4 ms·
This finding could be quite useful (LLM 1 refers to GPT-3.5 and LLM 2, GPT-4): "Reproducibility analyses revealed that highly reproducible answers were more li
by nopinsight 3y ago
This finding could be quite useful (LLM 1 refers to GPT-3.5 and LLM 2, GPT-4):
"Reproducibility analyses revealed that highly reproducible answers were more likely to be answered correctly than inconsistent answers by LLM 1 (66 of 88 [75.0%] vs 5 of 13 [38.5%]; P = .02), potentially indicating another marker of confidence of LLMs that might be leveraged to filter out invalid responses. The same observation was made with LLM 2, with 78 of 96 correct answers (81.3%) in those with high reproducibility vs 1 of 4 (25.0%) in answers with low reproducibility (P = .04)."
The paper submitted 50 independent queries to assess self-consistency. Exploring the trade-off graph between the number of queries and the effectiveness to filter likely invalid responses would be interesting.
- eightysixfour 3y agoMicrosoft research’s team took advantage of this fact by querying multiple times per question (with randomized option ordering) and choosing the most common answer. Here is the paper: https://www.microsoft.com/en-us/research/publication/can-generalist-foundation-models-outcompete-special-purpose-tuning-case-study-in-medicine/ https://www.microsoft.com/en-us/research/publication/can-gen...
- hlfshell 3y agoThis is driven by the idea that since LLMs are probabilistic engines during inference that, if the LLM "knows" or can "surmise" the correct answer, it will have a higher probability of being generated and thus self consistency methods can more reliably tease it out.