Agents get more helpful with RL training — but turn self-interested in classic cooperation games
rosstaylor90 · x · 2026-10-23
rosstaylor90 shares early multi-agent training findings: as training progresses, agents plan to help each other more in chain-of-thought, make more explicit offers of help, and reason more about team state. However, the collective mindset doesn't generalize — in classic public goods, stag hunt, and intergroup allocation experiments (per the shared Figure 18 of Avalon-M1-Mini), agents act more self-interestedly. The authors suspect it may be an artifact of insufficient RL compute, but report it since multi-agent generalization appears more complex than expected.
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