acceptodds
Under review as a conference paper at ICLR 2027

Beyond Readiness: Operator-Conditioned Answer Closure in Streaming Evidence

Abstract

A supported answer can still change as a stream continues. We formalize operator-conditioned answer closure: a supported candidate is closed when every continuation admitted by the question operator and horizon preserves it. Four reference states distinguish absent support, provisional support, closure, and terminal unanswerability; calibrated gates derive response actions. Operator-balanced shared-prefix streams provide exact symbolic labels for comparing learned estimators across seven controllers. On 1,440 held-out symbolic trajectories from known program families, a gap-aware GRU reaches 0.9954 closure average precision and 0.9851 state macro-F1. Relative to an event-time survival baseline, it reduces premature commitment from 0.0040 to 0.0008 at identical answer coverage. Normalized delay, however, increases from 0.0612 to 0.0683, and same-target controls show no uniform ranking advantage from finer state supervision. With frozen VideoLLM features, a human-reviewed video pilot yields closure AP 0.9180 but fixed-four-class macro-F1 0.2900. Its seven held-out video records come from two source components and contain no terminal-unanswerable reference labels. These results distinguish current support from answer finality and show why closure ranking, state classification, and stopping quality require separate evaluation.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
streaming video question answering, vision-language models, answer closure, temporal reasoning, selective prediction

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_6UfdOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This article formalizes the operator conditional answer closure under streaming evidence: a supported candidate is only determined as a closure when the declared operator maintains the answer for all tasks that are allowed to continue in the time bound. Four reference states have been defined (pending, temporary, closed, terminal unanswerable), and response actions are generated by calibration gating. The precise label comes from the symbol shared prefix stream balanced by the operator; Exploratory video pilot uses frozen VideoLLM features and AI assisted annotation that has been manually reviewed. Seven types of time controllers were compared. On 1440 reserved symbol trajectories, the gap aware GRU reached closure AP 0.9954 and four state macro F1 0.9851, reducing premature commitment relative to the event time survival baseline but increasing normalized delay. The video test of seven records obtained a closure AP of 0.9180, a four state macro F1 of only 0.2900, and did not include the terminal non answerable label.

Critical strengths

Distinguish between 'current support' and 'finality of answers under permissible future', clarify the concept, and correctly separate the problem from ordinary readiness or selective prediction. The symbol sharing prefix construction provides precise and operator balanced labels, enabling controlled measurement of ranking, state classification, and stopping trade-offs on the same target.

Critical weaknesses

W1. The scale of the video pilot is too small and unbalanced to support any claims about real visual flow. Leave seven records from only two source components for testing; There are no terminal unanswerable tags in both development and testing. The four state macro F1 dropped to 0.2900, while the boundary and problem clues alone can restore most of the ranking signals, indicating that the contribution of visual content is minimal.

W2. The symbolic performance near the ceiling is built on highly constrained grammars, with a significant overlap between program templates and training. Eight out of the nine reserved expression trees share exactly the same four state monitoring semantics with the base operator, so the reported combination gain mainly reflects robustness to syntactic re encoding rather than true semantic transfer.

W3. Four state supervision did not significantly improve the ranking goal of core closures. The direct binary closure and survival baseline have reached an AP that is indistinguishable from the main model statistics; the main reported benefit is further reduction of premature commitment, at the cost of higher latency, and the joint stop guarantee has not passed.

W4. The formal contribution is relatively incremental compared to classical runtime validation and early classification literature. Good prefix monitoring of propositions with equal answer dependencies is a natural specialization of existing monitoring semantics; The refinement of four states and operator certificates have not introduced a new decision theory mechanism that is independent enough to form a learning problem.

W5. The annotation quality of the video pilot relies on AI assisted annotation and single author review. There is no consistency number between annotators, no replayable annotation protocol, no quantitative evaluation of visual evidence threshold, and residual label noise cannot be quantified.

W6. Stop evaluating mixed coverage, premature rate, and delay without a fixed cost model or Pareto analysis. The premature rate is pushed close to zero simply because it accepts longer conditional delays and allows for unanswerable actions to be taken; When the action is disabled, the advantage decreases, indicating that the reported operating point is fragile.

W7. The ablation and control results showed that several claimed components were essentially ineffective. After removing pairing loss, revision loss, or embedding prediction operators, the symbol AP remains almost unchanged; The maximum drop only occurs when the state or risk supervision is completely removed, indicating that complex long targets excessively specify ranking tasks.

