Eyes Can Deceive: Benchmarking Counterfactual Reasoning Abilities of Multi-modal Large Language Models

Yian Li, Wentao Tian, Yang Jiao, Jingjing Chen, Yu-Gang Jiang

Abstract

Counterfactual reasoning, a crucial manifestation of human intelligence, involves making presuppositions based on established facts and extrapolating potential outcomes. While existing multimodal large language models (MLLMs) exhibit impressive cognitive capabilities on standard benchmarks, their performance on counterfactual questions remains under-explored. We introduce the CounterFactual MultiModal (CFMM) reasoning benchmark (also known as MARS-Bench) to systematically assess these capabilities across six challenging tasks. Our findings reveal that existing MLLMs often "believe what they see," ignoring counterfactual presuppositions in favor of empirical intuition. To address this, we evaluate a wide range of MLLMs and propose potential avenues for improvement toward human-level intelligence.

Introduction: The Intuition Trap

Current MLLMs like GPT-4o and Qwen2.5-VL often rely on "empirical intuition"—statistical patterns from their training data—rather than actual logical deduction. This makes them susceptible to being fooled by hypothetical "what if" questions. For example, if asked, "If I painted this bus blue, what color would it be?" many models default to the original color seen in the image. This over-reliance on pattern matching prevents models from performing true compositional reasoning, which requires breaking a problem into logical, hypothetical steps.

The MARS-Bench (CFMM) Dataset

To systematically evaluate these failures, we developed the Multimodal Assumptive Reasoning Benchmark (MARS-Bench). The dataset consists of 1,022 images and 6,000 questions categorized into six dimensions: Count, Color, Size, Shape, Direction, and Common Sense. Each image is paired with a fact-based "basic" question and several "assumptive" (counterfactual) questions that require the model to mentally simulate a change to the scene.

A sample image from the benchmark used to evaluate reasoning about object properties and spatial directions in real-world scenes.

Methods: Active Deduction (AD)

The researchers proposed a two-stage training paradigm called Active Deduction (AD) to move models away from simple pattern matching:

  1. AD-SFT (Supervised Fine-Tuning): The model is trained on a curated dataset where complex questions are paired with structured reasoning paths wrapped in <think> tags. Crucially, simple questions are answered directly, teaching the model to only "think" when necessary.
  2. AD-RFT (Reinforcement Fine-Tuning): Using Group Relative Policy Optimization (GRPO), the model is refined based on format rewards (correct use of reasoning tags) and accuracy rewards (evaluated by a judge model like GPT-4o-mini).

This "Active" approach allows the model to dynamically adjust its cognitive effort based on the difficulty of the prompt, avoiding the unnecessary computational cost of forced reasoning on simple tasks.

Key Results

Experiments demonstrate that the Active Deduction framework significantly closes the reasoning gap for open-source models:

  • Massive Accuracy Gains: On MARS-Bench assumptive questions, the Qwen-AD 7B model achieved an accuracy score of 495.2, a staggering 76.6-point improvement over the baseline Qwen2.5-VL 7B (418.6).
  • No "Reasoning Tax": Unlike models forced to "think" for every prompt, Qwen-AD maintained or improved its performance on standard, non-assumptive benchmarks like OCRBench and LLaVABench.
  • Deduction over Intuition: By training the model to decompose problems (e.g., identifying a direction before simulating a U-turn), the model moves toward human-like deduction.
An example of a scene requiring common-sense reasoning about biological interactions, a key category in the CFMM benchmark.

Conclusion and Future Work

While the Qwen-AD series significantly improves robustness, challenges remain. The current reward system relies heavily on the biases of judge models like GPT-4o-mini, and extremely complex, multi-layered logic scenarios still pose a challenge for 7B-parameter models. Future research will focus on expanding data diversity to include video and 3D environments to further refine MLLM intelligence.

Summary updated

Abstract

Counterfactual reasoning, as a crucial manifestation of human intelligence, refers to making presuppositions based on established facts and extrapolating potential outcomes. Existing multimodal large language models (MLLMs) have exhibited impressive cognitive and reasoning capabilities, which have been examined across a wide range of Visual Question Answering (VQA) benchmarks. Nevertheless, how will existing MLLMs perform when faced with counterfactual questions? To answer this question, we first curate a novel \textbf{C}ounter\textbf{F}actual \textbf{M}ulti\textbf{M}odal reasoning benchmark, abbreviated as \textbf{CFMM}, to systematically assess the counterfactual reasoning capabilities of MLLMs. Our CFMM comprises six challenging tasks, each including hundreds of carefully human-labeled and GPT-generated counterfactual questions, to evaluate MLLM's counterfactual reasoning capabilities across diverse aspects. Through experiments, interestingly, we find that existing MLLMs prefer to believe what they see, but ignore the counterfactual presuppositions presented in the question, thereby leading to inaccurate responses. Furthermore, we evaluate a wide range of prevalent MLLMs on our proposed CFMM. The significant gap between their performance on our CFMM and that on several VQA benchmarks indicates that there is still considerable room for improvement in existing MLLMs toward approaching human-level intelligence. On the other hand, through boosting MLLMs performances on our CFMM in the future, potential avenues toward developing MLLMs with advanced intelligence can be explored.
BibTeX
@inproceedings{li2024eyes,
  author = {Yian Li and Wentao Tian and Yang Jiao and Jingjing Chen and Yu-Gang Jiang},
  title = {{Eyes Can Deceive: Benchmarking Counterfactual Reasoning Abilities of Multi-modal Large Language Models}},
  booktitle = {ACM Multimedia},
  year = {2024},
}
Figure