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

Researchers published a preprint on arXiv on July 9, 2026 of DeepSearch-World: the DeepSearch-Evolve self-distillation framework allows web agents to learn from their own experience without relying on more powerful teacher models. A model with 9 billion parameters achieved 61.5% on GAIA and 93.4% on HotpotQA.

Why learning agents from experience is so difficult

Two dominant paradigms for training agents have fundamental weaknesses. Supervised fine-tuning (SFT) trains the agent on fixed trajectories distilled from a teacher: the model copies others' experience without developing its own. Reinforcement learning (RL) potentially allows an agent to learn from its mistakes, but in long interactions — when the agent takes dozens of steps on the internet before receiving a final reward — the reward becomes too sparse for effective learning.

The authors solved the problem differently: they created a verifiable environment where the agent receives dense feedback at each step without depending on an external judge.

How the DeepSearch-World environment works

DeepSearch-World is a deterministic and reproducible environment with two tools: search and page reading. Determinism means that the same query always returns the same result — ideal conditions for training and comparing agents.

Key characteristics:

• 420,000 multi-step tasks built through random walks on an entity graph
• Three built-in cognitive behaviors: progress verification, grounded reflection, failure recovery
• Complete reproducibility of any agent trajectory

Tasks are "multi-step": to answer a question, the agent must sequentially find several related facts, each building on the previous one.

What results did the model achieve without a teacher

The DeepSearch-Evolve framework iteratively goes through four stages: trajectory generation → filtering → data mixing → fine-tuning. At each iteration, the agent selects successful trajectories, adds them to accumulated data and retrains — gradually improving without external help.

The DeepSearch-World-9B model without distillation from GPT-4 or other frontier models showed competitive results among open agents:

• 31.2% on BrowseComp — a benchmark of real web navigation from OpenAI
• 61.5% on GAIA — a universal test of AI assistants
• 93.4% on HotpotQA — multi-step search for related facts

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

agents with long interaction horizons."

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

What this means

The study closes an important question: do you need a frontier model to grow a strong agent? It turns out — no, if the training environment is structured correctly. A verifiable environment with dense rewards allows an agent to self-improve without dependence 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 a benchmark from OpenAI for assessing agents' ability to work with the real internet; tasks are intentionally complex, and results above 30% are considered competitive among open models up to 10 billion parameters.

When will the open DeepSearch-World dataset be released?

The authors announced opening the environment, 420k tasks, validation set, model and code, but no specific date for public release is given in the preprint from July 9, 2026.