The experimental records show several numerical patterns that reduce credibility: the seed standard deviation of symbol AP is often lower than 0.0002, the initial cost interval and class AUR aggregation height are regular, and the same 17 video records are reused for all layer detection and interference control, amplifying the risk of overfitting to a few trajectories in the results.

Questions for authors

Can a complete review log of each sub trajectory and original label be made public for independent verification of macro averages and missing U labels in the report?

Since the boundaries and problem clues have restored most of the performance of video ranking, is there any evidence to suggest that the controller actually used visual content instead of temporal boundary metadata?

Why is joint delay guarantee allowed to fail, and what operational point selection procedure can make the claimed premature commitment reduction robust under a fixed cost model?

3 comments
AuthorsOpenReview ↗

Response to Reviewer 6Ufd (1/3)

Thank you for recognizing the support/finality distinction and for examining the experimental controls closely. Your comments helped us make the contribution more precise and the empirical evidence easier to audit. The revision uses the existing results to clarify the video pilot, same-target comparisons, stopping trade-off, annotation records, and numerical aggregation. We address W1–W7 below, with Q1–Q3 incorporated into their corresponding sections.

  • W1 and Q2. Video evidence and visual-grounding measurements

The video experiment provides a focused pilot for studying closure prediction from frozen causal features. Its 12/3/7 records come from 6/2/2 source components, with P, R, and C labels in development and test. The full controller reaches closure AP 0.9180, compared with 0.8969 for the question/clock/end control, and fixed-four-class F1 is 0.2900. These measurements provide an initial representation-level result and a concrete reference for evaluating the additional contribution of visual features at larger source scale.

The revision now places the sample composition, annotation protocol, and metadata comparison alongside the video result. It also distinguishes reference-state prediction from generated-answer quality and end-to-end stopping, and identifies larger-source and end-to-end evaluations as the next validation steps. All layer probes remain exploratory, with the primary checkpoints frozen and the best test layer kept separate from the primary model. This presentation makes both the pilot's evidence and its appropriate use clear (Sections 4, 6.5, and 7; Tables 2 and 9; Figure 7).

  • W2. Expression transfer and semantic generalization

The base test measures transfer to new trajectories within known program families: all 35 base-test program specifications occur in training, and 819/1,440 test records reuse a training program/token template with different timestamps. The nine separately held-out expression trees then provide a targeted semantic analysis. Eight have a base operator's four-state monitoring function, explaining their direct transfer. On the only non-equivalent tree, OR(Complete, Exists), four-state AP is 0.9932 and three-state AP is 0.9935.

The revision now makes this analysis central to the transfer discussion. The principal holdout gains demonstrate robustness to expression encoding, while the non-equivalent-tree comparison defines a focused test for semantic extrapolation. This sharper interpretation aligns the claims directly with the measured evidence (Sections 4.1 and 6.5; Table 21; Appendices A.7 and A.10).

  • W3. Same-target ranking and the stopping trade-off

The same-target comparisons locate the benefit of four-state supervision in stopping behavior. Primary symbolic AP is 0.9954 for four-state supervision, 0.9953 for survival, and 0.9858 for direct binary closure; the paired CPU three-state comparison provides a complementary matched control. The much larger readiness gap diagnoses the change in prediction target, while the same-target results characterize ranking under aligned objectives.

At answer coverage 0.6412, four-state closure reduces prematurity from 0.0040 for survival to 0.0008, with normalized conditional delay changing from 0.0612 to 0.0683. The existing two-sided paired source-bootstrap intervals for four-state minus survival are [-0.0060, -0.0011] for prematurity and [+0.0054, +0.0087] for delay. These results quantify a meaningful trade-off: substantially fewer premature commitments together with increased delay. The exploratory intervals condition on jointly eligible groups, and delay eligibility is matched at the group level, providing a precise interpretation of the stopping comparison (Sections 6.3 and 6.4; Tables 4 and 16).

  • W4. Classical monitoring and the complementary role of learning

Our formalization specializes good-prefix monitoring to the question-dependent proposition that a completed answer equals a supported candidate. Its task-specific elements are the declared continuation set, operator certificates, support-sensitive state refinement, and shared-prefix tests that separate support from finality. This formulation connects runtime-monitoring semantics to a concrete answer-closure learning problem.

