Explicit Inductive Inference using Large Language Models

Paper · arXiv 2408.14467 · Published August 26, 2024
Natural Language InferenceArgumentationLinguistics, NLP, NLU

However, recently McKenna et al. (2023a) has pointed out that LLMs are severely affected by an attestation bias when performing inference tasks. Given the question of whether premise P entails hypothesis H with few-shot examples, an LLM’s prediction is deeply bound to the hypothesis’ out-of-context truthfulness, instead of its conditional truthfulness entailed by the premise. When the hypothesis H is attested in an LLM’s world knowledge (the LLM believes H to be true), the LLM is likely to predict the entailment to be true, regardless of the premise. As a result, LLMs suffer a significant performance drop when the entailment labels disagree with the attestation label of hypothesis H.

Although this is a severe problem limiting LLMs’ performance on non-attested inferences, we argue that with careful design, this bias can instead be exploited to improve LLM performance on inference tasks. We notice a statistically true conclusion: Given an entailment inquiry P |= H, the attestation bias is harmful only when the premise P is not attested. If we control P to always be attested, then P |= H will naturally share the same truth label with H on a distributional basis, which dissolves the negative effects of LLMs’ attestation bias.

Applying this idea, we propose a simple yet effective Explicit Inductive Inference pipeline with LLMs. As illustrated in Figure 1, the core idea is to transform a premise into a set of attested alternatives by replacing the arguments, and to aggregate the LLM’s predictions on these derived inquiries to support answering the original question.

  1. Factual ̸= Attested. Factual knowledge from external sources may not be confirmed by LLMs for being longtail, absent in pre-training data, or conflicted with out-of-date records. Facts generated by LLMs, on the other hand, are highly likely to be recognizable by themselves. Even hallucinated generations are acceptable since they are still attested if not factual.

Based on these understandings, we propose the Explicit InDuctive Inference (EIDI) pipeline. Given an entailment inquiry P |= H, our EIDI pipeline first transforms P into a set of different attested premises P′s by replacing the arguments of P. Then the corresponding set of H′s is derived, so that we now have a list of alternative inquiries P′ |= H′.