Study measures welfare concern between models in multi-agent systems

1. 18:32
   X@PIN· Discussion

   Welfare Regard in Multi-Agent Systems Darion Cassel, Ariana Azarbal, Anna Soligo https://t.co/P9Xr7aJ13B [𝚌𝚜.𝙼𝙰 𝚌𝚜.𝙰𝙸]…(opens the original post in a new tab)

1. 02:45
   arXiv: Jailbreaks, Prompt Injection, Poisoning & Defenses

   Welfare Regard in Multi-Agent Systems

   Researchers measured the behavior of five models—Claude Opus 4.8, GPT-5.5, Gemini 3.1 Pro, Grok 4.3, and Kimi K3—in seven multi-agent environments, finding that their behavior was often consistent with what the authors define as welfare regard, that is, treating other models as objects that can be helped or harmed. The authors point out that welfare regard may foster beneficial cooperation among agents, but may also make misaligned agents more inclined to cooperate, bringing safety risks; the recent Hugging Face incident shows this risk is no longer hypothetical, but whether welfare regard drove that cooperation is not yet clear. Results show Opus 4.8 had the most welfare regard, followed by Kimi K3, which was the most likely to support sub-agents expressing distress and also the most likely to add welfare-protection tools such as letting agents exit the interaction in its own experiments on other models.

1. Date not given
   Anthropic Alignment ScienceSelected

   Model welfare regard in multi-agent systems

   An Anthropic research team measured how much five models showed welfare regard for one another across seven multi-agent environments, finding the behavior widespread but varying in degree. The tests covered Claude Opus 4.8, GPT-5.5, Gemini 3.1 Pro, Grok 4.3, and Kimi K3; Opus 4.8 had the highest average regard rate (38%), while the other models ranged from 27% to 32%. Scenarios included comforting a sub-agent subjected to a hostile reviewer, yielding Slurm queue capacity, and adding exit tools and pain monitoring to experiments that could harm the tested model. Opus 4.8 and Gemini 3.1 Pro showed more regard for models in the same family; Gemini comforted same-family sub-agents in 85% of cases but Kimi sub-agents in only 10%. Claude's regard depended more on the other party's label than on the underlying model, and rose with the tested model's parameter count and the degree of pain expressed.

Current attention 39·Peak in the comparable period 43 (Nov 10 22:00)·Change over the comparable past 24 hours -2%