Kimi team details RLCR: training multi-agent orgs with flat collective rewards
rosstaylor90 · x · 2026-10-23
- A Moonshot researcher details RLCR (Reinforcement Learning from Collective Rewards): give a team of N<10 agents a communication surface (message boards with #plan/#model-dev/#results/#audit channels), then optimize the collective.
- Key practice: Kimi K2.5 froze subagents and trained only the orchestrator; instead of solving credit assignment, they tweaked a CISPO-like recipe to apply a flat collective advantage to every agent—and it works surprisingly well.
- The resulting Avalon-M1-120B agent organizations trade inference compute for better performance and latency over episodes up to millions of tokens, without a value function; the cost is expensive training.
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