acceptodds
Under review as a conference paper at ICLR 2027

Do VLMs Fail at Causal Reasoning or Output Format Production?

Abstract

Evaluations of vision-language model (VLM) causal reasoning typically rely on a single output format, making it difficult to distinguish failures of format production from failures of reasoning. We introduce a diagnostic framework that combines structure-first generation, multi-format testing across four structural representations, three levels of prompt scaffolding, and score-compliance pairing, where content quality is reported alongside format compliance. Across six VLMs spanning 2023–2025 architectures, we find that the structural expression of causal reasoning is highly sensitive to format and prompt support. The same model reaches 97% compliance for numbered sequences but below 1% for arrows, while removing scaffolding yields compliance drops of 25–91 points. For five of the six models, structural and textual quality converge within the same response once the format barrier is removed. Targeted fine-tuning further helps distinguish limitations arising from insufficient training signal from those that persist under the tested architecture.

Primary area:
datasets and benchmarks
Keywords:
evaluation methodology, vision-language models, causal reasoning, structured generation, format robustness, scaffolding dependence, benchmark design, plausibility-faithfulness gap, diagnostic framework

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_53NjOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes an evaluation framework for vision-language models (VLMs) to determine whether their failures on structured causal-reasoning outputs stem from a lack of causal reasoning or from difficulty producing the requested output format. The framework compares structured outputs against free-form text, tests four structural formats, and varies the level of prompt scaffolding. It also applies targeted fine-tuning to examine whether the identified failures reflect data limitations or architectural limits. Six VLMs are evaluated, with outputs scored by two LLM judges and validated against human-annotated response

Critical strengths
  • Clear presentation and well-motivated problem. The paper is well organized and easy to follow. It targets an important and under-examined problem: whether failures in structured causal-reasoning evaluations reflect a lack of reasoning or a failure to produce the requested format. The proposed framework is conceptually simple, yet it covers the main confounding dimensions (output format and prompt scaffolding) systematically.

  • Concrete and well-designed experimental setup. The experiments are specified in detail: four formats, three scaffolding levels, two LLM judges with human validation, and a rule-based failure taxonomy. The choice of models is clearly justified, spanning five training approaches across a 2023–2025 capability range. The selection of LLaVA-CoT and MiniGPT-4 for fine-tuning is likewise explained in terms of architectural contrast on a shared blind spot. Overall, the setup is clean and easy to understand.

Critical weaknesses
  • Validation of the framework. The human study validates the LLM judges, not the framework’s ability to separate format production from causal competence. Existing benchmarks (CELLO, MM-CoT, CausalVLBench) are discussed only qualitatively, with no empirical comparison. The novelty of score-compliance pairing over Long et al. (2025), who already disentangle format compliance from task accuracy, should also be clarified.

  • Limited model range. Only open-source 2023–2025 models are evaluated. Layers 1–3 require only inference, so applying them to recent API-based models (GPT, Gemini, Claude) seems feasible. This would test whether the reported format sensitivity persists in stronger models, which the authors themselves leave open.

Questions for authors
  • Have the authors evaluated any of the six models on existing causal-reasoning benchmarks (e.g., CELLO, CausalVLBench)? Do the model rankings from the proposed framework agree with them?
  • How does score-compliance pairing differ methodologically from the format-compliance/accuracy disentanglement proposed by Long et al. (2025)?
  • The decoding configuration (greedy vs. sampling, temperature, top-p) is not reported for either the evaluated VLMs or the LLM judges. Since format compliance is the central measure, did the authors test whether the results, especially Cross-Format Substitution rates, are sensitive to the decoding strategy?
2 comments

We thank the reviewer for the careful reading. We address the questions on validation, model coverage, and reproducibility.

W1. We conducted two new analyses to validate the framework's separation of format from content quality, beyond the judge reliability established by the human study (r=0.84 for structure).

First, the failure taxonomy provides diagnostic classification independent of the LLM judges. At C3, Cross-Format Substitution accounts for 95.67% of MiniGPT-4's arrow responses (Table 3). These are not reasoning failures. The model generates valid causal content but in numbered-list notation instead of arrows. A compliance score alone classifies all of them as failures without distinguishing incorrect causal relationships from structures expressed in another format. The taxonomy recovers this distinction.

