acceptodds
Under review as a conference paper at ICLR 2027

EOR: Evidence-Obligation Reasoning for Verification-Driven Long-Video Question Answering

Abstract

Long-video question answering requires reasoning over sparse and temporally dispersed evidence. A central challenge is that retrieving video segments relevant to a question does not necessarily provide sufficient evidence to justify an answer. Existing retrieval-based methods mainly determine what to inspect from question relevance or intermediate reasoning cues, while recent hypothesis-verification approaches make evidence acquisition more goal-directed but typically do not maintain explicit, persistent verification requirements that govern subsequent reasoning.We propose EOR, an Evidence-Obligation Reasoning framework that formulates long-video QA as iterative hypothesis verification. Each candidate hypothesis is associated with persistent evidence obligations, i.e., observable conditions that should be established before the hypothesis is sufficiently grounded. Their states are maintained across reasoning rounds, allowing unresolved, contradicted, and discriminative requirements to determine what evidence should be inspected next. EOR uses a lightweight temporal event graph to localize candidate regions, verifies the corresponding source-video evidence, and allocates verification effort across a compact set of complementary hypotheses rather than repeatedly inspecting redundant candidates. For open-ended QA, accumulated verified evidence can further revise an inadequate hypothesis space instead of merely re-ranking a fixed set of initial candidates. We evaluate EOR in both multiple-choice and open-ended long-video question answering, examining answer quality, evidence-seeking efficiency, and robustness to incomplete initial hypotheses.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Long-video question answering, Video understanding, Multimodal reasoning, Evidence-guided retrieval, Hypothesis verification, Open-ended video QA

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_3LdFOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This article re models long video Q&A as an iterative hypothesis testing process and proposes an EOR framework. The core idea is to break down each candidate answer into several "ongoing evidence obligations" that need to be confirmed item by item in the source video, and maintain three states of pending/satisfied/refuted for each obligation, which are persistently retained between inference rounds, thus making the question of "whether the evidence is sufficient" explicit, rather than staying at "whether the retrieved fragments are relevant". The framework introduces a lightweight time event graph on the retrieval side to locate candidate intervals, and then returns to the source video to determine the obligation state using validators; And use a hypothesis selection mechanism that covers perception to allocate verification costs within a limited verification budget, and stop after meeting the preset grounding conditions; For open-ended questions, it is also allowed to dynamically evolve the hypothesis space driven by evidence.

The paper attempts to isolate the marginal contribution of the "persistent verification state" through a progressive ablation chain (from pure retrieval, to hypothesis retrieval, to instantaneous verification, and then to the complete framework), and verifies its effectiveness on two types of long video Q&A tasks: multiple-choice and open. Overall, its contribution mainly remains at the level of problem modeling and process design, while the empirical evidence supporting these claims is relatively weak.

Critical strengths

Witness the forefront clearly retaining complementary bottleneck profiles, rather than collapsing progress to a single incumbent; This design choice is highly consistent with the minimum margin structure achieved jointly, and is consistently reflected in the early hit behavior reported. Separate language proposals from posterior risk screening to keep the LLM output at a modifiable decision level and avoid contaminating the surrogate model with pseudo observations.

Critical weaknesses

This paper should be rejected.

W1. Deadliest: There is no positive comparison with the most relevant methods. The entire article repeatedly identifies "Think, Then Verify" as the "closest competitor" and claims that the difference between the two lies in "clue level goals vs. persistent obligations". It is unacceptable for an article to use "validation driven" as a selling point and clearly state that it is the primary benchmark, without any quantitative comparison of the strongest and most relevant baselines. Readers have no way of determining whether the benefits come from the new state of 'persistent obligations' or just from a few extra verification calls. This deficiency directly undermines the entire proposition.

W2. Headline numbers are abnormal and lack comparability evidence. The backbone model used in the framework has a limited scale, and the reported performance has shown an unusually large leap compared to the training free methods of the same backbone, even surpassing systems based on stronger closed source large models. Such a significant increase should have been accompanied by an extremely strict "evidence ledger", but on the contrary: the core selling point of the paper is "comparable video access budget/evidence acquisition efficiency", but it does not report the actual video frame rate, retrieval token, average access frame rate per call, or wall clock time of any method from beginning to end. On the premise that the access cost is completely unquantifiable, it cannot be ruled out that high accuracy is achieved through more aggressive video access, and "comparable budget" has become a verbal statement without data support. In addition, a baseline gear number in the ablation table is identical to an external method number in the main table, suggesting direct appropriation rather than independent reproduction, further weakening the credibility of the baseline.