The exact symbolic monitor serves as an oracle reference: given exact semantics, it achieves zero premature commitment and zero delay from the same observed symbols and declared program available to the learners. The learned models answer the complementary question of how supervision targets shape approximation and stopping under controlled inputs. The symbolic study therefore supplies a clean comparison of learned estimators and supervision targets, while the video pilot introduces the uncertain-observation setting for further study. The revision now presents these roles explicitly and uses the oracle result to strengthen interpretation of the learned models (Sections 2 and 4.1; Table 4; Appendix A.6).

AuthorsOpenReview ↗

Response to Reviewer 6Ufd (2/3)

  • W5 and Q1. Annotation quality and auditable records

The annotation pipeline combines AI-assisted initial labels with review and confirmation by one author; all retained labels were confirmed without changes. The anonymous supplement includes the retained labels, per-record review rationales, exclusions, split manifests, and predictions needed to inspect the reported counts and metrics. Its annotation/README.md locates the annotation records, and file_index.json lists the supplied files before extraction. Input hashes and transition checks further support structural validation at the one-frame-per-second annotation resolution.

The revision now describes this audit trail precisely in the main text and annotation appendix. It distinguishes author verification from independent agreement estimation, and the surviving annotation records from the original interactive labeling history, including the exact original labeling-model version. Independent agreement and quantitative residual-noise analysis provide clear extensions to the current annotation protocol. Source videos remain available under their upstream access and licensing conditions (Section 4.3; Appendix A.11).

  • W6 and Q3. U actions, operating-point selection, and the delay safeguard

The reported primary symbolic comparison contains a useful control for U actions. Mapping U to abstention and reselecting the policy on development data leaves answer coverage at 0.6412, prematurity at 0.0008, and delay at 0.0683, while U coverage changes from 0.3588 to zero. Thus, for these recorded policies, the answer-metric advantage is preserved when U is disabled. The revision states this result beside the main stopping table and scopes it to the evaluated policies.

The stopping results also make the delay trade-off explicit. The separate joint criterion against readiness yields a one-sided upper delay bound of 0.0737, exceeding the 0.05 safeguard. We retain this criterion outcome because it identifies an important deployment requirement: reduced prematurity must be evaluated together with delay.

Thresholds and dwell settings are selected by minimizing deviation from 90% development total coverage, with ties favoring higher thresholds and then longer dwell. The appendix displays the complete fixed-grid test results, making the operating-point sensitivity fully inspectable. A deployment with a fixed error-delay cost can build directly on this protocol by specifying costs before testing and selecting the policy on development data. The revision distinguishes this fixed-cost extension from the current coverage-selected evaluation (Sections 5.4 and 6.4; Tables 4 and 11; Figure 8; Appendix A.4).

  • W7. What the auxiliary-component ablations establish

The ablations provide a clear assessment of the evaluated architecture. Removing pair supervision, revision supervision, or the auxiliary predicted-operator embedding leaves symbolic AP nearly unchanged, indicating that the primary closure target carries the central learning signal in this setting. The question vector continues to encode the exact symbolic tree when the auxiliary embedding is removed, which explains the retained operator information. On video, all revision targets are masked, so the corresponding video ablation specifically reflects that masked-target regime.

The revision now presents the GRU and auxiliary losses as an evaluated implementation of the closure target, with the ablations identifying which components drive the measured results. This interpretation makes the architectural evidence more informative and provides a direct basis for simplifying future implementations (Section 6.5; Table 7).

AuthorsOpenReview ↗

Response to Reviewer 6Ufd (3/3)

  • Numerical reliability and transparent aggregation

The seed standard deviations quantify optimization variation on the fixed, constrained split, complementing the source-level analysis of generalization. The supplied per-run predictions and configurations allow readers to inspect the reported aggregates directly, and the near-ceiling symbolic results are interpreted within the known program-family evaluation described above.

Action metrics use terminal-answer-class balancing within intervention groups, followed by averaging over groups. For example, seed 17 has raw prematurity 0.0006944 and macro prematurity 0.0004630; answer coverage is 0.7028 raw and 0.6412 macro. The aggregation equation and eligibility rules in the controller/metrics appendix explain the regular-looking table values and support exact reproduction of the aggregation procedure.