We also ran a new compliant-only analysis, restricting L3 C3 structural scores to format-compliant responses:

ModelAdj. (compliant-only)ArrowSemicolonNumbered
LLaVA-NeXT1.996.755.996.83
LLaVA-CoT4.727.215.447.40
MiniGPT-42.356.644.715.75
Qwen3-VL6.007.867.798.19

The results show that even among compliant responses, adjacency content quality remains well below linearized content quality for every model. The gap is in the causal content, not in judge agreement or compliance rates.

On CELLO, MM-CoT, and CausalVLBench, we have not run direct comparisons. These benchmarks evaluate causal reasoning using a single structural representation per task. Rankings from such benchmarks may reflect format-following ability alongside causal-content quality. The MiniGPT-4 example above illustrates why. A comparison showing whether models that rank differently on CELLO also show different compliance-content profiles in our framework would be informative. We will discuss this in the camera-ready version.

W2. We agree that API-based models are feasible for Layers 1-3 and would strengthen the generalizability claim. Qwen3-VL provides an initial test: at C3, its compliance reaches 97.18% for arrows, 97.77% for semicolons, and 97.77% for numbered lists, but falls to 44.24% for adjacency lists, with Cross-Format Substitution accounting for 53.02% of adjacency responses (Table 3). Improved linearized adherence does not eliminate difficulty with parent-child representations. We do not extend these rates to untested frontier models. We will note API model evaluation as future work in the camera-ready version.

Q1. Addressed under W1. We have not run the six models on CELLO or CausalVLBench and will discuss the relationship in the camera-ready version.

Q2. Long et al. (2025) study the separation between task accuracy and format compliance for LLM classification tasks (multiple-choice, true/false, extraction) with binary compliance measures. We build on this distinction through experiments on VLM causal-response generation. Our framework adds structure-first generation, continuous 0-10 quality scoring for both structure and text, a five-category failure taxonomy, and fine-tuning experiments. The taxonomy captures phenomena specific to generative output. Cross-Format Substitution has no analog in classification settings where the output is a label. (We note that the first author's family name is Do. We will correct the citation in the camera-ready version.)

Q3. Inference configurations are as follows. LLaVA-RLHF uses temperature=0.2, top-p=0.6. MiniGPT-4 uses temperature=0.1, top-p=0.9. Qwen3-VL uses checkpoint defaults of temperature=0.7, top-p=0.8, top-k=20. LLaVA-NeXT, LLaVA-CoT, and CogVLM2 use generation defaults without explicit sampling overrides. All models use max_new_tokens=512. We did not standardize decoding settings across models, and we have not tested alternative configurations. Cross-Format Substitution rates may depend partly on decoding strategy, and we will note this as a limitation alongside the full configuration table in the camera-ready version.

Thank you for the authors' response. The response addressed my concerns and I will maintain my rating.

Reviewer_nP7cOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper asks whether vision-language models that fail to produce causal structures lack causal reasoning or only fail to produce the requested output format. It uses 1,000 MS-COCO validation images with nine GPT-4o-generated questions each, covering the associational, interventional and counterfactual levels of Pearl's hierarchy. Six VLMs released between 2023 and 2025 are asked to write a causal chain first and then a textual explanation. The chain is requested in four formats (arrows, semicolons, numbered lists and adjacency lists) under three levels of prompt scaffolding, ranging from full instructions with worked examples to a minimal task description. Two LLM judges score content quality independently of format. A rule-based parser measures format compliance and sorts failures into five categories, and the paper reports scores and compliance together. Compliance depends strongly on the format, ranging from 97% for numbered lists to below 1% for arrows within the same model. Removing scaffolding lowers average compliance for every format except numbered lists. When compliance is high, structural and textual scores are close for five of the six models, and adjacency lists remain the hardest format, even for Qwen3-VL. Finally, LLaVA-CoT and MiniGPT-4 are fine-tuned on GPT-4o-generated data, first on a single format and scaffolding level and then on all twelve combinations. The authors conclude that both fragilities reflect a lack of training data for LLaVA-CoT, whereas MiniGPT-4's format capacity is limited by its architecture.

Critical strengths

S1. A large, systematic and transparent evaluation. The design crosses six models, four formats and three scaffolding levels, giving 54,000 Text-Only and 648,000 Structure-Text responses. The appendix reports per-judge results, the complete prompt templates and human scores. The limitations section is candid about the prompt asymmetry for adjacency lists, the missing non-causal control, and the partial evaluation after multi-format training.

