DeepSearch-Evolve: a 9B-parameter web agent trains without a teacher and scores 61% on GAIA

Researchers published the DeepSearch-World preprint on arXiv on July 9, 2026. Its self-distillation framework, DeepSearch-Evolve, lets web agents learn from their own experience without stronger teacher models. The 9-billion-parameter model scored 61.5% on GAIA and 93.4% on HotpotQA.

Why learning from experience is so hard for agents

The two dominant paradigms for training agents have fundamental weaknesses. Supervised fine-tuning (SFT) trains an agent on fixed trajectories distilled from a teacher: the model copies someone else's experience instead of building its own. Reinforcement learning (RL) could in principle let an agent learn from its mistakes, but over long interactions, where the agent takes dozens of steps across the internet before receiving a final signal, the reward is too sparse for efficient learning.

The authors took a different approach: they built a verifiable environment in which the agent gets a dense feedback signal at every step, with no dependence on an external judge.

How the DeepSearch-World environment works

DeepSearch-World is a deterministic, reproducible environment with two tools: search and page reading. Determinism means the same query always returns the same result, which makes for ideal conditions to train and compare agents.

Key features:

• 420,000 multi-step tasks built through random walks over an entity graph
• Three built-in cognitive behaviors: progress verification, grounded reflection, and recovery from failures
• Full reproducibility of any agent trajectory

The tasks are "multi-step": to answer a question, the agent must find several related facts in sequence, each of which builds on the previous one.

What the model achieved without a teacher

The DeepSearch-Evolve framework iterates through four stages: trajectory generation → filtering → data mixing → further training. In each iteration, the agent selects its successful trajectories, adds them to the accumulated data and retrains, improving gradually without outside help.

Without distillation from GPT-4 or other frontier models, DeepSearch-World-9B posted competitive results among open agents:

• 31.2% on BrowseComp, OpenAI's benchmark for real-world web navigation
• 61.5% on GAIA, a general-purpose test for AI assistants
• 93.4% on HotpotQA, multi-step retrieval of linked facts

"Verifiable environments open the way to scalable self-evolution for

agents with long interaction horizons."

The authors plan to release the environment, the pool of 420,000 tasks, a validation set, the model checkpoint and the full code.

What this means

The study addresses an important question: do you need a frontier model to grow a strong agent? It turns out you do not, provided the training environment is well designed. A verifiable environment with dense rewards lets an agent improve itself without relying on closed commercial models, an important step toward open, scalable next-generation web agents.

Frequently asked questions

What is BrowseComp, and why is 31.2% a good result?

BrowseComp is an OpenAI benchmark that measures how well agents handle the real internet. Its tasks are deliberately hard, and a score above 30% is considered competitive among open models of up to 10 billion parameters.

When will the DeepSearch-World open dataset be released?

The authors announced they will open the environment, the 420,000 tasks, the validation set, the model and the code, but the preprint dated July 9, 2026 gives no specific public release date.