Language Models Encode the Contextual Truth of Propositions

2026-08-03

AI Summary

This study investigates how large language models represent the truth value of propositions in contexts that rely on contextual evidence rather than world knowledge, and how such representations are influenced by interlocutors’ assertions. Leveraging linear probing in activation space, causal steering interventions, and multi-turn collaborative vision-language dialogue data, the work reveals—for the first time—a stable linear representation of contextual truth values within the model. It further distinguishes two types of output-invisible conformity: superficial agreement, where internal truth representations remain unchanged, and deep alignment, characterized by a crossing of the truth-value decision boundary. The findings show that propositions near the decision boundary are more susceptible to social influence, and when explicitly agreeing with false claims, models cross the truth-value boundary 2.59 times more frequently than during implicit conformity.

Abstract

Prior work has shown that LLMs encode the truth of factual propositions along linear directions in activation space. It's unclear how these representations extend to contextual truth: propositions whose truth is determined by in-context evidence rather than world knowledge. We show that LLMs maintain a linear representation of contextual truth that persists across structurally different output policies, even when the output doesn't require the model to determine a proposition's truth, and show causal evidence via steering experiments. Using the transcripts from a collaborative vision-language task that requires two LLMs to maintain a shared common ground, we show that truth representations of a proposition are significantly swayed by partner assertions about that proposition, even when the LLM has enough evidence to determine its truth. We find evidence that propositions near the decision boundary are more susceptible to having their truth shifted through partner assertions. Separating representation from output distinguish two forms of sycophancy that output behavior alone cannot: the model may accommodate a false proposition while continuing to represent it as false, or shift its representation across the boundary. The latter is 2.59x more common when the model agrees by restating the false claim explicitly than when it agrees implicitly.

Problem

• contextual truth
• language models
• truth representation
• sycophancy
• common ground

Innovation

• contextual truth
• linear representation
• activation space
• sycophancy
• steering experiments