Recursive Harness Self-Improvement
RHI tunes the agent harness itself, using its own revision history to improve task traces and cut inference cost.
The paper treats harnesses as data-generating parts of future model training, not just runtime scaffolding. Its Recursive Harness Self-Improvement method rewrites a prompt-level agent loop through pairwise feedback over prior versions. On 30 synthetic ML research tasks, a few iterations lifted low-reasoning agents above the max-reasoning setting while reducing inference cost by up to 60%. The authors attribute the gains mainly to better task-specific context management and inter-agent information flow. HF Daily Papers' note
The paper treats harnesses as data-generating parts of future model training, not just runtime scaffolding. Its Recursive Harness Self-Improvement method rewrites a prompt-level agent loop through pairwise feedback over prior versions. On 30 synthetic ML research tasks, a few iterations lifted low-reasoning agents above the max-reasoning setting while reducing inference cost by up to 60%. The authors attribute the gains mainly to better task-specific context management and inter-agent information flow. HF Daily Papers' note
score 5