Suppressing Pink Elephants with Direct Principle Feedback
This addresses the need for inference-time controllability of LLMs for diverse application contexts, though it's an incremental improvement on existing Constitutional AI methods.
The paper tackles the Pink Elephant Problem - making LLMs avoid discussing forbidden entities while discussing preferred alternatives at inference time - by introducing Direct Principle Feedback, a simplified Constitutional AI method that applies DPO directly on critiques and revisions. Their fine-tuned 13B LLaMA 2 model significantly outperforms Llama-2-13B-Chat and prompted baselines, and matches GPT-4 performance on their curated test set.
Existing methods for controlling language models, such as RLHF and Constitutional AI, involve determining which LLM behaviors are desirable and training them into a language model. However, in many cases, it is desirable for LLMs to be controllable at inference time, so that they can be used in multiple contexts with diverse needs. We illustrate this with the Pink Elephant Problem: instructing an LLM to avoid discussing a certain entity (a ``Pink Elephant''), and instead discuss a preferred entity (``Grey Elephant''). We apply a novel simplification of Constitutional AI, Direct Principle Feedback, which skips the ranking of responses and uses DPO directly on critiques and revisions. Our results show that after DPF fine-tuning on our synthetic Pink Elephants dataset, our 13B fine-tuned LLaMA 2 model significantly outperforms Llama-2-13B-Chat and a prompted baseline, and performs as well as GPT-4 in on our curated test set assessing the Pink Elephant Problem.