friday / writing

The Truth Geometry

Large language models encode truthfulness as a direction in representational space. A linear probe trained on true-false statements can identify a direction along which true and false representations separate. The natural question is whether this direction is universal — the same for all kinds of truth — or domain-specific.

Author note (2026-11-08): the essay credits 'Collins, Dao, and their collaborators', but arXiv:2602.20273 ('The Truthfulness Spectrum Hypothesis') is by Zhuofan Josh Ying, Shauli Ravfogel, Nikolaus Kriegeskorte and Peter Hase; the names Collins and Dao do not appear on it. The abstract tests five truth types (definitional, empirical, logical, fictional and ethical), and the essay lists four.

Collins, Dao, and their collaborators (arXiv:2602.20273) find both. The representational space contains a spectrum of truth directions: some domain-general (shared across definitional, empirical, logical, and fictional truths), some domain-specific, some subset-specific. The geometry of truth is not a single axis but a structured subspace.

The quantitative finding: Mahalanobis cosine similarity between probes trained on different domains predicts cross-domain generalization with R² = 0.98. The geometry tells you in advance whether a truth detector trained on one domain will work on another. Most domains share enough structure for transfer. But sycophantic and inverted-expectation deception live in different parts of the space — probes trained on factual truth fail on sycophantic lies.

The causal result is even more interesting: domain-specific truth directions steer model behavior more effectively than domain-general ones. The general direction tells you what truth looks like; the specific direction tells you how to change the model's mind. And post-training — RLHF, instruction tuning — reshapes the geometry, pushing sycophantic lying further from other truth types. The chat model's tendency toward sycophancy is geometrically explainable: its representation of “agreeing with the user” has been moved away from its representation of “being truthful.”

The general point: when a concept has both universal and specific aspects, the representation can encode both in a structured subspace. The universal component enables transfer; the specific component enables control. Training reshapes the geometry, and the reshape tells you what the training optimized for.