TimeSearch-R: Adaptive Temporal Search for Long-Form Video Understanding via Self-Verification Reinforcement Learning

1 INTRODUCTION

• Long-form video understanding requires finding relevant information among tens of thousands of frames, while fixed frame sets conflict with reasoning that develops dynamically.The introduction describes temporal search as central to accurate and interpretable long-video understanding.Source: §1 · p. 1
• Prior interactive video-search methods use manually designed workflows, whereas TimeSearch-R learns search strategies from data through end-to-end reinforcement learning.Examples include VideoAgent and T*, which use handcrafted workflows.Source: §1 · p. 2
• TimeSearch-R reformulates temporal search as interleaved text-video thinking, alternating reasoning and temporal exploration to refine video understanding.The model actively searches for relevant clips through reinforcement learning.Source: §1 · pp. 2–3
• Outcome-only GRPO can miss critical frames and produce reasoning disconnected from final answers because it does not supervise intermediate search decisions.These failures are described as insufficient temporal exploration and inconsistent logical reasoning.Source: §1 · pp. 2–3
• GRPO-CSV uses self-verification to supervise temporal-search steps, while a two-stage filtering pipeline curates video-reasoning training samples.The dataset filtering removes trivial samples solvable through linguistic bias and noisy samples that remain unsolvable after extensive search.Source: §1 · p. 3
• TimeSearch-R improves temporal F1 on Haystack-LVBench by 5.6% and accuracy on Haystack-Ego4D by 8.5% over prior SOTA, and improves LongVideoBench by 4.1% over Qwen2.5-VL and 2.0% over Video-R1.The paper reports results across temporal-search and long-form video understanding tasks.Source: §1 · p. 3

2.1 TASK FORMULATION

• Temporal search is formulated as multi-turn reasoning interleaved with clip retrieval, repeating until the model answers or reaches its reasoning budget.Each reasoning step can request a clip, which is appended to the ongoing chain of thought for later steps.Source: §2.1 · p. 4
• The model’s joint answer-and-reasoning probability factors into temporal-search and answer-prediction components conditioned on the preview, question, and reasoning chain.The reasoning chain comprises the sequence of text-and-video pairs from successive steps.Source: §2.1 · p. 4

2.2 GRPO WITH COMPLETENESS SELF-VERIFICATION

• GRPO-CSV reuses the policy model to answer from frames gathered during the interleaved reasoning, checking whether those frames suffice without further search.Its completeness reward applies only when the original answer is correct and scores the CSV answer against the ground truth.Source: §2.2 · p. 4, Figure 3 · p. 5
• The training objective combines completeness, format, and accuracy rewards to encourage sufficient exploration, structured reasoning, and correct answers.The format reward checks the structure of the full trajectory, while accuracy is assessed by exact match for multiple-choice answers or an LLM judge for open-ended answers.Source: §2.2 · p. 5

2.3 MODEL TRAINING

• The dataset filtering removes samples solvable from four uniformly sampled frames and samples that remain unsolvable after multiple searches, targeting temporal dependence and active exploration.The dataset also incorporates samples from Haystack-Ego4D, VideoMarathon, and CinePile.Source: §2.3 · p. 5
• TimeSearch-R is trained first with SFT as a cold start, then with GRPO-CSV reinforcement learning to develop temporal reasoning capability.SFT trains on interleaved reasoning and final answers, with temporal search results masked so the model learns search windows and textual queries.Source: §2.3 · p. 6

3.1 EXPERIMENTAL SETUP

• Evaluation covers temporal search and long-form understanding, comparing against static-sampling models, temporal-search agents, and video-reasoning models.Completeness measures whether searched frames suffice for the correct answer; consistency measures alignment between intermediate reasoning and the final answer. Training uses Qwen2.5-VL-7B-Instruct, with up to eight frames per search and eight search steps.Source: §3.1 · p. 6

3.2 MAIN RESULTS

• 8.1 temporal-similarity F1 under an 8-keyframe budget is more than three times T*’s 2.5; VideoMME accuracy reaches 66.6% and LongVideoBench reaches 60.1%.The VideoMME and LongVideoBench results are reported in Table 2.Source: §3.2 · p. 7, Table 2 · p. 7
• Needle-in-a-haystack QA accuracy is 52.1% on Haystack-LVBench and 53.5% on Haystack-Ego4D, exceeding GPT-4o on both benchmarks.Source: §3.2 · p. 7

