Poster
Second Thoughts are Best: Learning to Re-Align With Human Values from Text Edits
Ruibo Liu · Chenyan Jia · Ge Zhang · Ziyu Zhuang · Tony Liu · Soroush Vosoughi
Hall J (level 1) #925
Keywords: [ alignment ] [ human-AI interaction ] [ social impact ] [ human values ] [ AI safety ]
We present Second Thoughts, a new learning paradigm that enables language models (LMs) to re-align with human values. By modeling the chain-of-edits between value-unaligned and value-aligned text, with LM fine-tuning and additional refinement through reinforcement learning, Second Thoughts not only achieves superior performance in three value alignment benchmark datasets but also shows strong human-value transfer learning ability in few-shot scenarios. The generated editing steps also offer better interpretability and ease for interactive error correction. Extensive human evaluations further confirm its effectiveness.