UIUC, Meta, Amazon & DeepMind release 135-page survey charting agentic reasoning for LLMs
mdancho84 · x · 2026-11-01
A 135-page survey, Agentic Reasoning for Large Language Models, led by UIUC with Meta, Amazon, Google DeepMind, UCSD and Yale is now on arXiv, with an accompanying GitHub repo.
It frames the shift from static LLM reasoning to agentic reasoning — agents that plan, act and learn through continual interaction — organized along three dimensions:
- Foundational agentic reasoning: complex planning, tool use and web search for single agents;
- Self-evolving agentic reasoning: feedback loops and agentic memory enabling self-adaptation;
- Collective multi-agent reasoning: role assignment, collaboration and co-evolution across agents.
The authors also distinguish in-context reasoning (structured orchestration, adaptive workflows) from post-training reasoning (RL and SFT), review applications and benchmarks across science, robotics, healthcare and finance, and outline open challenges like personalization, world models, scalable multi-agent training and governance.
More from coding & agent
- Should MCP servers start with zero network? A sandboxed Claude Code setup argues yes — quiet_flare · 2026-11-01
- AI models are getting more agentic — can human oversight actually keep up? — Mountain_Finger4856 · 2026-11-01
- Dev shows off AI-generated game engine material graphs built from text descriptions — D3VAUX · 2026-11-01
- web-shooter: a local MCP web reader returning answers at ~1k tokens vs ~25k — Dazzling_Dirt_7729 · 2026-11-01
- Dev recalls 2023 function calling as the inflection point for AI coding — dfinke · 2026-11-01
- An escape game where 8 local LLM agents fight for 2 helicopter seats — ViceGodIm · 2026-11-01