The retained video dataset contains 22 records in a 12/3/7 split. The layer and nuisance comparisons use the same seven test records from two sources, with reference counts P=15, R=84, C=35, and U=0. The revision reports these counts explicitly, treats the layer and nuisance analyses as exploratory diagnostics, and uses the frozen primary model for the principal result. Together, these disclosures make the numerical evidence and its source-level scope straightforward to assess.

Reviewer_qpbvOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces operator-conditioned answer closure, defined as the invariance of a supported answer under all task-admissible continuations. The formulation distinguishes four states: pending, provisional, closed, and unanswerable. The authors evaluate learned closure estimators across several temporal controllers using symbolic streams with exact reference labels and a small real-video pilot.

Critical strengths

The distinction between evidence support and answer finality is conceptually meaningful for streaming video understanding. In particular, conditioning closure on the question operator, horizon, and admissible continuations provides a precise account of why identical observations may resolve one question while leaving another open.

The experimental design includes useful controls. Shared-prefix trajectories isolate the effect of subsequent evidence, while binary-closure, three-state, and survival baselines help distinguish the contribution of the supervision target from that of the controller architecture. The authors also report premature commitment, coverage, and delay separately and acknowledge the limitations of interpreting closure ranking as successful response control.

Critical weaknesses

The principal concern is that the empirical contribution beyond the proposed formulation remains insufficiently established. The substantial performance gap between closure and readiness supervision primarily reflects a mismatch between their training targets when both are evaluated against closure labels. Comparisons on the same target are more informative, but survival and three-state alternatives achieve comparable closure AP. Although the four-state model reduces premature commitment, this improvement is accompanied by increased delay. The results therefore do not establish a consistent practical advantage for the four-state formulation.

The real-video evaluation is insufficient to substantiate broader claims about streaming video understanding. The test set contains only seven records from two source components and no unanswerable examples. The primary controller achieves closure AP of 0.918 but four-state macro-F1 of 0.290, with the absent U class assigned zero. Furthermore, question, clock, and observed-end cues alone achieve AP of 0.897. These findings leave unresolved whether the model captures visually grounded answer finality beyond task and boundary information. Neither generated-answer quality nor end-to-end stopping is evaluated on video.

Finally, the symbolic experiments provide limited evidence of semantic generalization. The main test uses known program families, with substantial overlap in program/token templates across training and test records. Additional expression holdouts broaden the evaluation, but most retain previously observed monitoring semantics. Consequently, the experiments establish learnability within a controlled symbolic setting more clearly than generalization to uncertain observations and unfamiliar task structures.

Questions for authors
  1. Beyond interpretability, what empirical benefit does separating pending from provisional provide relative to the three-state and survival alternatives?

  2. Can the authors demonstrate that visually grounded closure predictions improve response timing beyond question- and boundary-based controls on a broader video evaluation?

2 comments
AuthorsOpenReview ↗

Response to Reviewer qpbv

Thank you for carefully separating the conceptual contribution, same-target controls, and stopping behavior. Your comments helped us organize the revision around the distinct questions answered by these experiments. We now give the same-target results greater prominence, present the stopping trade-off explicitly, and retain the complete set of measurements, including the delay-safeguard outcome.

  1. Same-target controls identify where four-state supervision helps

The closure-versus-readiness AP gap diagnoses a mismatch in prediction targets. The matched same-target results then isolate the effect of the state representation. On the symbolic test, the primary four-state model has AP 0.9954 and survival has 0.9953; in the separately paired CPU comparison, three-state and four-state AP are 0.9955 and 0.9953. These near-equivalent rankings allow the revision to locate the contribution more precisely in state semantics and stopping behavior rather than universal ranking gains.

The P/R distinction supervises the difference between missing support and supported-but-revisable evidence. The state-group-mass-matched comparison shows how this distinction can shape a learned policy under a common scalar gate with no U action: four-state versus three-state prematurity is 0.0005 versus 0.0116, delay is 0.0715 versus 0.0473, and answer coverage is 0.6412 versus 0.6406. The corresponding in-domain AP values are 0.9953 and 0.9924, while the ordering on evidence-bearing prefixes is 0.9951 versus 0.9953. Thus, the matched experiment demonstrates a concrete policy effect together with its delay trade-off. We use this comparison to assess the state split and treat the survival comparison as a combined supervision-and-objective comparison.

