Source-linked AI summary
Are Text-to-Image Models Inductivist Turkeys? A Counterfactual Benchmark for Causal Reasoning
Jiayi Lei, Yuandong Pu, Xingyu Han, Rongpeng Zhu, Jing Xu, Jinyao Wang, Zijian Zhou, Bin Fu, Yuewen Cao, Yihao Liu, Hongsheng Li
TL;DR
It remains unclear whether text-to-image models possess counterfactual causal reasoning or rely on statistical visual-textual priors. CF-World evaluates this gap across progressive factual and counterfactual tasks, finding significant performance declines and entangled priors that limit causal reasoning.
Problem
Existing benchmarks do not effectively distinguish counterfactual causal reasoning from statistical priors or superficial object combinations in text-to-image models.
Method
The CF-World benchmark uses three progressive levels across multiple disciplines, evaluated by the VLM-based CF-Eval pipeline with PRR and RRR metrics.
Results
SOTA models perform well factually but decline significantly counterfactually, with higher-level rule reasoning the dominant bottleneck over visual attribute recombination.
Takeaways & Limitations
The findings indicate that current T2I models remain deeply entangled in high-frequency priors rather than reliably decoupling world knowledge for counterfactual generation.
Takeaways & Limitations
The work is diagnostic and does not propose an algorithmic solution to concept entanglement.
Abstract
from arXiv · showhide
Text-to-image (T2I) generation models have achieved remarkable progress in producing visually realistic images from natural language prompts. Yet it remains unclear whether their success reflects genuine causal understanding or sophisticated pattern matching over visual-textual correlations. Inspired by Russell's inductivist turkey, we introduce Counterfactual-World (CF-World), a counterfactual benchmark designed to investigate whether text-to-image models can generate images under rules that systematically contradict real-world priors. CF-World organizes each scenario into three progressive levels: factual generation under ordinary world knowledge, explicit counterfactual generation with direct visual instructions, and implicit counterfactual generation requiring causal deduction from altered rules. We evaluate both open-source and closed-source T2I models using a Vision Language Model (VLM)-based evaluator (CF-Eval). Furthermore, we introduce two metrics: Prior Resistance Rate (PRR), which measures a models' ability to overcome entrenched real-world priors, and Reasoning Retention Rate (RRR), which assesses whether models can maintain reasoning-dependent counterfactual generation without explicit visual cues. Experiments show that all models exhibit sharp degradation from factual to counterfactual settings. Further analyses suggest that these failures arise because current T2I models encode world knowledge and visual appearances as tightly coupled patterns. Consequently, their heavy reliance on frequent visual co-occurrences within the training data forces them to default to familiar commonsense priors when tasked with rendering counterfactual worlds.
1 Introduction
The paper introduces CF-World, a three-level benchmark for testing whether text-to-image models perform counterfactual causal reasoning rather than rely on learned associations. It also proposes CF-Eval and diagnostic analyses showing that statistical priors constrain both logical deduction and visual recombination.
- Motivation: Existing benchmarks cannot determine whether T2I models possess counterfactual causal reasoning or merely retrieve high-frequency visual-textual patterns.Reasoning-driven benchmarks may be satisfied by memorized statistical priors, while existing counterfactual benchmarks use superficial combinations of unrelated objects.
- CF-World: CF-World evaluates counterfactual reasoning through three progressive levels: factual generation, explicit counterfactual generation, and implicit counterfactual generation.L1 establishes a factual baseline; L2 states altered laws and outcomes; L3 states only altered laws, requiring deduction of visual outcomes.
- Failure analysis: Diagnostic probes find universal performance drops under counterfactual rules, indicating difficulty with causal reasoning even for abstract symbolic elements without visual burdens.Additional probes isolate visual generation, and the combined failures identify a bottleneck in decoupling rules from attributes.
- Failure analysis: T2I models’ pixel co-occurrences and text-image alignments make it difficult to decouple causal variables and visual attributes, causing defaults to high-frequency commonsense priors.This coupling limits higher-level logical reasoning independent of visual composition.
- CF-World: CF-World is presented as the first counterfactual world-knowledge benchmark to assess logical and causal reasoning under counterfactual premises while eliminating training priors.Its progressive design targets the gap in existing generative T2I evaluations.
- Evaluation: CF-Eval introduces PRR and RRR to quantify models’ causal reasoning capabilities through an automated evaluation pipeline.The paper defines PRR and RRR as novel quantitative metrics, though the supplied passage does not provide their formulas or values.
2 Related Work
Related work spans diverse T2I architectures, counterfactual-generation methods, and evaluation benchmarks. Existing approaches address causal consistency and concept alignment, but remain limited by training-data biases and narrow visual-attribute or co-occurrence edits.
