DeepSearch-World: Deterministic Cognitive Agents

• DeepSearch-World is a deterministic, verifiable offline environment that supports process supervision, grounded error correction, and reproducible multi-hop search tasks.
• It leverages the Wikipedia hyperlink graph to generate 420K multi-hop QA tasks, ensuring structured reasoning and prevention of shortcut answers.
• DeepSearch-Evolve employs self-distillation, iterative fine-tuning, and scaffold-to-ReAct conversion to enhance long-horizon reasoning and reliable tool use.

DeepSearch-World is a deterministic, verifiable offline environment for training and evaluating long-horizon, tool-using cognitive agents. Introduced together with DeepSearch-Evolve, it provides reproducible search and page-reading tools, a pool of 420K multi-hop QA tasks constructed from entity-level random walks, and explicit support for progress verification, grounded reflection, and failure recovery. Within this setting, DeepSearch-Evolve iteratively performs trajectory generation, filtering, data mixing, and fine-tuning, and the resulting DeepSearch-World-9B reaches 31.2% on BrowseComp, 61.5% on GAIA, and 93.4% on HotpotQA without distillation from more capable models (Geng et al., 8 Jul 2026).

1. Conceptual basis and design objectives

DeepSearch-World is designed around four stated properties: controllability, reproducibility, process-level supervision, and grounded error correction. It is deterministic in the sense that the same input yields the same output, and it is verifiable in the sense that intermediate tool decisions and progress signals can be checked objectively rather than judged heuristically. The environment is offline rather than live-web, but its tools, action schemas, and APIs match those of live search/browse environments, so transfer is intended to remain feasible (Geng et al., 8 Jul 2026).

This design directly targets a training difficulty emphasized in the accompanying framework description: supervised fine-tuning depends on fixed teacher-distilled trajectories, whereas sparse-reward reinforcement learning provides weak supervision for long-horizon interactions. DeepSearch-World therefore shifts the emphasis from terminal-answer-only feedback to process-level feedback on search, evidence verification, navigation, reflection, and recovery. A common misunderstanding is to view it merely as a cached web corpus; more precisely, it is an environment engineered so that the search process itself becomes reproducible and inspectable (Geng et al., 8 Jul 2026).

2. Construction from the Wikipedia hyperlink graph

The environment is built over the Wikipedia hyperlink graph $\mathcal{G}=(\mathcal{V},\mathcal{E})$, where nodes are articles or entities and edges are hyperlinks. Multi-hop problems are produced by sampling entity-level random walks $\tau=(v_1,\ldots,v_H)$, so that questions inherit explicit hop structure from graph traversal rather than from post hoc annotation. For each path, an LLM generates a multi-step question whose entity mentions are obfuscated or fuzzified to prevent shortcutting and to force actual search and reasoning (Geng et al., 8 Jul 2026).

The construction pipeline includes a validation and feature-fuzzification stage intended to ensure that answers cannot be recovered verbatim and that solution requires hop-by-hop research. For each target entity appearing in the random walks, the full corresponding Wikipedia articles are downloaded and indexed locally, yielding a corpus described as totaling approximately $10$ million articles. Retrieval is implemented with BM25 for search and SQLite for article access. The resulting task pool contains 420,000 unique multi-hop QA instances covering diverse reasoning patterns (Geng et al., 8 Jul 2026).

This architecture makes the environment simultaneously structured and open-ended in a limited sense. The graph backbone guarantees traceable reasoning paths, while obfuscation and local indexing prevent trivial lookup. A plausible implication is that DeepSearch-World occupies an intermediate regime between synthetic task generators with closed symbolic structure and live-web systems with unstable retrieval surfaces.

3. Supported cognitive behaviors

DeepSearch-World is not only a document collection but an environment explicitly organized around agentic cognitive behaviors. For each question $q_i$ with ground-truth entities $\mathcal{T}_i$, the environment maintains a discovered set $\mathcal{S}_t$ of already verified entities. Each tool observation is checked deterministically against this state, so subgoal completion can be monitored without LLM-based judging. This is the basis of its progress verification mechanism (Geng et al., 8 Jul 2026).

Grounded reflection is implemented through staged feedback after failed actions. Initial failures trigger generic hints such as revising a query; repeated failures trigger stronger signals, including canonical entity names or paraphrased clues. Failure recovery follows from this combination of deterministic progress tracking and reflection feedback, allowing the agent to retry, backtrack, or revise plans instead of merely accumulating unsuccessful steps (Geng et al., 8 Jul 2026).

The environment also scaffolds working memory and planning by maintaining explicit progress-state fields: completed_list, todo_list, experience, and information. These fields encode completed subgoals, pending tasks, lessons from failures, and extracted evidence. In effect, DeepSearch-World treats memory maintenance and self-correction as first-class components of the task interface rather than optional prompt engineering. This suggests that the environment is meant to teach not only retrieval but also persistent search-state management.

4. DeepSearch-Evolve and self-distillation

DeepSearch-Evolve is the self-distillation framework built on top of DeepSearch-World. Its loop consists of trajectory generation, trajectory filtering, data mixing, scaffold-to-ReAct conversion, and supervised fine-tuning. The current agent policy $\pi_{\theta_R}$ runs in the environment using a scaffold teacher with explicit planning, working memory, and reflection, and produces tool-use trajectories containing steps, rationales, and outcomes. Only trajectories with correct final answers are retained, and an additional LLM-based quality check by Qwen3.5-9B removes redundant or illogical traces (Geng et al., 8 Jul 2026).

