News summarySynced (Jiqizhixin)Nov 1, 2026, 11:17 AM
Google-Led Researchers Release a Multi-Agent Conflict Benchmark
A research team from Google and several universities jointly released CLASHarness. [[fact:fact-p6kxnnleu]] The benchmark evaluates 15 mainstream models in multi-agent conflict scenarios. [[fact:fact-0xhhaudwd]] Its dataset covers 20 domains, 600 executable multi-turn conflict tasks, 7,970 4K visual-evidence images, and 3,284 required checkpoints. [[fact:fact-4h725osf8]] Under the standard 20-step limit, Claude Opus 5 and GPT-5.6 Sol ranked highest, with pass rates of 25.2% and 24.5%, respectively. [[fact:fact-9c4u2mwul]] With a higher step allowance, GPT-6 Astra led with a 28.7% pass rate. [[fact:fact-j8bvu8nf8]]