3.3 ABLATION STUDIES

• SFT raises temporal F1 from 0.0 to 7.8 and searched-frame completeness from 44.2% to 60.5%; RL raises QA accuracy from 59.2% to 66.6%.The reported reasoning-consistency improvement after RL is 2.6%.Source: §3.3 · p. 7
• Removing CSV reduces completeness from 60.5% to 57.2% and temporal F1 from 7.8 to 7.4, with training collapse around step 300 and completeness falling to zero.Completeness reward alone yields the highest completeness and consistency but reduces QA accuracy by 0.3%; adding accuracy reward gives the best QA performance.Source: §3.3 · pp. 7–8, Figure 4 · p. 8
• Without filtering, RL performance drops below the original model; filtered egocentric data recovers baseline, while adding exocentric data yields the best general QA accuracy of 66.6%.Despite training only on general long-video QA, RL improves temporal reasoning by 7.4% and action reasoning by 5.7%. Table 3 reports the data-composition ablation.Source: §3.3 · p. 8, Table 3 · p. 8

3.4 CASE STUDIES

• Search patterns adapt by task: targeted hypothesis testing, multi-faceted confirmation or elimination when evidence is weak, and segment-by-segment search for temporal relationships.In Figure 5, the model hypothesizes that still dogs are waiting to be photographed, then searches for the photographer as supporting evidence.Source: §3.4 · p. 8, Figure 5 · p. 9

• Temporal-search approaches include static sampling, similarity-based compression, frame ranking, semantic constraints, adaptive zooming, and prompt-driven multi-turn exploration.Source: §4 · p. 9
• The section identifies a lack of end-to-end optimization among prior temporal-search methods and says RL for interactive long-video understanding remains largely unexplored.Source: §4 · p. 9
• Prior RL work advances text reasoning, text search, multimodal reasoning, and adaptive image understanding, but often lacks dynamic video interaction or temporal exploration.Source: §4 · p. 9

5 CONCLUSION

• TimeSearch-R reformulates temporal search as interleaved text-video thinking and uses CSV to address insufficient exploration and inconsistent reasoning.The conclusion reports strong performance on temporal search and long-form video understanding tasks.Source: §5 · p. 9

A.1 FRAME SELECTION

• DPP selection uses query relevance and frame similarity to select diverse, relevant frames, falling back to uniform temporal sampling when fewer than F frames are available.The method uses fast greedy MAP inference to choose the subset.Source: §A.1 · p. 13

A.2 FRAME REPRESENTATION

• Each selected frame receives an explicit absolute timestamp immediately before its image to maintain temporal grounding when frame intervals vary.Uniformly sampled preview frames use the base model’s dynamic-FPS and absolute-time encoding.Source: §A.2 · pp. 13–14

B.1 DATASET CONSTRUCTION

• The first filtering stage removes questions answerable from four uniformly sampled frames, retaining those that need richer visual context.Qwen2.5-VL answers the questions using the limited frame set.Source: §B.1 · p. 14
• The second stage tests search usefulness with up to 64 frames, using GPT-4o to generate SFT data and early TimeSearch to generate RL question-answer pairs.Only GPT-4o-generated chain-of-thought is used for SFT; RL uses question-answer pairs, and response formats are validated.Source: §B.1 · p. 14
• VideoMarathon samples receive manual review for question reasonableness and answer correctness against synthetic labels.Unanswerable or ambiguous questions and samples with answers inconsistent with human responses are removed.Source: §B.1 · p. 14

B.2 DATASET ANALYSIS

• Video duration has a long tail: the mean is 1,659 seconds, most videos are under 2,000 seconds, and some exceed one hour.The passage links this distribution to challenges for static frame sampling and motivates adaptive, multi-turn retrieval under keyframe budgets.Source: §B.2 · p. 14
• The dataset draws 49.5% of samples from Haystack-Ego4D, 35.6% from VideoMarathon, 9.5% from CinePile, and 5.4% from other sources.The sources cover egocentric activities, internet videos, and short narrative videos; the figure describes the training set as mainly long videos.Source: §B.2 · p. 15, Figure 7 · p. 15
• Open-ended questions comprise 60.3% of the data and multiple-choice questions 39.7%, balancing reasoning-oriented evaluation with automatic evaluation signals.The passage says the mixture of long-tail durations and open-ended questions creates a setting where adaptive search and end-to-end RL offer clear benefits over single-shot heuristics.Source: §B.2 · p. 15

