VideoResearchAgent: Grounded Task Synthesis and Sim-to-Real RL for Open-Web Video Research

1 INTRODUCTION

• Open-web video research requires discovering relevant videos, navigating temporal content, and verifying answers from visual evidence rather than relying on supplied videos or text alone.Existing deep-research agents predominantly use text and static images, while video reasoning systems typically assume relevant videos are already provided.Source: §1 · pp. 1–2
• VideoResearchAgent combines timestamped-evidence task synthesis, a field-aligned local simulator, and on-policy training while preserving deployment-facing video-search and watch interactions.The simulator accelerates video search by 34.5–64.6×.Source: §1 · pp. 2–3
• VideoResearchAgent is the only compared system combining open-web video search, temporal navigation, cross-video evidence, task synthesis, and policy training.Table 1 compares representative deep-research and video agents and defines its open-training-data indicators.Source: Table 1 · p. 3
• 40.48% accuracy on Video-BrowseComp with RDR-GRPO, comparable to Gemini-3-Flash-Preview, while cumulative API-token usage falls by 74.9% relative to the untrained model.Controlled comparisons also show stronger live-web improvement with retrieval randomization than with fixed-retrieval GRPO.Source: §1 · p. 3
• Retrieval-Domain-Randomized GRPO varies rankings, distractors, metadata, and result structure to reduce reliance on fixed simulator patterns and support transfer to live-web search.RDR-GRPO integrates retrieval randomization with Group Relative Policy Optimization while preserving the deployment-facing action space.Source: §1 · p. 3

2 RELATED WORK

• Search-agent research covers iterative retrieval, planning, evidence synthesis, RL, long-horizon interaction, and synthetic trajectories in text environments (Jin et al., 2025; Li et al., 2025c;b; Zhou et al., 2026; Du et al., 2026b;a; Chu et al., 2026).Multimodal extensions study visual webpages and interleaved textual and visual retrieval (Li et al., 2025d; Zhang et al., 2026b; Hou et al., 2026; Wei et al., 2026).Source: §2 · p. 4
• VideoResearchAgent trains a policy for the open-web search–watch–verify interaction rather than evaluating only a fixed research scaffold (Liang et al., 2025).Video-BrowseComp highlights cases where evidence is available only within an open-web video, making page discovery or metadata reading insufficient.Source: §2 · p. 4
• Video reasoning work improves temporal representation, evidence selection, and grounding, while active-tool systems select segments or observations from supplied videos or fixed collections (Zhang et al., 2024; Chen et al., 2025; 2026; Luo et al., 2026; Shen et al., 2026; Feng et al., 2025; Wang et al., 2026; Li et al., 2025a; Wang et al., 2025; Zhang et al., 2025a; Lin et al., 2026; Zhao et al., 2026).These systems do not address open-web source discovery together with timeline navigation and cross-video verification under noisy retrieval.Source: §2 · p. 4

3.1 PROBLEM DEFINITION

• Open-web video research is modeled as sequential interaction in which agents discover sources, navigate temporal content, and verify visual evidence to maximize final-answer accuracy.Source: §3.1 · p. 4

3.2 CONTROLLABLE VIDEO-GROUNDED TASK SYNTHESIS

• Tasks are built from retrievable multi-hop evidence chains, then generated and filtered to require frame evidence rather than closed-book or text-only shortcuts.Source: §3.2 · pp. 4–5, Figure 2 · p. 5
• The synthesis pipeline expands from seed queries and anchor videos into timestamped, multi-hop visual evidence paths across candidate videos.Source: §3.2 · p. 5
• Generated questions require iterative video search and multimodal reasoning, with answers grounded in a target entity, supporting video, timestamp, and visual observation.Source: §3.2 · p. 5

3.3 SUPERVISED FINE-TUNING FOR COLD-START

• Cold-start demonstrations are teacher trajectories screened for coherent, evidence-grounded interaction, followed by full-parameter Qwen3.5-4B fine-tuning with the vision encoder and multimodal aligner frozen.Source: §3.3 · p. 6

3.4 FIELD-ALIGNED LOCAL SIMULATOR CONSTRUCTION

• The local simulator replaces slow, brittle video calls with a deployment-aligned corpus of 43,751 videos and 1.39 million keyframes while preserving live API schemas.Source: §3.4 · p. 6

3.5 RETRIEVAL-DOMAIN-RANDOMIZED RL

• Because clean simulated retrieval can encourage shortcuts that transfer poorly, RDR varies result order, visible metadata, and candidate availability while retaining valid, watchable videos.Source: §3.5 · p. 6
• RDR-GRPO routes video retrieval to a separate local RL task set while keeping general web search live, and rewards semantic answer correctness and valid answer formatting.Source: §3.5 · pp. 6–7

4.1 EXPERIMENTAL SETUP

• Video-BrowseComp contains 210 open-web video-search tasks across three difficulty levels; evaluations use pass@1 accuracy with at most 50 interaction turns.Source: §4.1 · p. 7

