- Rating
- 3
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper investigates evidence-seeking collapse in agentic long-video understanding, where outcome-only RL may improve answer accuracy while reducing active evidence acquisition. To address this issue, this paper proposes VEPO, which uses a frozen Evidence Reward Model (ERM) to assess the sufficiency of accumulated search-retrieved visual observations. The resulting signal is used for trajectory-level rewards and action-local credit assignment, together with dynamic group filtering. Experiments across four long-video benchmarks demonstrate improvements in answer accuracy and evidence-supported correctness.
Critical strengths
-
This paper investigates a meaningful problem in agentic video understanding: improving answer accuracy does not necessarily imply effective evidence acquisition. The analysis of training dynamics provides useful motivation for optimizing evidence sufficiency rather than merely rewarding final-answer correctness or tool usage.
-
The proposed method is well motivated. The ERM evaluates accumulated visual evidence, while trajectory-level rewards and action-local credit assignment encourage the acquisition of useful evidence. This provides a reasonable approach to improving evidence-seeking behavior during RL training.
-
The experiments are relatively comprehensive and demonstrate meaningful improvements over SFT and standard GRPO across four benchmarks. The paper also includes component ablations, cross-backbone evaluation, inference efficiency analysis, and independent evidence assessments using another model and human annotators, providing substantial empirical support for the proposed approach.
Critical weaknesses
The methodological advantages over existing evidence-aware approaches are not sufficiently established. Prior methods such as TimeSearch-R[1] have already explored rewarding the completeness of retrieved video evidence. While VEPO introduces a separately trained ERM and action-local credit assignment, the main controlled comparisons focus on outcome-only GRPO and generic tool-use rewards. The component ablations demonstrate their usefulness, but it remains unclear how much additional benefit VEPO provides over existing evidence-aware reward designs under comparable training conditions.
[1] TimeSearch-R: Adaptive Temporal Search for Long-Form Video Understanding via Self-Verification Reinforcement Learning, ICLR 2026
Questions for authors
-
How would the evidence-support evaluation change if the initial overview frames were included? In particular, what proportion of correct predictions classified as unsupported could already be justified by the overview? Would the evidence-seeking collapse analysis and the improvements in supported correctness remain consistent?
-
The main experiments are conducted on the relatively outdated Qwen2.5-VL-7B backbone. Although the appendix includes results on Qwen3-VL-4B, could the authors further validate VEPO on a more recent and stronger model, such as Qwen3.5-9B, to assess whether its improvements over standard GRPO still hold?
5 comments
Author Response (W1), Part 2: Comparison with Evidence-Aware Rewards
W1 (continued): Controlled Comparison against Existing Evidence-Aware Reward Designs
- Controlled reward comparison. To strictly isolate the reward design, we hold the Qwen2.5-VL-7B SFT model, RL data, rollout budget, and native tool interface fixed, and replace VEPO's reward with the TimeSearch-R-style re-answer reward (Tab. R1b, c). The two settings differ only in whether search is supervised through re-answer correctness or through VEPO's separate semantic-evidence and answer signals.
(b) Final Accuracy under the Same Training Setup
| Reward | MLVU | VideoMME | LongVideoBench | LVBench |
|---|---|---|---|---|
| TimeSearch-R-style | 67.3 | 62.7 | 61.3 | 45.8 |
| VEPO | 68.0 | 64.4 | 63.0 | 47.5 |
(c) LVBench Evidence Acquisition under the Same Training Setup
| Reward | Acc. | Avg. Calls | ||
|---|---|---|---|---|
| TimeSearch-R-style | 45.8 | 21.9 | 31.9 | 2.24 |
| VEPO | 47.5 | 36.3 | 39.1 | 3.18 |
- Tool-use and retrieval controls. On LVBench, 65.5% of TimeSearch-R's calls request the entire video, confirming that semantic frame selection is largely performed by the external SigLIP retriever rather than the policy's temporal search. We therefore restore the original SigLIP retriever and separately cap VEPO at one or two search calls (Tab. R1d).
