The Direction of Ignorance: LLMs Encode and Temper a Unigram Prior
Abstract
What does a language model predict when it has few clues?The answer lies in its unembedding geometry: a single direction of the unembedding matrix encodes the unigram distribution of the training corpus, which serves as the Bayesian prior the model falls back on when uncertain. This structure, termed the direction of ignorance, appears across all model families examined (Llama, Qwen, Gemma, and Pythia), ranging from 0.4B to 405B parameters. Projecting the final prediction state onto this direction yields a per-token prior loading factor λ, which declines steadily as the context becomes more informative. Formally, the same projection decomposes the prediction state into two orthogonal vectors interpretable as the two components of a tempered Bayesian update: a unigram prior raised to the exponent λ and a context-driven likelihood. This geometric probabilistic connection makes λ meaningfully comparable across model sizes and families. Larger models generally exhibit lower prior reliance, and reasoning models exhibit lower reliance than their base counterparts. Finally, we show that λ serves as a causal, inference-time control: reducing it can promote context-driven continuations and improve Llama-3.2-3B-Instruct on GSM8K, whereas amplifying it or applying matched perturbations along random directions degrades performance.
Related
- TEMPER: Temporal Encoder-Masked Probabilistic Ensemble Regressor for Time-Series Forecasting
- Formalizing Fefferman's Official Periodic Navier-Stokes Problem (B) in Lean 4: Global Linear Strain Exclusion on T3, Biot-Savart Direction Depletion Identities, and Conditional Reductions via Beale-Kato-Majda
- Dimensional reduction associated with unchanged direction trajectories in flat fiber Robertson Walker spacetimes
- Functions determining few direction: the final missing case
- Generic factorizations and alternating direction method for fixed-rank approximation of matrix polynomials
BibTeX
@misc{liu2026direction,
title = {{The Direction of Ignorance: LLMs Encode and Temper a Unigram Prior}},
author = {Toni J. B. Liu and Jiajun Bao and Gurbir Arora and Mitchell Ostrow and Yizhou Liu and Raphaël Sarfati and Nicolas Boullé and Christopher J. Earls},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2609.02959},
eprint = {2609.02959},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2609.02959},
}
RIS
TY - UNPB
TI - The Direction of Ignorance: LLMs Encode and Temper a Unigram Prior
AU - Toni J. B. Liu
AU - Jiajun Bao
AU - Gurbir Arora
AU - Mitchell Ostrow
AU - Yizhou Liu
AU - Raphaël Sarfati
AU - Nicolas Boullé
AU - Christopher J. Earls
PY - 2026
DA - 2026/10/29
PB - arXiv
DO - 10.48550/arxiv.2609.02959
UR - https://arxiv.org/abs/2609.02959
AB - What does a language model predict when it has few clues?The answer lies in its unembedding geometry: a single direction of the unembedding matrix encodes the unigram distribution of the training corpus, which serves as the Bayesian prior the model falls back on when uncertain. This structure, termed the direction of ignorance, appears across all model families examined (Llama, Qwen, Gemma, and Pythia), ranging from 0.4B to 405B parameters. Projecting the final prediction state onto this direction yields a per-token prior loading factor λ, which declines steadily as the context becomes more informative. Formally, the same projection decomposes the prediction state into two orthogonal vectors interpretable as the two components of a tempered Bayesian update: a unigram prior raised to the exponent λ and a context-driven likelihood. This geometric probabilistic connection makes λ meaningfully comparable across model sizes and families. Larger models generally exhibit lower prior reliance, and reasoning models exhibit lower reliance than their base counterparts. Finally, we show that λ serves as a causal, inference-time control: reducing it can promote context-driven continuations and improve Llama-3.2-3B-Instruct on GSM8K, whereas amplifying it or applying matched perturbations along random directions degrades performance.
ER -