Developer Tools

TTHE: New method evolves LLM agent programs at test time for 10-25% gains

Optimizing agent harnesses during evaluation with no labels or retraining

Deep Dive

Researchers introduce Test-Time Harness Evolution (TTHE), a method that evolves the executable program (harness) surrounding a frozen LLM during evaluation using only unlabeled execution traces. Across text-to-SQL, competitive programming, software engineering, data-science coding, and agentic tool-use tasks, TTHE improves fixed ReAct-style baseline harnesses, yielding persistent, inspectable improvements rather than a pre-searched workflow or per-query retries. The solver, proposers, and judge are different roles and harnesses around the same frozen LLM, so all adaptation occurs through changes to the surrounding program, without gold labels or weight updates.

Key Points
  • TTHE evolves the agent's harness (context, tools, error handling) at test time using only unlabeled execution traces — no gold labels or weight updates.
  • Outperforms fixed ReAct baselines by 10-25% across 5 domains: text-to-SQL, competitive programming, software engineering, data-science coding, and tool-use.
  • All adaptation occurs in the surrounding program — the LLM itself stays frozen, enabling practical deployment without retraining.

Why It Matters

LLM agents can now continuously improve their own decision-making programs during deployment, adapting to new tasks without expensive retraining.

📬 Get the top 10 AI stories daily