ASCENT: online test-time training lets LLM agents self-distill verified experience into weights

Haodong Lu · hf · 2026-10-06

Introduces Online Agentic Test-Time Training (OaTTT): an LLM agent trains its own weights on deployment-time execution trajectories, with the trajectory plus its verification result as the only learning signal, persisting across a stream of tasks.
Directly imitating or reinforcing single-attempt tokens destabilizes the policy; ASCENT instead self-distills verified experience — a frozen copy of the LLM receives the trajectory as privileged information, predicts next-token distributions with hindsight, and distills them into persistent LoRA fast weights, with no external reference solution or stronger teacher.
Invalid-action turns are further removed to distill enhanced privileged experience.
Across ALFWorld, WebShop, and AppWorld at varied model scales, ASCENT improves task success and interaction efficiency as experience accumulates, outperforms online adaptation methods, and transfers to held-out scenes — consolidating verified experience into weights without a separate training phase or memory retrieval.

More from coding & agent