Reproduce “Red Teaming Language Models for Processing Contradictory Dialogues” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Xiaofei Wen, Bangzheng Li, Tenghao Huang, Muhao Chen
This research introduces a Red Teaming framework to enhance large language models' ability to detect, explain, and modify self-contradictory statements within dialogues. The approach leverages a novel synthetic dataset of 12,387 dialogues with machine-generated contradictions and detailed explanations. Fine-tuned models achieved up to 97.4% accuracy in contradiction detection, and the framework reduced the re-detection rate of contradictions in modified dialogues to 3.85% when using generated explanations as guiding prompts.
Key insights
- Vanilla LLMs demonstrate surprisingly poor inherent performance in detecting and explaining dialogue contradictions, significantly underperforming fine-tuned models.
- Instruction-tuning with specialized data dramatically improves LLMs' ability to accurately detect contradictions and generate high-quality, consistent explanations.
- Feeding explicit, generated explanations back into the modification process as guiding prompts substantially enhances the LLM's capability to resolve logical inconsistencies in dialogues.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.