W3. There is inherent tension and even logical conflict in the core ablation data. Compared to its simplified variants, the complete framework has significantly reduced verification calls, but the coverage of obligations confirmed as "fulfilled" has significantly increased, resulting in a synchronous increase in accuracy. There is less verification cost and more obligation to confirm. In the absence of mechanism level evidence to explain how to confirm more obligations with fewer calls, this is more like a loosely triggered "satisfied" judgment rather than a true coverage improvement. The author did not provide any explanation on this matter. At the same time, there is a lack of consistent reconciliation between the marginal revenue caliber presented by the budget curve and the "extremely few calls reaching high levels" described in the ablation table, and multiple numerical calibers are inconsistent with each other.

W4. There is no reliability assessment for the method foundation (obligation decomposition and support/future relationship labels). The relationship label is specified and cached by the model at once during creation, and does not determine whether the video truly contains this condition. This means that subsequent discriminative power calculations, coverage statistics, landing criteria, and even final ranking are all based on these initial labels; Once the label is incorrect, the entire production line runs on the wrong foundation. The author did not make any manual annotations or quality statistics on whether obligations are truly verifiable, whether they are just restatements of answers, and whether relationship labels are accurate, and did not even report the accuracy of obligation decomposition. This is the first domino of reliability, but it is completely hanging in the air.

W5. Validator and generator share the same origin, essentially being the self sealing of the model. The event clues, obligation decomposition, verification, and final implementation are all completed by the same model, and each event only samples a very small number of frames; The validator reports its confidence level, and the landing threshold is set quite loosely. Allowing a model with limited abilities to disassemble problems, search for evidence, and determine that the "evidence has been satisfied" on its own will inevitably result in extremely high noise levels in its "satisfied" label. The so-called 'high coverage of verification obligations' is likely due to overconfidence and incorrect stamping, rather than truly covering decisive evidence. The paper has zero ablation on the accuracy, precision/recall, and how the results change after switching to a stronger validator, and the entire commitment of "validation driven" is built on an unverified and highly likely unreliable validator.

W6. The landing criteria are too loose, making it almost impossible to intercept erroneous assumptions. The coverage rate and weighted evidence threshold required for landing judgment are both very low, and the scoring is conducted in the low resolution interval, resulting in a very small margin for distinguishing between the two candidate hypotheses; A mistaken assumption only needs to gather a very small number of obligations that have been misjudged as' fulfilled 'to pass. The paper neither provides a false positive rate for ground detection nor proves that this threshold can prevent 'confident incorrect answers'.

W7. Promise only, not fulfilled. The main text explicitly states that the triggering frequency of this fallback mechanism will be reported in the experiment, but there is no fallback frequency number throughout the main text and appendix. The absence of indicators in the commitment report itself is an issue of academic rigor.

W8. Multiple long video benchmarks for motivation naming have not been established, and there is no evidence of generalization. The abstract and introduction are motivated by the "difficulty" of multiple mainstream long video benchmarks, but the actual measurement only covers one multiple-choice dataset (and only its subset) and one open-ended dataset. Motivation and verification are severely disconnected; The sample size is relatively small, and there are no standard deviations/multiple runs/significance tests in the entire text. Without a confidence interval, the extent of improvement is not convincing.

W9. "Efficiency" advocates no ledger, and the active set strategy is not compared with the naive control. The greedy assumption of coverage perception is not compared with simple strategies such as ranking by score, randomness, and polling, which cannot prove that the benefits come from "coverage perception" rather than "leaving a few more assumptions arbitrarily". The so-called evidence acquisition efficiency does not have token/frame/delay ledger, and cannot evaluate the true inference cost.

W10. The evidence for the open-ended question "Hypothesis Evolution" is weak, and the reliability of the evaluator has not been verified. The complete evidence of the evolutionary mechanism consists of only two extremely small differences in numbers, without analysis of the triggering rate of evolution, the accuracy of new candidates, and the quality of the hypothesis pool before and after evolution. More importantly, if an open-ended question is scored by another large model acting as a judge without reporting consistency with manual evaluation, then the credibility of the open-ended question indicator itself is questionable.

W11. Lack of reproducibility and sensitivity. A large number of thresholds are fixed on both datasets, but there is no sensitivity analysis for key thresholds that have a significant impact on the results, and the paper does not promise code/data release. Highly customized systems are almost impossible to independently reproduce.

Questions for authors

(W1, most importantly) Please provide a direct quantitative comparison between EOR and the most relevant competitive methods under the same backbone and video access budget; If there is no such comparison, how can the net gain of "persistent obligation" relative to clue level verification be proven?

(W2) Please provide the actual video access cost ledger for each method (average access frame rate, token retrieval, average frame rate per verification call, end-to-end wall clock time), and indicate whether the headline results were obtained under true "comparable access".

(W3) Please explain the mechanism of "reducing verification calls and increasing coverage of fulfilled obligations" in ablation, and provide a decomposition of "how many obligations are confirmed on average per call" to prove that this is not triggered by excessive relaxation of "fulfilled".

(W4) Please provide a manual evaluation of the quality of obligation decomposition: the proportion of "verifiable/non answer restatement" obligations, and the accuracy of support/future/neutral relationship labels relative to manual annotation.

