acceptodds
Under review as a conference paper at ICLR 2027

Readout Fidelity Bottlenecks Self-Verification in Unified Multimodal Models

Abstract

Unified multimodal models combine image generation and visual understanding, allowing them to evaluate and select among their own generations without an external verifier. Yet self-verified Best-of-N recovers only a small fraction of the quality available in candidate pools. We identify readout fidelity as an independent bottleneck: useful candidate preferences can already exist in verifier evidence but are compressed by discrete Yes/No decoding. Across seven models, discrete readout produces extensive score collisions and can even underperform random selection despite substantial oracle headroom. We model discrete verification as finite observations of an underlying Yes/No preference, with a formulation agnostic to how the verifier evidence is produced. In the direct-answer setting, repeated stochastic observations converge to the pre-decoding logit margin, while full stochastic decoding closely tracks the predicted finite-observation behavior. We use Continuous Best-of-N (C-BoN) as a controlled intervention that directly exposes this limiting preference while keeping the remaining selection pipeline fixed. Across seven models and four benchmarks, C-BoN improves all 28 settings without additional models, training, generation, or verifier calls, and enables stronger Best-of-N scaling. Larger continuous margins also yield more reliable ordering within discrete ties, while recovered ordering tracks downstream gains. These results establish readout fidelity as an information bottleneck in self-verification and show that verifier evidence can contain richer preferences than discrete verdicts reveal.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Unified Multimodal Models, Self-Verification, Test-Time Scaling

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_HBexOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates self-verified Best-of-N selection in unified multimodal models, where a model evaluates and selects among its own generated images using its visual understanding capabilities. The authors argue that the limited effectiveness of self-verification does not necessarily indicate insufficient verification capability, but may instead arise from information loss when continuous preferences are converted into discrete Yes/No judgments. They refer to this as the readout fidelity bottleneck.

As a controlled intervention, the authors introduce Continuous Best-of-N (C-BoN), which directly uses pre-decoding Yes/No logit margins instead of decoded discrete verdicts, while keeping the candidate pools, verification questions, score aggregation, and selection procedure unchanged. They further formulate a finite-observation model showing that repeated stochastic observations converge to the underlying marginal Yes/No preference, which corresponds to the pre-decoding logit margin in the direct-answer setting. Experiments across seven unified multimodal models and four text-to-image benchmarks show improvements over discrete verification in all 28 model–benchmark settings. Additional analyses examine score collisions, tie resolution, ranking reliability, and Best-of-N scaling.

Critical strengths

1. Well-controlled and extensive empirical evaluation. The paper presents a carefully controlled comparison between discrete and continuous readouts, keeping the candidate pools, verification questions, score aggregation, and selection procedure fixed. Experiments across seven models and four benchmarks demonstrate consistent improvements, providing convincing evidence that the readout mechanism itself can substantially affect self-verified selection. The breadth and consistency of these results are the main reason I have no major concerns about the soundness of the empirical claims.

2. Analysis beyond aggregate performance improvements. The authors provide several complementary analyses, including tie resolution, margin quantization, pairwise ranking reliability, and Best-of-N scaling. In particular, Table 4 shows that resolving discrete score ties accounts for 64–81% of the overall improvement, while Tables 5 and 7 show that larger continuous margin gaps correspond to more reliable candidate ranking across all four benchmarks. These findings provide useful evidence for the proposed bottleneck rather than relying solely on aggregate benchmark improvements.

Critical weaknesses

Limited conceptual and theoretical novelty beyond existing confidence-based verification. Although the systematic investigation of readout fidelity is valuable, the central insight that discretizing continuous confidence scores discards potentially useful ranking information is relatively intuitive, particularly in light of existing probability- or logit-based verification methods such as VQAScore and dMLLM-TTS. Notably, VQAScore also serves as the official GenAI-Bench evaluator in this paper (Appendix A.11), which indicates that continuous Yes-probability readout is already standard practice in this area. The authors explicitly acknowledge that a continuous Yes score is not new, and position their contribution around identifying and characterizing the readout bottleneck.

However, I remain uncertain whether the proposed conceptual framing establishes sufficiently new insights beyond the expected consequences of discretizing continuous preferences. The theoretical formulation in Section 3.3 is also relatively straightforward. Equations (5)–(7) rely on standard binomial/multinomial sampling and the strong law of large numbers, while Equation (8) expresses the limiting preference directly through the model logits. Although mathematically sound, these results do not appear to provide substantial additional theoretical understanding beyond formalizing the relationship between repeated discrete observations and continuous confidence scores. Moreover, the empirical validation of this model (Section 4.3, Table 2) is close to agreement by construction, at least for autoregressive verifiers. The predicted condition samples from the Yes/No/other distribution given by the verifier logits, while the observed condition decodes at temperature 1.0 from the same model, and Appendix A.5 itself states that, because the prompt requests an immediate Yes/No answer, the verdict distribution matches the softmax used in Section 3.3. The close agreement is therefore expected and does not constitute an independent test of the model.

