KLong: Training LLM Agent for Extremely Long-horizon Tasks

Dimensions overall score 8.0

GitHub Code Pulse

No public code linked for this paper yet.

• KLong (106B) surpasses Kimi K2 Thinking (1T) by 11.28% on PaperBench.result 90%
• The performance improvement of KLong generalizes to other coding benchmarks like SWE-bench Verified and MLE-bench.result 85%
• Trajectory-splitting SFT preserves early context, progressively truncates later context, and maintains overlap between sub-trajectories to handle long trajectories.method 90%
• Progressive RL schedules training into multiple stages with progressively extended timeouts to improve long-horizon task-solving capability.method 90%
• The Research-Factory pipeline generates high-quality training data by collecting research papers and constructing evaluation rubrics.method 85%
• The approach's reliance on RL introduces challenges in stability and efficiency during learning.limitation 80%
• The method depends on high-quality training data and evaluation rubrics, which may vary in availability and quality.limitation 80%
• KLong could replace traditional manual methods of academic replication and validation, offering a faster and potentially more accurate automated solution.market 70%

Competitive landscape

Competitor map is still being generated for this paper. Enable generation or check back soon.

Startup potential card