acceptodds
Withdrawn from ICLR 2027

Answer-Level Trust Selection for Physical Vision-Language Reasoning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Vision-language models (VLMs) can estimate physical quantities such as duration, speed, and acceleration from visual observations, but existing benchmarks primarily assess overall model performance against annotated ground truth. In deployment, a key question is whether an individual prediction can be trusted when its ground truth is unavailable. Self-consistency alone may fail to capture important failure modes: a VLM may produce stable-but-wrong estimates or rely on textual priors rather than visual evidence. We formulate answer-level selective prediction for quantitative physical reasoning and propose Answer-Level Trust Selection (ATS), a post-hoc, model-agnostic framework for accepting or rejecting individual VLM predictions. ATS requires no fine-tuning, auxiliary verifier, or access to the model's internal logits. Instead, it aggregates eight interpretable behavioral diagnostic scores derived from repeated queries and controlled interventions into a unified trust score. We evaluate ATS in depth on Qwen2.5-VL-7B and across 20 VLM backbones, examining selective performance, diagnostic behavior, and targeted failure modes. Our results show that intervention-based diagnostics help identify stable-but-wrong and prior-tracking predictions that repeated agreement alone may miss. However, improved failure-case rejection can come at the cost of lower retention of correct predictions. ATS therefore complements model-level capability evaluation with answer-level reliability assessment for quantitative VLM predictions. Code will be released upon publication.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
AI safety, LLM agents, uncertainty estimation, selective prediction, safe decision making, tool-using agents

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_dr5tOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper addresses the instance-level reliability of VLMs in quantitative physical reasoning, where ground-truth values are unavailable at deployment and standard self-consistency fails due to "stable-but-wrong" predictions and over-reliance on textual numerical priors. To overcome these limitations, the authors propose answer-level trust selection (ATS), a training-free, verifier-free, black-box selective prediction framework that operates without access to internal model probabilities. Using controlled probing around an initial prediction, ATS computes eight behavioral diagnostic signals and aggregates them against a validation-calibrated threshold to selectively accept or reject outputs. Evaluated across 20 VLMs on QuantiPhy, ATS demonstrates its effectiveness in filtering out unreliable numerical predictions and mitigating both stable-but-wrong and prior-tracking failure modes.

Critical strengths
  1. The paper formulates a well-motivated and useful problem by shifting the evaluation paradigm from aggregate benchmark performance to instance-level selective prediction in quantitative physical reasoning. Because ground-truth measurements are unavailable at deployment, determining when to trust or abstain from an individual numerical output is essential for safety-critical downstream applications. Furthermore, the core premise of actively probing models via controlled input interventions offers a compelling approach to uncover behavioral vulnerabilities that passive repeated sampling inherently misses.

  2. The proposed ATS framework features a portable, model-agnostic black-box architecture. By relying purely on inference-time input perturbations and behavioral consistency checks, ATS can be seamlessly applied across diverse VLM architectures.

  3. The empirical evaluation is extensive, spanning 20 VLMs across both proprietary LLMs and open-weight families of varying scales. In addition to benchmarking against standard self-consistency, generic uncertainty quantification baselines, and oracle rankings, the authors provide diagnostic ablations, weight-sensitivity analyses, and paired cross-model comparisons that validate the contributions of the different ATS variants.

Critical weaknesses
  1. On standard selective prediction metrics, ATS fails to demonstrate a consistent advantage over simpler baselines. On the primary Qwen2.5-VL-7B backbone, both Prior-weighted and Learned ATS underperform either SC-only or Generic UQ rank fusion across MRA@30%, MRA@50%, and AURC. Across all 20 evaluated VLMs, the average gains are marginal. Ablating the signature counterfactual diagnostic actually improves cross-model performance, raising average MRA@50% from 0.4525 to 0.4646 and reducing AURC from 0.5488 to 0.5433. The authors need to substantiate that behavioral probing yields genuine overall risk reduction and clarify when ATS is preferable as a general-purpose selective predictor.

  2. The paper's headline results on targeted error subsets ("stable-but-wrong" and "prior-tracking") suffer from severe definitional coupling between the evaluation slices and the selection criteria. Specifically, "stable-but-wrong" instances are defined by combining low ground-truth accuracy with high repeated-sampling stability, making SC-only's dismal 1.3% rejection rate on this subset a tautological artifact of slice construction rather than a fair reflection of its general error-detection capability. Similar problem can be observed in terms of "prior-tracking" subset. While these analyses confirm that ATS responds as engineered to its constituent features, they offer limited independent validation of selective prediction performance.

  3. The ostensibly impressive error rejection rates achieved by ATS on targeted subsets come at the cost of severe false rejections of correct predictions, calling into question the framework's practical utility. For instance, while Prior-weighted ATS rejects 95.8% of erroneous prior-tracking predictions, it retains only 11.1% of correct prior-tracking predictions. In contrast, Learned Generic UQ achieves a much more balanced operating point, rejecting 63.8% of errors while preserving 52.3% of correct predictions. Without evaluating selection utility under matched operating points, it remains unclear whether ATS's aggressive filtering provides a net benefit over indiscriminate abstention in practical deployments.

