Documentdetail
ID kaart
oai:arXiv.org:2410.11055
Onderwerp
Computer Science - Computation and... Computer Science - Artificial Inte...Auteur
Yao, Jihan Ding, Wenxuan Feng, Shangbin Wang, Lucy Lu Tsvetkov, YuliaCategorie
Computer Science
Jaar
2024
vermelding datum
23-10-2024
Trefwoorden
answers wrong-over-wrong wrong preferencesBeschrijving
In the absence of abundant reliable annotations for challenging tasks and contexts, how can we expand the frontier of LLM capabilities with potentially wrong answers?
We focus on two research questions: (1) Can LLMs generate reliable preferences among wrong options?
And if so, (2) Would alignment with such wrong-over-wrong preferences be helpful?
We employ methods based on self-consistency, token probabilities, and LLM-as-a-judge to elicit wrong-over-wrong preferences, and fine-tune language models with preference optimization approaches using these synthesized preferences.
Extensive experiments with seven LLMs and eight datasets demonstrate that (1) LLMs do have preliminary capability in distinguishing various shades of wrong, achieving up to 20.9% higher performance than random guess; (2) Alignment with wrong-over-wrong preferences helps LLMs to produce less wrong and sometimes even outright correct answers, while overall improving model calibration.
Yao, Jihan,Ding, Wenxuan,Feng, Shangbin,Wang, Lucy Lu,Tsvetkov, Yulia, 2024, Varying Shades of Wrong: Aligning LLMs with Wrong Answers Only
Choice Between Partial Trajectories: Disentangling Goals from Beliefs
agents models aligned based bootstrapped learning reward function model return choice choices partial
Adverse and serious adverse events incidence of pharmacological interventions for managing chronic and episodic migraine in adults: a systematic review
aes medications bta incidence migraine drugs disorders
