DeepSearch-Evolve: 9B web agent self-distills to 31.2% BrowseComp

The interesting part of a new paper out of a team led by Xinyu Geng and collaborators is not the leaderboard number, it is the setup that produced it. Their 9B-parameter web agent reportedly hits 31.2% on BrowseComp, 61.5% on GAIA, and 93.4% on HotpotQA, and it gets there by learning from its own rollouts rather than from a larger teacher model. The framework is called DeepSearch-Evolve and it lives inside a purpose-built training environment the authors call DeepSearch-World, described in the paper on arXiv.

The mechanic is worth spelling out. DeepSearch-World provides 420,000 multi-hop question-answering tasks derived from entity-level random walks, with reproducible search and page-reading so a trajectory can be replayed and checked. Training then cycles through trajectory generation, filtering, data mixing, and fine-tuning. Because the environment is deterministic and answers are verifiable, the model can grade its own attempts and keep only the ones that actually solved the task, which is what the paper means by self-distillation without a superior teacher.

Why this matters if you are not training agents yourself: the going assumption for capable web agents has been that you either pay a hosted frontier model at inference time, or you distill from one at training time. A recipe that skips the teacher entirely, and does it at 9B, is the kind of cost-curve shift that lets smaller labs and product teams train a decent research agent on their own hardware and their own data. It also gives open-model ecosystems a plausible path to close the gap with hosted agent products on the tasks that most resemble structured web research.

The honest caveat is that the strongest claim here rests on a training environment the authors built themselves, and the benchmark wins do not tell you how the agent behaves on messy real-world browsing where answers are not cleanly verifiable. What the paper does not give you is a head-to-head cost comparison against teacher-distilled agents of similar size, or evidence that additional evolution rounds keep paying off rather than plateauing. Treat the specific numbers as reported, not settled.

The forward-looking part is the release. The authors say they will publish the environment, the training data, the validation materials, the trained model, and the code. If that lands as promised, the interesting followups will come from other groups plugging the DeepSearch-Evolve loop onto their own base models and their own vertical corpora.