DeepSearch-World: A 9B Search Agent That Taught Itself, No Teacher
The default recipe for a good search agent right now is: take a small model, distill trajectories out of a frontier model, ship it. HKUST's DeepSearch-World, arXiv 2607.07820, does not do that, and a 9B model comes out at 31.2% on BrowseComp, 61.5% on GAIA and 93.4% on HotpotQA.
The trick is the environment. DeepSearch-World is a deterministic world for web agents where search and reading are reproducible: the same query returns the same results every time. That sounds like a small engineering detail and it is the entire contribution. The live web is nondeterministic, which means a rollout cannot be replayed, which means reward is noisy, which means RL on real web agents has been miserable. Freeze the world and reward becomes verifiable, and once reward is verifiable a model can bootstrap off its own successes.
On top of it they built DeepSearch-Evolve, a self-distillation framework, and 420K multi-hop QA tasks generated from entity-level random walks over the environment's entity graph. Random walks are a neat way to manufacture questions with a known answer path and a tunable number of hops, so difficulty is a dial rather than a dataset you have to buy. The environment also supports the behaviors that matter in long-horizon search and that supervised traces rarely contain: progress verification and failure recovery. The agent learns what to do when a lead is dead, because the environment lets it hit dead ends reproducibly.
The claim the authors make is the one to sit with: verifiable environments enable scalable self-evolution for long-horizon web agents, without distilling from a larger model. If that holds, the ceiling on a small search agent is not the best frontier model available to copy from. It is how good your gym is.
They are releasing the environment, the training data, the validation set, the trained model and the code. The environment is arguably the more valuable artifact, because a 9B model is a result and a reproducible web is infrastructure everyone else can train against. arxiv.org/abs/2607.07820
← Back to all articles
The trick is the environment. DeepSearch-World is a deterministic world for web agents where search and reading are reproducible: the same query returns the same results every time. That sounds like a small engineering detail and it is the entire contribution. The live web is nondeterministic, which means a rollout cannot be replayed, which means reward is noisy, which means RL on real web agents has been miserable. Freeze the world and reward becomes verifiable, and once reward is verifiable a model can bootstrap off its own successes.
On top of it they built DeepSearch-Evolve, a self-distillation framework, and 420K multi-hop QA tasks generated from entity-level random walks over the environment's entity graph. Random walks are a neat way to manufacture questions with a known answer path and a tunable number of hops, so difficulty is a dial rather than a dataset you have to buy. The environment also supports the behaviors that matter in long-horizon search and that supervised traces rarely contain: progress verification and failure recovery. The agent learns what to do when a lead is dead, because the environment lets it hit dead ends reproducibly.
The claim the authors make is the one to sit with: verifiable environments enable scalable self-evolution for long-horizon web agents, without distilling from a larger model. If that holds, the ceiling on a small search agent is not the best frontier model available to copy from. It is how good your gym is.
They are releasing the environment, the training data, the validation set, the trained model and the code. The environment is arguably the more valuable artifact, because a 9B model is a result and a reproducible web is infrastructure everyone else can train against. arxiv.org/abs/2607.07820
Comments