistic Interpretability · Aug 13
DreamX-Phi 1.0: Action-Conditioned Video World Model for Robotic Manipulation
Robot video world models look right but drift from commanded actions, making them useless for scoring candidate rollouts. DreamX-Phi wires SE(3) end-effector transforms directly into attention as relative rigid motions, splitting heads per arm — ranking first of 31 on WorldArena 2.0 Track 1 at 60.65 EWMScore-P.
Agents · Aug 13
Beyond Final Scores: A Systematic Evaluation of Agents for Long-Horizon AI Research and Development
Ranking long-horizon coding agents by final score hides why they fail. Decomposing 36 AutoLab runs across seven frontier models into framing, execution, and feedback-control reveals reliability — not peak skill — separates them: 0.237 spread on avg@3 vs just 0.122 on best@3. Fix the loop bottleneck, not the leaderboard.
Agents · Aug 11
ComBodied Agents: a New Paradigm of Human-Centric Agentic AI
Most agent work optimizes software or physical state; this position paper argues the missing paradigm optimizes the human over months, scoring preserved agency and capability growth instead of task completion. The concrete ask: instrument an intervention-conditioned personal world model where "do nothing" and "escalate" are first-class actions.
Never fall behind the literature again.
Free account. Follow topics, build your queue.
Sign up free
Browse topics
Video Generation 19
Computer Vision 45
Multimodal 35
Reinforcement Learning 23
Reasoning 24
ML 14
All topics →
© 2026 r*cap — research papers as 5-minute podcasts
Terms
Privacy