SearchVL-RL-8k: RL Dataset for Multimodal Search

• The paper presents SearchVL-RL-8k as a distinct RL dataset, separating online exploration from supervised imitation using 8,000 disjoint VQA instances.
• SearchVL-RL-8k is generated via a three-stage pipeline: Wikipedia path sampling, fuzzy entity rewriting, and source-anchor visual grounding.
• The dataset underpins multimodal search performance by enabling active perception, visual repair, and policy optimization using a fatal-aware GRPO algorithm.

Searching arXiv for the cited papers and closely related work to ground the article. SearchVL-RL-8k is the reinforcement-learning dataset introduced in "OpenSearch-VL: An Open Recipe for Frontier Multimodal Search Agents" (Chen et al., 6 May 2026). It is a curated set of 8,000 VQA instances sampled from a filtered and enhanced multimodal VQA pool and used for the RL stage of an agentic multimodal deep-search system. Within the OpenSearch-VL recipe, SearchVL-RL-8k is paired with SearchVL-SFT-36k, but the two serve distinct roles: the former supports online RL exploration, whereas the latter contains supervised multi-turn demonstrations. The paper defines the two subsets as disjoint, with the RL examples explicitly kept separate from the VQA instances used to synthesize the SFT data (Chen et al., 6 May 2026).

1. Definition within the OpenSearch-VL training recipe

SearchVL-RL-8k is not presented as an isolated benchmark. It is the RL subset in a broader open training recipe for multimodal deep search agents. The paper states that OpenSearch-VL “curate[s] two training datasets, SearchVL-SFT-36k for SFT and SearchVL-RL-8k for RL,” and further specifies that SearchVL-RL-8k consists of 8,000 VQA instances drawn from the filtered and enhanced VQA pool. The RL subset is “disjoint from the instances used to synthesize the SFT data,” which establishes a clean separation between imitation-style supervision and online policy optimization (Chen et al., 6 May 2026).

This distinction is structurally important. SearchVL-SFT-36k contains final expert rollouts used for supervised imitation, whereas SearchVL-RL-8k is reserved for the RL stage, where the agent samples trajectories in the tool environment and is optimized by group-relative policy updates. A plausible implication is that the authors intended the RL stage to refine exploration, recovery, and tool sequencing behaviors that are difficult to encode fully in static demonstrations.

2. Upstream VQA construction pipeline

The dataset is produced by a three-stage pipeline: Wikipedia path sampling, fuzzy entity rewriting, and source-anchor visual grounding (Chen et al., 6 May 2026). OpenSearch-VL casts English Wikipedia as a directed hyperlink graph and samples constrained random walks of length 2, 3, or 4 from eligible seed pages. Seed entities are stratified across five domains—Person, Building/Place, Location (non-hub), Organism, Artifact—and must satisfy several constraints: an infobox, at least one Wikimedia Commons image of at least 512×512, and in-degree in [50, 10,000]. The walk skips disambiguation and list pages, excludes non-article namespaces, dereferences redirects, skips cycles, and excludes hub nodes above the threshold.

A sampled path assigns semantic roles to nodes: anchor $v_0$, bridge nodes $v_1,\dots,v_{h-1}$, and answer node $v_h$. From the answer node, the system extracts a short answer and asks GPT-4o to synthesize a canonical question. That question is then rewritten from the farthest bridge backward by replacing entity names with fuzzy relational descriptors. The rewrite is accepted only if three invariants hold:

$$a(q_f)=a(q_t),\qquad |\mathcal{R}(q_f)|=1,\qquad \left(\bigcup_{j=0}^{h}\mathrm{aliases}(v_j)\right)\cap q_f=\emptyset.$$

The final stage grounds the anchor entity visually rather than the answer entity. The pipeline retrieves a representative image from Wikimedia Commons or the infobox, ranks candidates by CLIP similarity, and replaces the anchor mention with a visual referring expression. The paper identifies this “source-anchor” design as central because it ties the image to the start of the reasoning path, thereby preventing single-hop reverse-image shortcuts and forcing multi-hop visual-to-text reasoning (Chen et al., 6 May 2026). This suggests that SearchVL-RL-8k is designed not merely for retrieval, but for retrieval under anti-shortcut constraints.

3. Filtering, enhancement, and the relation to SearchVL-SFT-36k

After the Wikipedia-derived instances are produced, the paper merges them with LiveVQA, FVQA, and WebQA. A frozen Qwen3-VL-32B then applies a two-stage difficulty filter. The first stage removes examples answerable without tools; the second removes examples solvable with a single ImageSearch call. According to the paper, this filtering removes parametric-only answers, pure perceptual shortcuts, answer-coincident anchors, and one-hop bridge leakage (Chen et al., 6 May 2026).

The authors also add robustness-oriented perturbations. They select 10% of the filtered pool and apply controlled degradations—blur, downsampling, and perspective distortion—paired with the corresponding enhancement tools. This produces a post-filtering pool from which both supervised and RL subsets are derived. For the full VQA pool, the paper reports 36,592 high-quality expert trajectories with an average of 6.3 tool-invocation turns, which become the SFT corpus. SearchVL-RL-8k is then sampled from the same filtered and enhanced pool, but with the explicit requirement that the RL examples remain disjoint from the VQA instances used for SFT synthesis (Chen et al., 6 May 2026).