(W5/W6) Please report the accuracy of the validator's own judgment (accuracy/recall/F1 of support/future/uncertainty), and how the results change after switching to a stronger validator; And provide the false positive rate of landing judgment.

(W7) Please provide the promised but not given rollback triggering frequency and its impact on final performance in the main text.

(W8) Please reproduce the headline conclusion on at least one other long video benchmark and report the variance and significance of multiple runs.

(W10) Please report the consistency between automatic refereeing and manual evaluation, otherwise the open question indicators are unreliable.

(W11) Please explain the basis for the value of the key threshold and provide a sensitivity curve, and clarify whether the code is open source.

Reviewer_W6ztOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

EOR is a training-free framework for long-video question answering that maintains persistent evidence obligations for competing answer hypotheses. Each obligation records an observable condition and its verification status, allowing unresolved requirements to guide subsequent retrieval and source-video inspection. A temporal event graph supports retrieval, while evidence-coverage-aware hypothesis selection limits redundant verification. For open-ended questions, accumulated evidence can also revise the initial candidate pool. Experiments on EgoSchema and ActivityNet-QA report improved answer accuracy, with controlled ablations supporting the contributions of persistent verification states, active hypothesis selection, and hypothesis evolution.

Critical strengths
  1. A useful persistent verification state. Evidence obligations explicitly track unresolved requirements, distinguish missing evidence from contradiction, and support evidence reuse across hypotheses (Sections 3.1–3.3). This provides a clear mechanism for moving beyond relevance-driven retrieval.

  2. Controlled ablations support the design. Persistent obligations improve EgoSchema accuracy from 67.0% to 69.2%; EC-MAHS further raises it to 70.2% while reducing verification calls from 7.9 to 6.1 (Table 3). Hypothesis evolution also improves open-ended QA accuracy (Table 2).

Critical weaknesses
  1. Grounding reliability needs independent validation. Verified Obligation Coverage measures the system’s own verification states, without an independent audit or calibration of the confidence threshold γh=0.7\gamma_h=0.7 (Table 3; Appendix A.6). Such validation would help establish whether persistent states preserve reliable evidence rather than early verification errors.

  2. Efficiency reporting is incomplete. Verification-call counts do not capture graph construction, hypothesis-management overhead, or the amount of video processed per call (Section 4.3). End-to-end latency and total visual/token costs would better substantiate the practical efficiency benefit.

Questions for authors
  1. Do existing audits or error analyses indicate how often obligations marked as satisfied are actually supported by the video?
  2. Can existing execution logs provide end-to-end costs, including graph construction and hypothesis management, for EOR and the controlled baselines?
Reviewer_bVT7OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Considering that retrieving video segments relevant to a question does not necessarily provide sufficient evidence to justify an answer, the paper proposes EOR, an evidence-obligation reasoning framework that formulates long-video QA as iterative hypothesis verification. EOR decomposes candidate answers into observable evidence obligations and maintains their satisfied, contradicted, or unresolved states across reasoning rounds. These states guide source-video verification, while an evidence-coverage-aware active set allocates the verification budget across competing hypotheses. For open-ended questions, verified observations can also revise the candidate space. Experimental results based on Qwen2.5-VL-7B report 70.2% accuracy on the EgoSchema 500-question subset and 57.2% accuracy on ActivityNet-QA.

Critical strengths
  • The motivation is clear, and the proposed method is reasonable. Rather than repeatedly retrieving video content, the proposed EOR uses the evolving verification state to determine what evidence should be inspected next and allocates a limited verification budget across complementary hypotheses. The controlled experiments in Table 3 provide evidence of improved accuracy and efficiency.
Critical weaknesses
  • The comparisons leave uncertainty about the incremental contribution of the proposed EOR and the content of its advantage. The most related method, VideoHV-Agent, has already used verification states, additional evidence acquisition, and hypothesis refinement. However, the comparison between the two methods in Table 1 cannot demonstrate this point for different backbones used.
  • It seems that the claim of stronger evidential grounding is not sufficiently supported by existing experiments. Verified Obligation Coverage measures the extent to which model-generated conditions are judged satisfied by the verifier, but does not establish whether these conditions cover the decisive evidence needed to justify the answer. Although EOR improves QA accuracy on both datasets, the paper provides limited qualitative analysis explaining the source of these gains. Figures 1 and 2 illustrate the framework, but comparisons on actual benchmark examples would help show what decisive evidence a baseline misses, how persistent obligations guide EOR to obtain that evidence, and how it changes the final answer.
Questions for authors
  • The paper admits that EOR share the same verification-oriented perspective as VideoHV-Agent, but represents the reasoning state differently. What result most directly establishes the value of the proposed method? Fair comparisons between EOR and VideoHV-Agent, based on the same strong backbone, such as a new GPT.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.