S2. Fine-tuning as a diagnostic intervention. Using fine-tuning to ask whether a fragility can be removed by training signal is a good idea. The single-format stage produces a clear dissociation between format robustness and scaffolding robustness. Score–compliance pairing also shows that LLaVA-CoT keeps its content quality when asked for arrows, even though it writes them in semicolon notation.

Critical weaknesses

W1. The main findings follow largely from the scoring design and concern format following rather than causal reasoning. The judges score content "regardless of format compliance" (line 199). Similar content scores across formats with very different compliance are therefore built into the protocol, for example MiniGPT-4's 5.48 for arrows at 0.58% compliance (line 311). The prompt also asks the text to explain "the reasoning based on the chain" (line 983), so near-zero CDD at high compliance is expected and does not test faithfulness. Without a non-causal control (line 1685), the format effects cannot be attributed to causal reasoning, and format sensitivity is already well documented for LLMs (Long et al., 2025; Tam et al., 2024). The headline gaps come from older models such as MiniGPT-4, while Qwen3-VL, the only 2025 model, keeps above 97% compliance on all linearized formats even at minimal scaffolding.

W2. The fine-tuning conclusions about architecture are confounded. MiniGPT-4 trains only its linear projection with the LLM frozen, whereas LLaVA-CoT trains LoRA on the LLM decoder and the vision encoder (Table 6, lines 1242–1244). MiniGPT-4's persistent adjacency-list limit may therefore reflect what is trainable rather than its architecture, contrary to line 1247, and the comparison rests on two models with one run each. After multi-format training, scaffolding robustness was evaluated only for semicolons (line 1671). The L1 text gain, presented as a "general reasoning benefit" of structural training (line 1611), may instead come from the text-only L1 answers from GPT-4o that are part of the training data (line 1187).

Questions for authors
  1. Controls for the central claim. Can you add a matched non-causal control in the same four formats, and report content scores for format-compliant outputs only?

  2. Judge and human agreement. How well do the two judges agree on individual responses, how many human annotators were involved, and how well did they agree with each other?

1 comment

We thank the reviewer for the careful engagement and provide additional analyses addressing the questions.

W1. We ran a compliant-only analysis on L3 C3 structural scores, restricted to format-compliant responses only. For adjacency lists, compliant-only scores fall well below any linearized format across every model:

ModelAdj. (compliant-only)ArrowSemicolonNumbered
LLaVA-NeXT1.996.755.996.83
LLaVA-CoT4.727.215.447.40
Qwen3-VL6.007.867.798.19

Models that successfully output the adjacency format still generate lower-quality causal content in it. For linearized formats, compliant-only scores narrow but are not identical (e.g., LLaVA-CoT semicolon 5.44 vs arrow 7.21). Overall linearized score gaps are driven primarily by compliance failure, not content quality differences. Structural complexity creates a genuine content quality boundary even when format compliance is achieved. We will add the full compliant-only table in the camera-ready version.

On CDD, we do not interpret low Causal Depth Disparity as evidence of faithful internal reasoning. The convergence finding rests on the compliant-only analysis, not on CDD in isolation. We will clarify this in Section 5.

On the non-causal control, non-causal conditions would determine whether the observed failures are specific to causal tasks. Our evaluation addresses a different question: whether format failure, as opposed to reasoning deficit, drives low structural scores in causal VLM benchmarks. A causal benchmark requiring a single structural representation may penalize format failures even when the response contains meaningful causal relationships. Cross-format and cross-scaffolding variation are internal contrasts within a single domain. Matched non-causal controls would help determine whether these format-dependent failures extend beyond causal tasks, a question we will identify as a direction for future work.

On prior format sensitivity work, the contribution is not format sensitivity per se. The structure-first protocol requires the model to record causal relationships in a specified structure before any text explanation. Score-compliance pairing measures content quality and format compliance as two separate dimensions on a single generative response. The failure taxonomy captures five distinct types, including Cross-Format Substitution, where a model outputs valid structure in the wrong format. Long et al. (2025) and Tam et al. (2024) address LLM classification with binary compliance. Our setting is generative, quality scores are continuous, and Cross-Format Substitution has no analog in binary frameworks.

On model range, Qwen3-VL does not show the 97%-to-<1% compliance collapse on linearized formats. We will make this explicit in Section 5. The adjacency finding persists: Qwen3-VL compliant-only adjacency scores (6.00) remain well below its linearized scores (7.79-8.19), and adjacency compliance drops from 97.9% to 44.2%.