Consequently, I find the conceptual and theoretical contributions to be incremental relative to existing confidence-based verification approaches. This is the primary reason for my recommendation of weak rejection.

Minor Presentation Issues. There are several minor presentation inconsistencies, including citation formatting where narrative-style citations appear without appropriate punctuation or parentheses, redundant cross-reference wording such as “Equations equation 5 and equation 6,” and inconsistent hyphenation of terms such as “predecoding” and “pre-decoding.” These issues are straightforward to correct and do not materially affect my overall assessment.

Questions for authors

1. Novelty and significance (could affect my rating). Compared with existing probability- or logit-based verification methods, what specific new or non-obvious insights does the proposed readout fidelity analysis establish? In particular, beyond the agreement reported in Table 2, which is largely expected by construction for autoregressive verifiers, does the finite-observation framework provide additional testable predictions or practical guidance? A prediction that does not follow directly from standard probabilistic sampling and is confirmed experimentally would substantially strengthen the contribution and could change my assessment.

2. Continuous scoring alternatives (clarification). Appendix A.3 shows that positive-only log odds slightly outperforms C-BoN on GenEval (78.51 vs. 78.20). Given that negation pairing is designed to cancel a uniform Yes bias (Appendix A.4), could the authors clarify whether it remains necessary or beneficial under continuous readout? Is the relative performance of these scoring formulations consistent across the other three benchmarks?

3. Masked diffusion verifiers (could affect my rating). For LaViDa-O and LLaDA-o, how are the pre-decoding Yes/No logits obtained, for example, at which answer position and denoising step? For these models, is the agreement between predicted and observed recovery in Table 2 non-trivial, and how do the authors explain the larger deviations for some models (e.g., LaViDa-O at m = 2: 82.17 vs. 84.63)? A clear answer could show that the finite-observation model captures behavior beyond the autoregressive case.

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

This paper investigates readout fidelity as a bottleneck in self-verified Best-of-N image generation for unified multimodal models. The authors show that discrete Yes/No judgments can discard useful candidate-ranking information through score collisions. They propose Continuous Best-of-N (C-BoN), which replaces discrete verdicts with pre-decoding Yes/No logit margins without additional training or verifier calls. A finite-observation framework is introduced to explain the relationship between stochastic verification and continuous preferences. Experiments across seven models and four benchmarks demonstrate consistent improvements in candidate selection and Best-of-N scaling.

Critical strengths
  1. Clear problem identification. The paper identifies readout fidelity as an important but underexplored bottleneck in self-verification. The controlled comparison between discrete and continuous readouts provides a convincing explanation for why existing Best-of-N methods may fail to exploit available candidate quality.
  2. Extensive experimental validation. The evaluation covers seven unified multimodal models and four benchmarks, with improvements across all 28 model-benchmark combinations. The consistent results across different model architectures support the generality of the observation.
  3. Detailed theoretical and empirical analysis. The finite-observation formulation, tie-breaking analysis, ranking reliability experiments, and Best-of-N scaling results provide useful insights into the underlying mechanism. The method also requires no additional training or verifier calls, making it practically attractive.
Critical weaknesses
  1. Limited methodological novelty. The core technique of replacing discrete Yes/No predictions with continuous logit-based confidence scores is relatively straightforward and closely related to existing approaches such as VQAScore and dMLLM-TTS. Appendix A.3 further shows that simpler continuous scoring variants achieve comparable or slightly better performance, making the unique methodological contribution of C-BoN less compelling.
  2. Potentially insufficient discrete baselines. The main discrete baseline relies on only three similar verification questions and deterministic tie-breaking, which naturally leads to frequent score collisions. Although the authors investigate question counts and tie-breaking strategies, stronger discrete verification baselines with more diverse questions or finer-grained judgments are not sufficiently explored.
  3. Limited validation in more general verification settings. Most experiments focus on direct Yes/No verification, while the evaluation of reasoning-based verification is substantially narrower. Moreover, several benchmark improvements are relatively small, and seven of the 28 comparisons are not statistically significant after multiple-testing correction. These limitations weaken the broader generalization claims.
Questions for authors
  1. Novelty relative to existing methods: What is the fundamental methodological distinction between C-BoN and existing confidence-based verification approaches such as VQAScore and dMLLM-TTS? Given that positive-only log odds outperform C-BoN on average in Appendix A.3, what is the specific advantage of the proposed formulation?
  2. Robustness against stronger baselines: How does C-BoN compare with discrete verification using more diverse questions or a substantially larger number of verification judgments? Would the performance advantage remain significant when the discrete baseline produces fewer score collisions?
  3. Generalization beyond direct verification: Can the authors provide more comprehensive evaluations of continuous readout in reasoning-based verification, particularly across different model architectures? The CoT-V experiment in Appendix A.6 is encouraging, but broader evidence would help establish whether the same bottleneck persists in more complex verification pipelines.
Reviewer_TewxOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies the loss of ranking information when a unified multimodal model verifies its generated images through discrete Yes/No answers. Continuous Best-of-N reads the Yes-versus-No log-probability ratio before answer decoding, while keeping the candidate images, verification questions, and aggregation rule fixed. The paper compares the two readouts across seven models and four benchmarks, and analyzes discrete ties, ranking reliability, and selection performance.

