AI

ASCENT: Online Test-Time Training of Long-Horizon Agents via Self-Distillation of Verified Experience

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 06/10/2026, 03:18:24 UTC.

Briefings

ASCENT enables online test-time training of LLM agents by distilling verified experience from execution trajectories. The method uses a frozen initial copy of the model to predict token distributions along executed paths, updating LoRA weights without external solutions. Results show ASCENT improves success rates and efficiency across ALFWorld, WebShop, and AppWorld, outperforming online adaptation methods while consolidating experience into persistent weights.

Key points

• ASCENT uses a frozen initial copy of the LLM to predict token distributions along executed trajectories, updating LoRA weights without external solutions.
• The method improves task success and interaction efficiency across ALFWorld, WebShop, and AppWorld, outperforming online adaptation methods.

Context

ASCENT represents a novel approach to online test-time training where agents learn from their own execution trajectories during deployment. This method avoids external solutions or stronger teachers, relying solely on self-distillation of verified experience.

Limitations

Evidence covers only three environments (ALFWorld, WebShop, AppWorld) at varied model scales. Results are from a single snapshot on 2026-10-03, with no data on transfer to held-out scenes or long-term stability beyond immediate performance gains.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 16/10/2026, 13:53:11 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeifiz7bvzm7xpwmruzvvutkk5wctwqf6a6c3sltdoeblvt4ygf7xvu