ClueFlow: Learning Long-Horizon Search through Attributed Research-State Transitions

Key points

• Long-horizon search requires agents to acquire relevant evidence and integrate it into an evolving research state.1
• ClueFlow constructs Source-State-Final attribution chains that connect evidence-acquiring browser actions, Evidence and Candidate updates, and the final answer supported by those updates.1
• We evaluate ClueFlow across two Qwen backbones and four search benchmarks spanning live-web and fixed-corpus settings, observing consistent gains from SFT and further improvements from PPO.1
• These results demonstrate that state-transition attribution enables efficient supervision and effective credit assignment for long-horizon search agents.1

Sources (1)

• [1]ClueFlow: Learning Long-Horizon Search through Attributed Research-State Transitions
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 19, 08:07 AM

  We introduce ClueFlow, which learns long-horizon search from attributed research-state transitions.

  Long-horizon search requires agents to acquire relevant evidence and integrate it into an evolving research state.

Extractive summary: sentences quoted from the sources.