Dimensions overall score 7.0

GitHub Code Pulse

No public code linked for this paper yet.

• Meta-Harness improves online text classification accuracy by 7.7 points over the state-of-the-art context management system ACE while using 4x fewer context tokens.result 95%
• On retrieval-augmented math reasoning, a single harness discovered by Meta-Harness improves accuracy on 200 IMO-level problems by 4.7 points on average across five held-out models.result 95%
• On agentic coding, harnesses discovered by Meta-Harness surpass the best hand-engineered baselines on TerminalBench-2.result 90%
• Meta-Harness provides orders-of-magnitude more context per artifact evaluation compared to prior text optimization methods.technical 85%
• Access to raw execution traces is the key ingredient enabling effective harness search in Meta-Harness.technical 90%
• Meta-Harness can discover harnesses across a broad range of the accuracy-context trade-off frontier by expressing joint preferences rather than a single scalar objective.method 80%
• Meta-Harness outperforms the next best method by 2.9 points on average across nine out-of-distribution text classification datasets.result 90%
• Changing the harness around a fixed LLM can produce a 6× performance gap on the same benchmark.technical 85%

Competitive landscape

Competitor map is still being generated for this paper. Enable generation or check back soon.

Startup potential card