WebShaper: Agentically Data Synthesizing via Information-Seeking Formalization¶

🕒 Published (v1): 2026-01-01 · Source: ICLR · Venue: ICLR 2026 · link

Why this paper was selected

Agentic synthesis of information-seeking training data via formalization

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

TL;DR¶

WebShaper is a formalization-driven data synthesis framework for training web information-seeking (IS) agents. It formalizes IS tasks using set-theoretic Knowledge Projections (KPs) and an agentic Expander that iteratively grows questions in complexity while grounding retrieval in the formal structure. Models trained on the synthesized dataset achieve state-of-the-art open-source performance on GAIA and WebWalkerQA.

Problem¶

Existing IS training data synthesis methods are information-driven: they freely scrape web content first, then prompt an LLM to generate question-answer pairs from it. This causes two failure modes: (1) the LLM may misread the information structure, producing reasoning structures or answers inconsistent with the collected facts; (2) uncontrolled retrieval yields redundant, homogeneous information structures that limit task diversity. High-quality IS training data is scarce and hard to crowdsource, making scalable synthesis critical.

Method¶

WebShaper formalizes IS tasks as compositions of Knowledge Projections (KPs). A KP is defined as \(R(V) = \{u \mid \exists v \in V,\, (u,v) \in R \text{ or } (v,u) \in R\}\) for relation \(R\) and entity set \(V\). Tasks are expressed as recursive intersections of KP R-unions:

\[T = \bigcap_{i=1}^{p} \left(\bigcup_{j} R_i(S_{i,j})\right)\]

KPs are encoded as triplets [X, r, S] (variable, relation, constant/variable), with R-union collapsed via the distributive law \(R(S_1)\cup R(S_2)=R(S_1\cup S_2)\), enabling a flat list-of-triplets representation.

Synthesis pipeline: 1. Seed construction — random walks over an offline Wikipedia graph; an LLM generates QA pairs grounded in visited articles; 5-rollout filtering retains 18k valid seeds. 2. Agentic Expansion — a ReAct-based Expander agent iteratively expands a seed question \(q^l(T)\) into \(q^{l+1}(T)\) using three tools: Search (Google queries with temporal filtering), Summarize (URL scraping to construct union constant sets), and Validate (checks formal consistency and non-triviality). Expansion follows a layer-wise strategy: the Expander identifies all leaf constants in the KP graph and replaces each with a new sub-question, avoiding redundancy (constants linked to other constants) and reasoning shortcuts (constants adjacent to the target). 3. Trajectory construction — 5 ReAct rollouts per question (Search/Visit/Answer actions); incorrect, hallucinated, or repetitive trajectories pruned; 5,000 clean trajectories retained. 4. Training — SFT (observation tokens masked from loss) followed by GRPO-based RL.

Key Contributions¶

Set-theoretic formalization of information-seeking tasks via Knowledge Projections with R-Union and Intersection operations, providing the first principled IS task specification analogous to Lean 4 for math.
Layer-wise expansion strategy that prevents redundancy and reasoning shortcuts by operating on leaf constants of the KP graph rather than arbitrary nodes.
Agentic Expander with Search/Summarize/Validate tooling that autonomously retrieves, validates, and integrates new KPs during synthesis.
WebShaper dataset (18k seed → 5k filtered trajectories) and trained models achieving SOTA among open-source IS agents.

Results¶

GAIA benchmark (Pass@1, open-source): - WebShaper (Qwen-2.5-72B): 60.1 avg vs. WebSailor 55.4 (+4.7), WebDancer 51.5, WebThinker-RL 48.5 - WebShaper (QwQ-32B): 53.3 avg vs. WebDancer 50.5, WebThinker-RL 48.5 - WebShaper (Qwen-2.5-32B): 52.4 avg vs. WebDancer 40.7 - Only open-source method exceeding 60 points; approaches OpenAI Deep Research (74.3)

WebWalkerQA benchmark (Pass@1): - WebShaper (Qwen-2.5-72B): 52.2 avg vs. WebDancer 49.7, WebThinker-RL 46.5 - WebShaper (QwQ-32B): 52.2 avg vs. WebSailor 52.2 (tied best)

SFT data comparison (GAIA, 5k samples each): - WebShaper vs. best baseline per backbone: Qwen-2.5-32B 43.6 vs. E2HQA 39.8; Qwen-2.5-72B 45.6 vs. E2HQA 44.6; QwQ-32B 53.3 vs. WebWalkerQA/E2HQA 45.6

RL gains over SFT: - Qwen-2.5-32B GAIA: +8.8 pts; Qwen-2.5-72B GAIA: +14.5 pts - Qwen-2.5-32B WebWalkerQA: +7.7 pts; Qwen-2.5-72B WebWalkerQA: +6.8 pts

Formalization ablation: formal-language synthesis (FL) outperforms natural-language synthesis (NL) across all backbones; layer-wise expansion (L) outperforms sequential (S) and random structures.

Limitations¶

Final dataset is only 5,000 trajectories after filtering from 18k seeds; scale is modest compared to RL-heavy open-domain pipelines.
RL gains on QwQ-32B were not significant; only SFT results are reported for that backbone—the formalization's benefit for RL may be backbone-dependent.
Seed construction relies on Wikipedia random walks, potentially under-representing recent or non-encyclopedic knowledge domains.
Expander uses Google Search, introducing dependency on external search availability and potential temporal biases via filter_year.
Formal KP representation assumes entity-relation structure; tasks with ambiguous or latent relations may be poorly captured.

Relevance to Agentic AI / LLM Agents¶

WebShaper directly advances the data foundation problem for agentic IS systems—a bottleneck that limits open-source deep research agents from matching proprietary systems like OpenAI Deep Research. The formalization-driven synthesis paradigm offers a principled alternative to the prevalent information-first approaches (WebDancer, TaskCraft, WebWalkerQA), demonstrating that encoding task structure before retrieval yields more consistent and diverse training trajectories. The agentic Expander component is itself an instance of tool-using autonomous agents applied to meta-level data generation, a pattern increasingly relevant for scalable agent training. The layer-wise expansion strategy generalizes beyond IS to any agent domain where avoiding reasoning shortcuts in synthetic curricula is critical.