AI Summary: KLong addresses the 'memory rot' that occurs when agents attempt tasks lasting several hours. It uses a unique training method that breaks long trajectories into manageable segments during fine-tuning, allowing the model to learn intermediate checkpoints.
Current LLM agents frequently fail in tasks requiring hundreds of steps due to error accumulation and context overflow. We introduce KLong, an agentic framework that utilizes 'Trajectory-Splitting Supervised Fine-Tuning' (TS-SFT) and progressive reinforcement learning to master extremely long-horizon tasks. KLong (106B) achieves state-of-the-art results on PaperBench, outperforming models ten times its size by maintaining high reasoning fidelity across thousands of tokens. This work provides a scalable method for training agents that can independently conduct scientific research and complex project management.
Share your opinion to help other learners triage faster.
Write a reviewInvite someone by email to share an invited review for KLong: Training LLM Agents for Extremely Long-horizon Tasks.