Meta-Harness: End-to-End Optimization (Updated)

Updated results page for Meta-Harness (Yoonho Lee et al.). Extends original page with new benchmarks on text classification, math reasoning, and agentic coding.

Key New Results

  • Text Classification: 48.6% vs ACE’s 40.9% (+7.7 points) using 4x fewer context tokens on GPT-OSS-120B
  • Math Reasoning: +4.7 points average across 5 held-out models on IMO-level problems, with a single discovered retrieval harness
  • TerminalBench-2: 76.4% on Opus 4.6 (#2 overall), 37.6% on Haiku 4.5 (#1 overall)
  • Matches OpenEvolve/TTT-Discover accuracy with 10x fewer evaluations

What Makes It Different

  • 10M tokens of diagnostic context per optimization step vs ≤26K for all prior methods
  • Proposer (Claude Code) navigates a full filesystem of prior candidates’ source code, execution traces, and scores via grep/cat
  • The key insight: harness failures are hard to diagnose without raw execution traces; compressing to scores/summaries loses critical information
  • Filesystem-based interface means the agent reads what it needs, not a fixed prompt window

Comparison Table

MethodHistoryMtok/iter
Self-RefineLast0.001
OPROWindow0.002
AlphaEvolveWindow0.022
TTT-DiscoverWindow0.026
Meta-HarnessFull10.0

Takeaways

  • The “give the optimizer everything” approach works dramatically better than compression
  • This validates the general principle: more context > better prompts, even for optimization
  • Transfer across models: a harness discovered with one model improves others
  • TerminalBench-2 results show practical agent improvement from harness optimization

Connections