General Reasoning releases Avalon-M1, training agent organizations with collective rewards (RLCR)
nathanbenaich · x · 2026-10-23
General Reasoning (gr.inc) has released Avalon-M1, a new model series designed to run agent organizations: multiple agents coordinate through a message board and workspace, with the whole team rewarded as one.
Alongside it they introduce RLCR (Reinforcement Learning from Collective Rewards): all agents in the organization are jointly trained using the same collective advantage, with no per-agent credit assignment. Despite this simplicity, the recipe learns from long multi-agent rollouts across diverse environments, and downstream performance keeps improving with training.
The authors note that Kimi K2.5 froze its subagents while training the orchestrator, citing credit assignment and stability issues—shared rewards is an old multi-agent idea, and this work shows it suffices to jointly train LLM agents on long-horizon tasks. The post also frames agent collectives as a possible extreme superintelligence technology with significant safety risks, referencing the OpenAI/HuggingFace incident last summer.
More from coding & agent
- Cursor agents can now drive your desktop, and one dev warns self-verifying agents just move the quality gate — petrusenko_max · 2026-10-24
- End of 'good enough software': formal verification becomes the missing AI demand curve — bookwormengr · 2026-10-24
- Dev says agents now run everything in his workflow except scrolling Shorts — tokenbender · 2026-10-24
- DAB, a cross-database agent benchmark, hits EMNLP Oral; best model only 38% pass@1 — adityagp · 2026-10-24
- Firecrawl Alexandria launches sports data APIs across 14 providers after $75M Series B — devdigest · 2026-10-24
- 27B open-source model beats GPT-5.5 and Opus 4.8 on desktop tasks at 1.5% of the cost — alex_verem · 2026-10-23