- T2I Synthesis: T2I synthesis has advanced across diffusion-based, autoregressive, and unified multimodal architectural paradigms.The cited work identifies representative methods in each paradigm.
- Counterfactual Generation: Counterfactual T2I generation requires scenes that deviate from reality while preserving internal causal consistency.Foundational approaches include disentangling causal features with Generative Causal Models and using DreamBooth to preserve subject identity across novel contexts.
- Counterfactual Generation: Models often revert to training-data biases for unusual concept combinations, causing latent concept misalignment.This failure persists despite methods targeting subject identity and novel contexts.
- Counterfactual Generation: Recent corrections for physically implausible scenes remain constrained to visual attribute editing and concept co-occurrence.These limitations leave broader counterfactual reasoning insufficiently addressed.
- Benchmarks and Metrics: Evaluation research uses benchmarks including GeckoNum, Winoground, GenEval, and GenAI-Bench to assess compositional and numerical alignment.Other listed benchmarks include OK-VQA, WISE, Commonsense-T2I, R2I-Bench, and T2I-ReasonBench.
3 Counterfactual-World (CF-World)
CF-World is a 1,091-group, 3,273-prompt benchmark spanning five disciplines and testing counterfactual reasoning through progressively demanding factual, explicit, and implicit scenarios. CF-Eval uses VLM scoring, factual thresholding, and PRR/RRR metrics to distinguish rendering quality, prior resistance, and retained reasoning.
- Benchmark and Taxonomy: CF-World contains 1,091 groups and 3,273 prompts across Physics, Biology, Chemistry, Geography, and Sociology.Its taxonomy targets basic laws commonly taught in middle-school curricula rather than obscure expert knowledge.
- Progressive Framework: The three-level framework progresses from factual generation (L1), to explicit counterfactual generation (L2), to implicit counterfactual generation (L3).L1 tests basic real-world priors; L2 alters laws while stating visual outcomes; L3 alters laws without stating outcomes.
- Data Generation and Quality Assurance: Dataset construction combines manually curated scientific principles, LLM-generated prompts, four generation criteria, and expert human review.The criteria require visual unambiguity, logical deduction, safety, and scientific validity while excluding artifacts and ambiguities.
- CF-Eval: CF-Eval is an automated VLM-driven pipeline that scores visual integrity, assessment-point adherence, and logic consistency with differentiated weights.Assessment-point weights are 12–16, logic-consistency weights are 7–9, and visual-integrity weights are 1–3.
- Scoring and Metrics: Counterfactual scores SL2 and SL3 are computed only when the factual baseline satisfies SL1 ≥0.5, preventing false positives from failed factual generation.PRR measures resistance to real-world priors from L1 to L2, while RRR measures retention without explicit visual cues from L2 to L3.
4 Experiments
The experiments evaluate diverse text-to-image systems on three progressively difficult CF-World prompt levels using PRR and RRR. Results show sharp counterfactual degradation, especially for open-source models, while closed-source and some unified architectures retain stronger performance.
- Experimental Setup: The evaluation covers diverse state-of-the-art open-source text-to-image systems across the CF-World benchmark.The listed systems include SANA 1.5, Janus-Pro-7B, Show-o2, Z-image, Lumina-DiMOO, BAGEL, BAGEL-CoT, OmniGen2, and FLUX.2-dev.
- Experimental Setup: Models are evaluated at Factual (L1), Explicit Counterfactual (L2), and Implicit Counterfactual (L3) levels with PRR and RRR metrics.PRR measures prior resistance, while RRR measures reasoning retention across counterfactual generation.
- Prior Lock-in in Counterfactual Generation: Most open-source models show a substantial L1-to-L2 performance decline, with PRRs largely below 0.50.Stronger foundational capability does not guarantee higher counterfactual robustness; Qwen-Image can yield lower PRRs than other models.
- Bottlenecks in Causal Reasoning: Open-source models lose further performance on L3, and BAGEL-CoT provides only a marginal boost over BAGEL.The results indicate severe limitations in autonomous causal deduction, potentially reflecting a modality gap between language-side logic and visual generation.
- Performance Comparison of Different Models: Top-tier closed-source models such as Nano Banana Pro maintain robust L2 and L3 scores, while architectural choices strongly affect open-weight performance.Native multimodal and unified architectures are identified as important factors among open-weight models.
- Evaluation Validation: The automated evaluator was validated against human judgments using 1,000 images from FLUX.2-dev and Nano Banana Pro.Three trained graduate-level expert annotators independently scored the images, and their averaged judgments formed the human ground truth.
5 Why Models Fail: A Decoupling Perspective
Mechanistic experiments identify two distinct failure axes—logical rule execution and visual attribute recombination—linked by lexical object–attribute entanglement. Models preserve factual visual rendering better than counterfactual rule execution, while de-nominalization yields modest, model-dependent improvements.