(d) LVBench under Controlled Search and Retrieval Settings
| Method | Retriever | Avg. Calls | Acc. | |
|---|---|---|---|---|
| TimeSearch-R (w/o SigLIP) | None | 1.56 | 38.4 | 17.3 |
| TimeSearch-R + SigLIP | SigLIP | 1.41 | 41.4 | 23.8 |
| VEPO (1 Call) | None | 1.00 | 45.3 | 24.2 |
| VEPO (2 Calls) | None | 2.00 | 46.6 | 31.8 |
| VEPO | None | 3.18 | 47.5 | 36.3 |
| VEPO + SigLIP | SigLIP | 3.24 | 47.9 | 37.1 |
With the policy, data, tools, and interaction budget fixed, VEPO achieves higher accuracy across all four benchmarks and boosts evidence-supported correctness on LVBench from 21.9% to 36.3%, while preserving stronger temporal recall and active search. Moreover, TimeSearch-R remains below VEPO even with its SigLIP retriever restored and with VEPO restricted to a single native search (45.3% vs. 41.4%), while adding SigLIP to VEPO yields only a marginal 0.4-point gain.
Takeaway: TimeSearch-R derives search credit from re-answer correctness and delegates semantic frame selection to an external retriever. VEPO learns semantic visual evidence acquisition within the policy, credits the individual searches that improve it, and achieves stronger evidence-supported performance under controlled training and retrieval settings.
Author Response (Q1): Initial Overview in Evidence Evaluation
Q1: Including the Initial Overview in Evidence Evaluation
Across long-video benchmarks, the 64-frame overview covers only about 1% of candidate frames on LVBench and 2% on CG-Bench, providing coarse global context without replacing active search for localized evidence. Including both the overview and retrieved observations in the evidence evaluation under both our ERM and the independent Gemini 2.5 Flash judge (Tab. R2a) raises supported correctness across all methods while preserving VEPO's clear lead (36.3% → 40.2% under ERM and 31.8% → 35.5% under Gemini).
Tab. R2: LVBench Evidence Support with the Initial Overview
(a) by Evidence Input
| Method | Acc. | Retrieved (ERM) | Overview + Retrieved (ERM) | Retrieved (Gemini) | Overview + Retrieved (Gemini) |
|---|---|---|---|---|---|
| Conan | 39.7 | 0.0 | 17.4 | 0.0 | 15.6 |
| LongVT | 29.4 | 11.8 | 16.3 | 10.5 | 14.6 |
| VideoTemp-o3 | 36.8 | 13.6 | 20.8 | 11.8 | 18.4 |
| Video-MTR | 40.4 | 14.1 | 22.0 | 12.8 | 19.8 |
| TimeSearch-R + SigLIP | 41.4 | 23.8 | 28.5 | 20.6 | 25.1 |
| SFT | 43.8 | 21.5 | 30.2 | 18.9 | 26.8 |
| Outcome-only GRPO | 42.3 | 19.2 | 27.9 | 16.8 | 24.7 |
| VEPO | 47.5 | 36.3 | 40.2 | 31.8 | 35.5 |
To directly quantify the proportion requested by the reviewer, we examine correct predictions judged unsupported under Retrieved and evaluate how often the initial overview alone contains sufficient evidence under both our ERM and Gemini 2.5 Flash (Tab. R2b).
(b) Overview Support among Retrieved-Unsupported Correct Predictions
| Method | Supported by Overview Alone (ERM, %) | Supported by Overview Alone (Gemini, %) |
|---|---|---|
| VEPO | 21.3 | 19.1 |
| TimeSearch-R + SigLIP | 22.2 | 19.7 |
| Outcome-only GRPO | 24.1 | 20.8 |
| LongVT | 24.4 | 21.2 |
| SFT | 27.5 | 24.1 |
| TimeSearch-R (w/o SigLIP) | 28.0 | 24.5 |
| Video-MTR | 28.9 | 25.0 |
| VideoTemp-o3 | 29.3 | 25.6 |
| Conan | 43.7 | 39.3 |
Among VEPO's retrieved-unsupported correct predictions, 21.3% (ERM) and 19.1% (Gemini) are supported by the overview alone, the lowest proportions across all compared methods. This confirms that excluding the overview in the main diagnostic does not preferentially favor VEPO.
