# C2PO: Diagnosing and Disentangling Bias Shortcuts in LLMs

arXiv: https://arxiv.org/abs/2512.23430 · HTML: https://mlsanity.com/paper/2512.23430/

## Key facts

- Cited by 5 papers in the map (data through 14 Oct 2026).
- Builds on 25 of its 37 references that are in the map.
- First posted on arXiv 29 Dec 2025.

## Builds on (most cited first, 25 of 25)

- [Training language models to follow instructions with human feedback](https://mlsanity.com/paper/2203.02155/) (2022)
- [Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena](https://mlsanity.com/paper/2306.05685/) (2023)
- [Direct Preference Optimization: Your Language Model is Secretly a Reward Model](https://mlsanity.com/paper/2305.18290/) (2023)
- [Measuring Massive Multitask Language Understanding](https://mlsanity.com/paper/2009.03300/) (2020)
- [A Broad-Coverage Challenge Corpus for Sentence Understanding through Inference](https://mlsanity.com/paper/1704.05426/) (2017)
- [Right for the Wrong Reasons: Diagnosing Syntactic Heuristics in Natural Language Inference](https://mlsanity.com/paper/1902.01007/) (2019)
- [BBQ: A Hand-Built Bias Benchmark for Question Answering](https://mlsanity.com/paper/2110.08193/) (2021)
- [Contrastive Preference Optimization: Pushing the Boundaries of LLM Performance in Machine Translation](https://mlsanity.com/paper/2401.08417/) (2024)
- [Self-Diagnosis and Self-Debiasing: A Proposal for Reducing Corpus-Based Bias in NLP](https://mlsanity.com/paper/2103.00453/) (2021)
- [Prompting GPT-3 To Be Reliable](https://mlsanity.com/paper/2210.09150/) (2022)
- [UnQovering Stereotyping Biases via Underspecified Questions](https://mlsanity.com/paper/2010.02428/) (2020)
- [An LLM can Fool Itself: A Prompt-Based Adversarial Attack](https://mlsanity.com/paper/2310.13345/) (2023)
- [Group Robust Preference Optimization in Reward-free RLHF](https://mlsanity.com/paper/2405.20304/) (2024)
- [Self-Debiasing Large Language Models: Zero-Shot Recognition and Reduction of Stereotypes](https://mlsanity.com/paper/2402.01981/) (2024)
- [SCANS: Mitigating the Exaggerated Safety for LLMs via Safety-Conscious Activation Steering](https://mlsanity.com/paper/2408.11491/) (2024)
- [Binary Classifier Optimization for Large Language Model Alignment](https://mlsanity.com/paper/2404.04656/) (2024)
- [FairSteer: Inference Time Debiasing for LLMs with Dynamic Activation Steering](https://mlsanity.com/paper/2504.14492/) (2025)
- [Unlearning Backdoor Attacks for LLMs with Weak-to-Strong Knowledge Distillation](https://mlsanity.com/paper/2410.14425/) (2024)
- [Seeing Sound, Hearing Sight: Uncovering Modality Bias and Conflict of AI models in Sound Localization](https://mlsanity.com/paper/2505.11217/) (2025)
- [IPO: Your Language Model is Secretly a Preference Classifier](https://mlsanity.com/paper/2502.16182/) (2025)
- [Co2PT: Mitigating Bias in Pre-trained Language Models through Counterfactual Contrastive Prompt Tuning](https://mlsanity.com/paper/2310.12490/) (2023)
- [Causal-Guided Active Learning for Debiasing Large Language Models](https://mlsanity.com/paper/2408.12942/) (2024)
- [Towards Reward Fairness in RLHF: From a Resource Allocation Perspective](https://mlsanity.com/paper/2505.23349/) (2025)
- [P2P: A Poison-to-Poison Remedy for Reliable Backdoor Defense in LLMs](https://mlsanity.com/paper/2510.04503/) (2025)
- [Learning from Mistakes: Self-correct Adversarial Training for Chinese Unnatural Text Correction](https://mlsanity.com/paper/2412.17279/) (2024)

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

- [Self-Debias: Self-correcting for Debiasing Large Language Models](https://mlsanity.com/paper/2604.08243/) (2026)
- [Reward Bias Substitution: Single-Axis Bias Mitigations Redirect Optimization Pressure](https://mlsanity.com/paper/2605.27996/) (2026)
- [A Heuristic Perspective on Debiasing Language Models](https://mlsanity.com/paper/2608.00622/) (2026)
- [Towards a Universal Causal Reasoner](https://mlsanity.com/paper/2605.24873/) (2026)
- [Where Detectors Fail: Closing the Tail-Domain Gap with Expert-Guided Mutual Distillation](https://mlsanity.com/paper/2607.26555/) (2026)

## Abstract (arXiv)

Bias in Large Language Models (LLMs) poses significant risks to trustworthiness, manifesting primarily as stereotypical biases (e.g., gender or racial stereotypes) and structural biases (e.g., lexical overlap or position preferences). However, prior paradigms typically address these in isolation, often mitigating one at the expense of exacerbating the other. To address this, we conduct a systematic exploration of these reasoning failures and identify a primary inducement: the latent spurious feature correlations within the input that drive these erroneous reasoning shortcuts. Driven by these findings, we introduce Causal-Contrastive Preference Optimization (C2PO), a unified alignment framework designed to tackle these specific failures by simultaneously discovering and suppressing these correlations directly within the optimization process. Specifically, C2PO leverages causal counterfactual signals to isolate bias-inducing features from valid reasoning paths, and employs a fairness-sensitive preference update mechanism to dynamically evaluate logit-level contributions and suppress shortcut features. Extensive experiments across multiple benchmarks covering stereotypical bias (BBQ, Unqover), structural bias (MNLI, HANS, Chatbot, MT-Bench), out-of-domain fairness (StereoSet, WinoBias), and general utility (MMLU, GSM8K) demonstrate that C2PO effectively mitigates stereotypical and structural biases while preserving robust general reasoning capabilities.