CogniBench: A Legal-inspired Framework and Dataset for Assessing Cognitive Faithfulness of Large Language Models

TL;DR

CogniBench introduces a legal-inspired framework and dataset for evaluating the faithfulness of cognitive statements generated by large language models, addressing a gap in existing benchmarks that overlook inference-based hallucinations.

Contribution

The paper presents a novel framework inspired by legal evidence assessment and a large-scale dataset for detecting cognitive hallucinations in LLMs, along with an automatic annotation pipeline.

Findings

Developed CogniBench dataset revealing insightful statistics.

Created an automatic annotation pipeline for scalable dataset generation.

Facilitated training of detectors for factual and cognitive hallucinations.

Abstract

Faithfulness hallucinations are claims generated by a Large Language Model (LLM) not supported by contexts provided to the LLM. Lacking assessment standards, existing benchmarks focus on "factual statements" that rephrase source materials while overlooking "cognitive statements" that involve making inferences from the given context. Consequently, evaluating and detecting the hallucination of cognitive statements remains challenging. Inspired by how evidence is assessed in the legal domain, we design a rigorous framework to assess different levels of faithfulness of cognitive statements and introduce the CogniBench dataset where we reveal insightful statistics. To keep pace with rapidly evolving LLMs, we further develop an automatic annotation pipeline that scales easily across different models. This results in a large-scale CogniBench-L dataset, which facilitates training accurate detectors for both factual and cognitive hallucinations. We release our model and datasets at: https://github.com/FUTUREEEEEE/CogniBench