Because rollout generation and fine-tuning are asynchronous, DeepSearch-Evolve uses importance sampling with decay factor $\gamma=0.5$ to upweight newer verified data while preserving some earlier diverse experiences. The scaffolded traces are converted into standard ReAct format with >, <tool_call>, <tool_response>, and <answer> blocks. The internal composition of the reasoning block is given as

$<think>_t = P_t \oplus R_t \oplus A_t,$

where $P_t$ is the progress-state summary, $\tau=(v_1,\ldots,v_H)$0 is reflection rewritten as self-correction, and $\tau=(v_1,\ldots,v_H)$1 is the local action rationale (Geng et al., 8 Jul 2026).

The policy update is standard SFT over verified trajectories: $\tau=(v_1,\ldots,v_H)$2 For reference, the paper contrasts this with on-policy self-distillation: $\tau=(v_1,\ldots,v_H)$3 The stated preference is for SFT on verified, deterministic traces, because non-deterministic tool-use settings can make OPSD unstable. The loop is iterated for 11 rounds and can use up to the full 420K QA pool, enabling the model to improve by imitating its own strongest verified prior behavior rather than a stronger external teacher (Geng et al., 8 Jul 2026).

5. Empirical performance and training effects

DeepSearch-World-9B is evaluated on BrowseComp, BrowseComp-ZH, GAIA, HLE, xbench, and HotpotQA. The headline claim is that, without any stronger-model distillation, it achieves competitive performance relative to open-source agents and large gains over its Qwen3.5-9B-Instruct backbone. The reported gains over the backbone are $\tau=(v_1,\ldots,v_H)$4 on BrowseComp, $\tau=(v_1,\ldots,v_H)$5 on GAIA, and $\tau=(v_1,\ldots,v_H)$6 on HotpotQA (Geng et al., 8 Jul 2026).

Model	BrowseComp	GAIA	HotpotQA
Qwen3.5-9B-Instruct	7.4	23.9	45.3
DeepSearch-World-9B	31.2	61.5	93.4
WebSailor (7B)	6.7	37.9	92.8
Marco-DR (8B)	31.4	69.9*	—
MiroThinker-v1.0 (8B)	31.1	66.4	—
OpenAI Deep Research	51.5	67.4	—

The analysis also reports substantial changes in search behavior. DeepSearch-World-9B sustains $\tau=(v_1,\ldots,v_H)$7 rounds of tool use per task versus $\tau=(v_1,\ldots,v_H)$8 for Qwen3.5-9B-Instruct, makes more visit calls, and receives much higher judged scores for planning, memory, reasoning, and self-correction, reported as $\tau=(v_1,\ldots,v_H)$9 versus $10$0. Using the full 420K QA pool, rather than a smaller subset, improves generalization, reduces format errors, and stabilizes self-improvement across rounds. Reflection internalization during scaffold-to-ReAct conversion, together with rejection sampling and filtering, is reported as critical for transferring planning and recovery skills (Geng et al., 8 Jul 2026).

These results position DeepSearch-World as evidence that verifiable environments can support scalable self-evolution for long-horizon agents. The benchmark profile also suggests that gains are not limited to one class of task: BrowseComp emphasizes difficult web search, GAIA emphasizes real-world multi-step queries, and HotpotQA emphasizes multi-hop reasoning transfer.

6. Relation to the broader DeepSearch landscape

DeepSearch-World emerged within a rapidly expanding literature on deep-search agents, simulated environments, and search benchmarks. SearchGym likewise argues that live commercial Web APIs are prohibitively expensive and that static data snapshots can corrupt reward signals; it introduces a simulated world $10$1 with verifiable knowledge-graph edges and aligned documents, and reports that a Qwen2.5-7B-Base model trained in SearchGym surpasses the web-enhanced ASearcher baseline across nine benchmarks by an average relative margin of $10$2, with zero API cost for training (Zhang et al., 21 Jan 2026). OpenSeeker addresses a complementary bottleneck—data scarcity—by open-sourcing both model and data, training on 11.7k synthesized samples and reaching state-of-the-art performance across BrowseComp, BrowseComp-ZH, xbench-DeepSearch, and WideSearch (Du et al., 16 Mar 2026).

At the evaluation level, DeepSearchQA defines a 900-prompt benchmark for exhaustive answer-set construction across 17 fields and shows that even advanced agents struggle with systematic collation, de-duplication, entity resolution, and stopping criteria (Gupta et al., 28 Jan 2026). In enterprise retrieval, HERB introduces a heterogeneous benchmark with 39,190 artifacts and reports that even the best-performing agentic RAG methods achieve an average performance score of 32.96, with retrieval identified as the main bottleneck (Choubey et al., 29 Jun 2025). These benchmarks are diagnostically relevant because they expose failure modes that process-level supervision may help mitigate, particularly premature stopping, shallow retrieval, and reasoning over partial context.

DeepSearch-World differs from live-web SFT systems such as SimpleDeepSearcher, which synthesizes realistic trajectories directly from live commercial APIs and shows that SFT on 871 curated samples can outperform RL-based baselines (Sun et al., 22 May 2025). It also differs from purely open-data or planning-centric systems such as OpenSeeker, WeDas, HierSearch, and Flash-Searcher, which focus respectively on open training data, query-result alignment, hierarchical local/web orchestration, and DAG-based parallel execution (Yu et al., 7 Mar 2026, Tan et al., 11 Aug 2025, Qin et al., 29 Sep 2025). The distinctive claim of DeepSearch-World is that determinism and objective verification make self-distillation itself scalable and stable. A plausible implication is that it defines a specific research program within DeepSearch: not merely better retrieval or more open data, but self-improving long-horizon agents trained in an environment where intermediate progress is machine-checkable (Geng et al., 8 Jul 2026).