Your daily digest
Research papers, made simple to understand
The AI and machine learning papers people are reading most, each one rewritten in plain words: the problem, the idea, what was found and why it matters.
- AI Agents
- Multimodal & Vision
Capability-driven recursive self-improvement evolves specialized multimodal agent modules
The Trinity framework organizes diverse multimodal tasks by shared capabilities, evolving reusable Harness modules through recursive self-improvement and capability-aware routing, improving performance by up to 34 points on training and 26 points on unseen benchmarks, including real-world embodied tasks.
Paper: The Trinity:Capability-Driven Recursive Self-Improvement for Multimodal Agents · Weidong Tang, Ruiheng Wang, Jingxuan Zhang and 17 others
532 upvotes on Hugging Face6 min readRead the article
- AI Agents
- Reinforcement Learning & Alignment
Multiscale Policy Optimization improves credit assignment for long-horizon LLM agents
This paper introduces Multiscale Policy Optimization (MPO), which aligns credit assignment and policy updates at the turn level while refining token-level credit, achieving consistent gains over strong RL baselines on search, tool-use, and automation tasks with 4B and 30B models.
Paper: Multiscale Policy Optimization for Long-Horizon LLM Agents · Tao Ren, Hui Yang, Jia Deng and 16 others
373 upvotes on Hugging Face6 min readRead the article
- Multimodal & Vision
ChronoReplay improves spatial VLMs by balancing and replaying diverse spatial tasks
ChronoReplay trains a spatial vision-language model on heterogeneous spatial tasks by replaying chronologically ordered episodes with native answer formats and balanced supervision, boosting performance on three spatial benchmarks by up to 13.41 points under fixed observation access.
Paper: ChronoReplay: One Answer Interface Is Not Enough for Streaming Spatial Intelligence · Yihao Hu, Yitong Lin, Pan Wang and 8 others
346 upvotes on Hugging Face7 min readRead the article
Browse earlier papers
2 days of papers in November 2026