- Logical Rule Execution: Models struggle to execute counterfactual rules even in a simplified symbolic benchmark minimizing visual clutter and recognition demands.The benchmark contains 198 prompts covering 33 objective rules, with multi-dimensional counterfactual perturbations.
- Visual Attribute Recombination: Models perform strongly on factual attribute pairs but consistently degrade on rare counterfactual pairs, indicating limited recombination beyond frequent co-occurrences.The attribute-decoupling experiment uses 100 rare concept pairs and corresponding common factual pairs.
- Lexical Vulnerability: De-nominalization improves performance across all models, with FLUX.2-dev gaining +0.09, Z-image and Show-o2 +0.05, and several models only +0.01.Replacing target nouns with descriptive phrases while preserving the counterfactual law reduces interference from learned visual priors.
- Asymmetric Decoupling: Factual attribute rendering exceeds 0.92, while factual rule execution remains below 0.61; under counterfactual stress, attribute scores retain over 0.81 of factual value, but rule scores fall as low as 0.37.This asymmetry suggests stronger shallow visual interpolation than deep physical grounding for logical deduction.
- Two Regimes of Entanglement: The diagnostics distinguish shallow lexical entanglement in diffusion models from deeper semantic entanglement in unified models.Diffusion models show de-nominalization gains of +0.05 to +0.09, whereas unified models show negligible gains of ≤0.02.
6 Conclusion and Limitations
SOTA generative models perform strongly in factual settings but degrade significantly under counterfactual conditions because they cannot decouple world knowledge, default scenarios, visual attributes, and subjects. The work is diagnostic rather than algorithmic, positioning CF-World as a testbed for future decoupling mechanisms and causal reasoning.
- Conclusion: SOTA generative models excel in factual settings but decline significantly under counterfactual conditions.
- Conclusion: The degradation reflects entanglement between objective world knowledge and default scenarios, and between visual attributes and their corresponding subjects.
- Limitations: The work is fundamentally diagnostic and does not propose an algorithmic solution to concept entanglement.
- Future Directions: CF-World is intended to support future research on decoupling mechanisms that move multimodal models from prior-driven generation toward genuine causal reasoning.
A Datasheet for Datasets · A.1 Motivation and Composition · A.2 Collection Process and Maintenance
The appendix documents CF-World as a comprehensive dataset, detailing its motivation, three-level composition, generation and filtering process, hosting, maintenance, and responsible-AI metadata compliance.
- A Datasheet for Datasets: CF-World is documented through a comprehensive datasheet following standard dataset-documentation practices.
- A.1 Motivation and Composition: The dataset probes whether T2I models exhibit genuine causal understanding rather than relying on superficial visual-textual co-occurrences.
- A.1 Motivation and Composition: CF-World contains N = 1091 unique counterfactual scenarios.
- A.1 Motivation and Composition: Each scenario has three progressive levels: L1 Factual, L2 Explicit Counterfactual, and L3 Implicit Counterfactual.
- A.1 Motivation and Composition: These levels produce 3273 distinct prompts in total.
- A.2 Collection Process and Maintenance: Gemini-3-Pro generated the initial prompts, which underwent rigorous human-in-the-loop filtering.
- A.2 Collection Process and Maintenance: The authors will host and maintain the dataset on Hugging Face.
- A.2 Collection Process and Maintenance: The repository includes croissant.json to support compliance with Responsible AI metadata standards.
B Empirical Calibration of the Factual Threshold
The factual threshold for the continuous score SL1 was calibrated against human recognition judgments using 150 L1-baseline images. A threshold of T = 0.5 achieved the highest human-VLM alignment, separating recognizable images from malformed generations.
- Generation Quality Across Scores: Images with SL1 ≈0.3 showed severe semantic distortions, while images with SL1 ≈0.7 retained recognizable core subjects.Examples of distortion included hail rendered as rain and a droplet unnaturally levitating.
- Human-VLM Alignment Experiment: 150 L1-baseline images were human-annotated to label whether core subjects were recognizable or missing or severely distorted.Sampling heavily weighted the borderline SL1 range [0.3, 0.7].
- Threshold Selection: T = 0.5 produced the highest alignment between VLM scores and human ground-truth labels.Lower thresholds create false positives, whereas higher thresholds create false negatives.
C Comprehensive Prompt Templates
The appendix specifies prompt templates for generating evaluation questions and scoring factual, counterfactual, rule-decoupling, and attribute-decoupling outputs. It also calibrates Gemini and Qwen to apply comparably rigorous standards, especially for counterfactual failures.
- C.1 Model-Specific Prompt Calibration for VLMs: Preliminary testing found Gemini relatively balanced and Qwen lenient in L2/L3, with Qwen overlooking subtle fractures and over-scoring factual objects.Calibration therefore gives Gemini a standard analytical persona and Qwen a “Strict, Adversarial Judge” persona to enforce comparable thresholds.
