The authors applied a recursive self-improvement-inspired process at the harness layer, repeatedly testing candidate changes across increasingly numerous and diverse rollout environments.
Recursive harness search cuts coding-agent token use nearly in half
SoL-Pi combines four automatically selected harness mechanisms to preserve task performance while reducing token traffic and API costs.
Big Tech
Haozhe Liu · Tian Ye · Sensen Gao · Qihang Cao · Yitong Li · Mingchen Zhuge · +8 more
NVIDIA · NTU · MIT
Research Digest··2 min read
Thread:Agent Harness Optimization
Liu et al.
Why this paper
From NVIDIA and 2 others · Part of Agent Harness Optimization, now 67 papers
In one line
SoL-Pi reduces token traffic by 44.7-49.0% and API cost by about one third while matching Pi's performance on EdgeBench.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ·No stated limitations found
- ✓Reports numbers on named benchmarks
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§