News
Our paper “Adapting the Interface, Not the Model:Runtime Harness Adaptation for Deterministic LLM Agents” is open-sourced at Life-Harness
. Four-layer Harness improve cross-model performance by 88% without training.
. Four-layer Harness improve cross-model performance by 88% without training.Our paper “CLEANER:Self-Purified Trajectories Boost Agentic Reinforcement Learning” is accepted to ICLR 2026 SPOT Workshop! [Paper]; [Code]. CLEANER resolves the credit assignment dilemma in agentic RL by training on self-purified trajectories, achieving SOTA performance with just one-third of the training cost.