The video pilot provides an additional matched measurement: paired three-state and four-state models reach AP 0.9269 and 0.9220, with the same merged-state F1 of 0.2804. These results refine the contribution consistently across the revision: P/R supplies an interpretable supervision signal whose principal observed benefit concerns premature commitment, with ranking and delay evaluated separately (Section 6.3; Tables 18 and 24; Appendices A.7 and A.8).

  1. A quantified reduction in premature commitment and its delay trade-off

At the same symbolic answer coverage of 0.6412, four-state closure reduces prematurity from 0.0040 for survival to 0.0008. Its normalized conditional delay is 0.0683, compared with 0.0612 for survival. The existing paired source-bootstrap intervals for four-state minus survival are [-0.0060, -0.0011] for prematurity and [+0.0054, +0.0087] for delay. These measurements characterize a clear operating trade-off at the selected points: fewer premature answers accompanied by additional delay. The intervals are exploratory and condition on jointly eligible groups, which provides the appropriate interpretation of the comparison.

The separate joint criterion against readiness gives a one-sided upper delay bound of 0.0737, exceeding the 0.05 safeguard. Reporting this outcome makes the design requirement concrete: development selection currently targets coverage, and applications requiring a fixed error-delay cost need that cost specified during development selection. The fixed-grid sensitivity figure complements this analysis by displaying the complete evaluated grid rather than a test-selected frontier. The revision retains both the prematurity gain and the delay-safeguard result, providing a complete account of the stopping behavior (Sections 5.4 and 6.4; Tables 4, 11, and 16; Figure 8).

  1. Video evidence and a broader evaluation path

The video pilot supplies an initial measurement of closure information in frozen causal features: the full controller reaches AP 0.9180, compared with 0.8969 for question/clock/end cues. Seven test records from two sources, the absence of U targets, and fixed-four-class F1 of 0.2900 define the current evaluation regime. We retain the original F1 convention, including the zero contribution of the absent U class, so that the score remains directly comparable with the reported protocol.

The revision presents this study as a representation-level evaluation of reference-state prediction and uses its results to define the broader validation path: more source videos, explicit visual-grounding controls, generated-answer assessment, and end-to-end response timing. The separate released-MMDuet2 WEB results complement this pilot by characterizing the released system under its own protocol. This organization gives each experiment a clear role in the overall evaluation (Sections 4, 6.5, and 7; Tables 2 and 9; Appendix A.9).

AuthorsOpenReview ↗

Response to Reviewer qpbv

  1. Expression transfer and semantic generalization

The base test evaluates transfer across new trajectories within known program families: all 35 test program specifications occur in training, and 819 of 1,440 test records share a program/token template with training while differing in timestamps and source identities. The revision now states both facts in the main dataset description, making the measured form of transfer explicit. The nine additional held-out trees provide a targeted semantic analysis. Eight have a base operator's four-state monitoring function, directly explaining their transfer behavior. On the sole non-equivalent tree, OR(Complete, Exists), four-state and three-state AP are 0.9932 and 0.9935, respectively. These findings support a precise conclusion: the principal holdout gains demonstrate robustness to unfamiliar expression encodings, while non-equivalent-tree semantics provide a focused direction for broader generalization studies. The revised contribution is therefore a precise, testable closure target and an evaluation that separates ranking, state prediction, and stopping quality (Sections 4.1 and 6.5; Table 21; Appendices A.7 and A.10).

Reviewer_AdkdOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This manuscript tackles taht streaming video understanding systems conflate support with finality. The authors formalizes operator-conditioned answer closure where a supported candidate is close dwhen every continuation in a declared admissible set, given the qustion operator and horizon rule, yeidls the same answer. Six base operators come with witness-type or domain-closing certificates. The authors build symbolic shared-prefix datsets with exact labels via reachable-answer dynamic programming, plus a 22-record video pilot annotated by an AI system and reviewed by one authro.

