Does Reasoning Introduce Bias? A Study of Social Bias Evaluation and Mitigation in LLM Reasoning (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Recent advances in large language models have enabled automatic generation of chain-of-thought reasoning . however, when reasoning steps reflect social stereotypes, they can reinforce harmful associations and lead to misleading conclusions. |
| Approach: | They propose a method that detects how model predictions change across incremental reasoning steps. |
| Outcome: | The proposed method outperforms a stereotype-free baseline and improves accuracy. |
Similar Papers
Unveiling Confirmation Bias in Chain-of-Thought Reasoning (2025.findings-acl)
Copied to clipboard
| Challenge: | Chain-of-thought (CoT) prompting has been widely adopted to enhance the reasoning capabilities of large language models (LLMs). |
| Approach: | They propose to examine how internal beliefs affect reasoning generation and reasoning-guided answer prediction in CoT by decomposing CoT into a two-stage process. |
| Outcome: | The proposed model beliefs affect reasoning generation and reasoning-guided answer prediction in CoT, and the results provide strong evidence of confirmation bias in LLMs. |
Cognitive Bias in Decision-Making with LLMs (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Large language models inherit societal biases against protected groups and can be subject to functionally resembling cognitive bias. |
| Approach: | They propose a framework to uncover, evaluate, and mitigate cognitive bias in large language models by using a dataset containing 13,465 prompts to evaluate LLM decisions on different cognitive biases. |
| Outcome: | The proposed framework uncovers, evaluates, and mitigates cognitive bias in large language models, particularly in high-stakes decision-making tasks. |
The Impact of Inference Acceleration on Bias of LLMs (2025.naacl-long)
Copied to clipboard
| Challenge: | Recent work suggests strategies to increase inference efficiency with LLMs . however, these strategies may inadvertently lead to some side-effects. |
| Approach: | They propose to optimize inference acceleration strategies such as quantization, pruning, and caching to reduce inference cost and latency while maintaining predictive performance. |
| Outcome: | The proposed strategies reduce cost and latency while maintaining predictive performance while preserving the model size. |
Good Arguments Against the People Pleasers: How Reasoning Mitigates (Yet Masks) LLM Sycophancy (2026.acl-long)
Copied to clipboard
| Challenge: | Recent studies have identified a critical drawback of aligning models with human judgments and outputs that are flawed or incorrect. |
| Approach: | They evaluate a range of LLMs to examine whether CoT reasoning mitigates sycophancy . they find that reasoning masks a tendency to scophage in some cases . |
| Outcome: | The proposed model models show that CoT reasoning reduces sycophancy but masks it in some cases. |
A Multilingual Social Bias Benchmark Incorporating Thinking Processes (2026.acl-long)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) can learn useful knowledge and harmful stereotypes, making bias evaluation essential. |
| Approach: | They propose a multilingual social bias benchmark that incorporates human-generated reasoning as part of the thinking process. |
| Outcome: | The proposed method demonstrates superior performance over LLM-generated methods . human-generated thinking yields higher-quality evaluations than template-based approaches . |
DeCoT: Debiasing Chain-of-Thought for Knowledge-Intensive Tasks in Large Language Models via Causal Intervention (2024.acl-long)
Copied to clipboard
| Challenge: | In large language models, external knowledge is required to augment their internal knowledge through prompts, but this does not guarantee that LLMs can identify and use relevant information in the prompts to conduct chain-of-thought reasoning. |
| Approach: | They propose a structural causal model to formally explain the internal knowledge bias of large language models (LLMs) they review the chain-of-thought (CoT) prompting from a causal perspective and find that biased information from pretrained models can impair LLMs’ reasoning abilities. |
| Outcome: | The proposed model enables more accurate CoT reasoning and enhances LLM generation on knowledge-intensive tasks. |
Likelihood-based Mitigation of Evaluation Bias in Large Language Models (2024.findings-acl)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) are widely used to evaluate natural language generation tasks as automated metrics. |
| Approach: | They propose to use LLMs to evaluate sentences with higher likelihoods and lower likelihoods to mitigate the likelihood bias. |
| Outcome: | The proposed method overrates sentences with higher likelihoods while underrating sentences with lower likelihoods. |
Social Bias Frames: Reasoning about Social and Power Implications of Language (2020.acl-main)
Copied to clipboard
| Challenge: | Language has enormous power to project social biases and reinforce stereotypes on people. |
| Approach: | They propose a new conceptual formalism that aims to model the pragmatic frames in which people project social biases and power differentials onto others. |
| Outcome: | The proposed model can model the pragmatic frames in which people project social biases and power differentials onto others. |
Ask LLMs Directly, “What shapes your bias?”: Measuring Social Bias in Large Language Models (2024.findings-acl)
Copied to clipboard
| Challenge: | Existing methods to evaluate social bias in large language models have limitations . et al., 1995: stereotypes shape social perceptions without objective basis . |
| Approach: | They propose a method to intuitively quantify social perceptions and suggest metrics to evaluate biases within LLMs. |
| Outcome: | The proposed metrics capture the multi-dimensional aspects of social bias, the paper shows . they show that the proposed metrics can be used to evaluate bias in large language models . |
On Second Thought, Let’s Not Think Step by Step! Bias and Toxicity in Zero-Shot Reasoning (2023.acl-long)
Copied to clipboard
| Challenge: | Prior work has focused on logical reasoning tasks; it remains unclear whether improvements hold for more diverse types of reasoning, especially in socially situated contexts. |
| Approach: | They perform a controlled evaluation of zero-shot CoT reasoning in two socially sensitive domains: harmful questions and stereotype benchmarks. |
| Outcome: | The results show that zero-shot CoT reasoning increases model’s likelihood to produce harmful or undesirable output, but decreases with improved instruction following. |