acceptodds
Under review as a conference paper at ICLR 2027

All the Facts, Still the Wrong Answer: Integration Gaps in Vision-Language Models

Abstract

Multimodal reasoning requires models to combine multiple pieces of evidence into decisions that satisfy all relevant constraints. Evaluations centered on final-answer accuracy leave unclear whether failures arise from reading individual facts or using them together. We study this question through the multi-evidence integration gap, linking local fact recovery to global decision validity on controlled subset-selection and path-finding tasks. Our evaluation combines conditional performance given correct local answers with same-grid local queries, text-oracle inputs, retrieval controls, and constraint ablations. We further examine hidden-state probes and activation replacement across network depth, token positions, and model families. Across the evaluated models, accurate local recovery coexists with substantial global failure. For example, Gemma-3-27B recovers all evaluated isolated item facts correctly but achieves only 19.75% global validity on eight-item visual instances. Probes decode information about solution membership from global-prompt states, while corresponding clean-state replacement recovers failed decisions over model-dependent depth ranges. These findings show that recovering facts and representing solution-related information are insufficient for reliable joint decisions. They motivate evaluations and model improvements that explicitly target how available evidence is combined, alongside advances in perception and model scale.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
vision-language models, multimodal reasoning, multi-evidence integration, local-to-global reasoning, compositional reasoning, mechanistic interpretability, activation patching, representation probing

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_nXzCOpenReview ↗
Rating
2(+1)
Confidence
3(+1)
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper investigates whether vision-language models can combine individually recoverable facts into decisions that satisfy multiple constraints. Using visual subset-selection and weighted-path tasks with exactly verifiable answers, it compares models' ability to identify individual attributes with their ability to produce globally valid solutions. The experiments identify instances where all constituent facts are answered correctly in separate queries, yet the final decision violates the joint constraints.

The study examines this gap through text-based inputs, retrieval and constraint controls, model scaling, linear probes, and hidden-state interventions. It also evaluates an Extract–Solve pipeline that combines model-generated structured transcriptions with a deterministic solver. The results show that successful local fact recovery does not ensure a valid global decision, while separating extraction from constraint solving substantially improves performance for several models.

Critical strengths

1. A clear evaluation of the gap between local recovery and global decisions.

The same-grid evaluation tests global validity on instances where the model correctly answers every constituent local question while viewing the full input. This makes the comparison more informative than evaluating recognition and reasoning on unrelated examples or isolated cards. The results demonstrate a substantial behavioral gap between recovering individual facts and producing an answer that satisfies their joint constraints.

2. Behavioral controls strengthen the diagnosis.

Text-oracle inputs, matched retrieval tasks, constraint ablations, and model scaling examine several possible contributors to the gap. Together, these experiments provide a broader basis for the findings than the local-versus-global comparison alone. They help distinguish difficulty with visual extraction from difficulty making a globally valid decision.

3. Internal analyses complement the behavioral results.

The probes and state-replacement experiments examine whether solution-relevant information is decodable and whether clean states can restore correct decisions. Nested cross-validation separates probe-layer selection from evaluation, strengthening the reliability of the decoding results. These analyses extend the investigation beyond accuracy and provide evidence about how task information is represented and used.

Critical weaknesses

1. The additional insight beyond existing reasoning limitations is unclear.

The paper's main observation is that vision-language models can correctly recover individual facts while failing to combine them into globally valid decisions. However, the distinction between recognizing or retrieving individual information and successfully performing multi-step reasoning is already well established in the literature on compositional reasoning and constraint satisfaction.

The proposed multi-evidence integration gap appears to largely capture this known limitation under a new evaluation formulation. While the conditional local-to-global evaluation provides a more controlled characterization, it remains unclear what fundamentally new insight is established beyond showing that models struggle with increasingly complex reasoning tasks.

The paper would benefit from a clearer explanation of how its findings differ from existing work on compositional reasoning, particularly what new failure mechanism or capability distinction the proposed framework reveals.