  4. The statistical reliability of the empirical conclusions is constrained by the small sample size of the evaluation benchmark. Each of the 20 VLMs is evaluated on 159 questions from the QuantiPhy validation split; while the authors employ 15 random video-level splits to assess partitioning sensitivity, resampling a small fixed pool cannot substitute for an independent, large-scale test set. Compounded by unweighted macro-averaging across models regardless of subset size, the reported rates carry substantial unquantified variance. The authors might validate their claims on a larger benchmark.

Questions for authors

See weaknesses

Reviewer_RfK4OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies answer-level selective prediction for quantitative physical reasoning with vision-language models: given a VLM's numerical prediction, decide whether to trust or reject it. The authors identify two failure modes that consistency-based uncertainty misses, namely stable-but-wrong predictions and prior-tracking, where the model copies the textual prior instead of using visual evidence. They propose ATS, a black-box framework that probes the model under five controlled conditions (including prior-only and counterfactual-prior queries), computes eight behavioral diagnostic scores, aggregates them into a trust score, and applies event-family-specific thresholds. Experiments on QuantiPhy across 20 backbones show ATS rejects more targeted failures, but with reduced retention of correct cases and no aggregate improvement over self-consistency baselines.

Critical strengths
  1. The paper is easy to follow.
  2. The experimental reporting is detailed and rigorous.
Critical weaknesses

W1. High query cost limits practical utility. ATS requires 7 queries per example just to decide whether a single answer should be rejected.

W2. Limited technical contribution. The probing conditions (repeated sampling, removing the video, perturbing the prior, CoT prompting) are simple and largely standard black-box techniques, and the eight diagnostic scores are heuristic combinations of their outputs. The aggregation and thresholding pipeline is also straightforward.

Questions for authors

How should the selector distinguish appropriately incorporating a valid prior from blindly copying it?

Reviewer_9rYhOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates how to tell whether an individual numerical prediction from a vision-language model (VLM) can be trusted on physical reasoning tasks. The authors propose Answer-Level Trust Selection (ATS), a black-box framework that probes the model using repeated sampling, video removal, altered numerical priors, and prompt variations. From these probes, ATS computes eight behavioral scores and combines them to decide whether to accept or reject the model's original answer. The method does not require fine-tuning, access to logits, or an external verifier. The authors evaluate ATS across 20 VLMs on 159 examples from QuantiPhy, measuring the accuracy of retained answers at different coverage levels and analyzing failure modes such as confidently wrong predictions or high sensitivity to priors.

Critical strengths

S1. Practical problem formulation. Aggregate benchmark accuracy does not tell a user whether to trust a specific prediction in deployment. Framing this as a per-instance selective prediction (accept/reject) problem makes a lot of practical sense. Because ATS operates strictly in a black-box manner via API queries without altering the original answer, it can be easily applied to proprietary VLMs.

S2. Transparent and thorough empirical reporting. The evaluation across 20 models is extensive, and I appreciate that the authors include component ablations, weight sensitivity analyses, and confidence intervals. By reporting both the rejection of bad answers and the retention of good ones, the paper gives a clear picture of the coverage-accuracy trade-offs. The authors are also upfront about the regimes where their method brings limited gains.

Critical weaknesses

W1. Limited methodological novelty; ATS feels more like a heuristic scoring pipeline than a new framework. At its core, ATS is an ensemble of familiar probing techniques: self-consistency, modality ablation (removing the video), counterfactual priors, and prompt perturbations. While packaging these into a diagnostic suite for physical reasoning is neat, the paper does not make a strong case for what new technical insights this combination provides over standard consistency-based uncertainty estimation. For a scoring-and-filtering pipeline to stand out as a primary contribution, it needs to either catch failure modes that simpler heuristics miss or offer deeper analytical insights. Right now, the empirical results do not show that this specific combination is doing much heavy lifting.

W2. The empirical gains over simple baselines are marginal and do not justify the extra query cost. Despite evaluating 20 models, the actual improvements over simple self-consistency (SC-only) are tiny. At 50% coverage, ATS beats SC-only on mean MRA by just 0.23 percentage points, and at 70% coverage, it actually performs slightly worse. Looking at Table 15, the confidence intervals for all four mean differences cross zero, which makes it hard to claim that ATS offers a reliable improvement.

Furthermore, the ablation in Table 11 shows that dropping the counterfactual prior stability score actually improves all four average metrics, which calls into question the design of the 8-score combination. Finally, full ATS requires 7 model calls per example compared to 4 for SC-only. The paper lacks an equal-budget comparison (e.g., ATS with 7 calls vs. SC-only with 7 calls) and does not discuss the added latency or token costs. Given the negligible performance gap, it is hard to justify paying nearly double the inference cost in practice.

Questions for authors

1.Beyond engineering a combination of eight behavioral probes, what do you view as the core methodological novelty of ATS compared to standard consistency-based uncertainty estimation? Does the framework reveal any surprising analytical findings about VLM behavior that simpler checks miss?

2.Could you clarify how ATS compares to physical evaluation benchmarks like Physics-IQ in terms of intended use, data requirements, and compute cost? Since Physics-IQ evaluates generated videos while ATS filters numerical VLM outputs, what is the exact scope for a fair comparison between the two?

3.Have you considered using the ATS trust score as a reward signal for RL or preference optimization? Showing that optimizing for ATS scores actually improves downstream physical reasoning accuracy on a held-out test set (rather than just gaming the heuristic) would go a long way toward proving the practical value of these diagnostics.

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

Related papers

Discussion (0)

Sign in to comment.