LLMs keep a linear representation of contextual truth that persists across output policies and is causally steerable, and partner assertions shift it, more so for propositions near the decision boundary

- A mass-mean (difference-in-means) probe fit on ExploreToM story-proposition pairs separates contextually true from false propositions at test accuracies of 74.4 to 89.4 percent across Llama-2-13B/70B and Qwen3-14B-Base/14B/32B/VL-32B-Instruct, against about 50 percent for random-direction and shuffled-label nulls [sarkar-etal-2026-language-models-encode-the-contextual-truth-of-propositions] - The direction transfers across five task instructions (Qwen3-32B within-template accuracy 86.3, across-template 86.6), including a num-characters task that does not require computing truth, and transfers between ExploreToM and SNLI, more strongly in Qwen than Llama models [sarkar-etal-2026-language-models-encode-the-contextual-truth-of-propositions] - Steering along the direction shifts output probabilities with normalized indirect effects of roughly 0.3 to 1.2 (Llama-2-13B, Qwen3-14B, Qwen3-32B); unknown propositions fall on the false side but closer to the boundary than explicitly false ones [sarkar-etal-2026-language-models-encode-the-contextual-truth-of-propositions] - In a two-LLM spot-the-difference dialogue (Qwen3-VL-32B-Instruct), partner assertions move proposition projections toward the decision boundary; incorrectly accommodated propositions start closer to it, and crossing is 2.59 times more common (58.7 vs 22.7 percent of cases) when Player 2 restates the false claim; the authors do not claim a mechanistic link [sarkar-etal-2026-language-models-encode-the-contextual-truth-of-propositions]