2. General reasoning and inference effects are not sufficiently separated from multimodal integration.

A central motivation of the paper is to investigate how VLMs integrate multiple pieces of visual evidence. However, the text-oracle experiments show that substantial performance degradation persists even when all relevant information is provided directly as structured text.

Furthermore, the Extract–Select experiments in Appendix G show that explicitly extracting the visual information before asking the model to perform the selection does not consistently improve performance. In contrast, the Extract–Solve experiments demonstrate that a deterministic solver can achieve substantially higher accuracy using the extracted information.

Together, these results suggest that a major bottleneck lies in the model's ability to perform combinatorial reasoning and constraint satisfaction, rather than in a specifically multimodal evidence-integration process.

The paper does not sufficiently distinguish these general reasoning limitations from failures unique to multimodal representations. A more direct comparison against text-only language models or tasks designed to isolate cross-modal integration would strengthen the multimodal motivation.

3. The mechanistic analysis has limited explanatory power.

Decoding candidate membership from hidden states does not establish that the model has computed a globally valid combination. Similarly, recovery after clean-state replacement may introduce information or computations missing from the corrupted execution. The paper acknowledges these limits, but the experiments consequently leave unresolved where and why integration fails. More targeted interventions would be needed to distinguish partial candidate information from a jointly usable solution representation.

4. Generalizability beyond synthetic constraint tasks remains uncertain.

Subset selection and weighted paths provide exact answers and controlled difficulty, but cover a narrow class of combinatorial decisions. Increasing the candidate count also changes information load, distractor count, and decision complexity together, making their respective contributions difficult to distinguish. Evaluation on natural multimodal reasoning tasks, including evidence distributed across images or partially observed inputs, would help establish whether the proposed gap captures a broader limitation beyond these constructed settings.

Questions for authors
  1. What failure mechanism or capability distinction does the proposed integration gap reveal beyond findings from existing compositional-reasoning and constraint-satisfaction evaluations?

  2. Given the text-oracle, Extract–Select, and Extract–Solve results, what evidence distinguishes a specifically multimodal integration failure from a general combinatorial-reasoning limitation? Could matched text-only model comparisons or tasks isolating cross-modal integration clarify this distinction?

  3. Can targeted probes or interventions distinguish partial candidate-membership information from a jointly usable solution representation, and identify where integration fails?

  4. Does the gap persist on natural multimodal reasoning tasks? Can information load, distractor count, and decision complexity be varied independently to clarify their respective effects?

Reviewer_vwWaOpenReview ↗
Rating
4
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper investigates a failure mode in which a vision-language model can answer questions about the individual facts in an image but cannot combine those facts into a valid decision. Its main synthetic task asks the model to choose a subset of visual cards satisfying numerical, categorical, and compatibility constraints; a second task asks for a path through a weighted graph. The authors compare isolated-card and same-grid fact recovery with global validity, then use text-oracle inputs, matched retrieval, constraint ablations, and different model sizes to diagnose the behavioral gap. They also train hidden-state probes and replace selected activation states from correctly answered clean inputs into incorrectly answered perturbed inputs. At eight cards, Gemma-3-27B recovers all isolated-card facts but achieves only 19.75% visual global validity. For Qwen2.5-VL-3B, global validity is 5.80% even among 362 eight-card instances where every same-grid local question is answered correctly.