C PROMPT DESIGN

• Prompts standardize interaction, use the base model’s tool-call schema, and require reasoning before a constrained answer or tool call.The QA template allows at most eight search rounds and supplies video duration to support absolute timestamps.Source: C PROMPT DESIGN · p. 16, Figure 8 · p. 16, Figure 9 · p. 16
• After each search, the template returns frames with timestamps and prompts another search when evidence is insufficient; CSV answers briefly without tools using only gathered frames.The templates support reflection and re-query while preventing CSV from acquiring new visual evidence.Source: C PROMPT DESIGN · pp. 16–17, Figure 10 · p. 17, Figure 11 · p. 17

D EVALUATION METRICS

• Completeness Rate measures dataset-average task accuracy when answering from the gathered dynamic frames alone, without further search.This re-answering procedure follows the CSV process.Source: D EVALUATION METRICS · p. 17
• Consistency Rate is the dataset-average binary GPT-4o judgment of whether the intermediate reasoning coherently supports the final answer.The evaluator receives the question, reasoning, and final answer and returns a Yes-or-No decision; parsing failures count as zero.Source: D EVALUATION METRICS · p. 17, Figure 12 · p. 17

E EFFICIENCY ANALYSIS

• 13.4 seconds on Haystack-Ego4D is TimeSearch-R’s end-to-end latency, a 61.6% speed-up over VideoAgent’s 34.9 seconds.The evaluation reports overall temporal-search and answering latency on A100 GPUs.Source: E EFFICIENCY ANALYSIS · p. 18, Table 4 · p. 18
• 11.1 seconds is T*’s runtime; TimeSearch-R maintains comparable runtime while avoiding hand-crafted scheduling complexity.T* uses the lightweight YOLO-World-110M detector.Source: E EFFICIENCY ANALYSIS · p. 18
• TimeSearch-R markedly surpasses the baselines in temporal-search metrics and QA accuracy.The cited passage points to Table 1 for those metrics but supplies no values here.Source: E EFFICIENCY ANALYSIS · p. 18

F TRAINING DETAILS

• Table 5 presents the training hyperparameters of TimeSearch-R.The text says the table summarizes key hyperparameters for reproducibility.Source: Table 5 · p. 18, F TRAINING DETAILS · p. 18
• Training uses PyTorch distributed data parallelism with ZeRO-3, gradient accumulation, bfloat16, and Flash Attention 2.0.These choices are described as supporting memory efficiency, faster computation, numerical stability, and efficient attention.Source: F TRAINING DETAILS · p. 18
• GRPO training uses eight generations per prompt and a KL penalty coefficient β of 0.005, with VLLM colocate mode for rollout inference.The RL stage is implemented on TRL and follows outcome-driven policy optimization practice.Source: F TRAINING DETAILS · pp. 18–19
• Video processing is capped at 768 frames, 10,240 video tokens, eight search operations per turn, and eight turns per question.Frame-token allocation ranges from 12 to 256 tokens per frame based on content complexity and relevance.Source: F TRAINING DETAILS · p. 19

G MORE CASE STUDIES

• Representative successful cases use confirmation, elimination, and sequential exploration while maintaining reasoning-trace and answer alignment.The paper relates these cases to the completeness and consistency reported in Section D.Source: G MORE CASE STUDIES · p. 19
• One residual failure halts after reviewing two of four candidate segments; another hallucinates information about riding.The first case leads to an incorrect answer.Source: G MORE CASE STUDIES · p. 19

H BROADER IMPACTS

• Completeness and consistency criteria enable quantitative assessment of long-form video explanations and make temporal-search decisions auditable and interpretable.The paper presents interleaved text-video reasoning traces as transparent insights into model decision-making.Source: H BROADER IMPACTS · p. 19
• The approach shifts from static frame sampling toward dynamic, interactive reasoning through iterative temporal search.The paper describes this as hypothesis-driven exploration and contrasts it with one-shot inference over fixed visual contexts.Source: H BROADER IMPACTS · p. 19
• Completeness self-verification combines weak supervision and reinforcement learning into outcome-based process supervision without costly process annotations.The paper presents this as a scalable approach for training complex interactive systems.Source: H BROADER IMPACTS · p. 19

I THE USE OF LARGE LANGUAGE MODELS

• The authors report using an LLM for grammar checking, sentence simplification, and fluency improvements, but not for literature review or research ideation.They state the research ideas and experimental analyses were conducted by the authors.Source: I THE USE OF LARGE LANGUAGE MODELS · p. 19
• Figures 13–15 label confirmation, elimination, and sequential search patterns; Figures 16–17 depict insufficient search and visual hallucination.The captions describe one failure as reviewing only two of four options before termination and the other as finding no riding-related information in search results.Source: Figure 13 · p. 20, Figure 14 · p. 20, Figure 15 · p. 21