We then use GPT-5 mini to separate questions whose evidence is already visible in the overview (Overview-Sufficient, 27.7%) from those requiring additional search (Search-Required, 72.3%). In Tab. R3, existing agentic baselines rely mainly on the Overview-Sufficient subset (48.5%–65.5%) and drop sharply on the Search-Required subset (0.0%–17.4%; outcome-only GRPO likewise degrades relative to SFT on both subsets, 64.3% vs. 66.0% and 14.0% vs. 16.5%), whereas VEPO achieves the highest throughout (67.8% and 29.6%).
Tab. R3: LVBench by Overview Sufficiency
| Method | Acc. | Overview-Sufficient (27.7%) | Search-Required (72.3%) | Overall |
|---|---|---|---|---|
| Conan | 39.7 | 62.7 | 0.0 | 17.4 |
| LongVT | 29.4 | 48.5 | 4.0 | 16.3 |
| VideoTemp-o3 | 36.8 | 48.5 | 10.2 | 20.8 |
| Video-MTR | 40.4 | 65.5 | 5.4 | 22.0 |
| TimeSearch-R + SigLIP | 41.4 | 57.6 | 17.4 | 28.5 |
| SFT | 43.8 | 66.0 | 16.5 | 30.2 |
| Outcome-only GRPO | 42.3 | 64.3 | 14.0 | 27.9 |
| VEPO | 47.5 | 67.8 | 29.6 | 40.2 |
Takeaway: Most long-video queries require visual evidence beyond the initial overview, and VEPO's retrieved-unsupported correct answers rely least on the overview across all compared methods. VEPO remains strongest under both evaluation protocols and both judges because it actively acquires the visual evidence missing from the overview.
Author Response (W1), Part 1: Comparison with Evidence-Aware Rewards
We thank Reviewer spJn for the positive assessment and for finding the problem meaningful and the method well motivated.
TimeSearch-R:
Motivation and core distinction. The core difference is how two capabilities receive credit: acquiring semantically sufficient visual evidence and using that evidence to answer correctly. TimeSearch-R asks the policy to re-answer the question from the retrieved frames and rewards the search when the re-answer is correct. Search credit therefore still comes from answer correctness, so incomplete evidence can receive credit when the model guesses correctly. VEPO evaluates the semantic sufficiency of the acquired observations with the ERM and scores final-answer correctness separately. Replacing TimeSearch-R's self-verifier with an external model or self-consistency would not remove this coupling because the search would still be judged through another answer.
Native evidence acquisition. TimeSearch-R uses a text-conditioned SigLIP retriever built on a precomputed video-wide frame cache. The policy can submit a broad temporal range and leave query-conditioned semantic frame selection to the external retriever. Its CSV reward evaluates whether the returned frames support a correct re-answer, without directly crediting the policy for locating useful visual evidence. VEPO uses no external semantic frame selector. Its tool returns uniformly sampled frames from the temporal interval selected by the policy, so the policy must learn where to search, with semantic evidence credit directly supervising this capability.
Action-local evidence credit. TimeSearch-R produces a single CSV reward after the complete search trajectory, without identifying which search supplied the missing information. VEPO evaluates the accumulated evidence after each search and assigns additional credit when an action increases semantic evidence sufficiency. It therefore teaches the policy which searches fill the current evidence gap, while repeated or unhelpful searches receive no additional evidence credit.
Experimental validation:
Tab. R1: Controlled TimeSearch-R-Style Reward Analysis
(a) Measured Training Dynamics