What matters in AI.

Subscribe

Life-Harness lifts frozen agents with no change to model weights

With Qwen3-4B-Instruct, its mean accuracy was 28.7 points above Meta-Harness.

Researchers show Life-Harness, a runtime layer that turns agent errors into fixes. It does not change model weights. With Qwen3-4B-Instruct, it had a score of 28.7 points above Meta-Harness and 7.5 points above AHE.

Life-Harness vs rivals With Qwen3-4B-Instruct weights unchanged, Life-Harness beats Meta-Harness by 28.7 points and AHE by 7.5 points in mean accuracy across seven shared environments, with 8.5% and 68.4% fewer optimization tokens. Life-Harness vs rivals Qwen3-4B-Instruct, weights unchanged Meta-Harness AHE ACCURACY LEAD +28.7 points +7.5 points FEWER OPTIMIZATION TOKENS 8.5% fewer 68.4% fewer Mean accuracy, seven shared environments
With the Qwen3-4B-Instruct model frozen, Life-Harness beat two rival harnesses on mean accuracy across seven shared environments, using fewer optimization tokens.

Sources

  1. Adapting the Interface, Not the Model: Runtime Harness Adaptation for Deterministic LLM Agentsarxiv.org
AI MATTER · NEWS · AI MATTER · NEWS ·16 OCT2026

Posted

Tags

Learn the terms in this story

More in Agents

All Agents news