mlsanityData to 7 Oct 2026

Harness Annealing: Learning to Act with Less External Control

arXiv 2610.01235 ↗First posted 1 Oct 2026cs.CLcited by 0

Not on the map. Why?

Abstract (arXiv)

Language agents rely on external harnesses to track state, organize workflows, and verify answers. Beyond providing tools and information, these harnesses supply control decisions about what to investigate, whether to revise, and when to stop. Training on successful harness-supported trajectories can improve task performance while leaving these decisions dependent on runtime intervention. We ask whether harness-supported experience can also teach the model to make these decisions, allowing the division of control to change as the model learns. We call this objective harness internalization: learning to assume specified control responsibilities while retaining task performance after the corresponding support is withdrawn. We introduce HARNESS ANNEALING TRAINING (HAT), which combines explicit control supervision with a curriculum over teacher trajectories collected under progressively weaker harnesses. Experiments with 9B and 35B models on SWE-QA and SWE-QA-Pro evaluate every checkpoint under four deployment harnesses. Selected annealed checkpoints operating with tools alone achieve scores close to those of their respective starting checkpoints deployed with the full harness. The benefits vary with model scale and deployment configuration, and further annealing does not uniformly improve performance. These findings suggest that harness-supported experience can help reduce the runtime control required by a trained agent.

Key facts. No citing papers in the map yet (data through 7 Oct 2026). Builds on 19 of its 28 references that are in the map. First posted on arXiv 1 Oct 2026.

Builds on

19 of 28 references.

References · 19

  1. 2015Distilling the Knowledge in a Neural Networkcited by 10,076
  2. 2024SWE-agent: Agent-Computer Interfaces Enable Automated Software Engineeringcited by 1,628
  3. 2024LlamaFactory: Unified Efficient Fine-Tuning of 100+ Language Modelscited by 1,442
  4. 2023SGLang: Efficient Execution of Structured Language Model Programscited by 1,107
  5. 2025τ2-Bench: Evaluating Conversational Agents in a Dual-Control Environmentcited by 558
  6. 2024Automated Design of Agentic Systemscited by 360
  7. 2023AgentTuning: Enabling Generalized Agent Abilities for LLMscited by 319
  8. 2026Meta-Harness: End-to-End Optimization of Model Harnessescited by 248
  9. 2023FireAct: Toward Language Agent Fine-tuningcited by 247
  10. 2024Agent-FLAN: Designing Data and Methods of Effective Agent Tuning for Large Language Modelscited by 97
  11. 2026Skill-SD: Skill-Conditioned Self-Distillation for Multi-turn LLM Agentscited by 75
  12. 2025SWE-QA: Can Language Models Answer Repository-level Code Questions?cited by 37
7 more
  1. 2026HarnessX: A Composable, Adaptive, and Evolvable Agent Harness Foundrycited by 37
  2. 2026τ-Knowledge: Evaluating Conversational Agents over Unstructured Knowledgecited by 35
  3. 2025ATLaS: Agent Tuning via Learning Critical Stepscited by 29
  4. 2026Stop Comparing LLM Agents Without Disclosing the Harnesscited by 25
  5. 2026SWE-QA-Pro: A Representative Benchmark and Scalable Training Recipe for Repository-Level Code Understandingcited by 9
  6. 2026Harness-1: Reinforcement Learning for Search Agents with State-Externalizing Harnessescited by 6
  7. 2026Compiling Agentic Workflows into LLM Weights: Near-Frontier Quality at Two Orders of Magnitude Less Costcited by 4

Built on by

Not cited yet.