SailorFog-QA-V2: Advanced QA Dataset & Pipeline

• SailorFog-QA-V2 is an advanced QA dataset and methodology that uses structured knowledge graph sampling to generate context-rich, multi-hop QA pairs.
• It systematically expands uncertainty through diverse obfuscation techniques, incorporating ambiguous event references and distractors to challenge reasoning.
• The synthetic data primes reinforcement learning pipelines, leading to superior benchmark performance in multi-evidence synthesis and tool use.

SailorFog-QA-V2 is an advanced question-answering (QA) dataset and task generation methodology situated within the pipeline of large-scale agentic training for web-based information-seeking agents. Developed as a foundation for robust, uncertainty-reduction reasoning in complex environments, SailorFog-QA-V2 integrates structured knowledge graph sampling, expanded uncertainty techniques, and serves directly as the synthetic data substrate for reinforcement learning pipelines exemplified by WebSailor-V2. The methodology aims to bridge the persistent performance gap between open-source QA agents and proprietary systems by priming agents for multi-hop, multi-evidence synthesis and tool use under intrinsically high ambiguity (Li et al., 16 Sep 2025).

1. Structured Sampling from Dense Knowledge Graphs

SailorFog-QA-V2 departs from simplistic tree-structured or fact-centric QA sample generation. Starting from a seed entity, web tools are used to extract a combinatorially rich knowledge graph where entities and their relationships (edges) capture cyclic and non-tree structures. Subgraph extraction leverages random-walk algorithms for coverage, followed by Weisfeiler–Leman isomorphism filtering to ensure diversity of sampled subgraphs. This method results in QA pairs embedded in intricate context graphs, challenging agents to reason over relationship sets beyond acyclic or linear inference chains.

Component	Technique	Purpose
Knowledge Graph	Random Walk	Rich Coverage, Structure Diversity
Subgraph Verification	Weisfeiler–Leman	Non-isomorphic Filtering

2. Information Obfuscation and Uncertainty Expansion

A central design innovation is the systematic expansion of uncertainty types within sampled QA instances. Rather than restricting uncertainty to simple masking (e.g., anonymized names, redacted dates), SailorFog-QA-V2 introduces diverse forms such as ambiguous event references, compositional distractors, or omitted key relations. By obfuscating multiple facets of the question context and answer space, the methodology forces models to engage in contextual inference, latent variable disambiguation, and cross-source triangulation. This expands the reasoning requirements well beyond shallow retrieval or exact-match patterning.

3. Synthetic Data Generation for SFT Cold Start

Synthetic data generation in SailorFog-QA-V2 is tightly linked to its role in the agent training pipeline. Initial training trajectories are created using high-quality open-source models subject to rejection sampling: only QA sessions that reach correct answers, demonstrate requisite complexity (e.g., multiple tool calls), and respect token budget constraints are retained. This SFT cold start phase primes agents with essential tool-use fluency and complex reasoning traces. The quality and diversity of SailorFog-QA-V2 data are critical in bootstrapping agent policies, enabling meaningful exploration during subsequent reinforcement learning (Li et al., 16 Sep 2025).

4. Token-Level Reinforcement Learning with DUPO

The agentic RL phase employs a dual-environment strategy: a fast simulated Wikipedia corpus and a managed real-world environment with controlled tool deployment and fault tolerance. Duplicating Sampling Policy Optimization (DUPO) underpins token-level policy gradient optimization, maintaining sample diversity and mitigating overfitting to redundant trajectories. The RL objective is formalized as:

$$\mathcal{J}(\theta) = \mathbb{E}\left[\frac{1}{\sum_{i=1}^{G}|o_i|}\sum_{i=1}^{G}\sum_{t=1}^{|o_i|} \min\left( r_{i,t}(\theta)\hat{A}_{i,t},\, \operatorname{clip}\left(r_{i,t}(\theta),\, 1-\varepsilon_{low},\, 1+\varepsilon_{high}\right)\hat{A}_{i,t}\right)\right]$$

with the importance sampling ratio

$$r_{i,t}(\theta) = \frac{\pi_{\theta}(o_{i,t} \mid \text{context})}{\pi_{\theta_\text{old}}(o_{i,t} \mid \text{context})}$$

and advantage estimate

$$\hat{A}_{i,t} = {R_i - \text{mean}(\{R_i\}_{i=1}^{G})}$$

Reward blending avoids hacking, as in

$$R_i = 0.1 \cdot R_i^{\text{format}} + 0.9 \cdot R_i^{\text{answer}}$$

5. Reasoning Trajectory Formulation

QA sessions within SailorFog-QA-V2 are formulated as complete trajectories over T steps:

$$\mathcal{H}_T = ( \tau_0, a_0, o_0, ..., \tau_i, a_i, o_i, ..., \tau_T, a_T)$$

Each entry records the agent's "thought" $\tau_i$, action $a_i$, and observation $o_i$ at step i. Actions sample both thought and invocation from the policy conditioned on historical trajectory, $\pi(a_t, \tau_t \mid \mathcal{H}_{t-1})$, ensuring fidelity to multi-step reasoning objectives.

6. Benchmarking and Performance Outcomes

When SailorFog-QA-V2 is used as the synthetic data substrate within WebSailor-V2, substantial gains are observed on complex benchmarks. The RL-trained 30B-A3B agent achieves scores of 35.3 (BrowseComp–EN), 44.1 (BrowseComp–ZH), and 30.6 (HLE), surpassing all open-source competitors and even outperforming larger proprietary models such as DeepSeek-V3.1 (671B) (Li et al., 16 Sep 2025). Enhanced transferability is observed across related tasks (xbench–DeepSearch, GAIA), demonstrating the pipeline’s generality.

Benchmark	WebSailor-V2 Score	Proprietary Baseline
BrowseComp-EN	35.3	>30 (DeepResearch)
BrowseComp-ZH	44.1	>43
HLE	30.6	<30.6 (Open-source)

7. Significance and Future Prospects

The SailorFog-QA-V2 methodology, through structured graph-based sampling, uncertainty-rich QA pairs, and integration into closed-loop RL agents, provides a robust procedural foundation for advancing agent reasoning. Its impact is seen in closing the capability gap with proprietary agents in high-uncertainty, multi-step web reasoning. The approach sets a precedent for future dataset design—structural diversity, enriched uncertainty, and alignment with token-level RL objectives are necessary conditions for superhuman performance in open-source agents.

Further research paths include expanding uncertainty representations, refining sampling algorithms, and exploring more granular reward structures and environment paradigms. SailorFog-QA-V2’s principles can be extrapolated to other domains requiring agentic synthesis and adaptive evidence integration under extreme ambiguity (Li et al., 16 Sep 2025).