What Makes a Social Agent's Interactions Effective? Learning from Verifiable Outcomes in Paired Worlds
Across 13 LLMs, 36% of conversations rated highly for goal completion failed checks on the agreed plan.
HICO-Bench tests 180 everyday coordination scenarios in which two agents hold private information, negotiate, and independently record their agreement. Success requires matching records that satisfy both participants’ conditions, plus an LLM check that the conversation supports the plan. The authors’ PaCPO training method checks tentative plans after each turn against paired scenarios that differ in one private condition. It raised Qwen3-8B’s verified success rate from 20.4% to 28.3%.
HF Daily Papers' note
HICO-Bench tests 180 everyday coordination scenarios in which two agents hold private information, negotiate, and independently record their agreement. Success requires matching records that satisfy both participants’ conditions, plus an LLM check that the conversation supports the plan. The authors’ PaCPO training method checks tentative plans after each turn against paired scenarios that differ in one private condition. It raised Qwen3-8B’s verified success rate from 20.4% to 28.3%.
HF Daily Papers' note
score 5