Multi-Turn Agentic Scientific Literature Search via Workflow Induction

Researchers from the University of Illinois Urbana-Champaign, Together AI, the University of Pennsylvania, and Stanford introduce PAPER PILOT, a multi-turn scientific literature-search agent that treats retrieval as workflow induction rather than fixed query matching. Given an anchor paper and an evolving request, the agent builds an executable, type-consistent DAG from operators for keyword search, citation expansion, filtering, scoring, reranking, and evidence extraction, then edits that DAG in response to clarification feedback. Training combines supervised workflow imitation with IPO-style DPO preference optimization, using 2,723 anchor-query cases, 5,540 workflow-supervision examples, and 1,733 hard chosen–rejected pairs created by corrupting valid workflows. On a benchmark spanning five search directions, PAPER PILOT-9B improves the multi-turn Qwen3.5-9B toolset baseline from 58.0 to 77.0 Hit@5, from 47.5 to 59.4 MRR, and from 26.8 to 32.5 nDCG@10, while reducing workflow execution errors from 9.5% to 0%. It also reaches 74.7% human-session success and a 4.2 out-of-5 question-satisfaction score. The system is competitive with much larger models: GPT-5.4 with the PAPER PILOT toolset achieves 84.0 Hit@5, while OpenAI DeepResearch is evaluated as a one-shot baseline. The central result is that explicit, editable workflows translate user feedback into concrete retrieval operations, improving controllability, ranking quality, and interaction alignment.