4.3 LOCAL VIDEO RETRIEVAL SIMULATOR AND LIVE-BACKEND COMPARISON

• 34.5–64.6× faster video search and 5.3× faster cold watch-video calls are achieved by the local simulator; fully cached watch latency remains comparable.The simulator has 100% success rates across evaluated operations, compared with 73.3% and 80.0% for live calls.Source: Table 3 · p. 8, §4.3 · p. 8

4.4 RETRIEVAL DOMAIN RANDOMIZATION ABLATION

• RDR-GRPO test-set accuracy rises at +0.39 percentage points per 10 steps, versus +0.05 for fixed-retrieval GRPO.Source: Figure 4 · p. 8, §4.4 · p. 9

4.2 MAIN RESULTS

• GPT-5-mini overall accuracy rises from 15.71 to 31.43, while Qwen3-VL-8B-Think rises from 7.14 to 17.14 with VideoResearchAgent.For GPT-5-mini, L2 and L3 accuracy rise from 0.00 to 20.97 and 8.70, respectively.Source: §4.2 · p. 8
• 40.48% overall accuracy for Qwen3.5-4B after RDR-GRPO follows 30.00% before training and 36.67% after SFT, closely matching Gemini-3-Flash-Preview.The final 4B policy also exceeds untrained Qwen3.5-9B and Qwen3.5-27B by 7.62 and 5.72 points, respectively.Source: §4.2 · p. 8

4.5 BEHAVIOR ANALYSIS

• Video search and watch calls increase from 8.2% to 14.4% and 30.3% across Base, SFT, and SFT+RDR-GRPO, while API-token use falls 74.9%.Source: Figure 5 · p. 9, §4.5 · p. 9
• Post-training reduces maximum-turn-limit hits to 2.9%, from 60.0% for the base model, and reformulates 87% of invalid or low-information search queries.The base model repeats the exact same query in 35% of those cases.Source: §4.5 · p. 9

• VideoResearchAgent combines controllable task synthesis, a field-aligned local simulator, and RDR-GRPO to address data, training-efficiency, and transfer bottlenecks.Source: §5 · pp. 9–10
• On Video-BrowseComp, post-training raises Qwen3.5-4B accuracy from 30.00% to 40.48% while cutting cumulative API tokens by 74.9%.Source: §5 · p. 10
• The results show that compact VLMs can conduct accurate and efficient open-web video research when training targets temporal supervision, scalable interaction, and generalization.Source: §5 · p. 10

• RL relies mainly on terminal outcome rewards, limiting supervision for intermediate research decisions; resource constraints leave evaluation mainly on Qwen3.5-4B.Source: A LIMITATIONS AND FUTURE WORK · p. 14

B IMPLEMENTATION DETAILS

• The final RDR-GRPO policy uses a dense vision-language model with native image input and function calling, trained from multimodal trajectories with GRPO updates.Tool-returned frames are reinserted as visual tokens, and the reported policy is the checkpoint after 85 optimization steps.Source: B IMPLEMENTATION DETAILS · p. 14

B.1 TEACHER TRAJECTORY FILTERING

• The trajectory judge keeps coherent, productive, tool-relevant, evidence-grounded SFT records and does not treat synthesized ground truth as guaranteed correct.The prompt also asks the judge to assess structural correctness and avoid rewarding an answer merely for matching ground truth.Source: §B.1 · p. 14

B.2 SUPERVISED FINE-TUNING CONFIGURATION

• SFT uses full-parameter language-model updates while freezing the vision encoder and multimodal aligner, with a 196,608-token sequence limit.Examples exceeding the limit after multimodal tokenization are discarded.Source: §B.2 · p. 15, Table 4 · p. 15

C AGENT INTERFACE

• The field and local agents share four policy-facing tools and semantic outputs, with video tools returning image tokens alongside compact text.The policy terminates by emitting an answer span.Source: C AGENT INTERFACE · p. 16, Table 5 · p. 16

C.1 WEB_SEARCH

• web_search retrieves general-web entity clues and background context, especially when the target video’s title or channel is unknown.Its results contain titles, URLs, and snippets.Source: §C.1 · p. 16

C.2 SEARCH_YOUTUBE

• search_youtube is the primary video-discovery tool, returning ranked metadata for videos matching a text query.The implementation uses yt-dlp’s lightweight search, with results ordered by YouTube relevance ranking.Source: §C.2 · p. 16

C.3 WATCH_VIDEO

• watch_video extracts frames for visual analysis, using sparse whole-video sampling first and dense sampling within a selected time window for fine detail.Sparse mode returns the full transcript; dense mode supports closer examination of a region identified during sparse viewing.Source: §C.3 · pp. 16–17

C.4 VISUAL_GROUNDING

• visual_grounding detects prompted objects in video frames and returns labeled bounding boxes and annotated images for cross-validation against the original visual context.It is intended for precise localization, counting, or identification that may be difficult from sparse frames.Source: §C.4 · p. 17

