# Beyond Spurious Signals: Debiasing Multimodal Large Language Models via Counterfactual Inference and Adaptive Expert Routing

arXiv: https://arxiv.org/abs/2509.15361 · HTML: https://mlsanity.com/paper/2509.15361/

## Key facts

- Cited by 5 papers in the map (data through 14 Oct 2026).
- Placed in Multimodal reasoning, one of 17 directions in LLM reasoning.
- Builds on 35 of its 63 references that are in the map.
- First posted on arXiv 18 Sep 2025.

## Placed in

- LLM reasoning / [Multimodal reasoning](https://mlsanity.com/lr/multimodal-reasoning/)
- Multimodal LLMs / vision-language models / [MLLM training and visual instruction tuning](https://mlsanity.com/mllm/mllm-training-and-visual-instruction-tuning/)
- Mixture-of-experts / [Multimodal MoE models](https://mlsanity.com/moe/multimodal-moe-models/)

## Builds on (most cited first, 30 of 35)

- [Learning Transferable Visual Models From Natural Language Supervision](https://mlsanity.com/paper/2103.00020/) (2021)
- [Qwen2-VL: Enhancing Vision-Language Model's Perception of the World at Any Resolution](https://mlsanity.com/paper/2409.12191/) (2024)
- [InternVL: Scaling up Vision Foundation Models and Aligning for Generic Visual-Linguistic Tasks](https://mlsanity.com/paper/2312.14238/) (2023)
- [Expanding Performance Boundaries of Open-Source Multimodal Models with Model, Data, and Test-Time Scaling](https://mlsanity.com/paper/2412.05271/) (2024)
- [LlamaFactory: Unified Efficient Fine-Tuning of 100+ Language Models](https://mlsanity.com/paper/2403.13372/) (2024)
- [ZeRO: Memory Optimizations Toward Training Trillion Parameter Models](https://mlsanity.com/paper/1910.02054/) (2019)
- [Mitigating Object Hallucinations in Large Vision-Language Models through Visual Contrastive Decoding](https://mlsanity.com/paper/2311.16922/) (2023)
- [Winoground: Probing Vision and Language Models for Visio-Linguistic Compositionality](https://mlsanity.com/paper/2204.03162/) (2022)
- [Hallucination of Multimodal Large Language Models: A Survey](https://mlsanity.com/paper/2404.18930/) (2024)
- [Don't Just Assume; Look and Answer: Overcoming Priors for Visual Question Answering](https://mlsanity.com/paper/1712.00377/) (2017)
- [Counterfactual VQA: A Cause-Effect Look at Language Bias](https://mlsanity.com/paper/2006.04315/) (2020)
- [Causal Inference in Natural Language Processing: Estimation, Prediction, Interpretation and Beyond](https://mlsanity.com/paper/2109.00725/) (2021)
- [Towards Multimodal Sarcasm Detection (An _Obviously_ Perfect Paper)](https://mlsanity.com/paper/1906.01815/) (2019)
- [Counterfactual Samples Synthesizing for Robust Visual Question Answering](https://mlsanity.com/paper/2003.06576/) (2020)
- [Towards Causal VQA: Revealing and Reducing Spurious Correlations by Invariant and Covariant Semantic Editing](https://mlsanity.com/paper/1912.07538/) (2019)
- [FOIL it! Find One mismatch between Image and Language caption](https://mlsanity.com/paper/1705.01359/) (2017)
- [Towards Vision-Language Mechanistic Interpretability: A Causal Tracing Tool for BLIP](https://mlsanity.com/paper/2308.14179/) (2023)
- [A Comprehensive Review of Multimodal Large Language Models: Performance and Challenges Across Different Tasks](https://mlsanity.com/paper/2408.01319/) (2024)
- [SocialCounterfactuals: Probing and Mitigating Intersectional Social Biases in Vision-Language Models with Counterfactual Examples](https://mlsanity.com/paper/2312.00825/) (2023)
- [COCO-Counterfactuals: Automatically Constructed Counterfactual Examples for Image-Text Pairs](https://mlsanity.com/paper/2309.14356/) (2023)
- [MM-SpuBench: Towards Better Understanding of Spurious Biases in Multimodal LLMs](https://mlsanity.com/paper/2406.17126/) (2024)
- [Looking Beyond Text: Reducing Language bias in Large Vision-Language Models via Multimodal Dual-Attention and Soft-Image Guidance](https://mlsanity.com/paper/2411.14279/) (2024)
- [MMSD2.0: Towards a Reliable Multi-modal Sarcasm Detection System](https://mlsanity.com/paper/2307.07135/) (2023)
- [Towards Multi-Modal Sarcasm Detection via Hierarchical Congruity Modeling with Knowledge Enhancement](https://mlsanity.com/paper/2210.03501/) (2022)
- [Towards Multimodal Sentiment Analysis Debiasing via Bias Purification](https://mlsanity.com/paper/2403.05023/) (2024)
- [Robust Emotion Recognition in Context Debiasing](https://mlsanity.com/paper/2403.05963/) (2024)
- [CLMLF:A Contrastive Learning and Multi-Layer Fusion Method for Multimodal Sentiment Detection](https://mlsanity.com/paper/2204.05515/) (2022)
- [A Survey of Multimodal Sarcasm Detection](https://mlsanity.com/paper/2410.18882/) (2024)
- [Images Speak Louder than Words: Understanding and Mitigating Bias in Vision-Language Model from a Causal Mediation Perspective](https://mlsanity.com/paper/2407.02814/) (2024)
- [Aligning Multimodal LLM with Human Preference: A Survey](https://mlsanity.com/paper/2503.14504/) (2025)

## Built on by (most cited first, 5 of 5)

- [OmniMER: Auxiliary-Enhanced LLM Adaptation for Indonesian Multimodal Emotion Recognition](https://mlsanity.com/paper/2512.19379/) (2025)
- [QASA: Quality-Aware Semantic Augmentation for Robust Multimodal Sentiment Analysis](https://mlsanity.com/paper/2601.06870/) (2026)
- [UGID: Unified Graph Isomorphism for Debiasing Large Language Models](https://mlsanity.com/paper/2603.19144/) (2026)
- [Sensing Which Modality Matters: Evidence-Gated Regularization for Robust VLA Policies](https://mlsanity.com/paper/2609.03142/) (2026)
- [Beyond Text: LLM-Based Dimensional Emotion Evaluation in Multimodal Dialogue](https://mlsanity.com/paper/2609.39072/) (2026)

## Abstract (arXiv)

Multimodal Large Language Models (MLLMs) have shown substantial capabilities in integrating visual and textual information, yet frequently rely on spurious correlations, undermining their robustness and generalization in complex multimodal reasoning tasks. This paper addresses the critical challenge of superficial correlation bias in MLLMs through a novel causal mediation-based debiasing framework. Specially, we distinguishing core semantics from spurious textual and visual contexts via counterfactual examples to activate training-stage debiasing and employ a Mixture-of-Experts (MoE) architecture with dynamic routing to selectively engages modality-specific debiasing experts. Empirical evaluation on multimodal sarcasm detection and sentiment analysis tasks demonstrates that our framework significantly surpasses unimodal debiasing strategies and existing state-of-the-art models.