Skip to content
daily-hour-news·

🔬Paper: Train Long-Context LLMs on Agent Trajectories

TL;DR

A new arXiv paper, Agent Context Compiler (ACC), turns the long tool-call trajectories agents already produce into long-context training data. It targets the blind spot in standard agent fine-tuning, which masks tool responses and only learns turn-level tool selection.

A new arXiv paper, Agent Context Compiler (ACC), turns the long tool-call trajectories agents already produce into long-context training data. It targets the blind spot in standard agent fine-tuning, which masks tool responses and only learns turn-level tool selection.

Paper: Train Long-Context LLMs on Agent Trajectories — daily-hour-news

Key Points

1

Submitted to arXiv on 21 May 2026 (2605.21850)

2

Reuses agent trajectories instead of costly long-document curation

3

Targets the SFT blind spot where masked tool outputs go unused

4

Aims to integrate evidence scattered across many agent turns

Why It Matters

Free long-context training data harvested from agent runs could cut one of the bigger costs in building long-horizon agents.

Quick Facts

researchlong contextAI agentsLLM trainingarXivfine-tuning

Comments

Subscribe to join the conversation...

Be the first to comment

Enjoyed this article?

Get it daily. 7am. Free. Reads in 5 minutes.

Join 3,483 builders reading daily.

Also get