New Coding-Agent Framework Raises Claude-4.5-Opus Benchmark Scores

A new paper outlines a test-time scaling framework for agentic coding that converts rollouts into structured summaries of key hypotheses, progress and failure modes while discarding lower-signal details. The approach supports two inference-time methods: Recursive Tournament Voting, which uses iterative small-group comparisons for parallel selection, and Parallel-Distill-Refine, which conditions new rollouts on distilled summaries for sequential improvement.

According to the paper, the framework improved Claude-4.5-Opus on SWE-Bench Verified to 77.6% from 70.9% and on Terminal-Bench v2.0 to 59.1% from 46.9%. The authors said the results suggest effective test-time scaling for long-horizon coding agents depends on representation, selection and reuse rather than simply sampling more trajectories.