Position: It's Time to Optimize LLMs for Self-Consistency

2026-07-31

AI Summary

Current large language models often generate responses that are overly accommodating, logically incomplete, or confidently incorrect when guided by user inputs, primarily because prevailing training paradigms evaluate individual outputs in isolation and neglect consistency across responses to different but related inputs. This work establishes self-consistency as a central optimization objective for the first time, proposing a unified framework that subsumes capabilities such as adversarial robustness and factual coherence as special cases of consistency optimization. By introducing systematic analysis of cross-input response relationships and leveraging standard optimization tools, the framework enhances the model’s intrinsic consistency, thereby offering a principled theoretical pathway and methodological foundation for developing large language models that are logically rigorous, robust to perturbations, and trustworthy.

Abstract

Despite ever-increasing sophistication in language model (LM) pre- and post-training pipelines, many important failures persist: models overcondition on user framing ("sycophancy"), exhibit incomplete logical generalization, and produce confident but incorrect responses. We argue that these failures arise from a modeling assumption permeating all aspects of the pipeline: that behavior can be specified and evaluated independently on single-output pairs. Many model failures are difficult, if not impossible, to detect without reasoning about relationships between a model's responses across inputs. In this position paper, we propose self-consistency as a framework for understanding these failures. We first observe that a wide variety of techniques designed to improve specific aspects of LM behavior-targeting properties as diverse as adversarial robustness and factual coherence-can be understood as special cases of a common "consistency optimization" procedure and addressed with a standard set of optimization tools. We next outline a set of new model properties that could be achieved by optimizing for consistency, and conclude with a discussion of what it would mean to develop generally consistent LMs, including the capabilities they would enable and the objections they raise.

Problem

• self-consistency
• language models
• sycophancy
• logical generalization
• model failures

Innovation

• self-consistency
• consistency optimization
• language models
• behavioral alignment
• logical generalization