The relationship between the two datasets is therefore hierarchical and disjoint. They share the same upstream construction logic, but differ in downstream use. SearchVL-SFT-36k captures expert behavior in finalized trajectories; SearchVL-RL-8k exposes the agent to online exploration in the same problem distribution without reusing the same instances. In practical terms, this separation is intended to reduce train–optimization leakage between supervised warm-starting and RL policy improvement.

4. Tool environment and multimodal interaction structure

SearchVL-RL-8k is used in a multimodal environment that supports both retrieval and active visual preprocessing. The paper describes the following tool suite (Chen et al., 6 May 2026):

Tool	Function
TextSearch	web search + page reading + summarization
ImageSearch	reverse image / visual entity search
OCR	structured document parsing with text/layout labels
Crop	rectangular region extraction
Sharpen	deblurring/detail enhancement
SuperResolution	upscaling via EDSR
PerspectiveCorrect	document rectification

The environment is explicitly multimodal. Search tools can return text, while visual tools return images, and the active visual context accumulates these images over time. The agent is trained to compose tools into multi-step programs such as perspective_correct → crop → OCR or Crop → ImageSearch → TextSearch. The system prompt makes this operational policy explicit: “small text → crop; blurry → sharpen; tilted → perspective_correct,” and it mandates external validation via text_search when facts are not directly visible (Chen et al., 6 May 2026).

This environment design matters because SearchVL-RL-8k is not simply a pool of question–answer pairs for standard response optimization. It is the RL substrate for training agents that must decide whether to retrieve, repair, crop, read, or verify. A common misconception is that “search RL” in multimodal settings is mainly about web search. In OpenSearch-VL, the RL stage also covers active perception and visual repair, so the search problem includes both external knowledge acquisition and input conditioning.

5. Fatal-aware GRPO and trajectory-level reward design

The RL algorithm used with SearchVL-RL-8k is a multi-turn fatal-aware GRPO variant (Chen et al., 6 May 2026). For each prompt $(I_0,q)$, training samples a group of rollouts $\tau_i$ and assigns a composite trajectory reward

$$r(\tau)=r_{\text{fmt}(\tau)}\cdot\bigl[\alpha\,r_{\text{acc}(\tau)}+(1-\alpha)\,r_{\text{query}(\tau)}\bigr],\qquad \alpha=0.8.$$

Here $r_{\text{fmt}}\in[0,1]$ scores trajectory formatting, $r_{\text{acc}}\in\{0,1\}$ is a terminal correctness score judged by GPT-4o, and $r_{\text{query}}\in[0,1]$ is a GPT-5.4 process reward scoring semantic relevance, query progression, signal-to-noise, and cross-modal complementarity.

The defining feature is its treatment of cascading tool failures. A trajectory becomes fatal at the earliest step where 3 consecutive tool-execution errors begin. Tokens after that point are masked out by

$v_1,\dots,v_{h-1}$0

so only policy-emitted tokens from the valid prefix receive gradient. Group-normalized returns are computed over the full group, including fatal trajectories, and the algorithm introduces one-sided advantage clamping:

$v_1,\dots,v_{h-1}$1

This means fatal trajectories with negative relative return receive zero gradient, while fatal trajectories with useful prefixes can still be positively reinforced (Chen et al., 6 May 2026).

The appendix further states that this scheme is weakly informative-dominant over hard masking: it never trains on post-fatal suffixes, never penalizes viable prefixes, and can still reinforce useful pre-failure reasoning. In the context of SearchVL-RL-8k, the significance is direct. Tool-rich multimodal rollouts are vulnerable to partial failure, and the RL objective is designed to preserve learning signal from successful prefixes rather than discarding the entire trajectory.

6. Empirical role, release status, and relation to later deep-search recipes

SearchVL-RL-8k is the RL data source behind the performance gains reported for OpenSearch-VL. In the paper’s ablation on Qwen3-VL-8B, the training progression is SFT only = 64.6 average, Vanilla GRPO = 67.6, GRPO with hard masking = 67.7, fatal masking only = 69.1, and fatal masking + one-sided clamp (full method) = 71.8 average across the ablation benchmarks. The full method therefore outperforms vanilla GRPO by 4.2 points. In the main benchmark table, OpenSearch-VL-30B-A3B reaches 61.6 average versus 47.8 for the base agentic baseline, while OpenSearch-VL-8B reaches 56.6 (Chen et al., 6 May 2026).

These are system-level rather than isolate-only-attribution results, but the paper explicitly attributes the RL-stage improvements to the combination of the high-quality 8k RL subset and the fatal-aware GRPO recipe. The release plan is also explicit: the authors state that SearchVL-SFT-36k / SearchVL-RL-8k, model checkpoints, and training code will be released under permissive licenses. At the same time, they note several reproducibility constraints: search ranking drift, fetch failures, summarization hallucinations, reliance on proprietary GPT-4o/GPT-5.4 judges, and dependence on external APIs such as Serper and PaddleX OCR (Chen et al., 6 May 2026).

A closely related later paper, "SimpleSearch-VL: A Simple Recipe for Multimodal Agentic Deep Search" (Dai et al., 30 Jun 2026), addresses the same multimodal agentic deep-search setting but does not define a task or benchmark explicitly named “SearchVL-RL-8k.” Instead, it presents a smaller-scale RL recipe with 5,193 evidence-aware SFT trajectories and 1,995 RL prompts, a lightweight tool interface, evidence-verified reasoning, and Factorized Adaptive Rollout (FAR). This suggests that SearchVL-RL-8k should be understood less as a standalone public benchmark name than as a dataset component within a broader line of work on open multimodal deep-search agents.