Critical strengths

I found the analysis informative. The paper asks a concrete question and designs a reasonably clean intervention to answer it. Keeping the candidates and verification pipeline fixed makes it possible to examine the readout itself. The tie-resolution experiments are particularly useful: continuous scores recover distinctions among candidates with the same discrete score, and using them beyond tie-breaking provides further gains. This helps explain why self-verification can underperform even when the model has useful evidence.

The experimental work is also fairly complete. Gains are positive in all 28 model-benchmark combinations, with 21 remaining significant after Holm correction for multiple comparisons. The matched external-verifier comparison, alternative continuous scores, and chain-of-thought verification provide useful checks. I appreciate that the paper reports simpler continuous scores that can slightly outperform the chosen formulation, rather than presenting its particular score as uniquely effective.

Critical weaknesses
  1. The novelty is mainly in the controlled diagnosis. Continuous Yes-confidence scoring already appears in VQAScore and dMLLM-TTS, as the paper acknowledges. I think the empirical analysis adds value, but the theoretical interpretation should remain precise. Equation 7 concerns a log-count-ratio estimator of repeated judgments. Repeated averaging of the original discrete votes instead converges to a probability difference. These are different estimators, so the former should not be read as establishing that ordinary vote averaging recovers the proposed continuous score, or that this score is generally optimal for ranking image quality.

  2. Some implementation and evaluation details need clarification. Table 10 does not clearly identify the model and checkpoint used for the chain-of-thought comparison. The Yes/No token sets and the position at which probabilities are read should be documented for each understanding pathway, especially the masked-diffusion models. Please also give the exact aggregation formula for GenEval overall and clarify whether it averages task scores equally or averages correctness over prompts. These details matter for reproducing and interpreting the results.

I lean toward acceptance. The scoring operation itself is familiar, but the paper gives a useful account of where the discrete selection pipeline loses information. The controlled experiments are the main reason for my positive assessment.

Questions for authors
  1. Can the authors explicitly distinguish the repeated log-count-ratio estimator from averaging discrete votes, and delimit what the finite-observation result establishes about selection?
  2. Which model and checkpoint produced Table 10? Please specify the Yes/No token sets, readout positions, and any denoising-step choices for the different understanding pathways.
  3. How is GenEval overall aggregated? If prompt-level and task-level averages differ, please report both.
Reviewer_7HaKOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
3(+1)
Read review
Paper summary
  • studies self-verified best-of-N in unified multimodal models, where the model uses its understanding pathway to score multiple candidates produced by its generation pathway before selecting the best one for final output
  • they extend a discrete yes/no self-verification method (eg, "does the image match the prompt?"), having the self-verifiers produce scores via pre-decoding continuous yes/no logit margins instead of decoded binary yes/no output tokens
    • they claim that the discrete method suffers from a "readout fidelity bottleneck" --- ie, the verifier is properly able to distinguish between the N candidates but the fidelity of discrete scoring is too coarse and causes ties which forces an arbitratily choice between the ties
    • the continuous verifier scoring method effectively prevents the score ties from occurring
  • the continuous C-BoN self-verification method improves over the discrete method in their tested settings and shows better scaling with increasing candidate count N than the discrete method
Critical strengths
  1. the method shows consistent improvement over the discrete baseline in all tested settings, including a human study
  2. isolating the "readout fidelity" as distinct from the verifier quality is a useful framing
Critical weaknesses
  1. the core contributions / insights of the paper are both relatively minor and also not new
    • continuous scoring (P(yes)) for BoN self-verification has been previously studied (eg, they cite dMLLM-TTS $\S$2)
    • the gains appear to come from using any continuous score, not from C-BoN's specific form (§\SA.3, Tab. 9). in fact, the C-BoN method is not even the best continuous score in their tested settings. a simple method like the VQA Score should be compared against in the main paper
    • since the method is not novel, the value of the paper depends on the diagnosis + analysis
      • it seems like the contribution here is relatively minor, largely pointing out well-known downsides of discretization
  2. the $\S$3.3 "finite-observation model" is presented as a contribution, but it contains unnecessary formalism and adds little beyond the empirical results.
    • "repeated stochastic observations converge to the pre-decoding logit margin" is simply the law of large numbers applied to counting yes/no observations
    • Tab. 2 simply confirms that sampling from the logits matches sampling by decoding (modulo some sampling noise), which holds by design
Questions for authors
  1. can you add comparisons of VQA score P(yes) as a top-level baseline for all models and benchmarks?
  2. are you able to provide justification for why C-BoN should be preferred over other continuous scores? and why?
  3. what does $\S$3.3's "finite-observation model" predict beyond what is already implied by the law of large numbers already implies? is there any outcome in Tab. 2 would have falsified it?

(suggestion) the differences are not visible in the teaser figure radar plot, consider a table or bar chart of deltas instead

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.