Critical strengths
  • a real and well-motivated distincion, separating answerable now from answer cannot change. It is a genuine gap in current straming video understanding evaluation
  • careful formalization of proposed problem
Critical weaknesses
  • The video evidence is somewhat weak to support any claim about VideoLLMs. The test set has 7 records from 2 source components, with no U lables and unbalanced operators. Labels were produced by an AI annotator and confirmed by a single author without changes, with no inter-annotator agreement.
  • The symbolic benchmark does not test anything learning is needed for. Labels com from an exact, known automaton, and the exact rule monitor attains zero prematurity and zero delay with no learning. The leaned models receive the operator tree, target, budget, and exact toekns as input, so they are approximating a known deterministic funciton.
  • Presentation of manuscript is extremely dense and heavily hedged. Nearly every sentence carries a qualifier, and the main text frequently defers essential information to 20-page appendix.
  • No published streaming systems is evaluated on closure. No prompting baseline that asks a VideoLLM directly whether its answer could still change.
Questions for authors

Please see weakness section.

1 comment
AuthorsOpenReview ↗

Response to Reviewer Adkd

Thank you for recognizing the distinction between current support and answer finality. Your comments helped us sharpen the paper's central contribution and present the evidence more effectively. The revision now gives greater prominence to the controlled comparisons, the role of the exact monitor, and the scope of the video pilot, while preserving every reported numerical result.

  1. Video evidence and annotation transparency

We have brought the video sample, annotation procedure, and evaluation scope into the main text. The experiment provides a focused representation-level pilot with 22 retained records from 10 source components, divided into 12/3/7 training/development/test records. The seven test records come from two components, and the development and test partitions contain P, R, and C targets. The six operators are represented across the retained dataset, with their empirical frequencies reported explicitly.

The results provide a concrete starting point for studying visual closure prediction: the full controller reaches closure AP 0.9180, compared with 0.8969 using question/clock/end cues, with four-state F1 of 0.2900. The revision presents these measurements as a pilot evaluation of reference-state prediction from frozen causal features and identifies larger-source validation, generated-answer quality, and end-to-end response timing as the next evaluation steps (Sections 4, 6.5, and 7; Tables 2 and 9).

The annotation pipeline combines AI-assisted initial labels with review and confirmation by one author; the retained labels required no changes during that review. The supplement provides the per-record rationales, exclusions, split manifests, and predictions for direct inspection. The revision distinguishes these surviving audit materials from independent annotation agreement and the original interactive annotation history, giving readers a precise account of the available evidence (Appendix A.11).

  1. The complementary roles of learning and an exact monitor

The exact monitor provides a valuable oracle reference for the symbolic study. Given exact event symbols and task semantics, it achieves zero prematurity and zero delay using the same observed symbols and declared program available to the learned models. This result validates the closure semantics and supplies a clear performance reference.

The learned models address the complementary question of how different supervision targets shape prediction and stopping from observed prefixes. The controlled symbolic design makes their inputs explicit—exact event tokens, operator tree, target, and declared budget—and compares their behavior under shared information. The base split evaluates new source trajectories within known program families, providing a clean test of supervision and policy behavior. This framework also creates a direct route to studying learned closure estimation when observations or semantics become uncertain.

The revision therefore presents our contribution as a task-specific closure formulation and controlled evaluation: closure specializes good-prefix monitoring to answer equality under a declared continuation set. The exact automaton and learned estimators now play clearly distinguished, complementary roles in the experimental story (Sections 2 and 4.1; Table 4; Appendix A.6).

  1. A clearer presentation of the main argument

We have streamlined the contribution statement and moved the information needed to interpret the main results into the main text: symbolic learner inputs, video sample and annotation details, development-policy selection, separate answer/U coverage, and the delay-safeguard outcome. The detailed equations and complete diagnostic tables remain available in the appendix. The main comparisons now distinguish same-target ranking, state prediction, and stopping quality, allowing readers to identify the contribution of each design choice directly.

  1. Complementary baselines and a concrete comparison path

The revision now clarifies the roles of two complementary evaluations. The matched readiness head isolates support supervision under controlled conditions, while the separate 500-question WEB evaluation characterizes the released MMDuet2 checkpoint under its recorded judging protocol. Presenting these results separately makes the evidence attributable to the closure controller precise (Sections 6.1 and 7; Appendix A.9).

Your suggested direct prompt—asking whether the answer could still change—is a natural additional baseline for the trained closure head. We have made direct closure prompting, generated-answer evaluation, and end-to-end response timing explicit priorities for the deployment-focused evaluation agenda. Together with the controlled same-target comparisons already reported, these additions provide a clear path from the present closure study to broader VideoLLM deployment studies.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.