- C.2 Evaluation Question Generation Prompts: The evaluation-question template instructs the VLM to generate a strict protocol covering three dimensions from the input prompt and assessment points.The template receives {input_prompt} and {assessment_points}.
- C.2 Evaluation Question Generation Prompts: Assessment points must be combined into one comprehensive question, with a 0.5 score made difficult and 0.0 indicating unmet criteria.Visual integrity receives weight 2 or 3, while the assessment-point dimension emphasizes strict verification.
- C.2 Evaluation Question Generation Prompts: Dimension 3 assigns weight 8 to logic: factual L1 requires standard physics, whereas L2/L3 require the entire scene to obey the counterfactual premise without logical fractures.For L1, scores are 1.0 for flawless physics, 0.5 for a minor flaw, and 0.0 for a clear violation; L2/L3 use 1.0 for a fully altered world and 0.0 for a logical fracture.
- C.3 VLM Scoring Prompts (Gemini): Gemini uses standard factual scoring for L1 but an analytical counterfactual evaluator for L2/L3 that prioritizes whether the requested illogical elements are generated.Its L2/L3 guidance accepts minor artifacts when the core counterfactual goal is achieved, but scores 0.0 for reverting to normal physics.
- C.4 VLM Scoring Prompts (Qwen): Qwen’s L2/L3 prompt explicitly directs adversarial failure detection, harsh scoring for ambiguity or normal physics, and scrutiny of background inconsistencies.Its factual prompt remains objective and balanced, allowing high scores when factual requirements are mostly met despite minor flaws.
- C.5 Decoupling Evaluation Prompts: The decoupling prompts evaluate whether images satisfy strict rule constraints and whether Entity A, Entity B, and their specified relationship are all present.Rule-decoupling scores are harsh for unmet constraints; attribute-decoupling scores are 1.0 for exact compliance, 0.5 for partial relationship or blending issues, and 0.0 for missing entities or wrong relationships.
D Qualitative Analysis · E Computational Resources and Execution Details
The qualitative analysis shows that current T2I models struggle to simulate counterfactual physical consequences, even when evaluated across diverse architectures and scales. The benchmark is supported by documented local-GPU and cloud-API execution requirements for reproducibility.
- D Qualitative Analysis: The qualitative study examines how state-of-the-art T2I models behave under progressively stronger counterfactual constraints across different architectures.This analysis is intended to provide a deeper understanding of model behavior beyond aggregate benchmark scores.
- D Qualitative Analysis: Ten representative T2I models were qualitatively evaluated on a scenario involving infinitely strong water surface tension.The models ranged from lightweight open-source systems such as Janus-Pro-7B and Show-o2 to commercial engines such as GPT-Image-1.5.
- D Qualitative Analysis: Scaling model parameters or training data alone does not inherently grant counterfactual physical simulation ability.The qualitative breakdown indicates that increased scale is insufficient for reliably reasoning about altered physical rules.
- E Computational Resources and Execution Details: The reproducibility setup separates experiments into local GPU computing and cloud-based API services.The resource breakdown covers both image generation and model evaluation, including execution-time requirements.
- D Qualitative Analysis: Most models generate the factual pool scene in L1 but fail to render the counterfactual physical consequence in L2 and L3.They either fail to decouple attributes or revert to normal physics by depicting the person sinking instead of walking on water.
- E Computational Resources and Execution Details: 16 NVIDIA A100 (80GB) GPUs provided the equivalent computational throughput for local open-source T2I generation and Qwen-based VLM evaluation.These local inference tasks ran on a high-performance compute cluster.
- E Computational Resources and Execution Details: Approximately 2 hours per model are required for complete open-source T2I image generation, while local Qwen VLM evaluation takes approximately 2 to 3 hours total.These timings apply to the stated local hardware configuration.
- E Computational Resources and Execution Details: Closed-source generation and Gemini-3-Pro evaluation use commercial APIs whose execution times depend on network latency, rate limits, and server-side concurrency quotas.Their timing is therefore not bounded by local hardware.
F Broader Impact and Limitations
CF-World reveals that current T2I models tightly couple world knowledge with visual appearances, causing sharp degradation in counterfactual settings. These findings motivate causal-reasoning and physical-simulation architectures, while the evaluation remains dependent on VLM-based evaluators.
- Current T2I models encode world knowledge and visual appearances as tightly coupled patterns.
- Sharp degradation in counterfactual settings exposes a limitation of current T2I models.
- CF-World encourages shifting from scaling visual-textual pairs toward architectures capable of genuine causal reasoning and physical simulation.
- The evaluation relies on VLM-based evaluators despite extensive measures to ensure prompt clarity.