Interlock synthesizes completion graphs to train long-horizon agents
Original title: Interlock: From Local Correctness to Task Completion in Long-Horizon Agents
Fine-tuning two Qwen3.6 backbones on accepted trajectories lifts AutomationBench pass rates by 4.00-4.50 points.
Interlock represents long-horizon tasks as typed completion graphs linking evidence, decisions, and required effects, then compiles them into a workspace, request, and rubric. Reviews and an agent-as-judge keep trajectories that meet mandatory requirements. The authors fine-tune two Qwen3.6 backbones on accepted trajectories and report gains on three external benchmarks.
Unconfirmed: Training gains are author-reported, not independently confirmed.
- AutomationBench pass rates rise 4.00-4.50 percentage points.
- JobBench weighted scores rise 3.20-6.64 points.
- Harvey LAB criterion pass rates rise 3.89-4.16 percentage points.
- Each backbone is trained on its own accepted trajectories.
From the original article
1 min read at huggingface.co
Large language model agents increasingly undertake long-horizon tasks in complex environments characterized by fragmented evidence, partially completed states, and dependencies across resources and artifacts. Existing agents may satisfy individual requirements while failing to maintain these relationships, producing locally correct but incomplete outcomes.
By Linji Hao, Yanji He, Zihan Wang · 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.