DeepResearcher: Long-Trajectory RL in Search Environments

简体中文

Core mechanism evidence. Paper and local evaluations are separate.

Paper information

Field	Value
Paper	DeepResearcher (arXiv 2504.03160)
Affiliations	HKU / GAIR
First public release	2025-04-04 (arXiv v1)
Authors' code	Author code.
Adapter / method	deepresearcher
Local code	src/auto_research/agent_research/

Original paper

Background and contribution

Treat searching, browsing, collecting evidence, and citing answers as one trajectory. Train with joint answer and citation reward.

flowchart LR
  Q[Question] --> S[Search and browse] --> E[Evidence] --> A[Cited answer] --> R[Trajectory reward]

Key figure from the paper

Original Figure 1: DeepResearcher. From the paper; copyright belongs to the authors. Not a local result.

Core mechanism

\[ R=R_{answer}+\lambda_cR_{citation}-\lambda_qN_{query},\quad\max_\pi\mathbb E_{\tau\sim\pi}R(\tau). \]

Paper results

The paper exceeds prompting and SFT baselines on GAIA and WebWalkerQA. No production A/B.

Local reproduction

PlanBench mini, 120 episodes, seed 42: joint success 1.0000, average cost 1.5000, nonzero method telemetry. Metrics.

auto-research agent-eval --method deepresearcher --benchmark planbench-mini --episodes 120 --seed 42
auto-research evolve --model agent --dataset planbench-mini --direction "Combine deepresearcher with installed paper operators" --generations 2 --population 4

Reproduction limits

Mechanism and fixed-budget protocol only; no original LLM, multi-card training, private or real browsing environment, or full benchmark.