D.1 SYNTHESIZED TASK DISTRIBUTION

• 19.24% of tasks concern Music & Performing Arts, 15.85% Travel & Geography, and 13.70% Sports & Fitness; together they account for 48.79%.The remaining 51.21% spans ten categories.Source: §D.1 · p. 18

D.2 EXAMPLE SYNTHESIZED TASK

• At 147.60 seconds, the Hubble frame shows four fully visible rectangular solar-panel sections, reached by tracing JWST to Hubble and inspecting its 30th-anniversary retrospective.The text-only adversary could not recover the answer after five searches, while the cited frame provides the unique count.Source: §D.2 · p. 18

D.3 DATA CONTAMINATION ANALYSIS

• The audit compares all 210 evaluation tasks with filtered SFT trajectories and the full synthesized-task pool, using exact, lexical, semantic, and video-identifier checks.It also inspects nearest neighbors and manually reviews 88 higher-risk or source-linked candidates.Source: §D.3 · pp. 18–19
• No task-level contamination was detected in SFT or RL data, based on lexical, semantic, and metadata checks plus source matching where identifiers were resolvable.The audit found no shared target fact in manually reviewed candidates; answer-only collisions were not accompanied by question or source overlap.Source: §D.3 · p. 19

E LOCAL ENVIRONMENT DETAILS

• The local environment maps deployment-compatible video identifiers to local media and combines dense frame retrieval with sparse text retrieval before returning fused candidates.RL uses local video retrieval, decoding, and visual grounding while keeping general web retrieval live.Source: E LOCAL ENVIRONMENT DETAILS · p. 19

F RETRIEVAL-DOMAIN RANDOMIZATION CONFIGURATION

• Retrieval-Domain Randomization perturbs training observations to reduce reliance on the simulator’s high recall and stationary rankings and improve transfer to live retrieval.The intended response is to retrieval conditions that are cleaner and more predictable in simulation than on the live web.Source: F RETRIEVAL-DOMAIN RANDOMIZATION… · p. 20

F.1 DIFFICULTY SAMPLING

• Each search_youtube call samples a difficulty tier independently from a fixed categorical distribution.The tier sampling probabilities are specified in Table 6.Source: §F.1 · p. 20, Table 6 · p. 20

F.2 PERTURBATION PARAMETERS BY DIFFICULTY TIER

• Table 7 specifies RDR perturbation parameters by difficulty tier, including tier-specific probabilities for metadata jitter and structural noise.The supplied description does not state the parameter values.Source: Table 7 · p. 20, §F.3 · p. 20

F.3 FOUR TYPES OF PERTURBATION OPERATIONS

• RDR changes gold-video rank or removes it, promotes watchable hard negatives, corrupts displayed metadata, and varies result counts or returns empty results.These operations perturb candidate order, membership, displayed metadata, and result structure.Source: §F.3 · p. 20

F.4 SAMPLING GRANULARITY AND SAFETY CONSTRAINTS

• A new tier is sampled for every search call, while RDR never rewrites URLs and keeps promoted negatives resolvable by watch_video.The random seed combines a base seed, rollout instance identifier, and call index, so repeated searches can differ.Source: §F.4 · p. 21

G INTERACTION CONTROL ANALYSIS

• RDR-GRPO leaves fewer trajectories without answer-bearing evidence after 20 calls (90 →36) and has no abstentions or empty answers.Source: Table 8 · p. 21, G INTERACTION CONTROL ANALYSIS · p. 21
• 60.0% of base trajectories hit the cap and 50.5% contain a degenerate loop; SFT reduces tool calls per task from 35.8 to 17.7 while distinct calls stay nearly unchanged (17.9 to 17.2).Source: Table 8 · p. 21, G INTERACTION CONTROL ANALYSIS · p. 21
• Among trajectories still lacking answer-bearing evidence after 20 calls, only 1.1–11.1% acquire it later, although the base model issues 35.6% of its calls after that point.Source: G INTERACTION CONTROL ANALYSIS · p. 21
• Post-training reduces the 4B model’s 50-turn-cap rate to 2.9%, compared with 35.7% for Qwen3.5-9B, 23.6% for Qwen3.5-27B, 26.2% for Qwen3.5-Plus, and 27.6% for Gemini-3-Flash.Source: G INTERACTION CONTROL ANALYSIS · p. 21, Table 9 · p. 22
• After empty search results, SFT and RDR-GRPO reformulate 92% and 87% of failed queries and repeat fewer than 1%, while the base model repeats the identical query 35% of the time.Source: G INTERACTION CONTROL ANALYSIS · p. 21, Table 9 · p. 22

H TRAJECTORY EXAMPLE

• Figure A2 traces a successful run through the agent’s thoughts, tool calls, and observations, highlighting pivotal steps toward the correct answer.Source: Figure A2 · p. 23