W2. MiniGPT-4 trains only its projection layer, while LLaVA-CoT trains LoRA across the decoder and vision encoder. We will add this distinction to Section 4.2 and note that separating trainable parameter space from architectural capacity requires ablation beyond this study. The within-model findings are independent. MiniGPT-4 gains format robustness after SF training but loses scaffolding robustness. LLaVA-CoT shows the reverse. These dissociations are within-model contrasts under identical data and compute.

On L1 text quality, the reviewer correctly notes that L1 training data includes GPT-4o text answers (line 1187). The +0.63 MiniGPT-4 text improvement may reflect that exposure. We will revise the claim at line 1611.

Q1. The compliant-only analysis above addresses this. The table will appear in the camera-ready version.

Q2. Inter-judge Pearson correlation: 0.61 (structure), 0.70 (text). After excluding response pairs where either structural score is zero, the structural correlation is 0.60 across 71 conditions, confirming the agreement is not inflated by shared zero-score assignments. Human-judge correlation: 0.84 (structure), based on expert annotation of 6,480 responses. Two annotators evaluated disjoint subsets, and human-human agreement is not directly estimable from this design. We will clarify the procedure and report all correlations.

Inference settings range from near-greedy (temperature 0.1-0.2 for LLaVA-RLHF and MiniGPT-4) to moderate sampling (Qwen3-VL checkpoint defaults at temperature 0.7). LLaVA-NeXT, LLaVA-CoT, and CogVLM2 use generation defaults. Decoding was not standardized across models. Cross-Format Substitution rates may depend partly on decoding strategy. We will note this alongside the full configuration table in the camera-ready appendix.

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

This paper investigates whether failures in VLM causal-reasoning evaluations reflect weak reasoning or difficulty producing the required output format. It introduces a diagnostic framework that compares free-text answers with structure-first responses across four structural formats and three levels of prompt support, reporting content quality alongside format compliance. Experiments on six VLMs show that models can generate comparable causal content across formats while differing sharply in format compliance, with adjacency lists particularly sensitive to reduced guidance. Targeted fine-tuning demonstrates that training across diverse formats and prompting conditions can improve robustness, highlighting the need to distinguish reasoning errors from format-production failures.

Critical strengths
  1. The paper shows that low causal-reasoning scores can reflect difficulty expressing an answer in the requested format. Large compliance differences on identical questions demonstrate why capability judgments based on a single representation can be misleading.

  2. Reporting content quality alongside format compliance distinguishes incorrect causal relationships from omitted structures or substitution of another notation. Comparing structure and explanation within the same response further helps identify what the model successfully expresses despite failing the format requirement.

  3. The experiments reveal that scaffolding dependence varies substantially across representations: numbered sequences remain robust, while adjacency lists often require extensive guidance. Non-monotonic effects also show that adding examples does not uniformly improve performance, motivating evaluation across multiple prompting conditions.

  4. Single-format fine-tuning can preserve content quality while damaging adherence to other formats; diverse-format training largely resolves this problem for LLaVA-CoT. This provides concrete evidence that some apparent capability limitations arise from training exposure and suggests treating format robustness and scaffolding robustness as distinct training objectives.

Critical weaknesses
  1. Questions and reference structures are generated by GPT-4o from everyday photographs, and some hypothetical outcomes may be answered through common knowledge without examining the image. Without systematic image-removal controls or scenarios with independently verified causal relationships, high scores may reflect plausible causal storytelling rather than grounded intervention or counterfactual reasoning.

  2. “Format compliance” includes image grounding, so it is not purely a measure of syntactic adherence. Adjacency lists also introduce additional decomposition and ordering requirements beyond changing delimiters. Without matched non-causal generation and format-conversion controls, the experiments cannot determine whether their difficulty reflects causal reasoning, general structured generation, or instruction following.

  3. MiniGPT-4 updates only its projection layer, whereas LLaVA-CoT adapts substantially more components, including the language decoder. Multi-format training also reduces exposure to each individual condition. Persistent failures therefore cannot confidently establish an architectural ceiling: restricted adaptation, training dosage, and optimization remain competing explanations.

  4. The evaluation primarily covers a 2023–2025 capability range, and its strongest model already achieves over 97% compliance across the three linearized formats under minimal guidance. This suggests that much of the headline effect may diminish with improved instruction following. Testing larger, newer models is necessary to establish whether the framework reveals an ongoing causal-reasoning bottleneck or mainly characterizes formatting weaknesses in earlier systems.

