What LLMs explain is not what they believe: Evaluating explanation sufficiency under models' own input beliefs

2026-06-26

AI Summary

It remains unclear whether the free-text explanations generated by current large language models (LLMs) genuinely reflect their internal reasoning processes, particularly due to the absence of methods that evaluate explanatory adequacy under the model’s own beliefs. This work introduces the first formalization of “self-consistent sufficiency” and proposes SCSuff, an information-theoretic measure that assesses explanation quality without relying on external biases by generating counterfactual inputs based on the LLM’s own beliefs. Integrating conditional generation, hidden state analysis, and perturbation testing, experiments reveal that LLM-generated explanations are generally insufficient and only weakly correlated with model scale, accuracy, or output entropy. Notably, SCSuff aligns with perturbation-based findings and can be predicted from the hidden states of the final output token, offering a promising new direction for improving the reliability of model explanations.

Abstract

Large language models (LLMs) are increasingly deployed in high-stakes domains, where free-text explanations such as chain-of-thought and post-hoc rationales are used to justify model outputs. Yet it remains unclear whether these explanations are sufficient, i.e., if they contain enough information to explain the model's output-generating process. We generalize classical sufficiency from feature attributions to arbitrary explanations and prove that explanation sufficiency can change depending on the input distribution, which must be explicitly defined for LLM explanations. We propose using the LLM itself to generate alternative inputs conditioned on an explanation, capturing its beliefs about possible inputs. We formalize self-consistent sufficiency as a goal for free-text explanations and introduce an information-theoretic metric, SCSuff, that enables evaluation of free-text explanations without relying on predefined biases or shortcuts. Our experiments show that SCSuff agrees with targeted perturbation tests where applicable and demonstrate that explanation sufficiency can vary with the input distribution. We find LLM explanations are generally insufficient and weakly correlated with model size, accuracy, or output entropy. Analysis of final-token hidden states shows that top and bottom SCSuff scores can be predicted from internal representations, suggesting that SCSuff can guide detection and improvement of sufficient LLM explanations. The code for this paper is available at https://github.com/rajesh-lab/self-consistent-sufficiency .

Problem

• explanation sufficiency
• large language models
• input beliefs
• free-text explanations
• self-consistent sufficiency

Innovation

• explanation sufficiency
• self-consistent sufficiency
• large language models
• information-theoretic metric
• input beliefs