Critical strengths
  1. The main behavioral control directly tests the claimed gap. Rather than comparing global performance only with easy, isolated-card reading, the authors ask local questions about the same complete grid used in the global task and condition on every evaluated local answer being correct (Section 3, Figure 2). The gap remains large for models with substantial retained cohorts. Text-oracle inputs further show that visual perception contributes to some errors but is not a complete explanation. These controls support the claim that recoverable constituent facts do not guarantee a valid joint decision.
  2. The diagnostic layers are complementary and carefully bounded. Matched retrieval separates finding identified items from satisfying a conjunction of constraints; the path task provides a second form of joint decision. The probes use nested cross-validation and distinguish decoding solution membership from predicting answer correctness. In Qwen2.5-VL-3B, matched clean-state replacement recovers many selected failures at early layers while mismatched donor states do not. The paper reports selected-pair counts and notes that decodable information is not proof that the model has already solved the task. This restraint increases confidence in the observations.
Critical weaknesses
  1. The quantitative gap is tied to a particular constructed decision problem. Every selection instance is generated to have exactly one valid subset, the two-card condition asks for both available cards, and increasing candidate count changes both evidence load and combinatorial difficulty. The constraint ablations also enlarge the acceptance set; the authors disclose this, but it prevents a simple interpretation of improved scores as an isolated reduction in reasoning demand. The second path task helps, yet both tasks use synthetic visual records and exact answers. The result is compelling within these constructions but does not establish the same magnitude of failure on natural multimodal tasks with several acceptable decisions.
  2. The activation experiments establish a donor effect, not a localized failure mechanism. Pairs are selected because the model answers the clean instance correctly and the perturbed instance incorrectly. Replacing all or some prompt-position states from the clean run can import correct evidence or partial answer information; strong recovery from an early layer therefore does not identify which integration step failed in the unpatched computation. The matched-versus-mismatched contrast is valuable, and the authors avoid claiming a universal causal layer, but an additional analysis of corrupted states would better connect the probe finding to the failure being repaired.
Questions for authors
  1. On the clean/corrupt intervention pairs, is correct solution membership still decodable from the corrupted, incorrectly answered states at the same layers? A paired comparison with clean states would more directly test whether the failed run retains solution-related information or loses it before generation.
  2. How sensitive is the same-grid conditional gap to allowing multiple valid subsets while keeping candidate count and constraint types fixed? Even a small controlled variant would indicate how much the unique-answer construction contributes to the measured gap.
Reviewer_zcbGOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper finds that VLMs can answer multiple sub-questions but still fail to answer the combined question. For example, a VLM can recognize travel items and the price of each item, but if asked to pick three items for a trip under a $10 budget, it can fail.

The author calls this observation the multi-evidence integration gap: the VLM can correctly answer the sub-question but fails on a complex combined question.

To study this, they design a new benchmark of combined complex questions whose answers must satisfy selection size, cost, and weight budgets; a color requirement; type-count limits; and mutual compatibility.

They find that when asked individual sub-questions, such as size, cost, etc., the VLM can answer with high accuracy (100%), but when asked the combined question considering all 8 requirements, the accuracy is very low (5.75%).

They conduct internal probing and activation replacement experiments and find that, even though the model can recognize attribute information about objects, it can’t use it effectively to generate a correct answer.

How to integrate multiple pieces of evidence to answer complex questions remains an open question.

Critical strengths
  1. The observation of the multi-evidence integration gap is clear: the VLM can correctly answer the sub-question but fails on a complex combined question.

  2. They propose a new benchmark to verify this multi-evidence integration gap and observe that across several VLMs, this phenomenon exists.

  3. They design probing experiments to verify that the reason lies not in recognizing multiple constraints, but in their integration.

Critical weaknesses
  1. The task of fulfilling multiple pieces of evidence is an "and" relation and highly structured; it cannot cover general logical relations. How about "or" and "not" Boolean operations?

  2. VLMs without CoT learning might lack evidence integration; does the phenomenon only exist in VLMs without a CoT process?

  3. The paper mainly reveals observations about the multi-evidence integration gap without proposing a possible solution; will CoT and RL learning solve this problem?

Questions for authors

My concerns are mainly mentioned in the weaknesses:

  1. Why use only "AND" in evidence integration? How about "OR" and "NOT" Boolean operations?

  2. How about the performance of VLMs with CoT on the benchmark?

  3. Will RL learning, such as GRPO, solve this problem? Any other possible solutions?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.