Questions for authors
  1. How do you distinguish visually grounded causal reasoning from commonsense storytelling or general structured generation without image-removal and matched non-causal controls?

  2. Why attribute persistent fine-tuning failures to architectural limitations when the models have different trainable components and per-format training exposure is not controlled?

1 comment

We thank the reviewer and provide additional analyses addressing the concerns.

W1. The dataset contains 9,000 questions from 1,000 MS-COCO images. Manual inspection of 200 randomly sampled questions confirmed that over 95% were answerable from the images and assigned to the intended causal levels (Section 3). All comparisons use the same image-question pairs across output formats. Differences in compliance across formats reflect format-specific difficulty, not changes in question content or image difficulty.

We agree that image-removal controls would provide stronger evidence of visual dependence. The paper's primary analysis concerns format sensitivity under fixed inputs. Even when a response draws on commonsense knowledge, a failure to express causal relationships in the requested format can distort the interpretation of a single-format evaluation. The framework's value holds regardless of how much content originates from images versus language priors. We will state this boundary more clearly in the camera-ready version.

W2. The reviewer identifies three possible sources of adjacency difficulty: causal reasoning, general structured generation, and instruction following. Our current design does not distinguish among these explanations. The paper's claim is narrower: format production failures contribute to low structural scores, and different formats expose distinct failure patterns. We ran a compliant-only analysis, restricting L3 C3 structural scores to format-compliant responses:

ModelAdj. (compliant-only)ArrowSemicolonNumbered
LLaVA-NeXT1.996.755.996.83
LLaVA-CoT4.727.215.447.40
MiniGPT-42.356.644.715.75
Qwen3-VL6.007.867.798.19

Even among compliant responses, adjacency content quality remains well below linearized content quality for every model. Compliance failure alone does not fully explain the difficulty. Adjacency lists require explicit parent-child mappings and terminal node specifications (Section 3.2), imposing structural decomposition demands beyond changing delimiters. Matched non-causal controls would help determine whether these demands are specific to causal tasks or reflect broader difficulties with structured generation. We will identify this comparison as a priority for future work in the camera-ready version.

The reviewer correctly observes that our compliance criterion includes image grounding. The five-category taxonomy addresses this by separating Cross-Format Substitution (valid structure, wrong format) from Ungrounded Structure (structure not supported by the image). These categories are identified by rule-based parsing independent of the LLM judges.

W3. Both fine-tuning stages use the same dataset, training duration, learning rate, and optimization steps (Appendix A.5, Table 6). The adaptation methods differ: MiniGPT-4 updates its projection layer while LLaVA-CoT adapts LoRA on the decoder and vision encoder. Multi-format training also divides examples among twelve conditions, reducing per-condition exposure. We will revise the architectural discussion to describe outcomes under these training configurations, without claiming an intrinsic architectural ceiling. The within-model findings are independent of the cross-model comparison. After SF training, MiniGPT-4 gains format robustness but loses scaffolding robustness, while LLaVA-CoT gains scaffolding robustness but loses single-format advantage.

W4. At C3, Qwen3-VL reaches 97.18% compliance for arrows, 97.77% for semicolons, and 97.77% for numbered lists, but falls to 44.24% for adjacency lists. Cross-Format Substitution accounts for 53.02% of its adjacency responses (Table 3). High linearized compliance does not extend to adjacency lists, even for the strongest model in the study. The compliant-only adjacency score (6.00) also remains well below its linearized scores (7.86–8.19). Although evaluation on additional frontier models would help assess the generality of these findings, the within-model variation already supports evaluating multiple representations regardless of baseline compliance.

Q1. Our experiments measure the quality and format of generated causal responses under varying output requirements. They do not determine how much causal knowledge comes from visual evidence versus language priors. The fixed-input design isolates variation associated with format and scaffolding, while image-removal controls would help assess the extent to which responses depend on visual evidence.

Q2. We will remove claims of intrinsic architectural limitations from the camera-ready version. The supported conclusion is narrower. Under the tested configurations, fine-tuning substantially reduces LLaVA-CoT's format difficulties but leaves persistent failures in MiniGPT-4. Whether MiniGPT-4's limitation reflects the trainable parameter space, training dosage, or architecture requires further ablation.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.