Skip to content- Sources: primary, discussion
- Summary: MASS applies multi-agent self-supervision to Qwen3.6-27B in the qwen-code agent environment and reports 1.2x to 1.6x score per output token after two rounds on MLR-Bench, DSBench, ScienceAgentBench, and AstaBench. The same configuration reports 0.96x on Terminal-Bench 2.0 and 0.94x on SWE-bench Verified. The post publishes the regressions alongside the gains.
- Why it matters: A team reading this as a route to cheaper coding agents gets nothing from it, because the two software engineering benchmarks are the ones where the method falls below its own baseline.