ASCENT: Online Test-Time Training of Long-Horizon Agents via Self-Distillation of Verified Experience
ASCENT updates deployed LLM agents by self-distilling verified trajectories into LoRA weights at test time.
ASCENT is an online test-time training method that updates an LLM agent's weights from a single verified execution trajectory, without an external teacher or retrieved memory. A frozen copy of the initial model uses that trajectory as privileged hindsight and the resulting next-token distributions are distilled into persistent LoRA adapters; invalid-action turns are removed. Across ALFWorld, WebShop, and AppWorld at multiple scales, success and interaction efficiency improve as experience accumulates, including on held-out scenes.
- Updates persistent LoRA weights from one verified trajectory per task
- Frozen initial model supplies hindsight next-token targets
- Drops invalid-action turns before distillation
- Improves ALFWorld, WebShop, and AppWorld success and efficiency
From the original article
1 min read at huggingface.co
A large language model (LLM) agent solves long-horizon tasks through many reasoning-action turns, with one verification signal at termination. Deployed agents face streams of related tasks, making their trajectories a natural resource for improvement. In-context adaptation agents store reflections, memories, or skills as text, so reuse depends on retrieving the right experience and on a frozen policy executing it.
By Haodong Lu, Dong Gong · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.