BiasTrace: New Annotation Scheme Links Reasoning Behaviours to Biased Outputs in LLMs
A new annotation framework named BiasTrace has been developed by researchers to label reasoning behaviors in large language model (LLM) traces associated with biased results. This approach, outlined in a paper on arXiv (ID 2608.14161), fills a significant void in bias research by examining reasoning processes rather than merely focusing on end results. BiasTrace identifies specific behaviors related to bias, such as unfounded demographic assumptions, along with broader reasoning patterns that lead to bias, like overgeneralization. The authors contend that current methods fail to address the reasoning behaviors that contribute to biased outcomes. By methodically annotating these reasoning traces, BiasTrace seeks to improve the understanding of social biases in LLMs, which is vital for critical applications.
Key facts
- BiasTrace is an annotation scheme for labelling reasoning behaviours in LLM traces.
- It links reasoning behaviours to biased outcomes.
- The paper is available on arXiv with ID 2608.14161.
- It addresses the gap in bias research focusing on final outputs only.
- BiasTrace captures bias-specific behaviours like unsupported demographic assumptions.
- It also captures general reasoning patterns that may implicitly contribute to bias.
- The approach leverages recent advances in LLM reasoning.
- Existing methods overlook different reasoning behaviours that drive biased outcomes.
Entities
Institutions
- arXiv