Skip to content
AI News HubLIVE
Original source2 min read

The Geometry of Ignorance: LLMs Know When to Temper Bayesian Priors

Summary

An arXiv paper identifies a “direction of ignorance” in the unembedding matrix of LLMs that encodes the training-corpus unigram distribution as a Bayesian prior. Projecting the final prediction state onto this direction yields a per-token prior loading factor λ that declines as context becomes more informative, and decomposes the prediction into the two factors of a tempered Bayesian update. The structure appears across Llama, Qwen, Gemma, and Pythia (0.4B–405B parameters) and is causally active: adjusting λ steers predictions toward or away from the unigram prior in KL divergence.

SourcearXiv Machine LearningAuthor: Toni J. B. Liu, Jiajun Bao, Yizhou Liu, Gurbir Arora, Nicolas Boull\'e, Rapha\"el Sarfati, Christopher J. Earls
The Geometry of Ignorance: LLMs Know When to Temper Bayesian Priors
Report an error

The correction channel is not available yet. You can copy the article reference below for later.

Correction instructions
Read article

[Submitted on 2 Sep 2026]

Title:The Geometry of Ignorance: LLMs Know When to Temper Bayesian Priors

View a PDF of the paper titled The Geometry of Ignorance: LLMs Know When to Temper Bayesian Priors, by Toni J.B. Liu and 6 other authors

View PDF HTML (experimental)

Abstract:What does a language model predict when it has few clues? The answer lurks 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 --- which we term the \emph{direction of ignorance} --- appears in all four model families examined (\texttt{Llama}, \texttt{Qwen}, \texttt{Gemma}, and \texttt{Pythia}), ranging from 0.4B to 405B parameters. Projecting the final prediction state onto this direction yields a per-token \emph{prior loading factor} $\lambda$, which, empirically, declines steadily as the context becomes more informative. Formally, the same projection decomposes the prediction state into two orthogonal vectors that correspond exactly to the two factors of a tempered Bayesian update: a unigram prior raised to the exponent $\lambda$ and a context-driven likelihood. This geometric-probabilistic interpretation calibrates $\lambda$, making it meaningfully comparable across model sizes and families, with larger models generally exhibiting lower prior reliance in the high-context limit. Finally, we show that the direction of ignorance is causally active: raising or lowering $\lambda$ at the final prediction state steers the prediction toward or away from the unigram prior in KL divergence.

Subjects:

Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Computation and Language (cs.CL); Machine Learning (stat.ML)

Cite as: arXiv:2609.02959 [cs.LG]

(or arXiv:2609.02959v1 [cs.LG] for this version)

https://doi.org/10.48550/arXiv.2609.02959

arXiv-issued DOI via DataCite

Submission history

From: Jianbang Liu [view email] [v1] Wed, 2 Sep 2026 05:05:50 UTC (6,122 KB)

Full-text links:

Access Paper:

View a PDF of the paper titled The Geometry of Ignorance: LLMs Know When to Temper Bayesian Priors, by Toni J.B. Liu and 6 other authors

View PDF

HTML (experimental)

TeX Source

view license

Current browse context:

cs.LG

new | recent | 2026-09

Change to browse by:

cs cs.AI cs.CL stat stat.ML

References & Citations

NASA ADS

Google Scholar

Semantic Scholar

Loading...

Data provided by:

Bibliographic Tools

Bibliographic and Citation Tools

Bibliographic Explorer Toggle

Bibliographic Explorer (What is the Explorer?)

Connected Papers Toggle

Connected Papers (What is Connected Papers?)

Litmaps Toggle

Litmaps (What is Litmaps?)

scite.ai Toggle

scite Smart Citations (What are Smart Citations?)

Code, Data, Media

Code, Data and Media Associated with this Article

alphaXiv Toggle

alphaXiv (What is alphaXiv?)

Links to Code Toggle

CatalyzeX Code Finder for Papers (What is CatalyzeX?)

DagsHub Toggle

DagsHub (What is DagsHub?)

GotitPub Toggle

Gotit.pub (What is GotitPub?)

Huggingface Toggle

Hugging Face (What is Huggingface?)

ScienceCast Toggle

ScienceCast (What is ScienceCast?)

Demos

Demos

Replicate Toggle

Replicate (What is Replicate?)

Spaces Toggle

Hugging Face Spaces (What is Spaces?)

Spaces Toggle

TXYZ.AI (What is TXYZ.AI?)

Related Papers

Recommenders and Search Tools

Link to Influence Flower

Influence Flower (What are Influence Flowers?)

Core recommender toggle

CORE Recommender (What is CORE?)

IArxiv recommender toggle

IArxiv Recommender (What is IArxiv?)

Author

Venue

Institution

Topic

About arXivLabs

arXivLabs: experimental projects with community collaborators

arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.

Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.

Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.

Which authors of this paper are endorsers? | Disable MathJax (What is MathJax?)

Key points and analysis

Article intelligence

InvestorsAdvanced

Key points

  • A single direction in the unembedding matrix encodes the training-corpus unigram distribution and serves as the Bayesian prior LLMs fall back on when uncertain.
  • Projecting onto this direction gives a per-token prior loading factor λ that declines with context informativeness and maps to a tempered Bayesian update.
  • The structure appears across Llama, Qwen, Gemma, and Pythia models from 0.4B to 405B parameters, with larger models relying less on the prior in high-context settings.
  • Causal experiments show that raising or lowering λ steers predictions toward or away from the unigram prior in KL divergence.

Highlights and analysis are generated automatically and may contain errors. Check the original source.