Pith. sign in

REVIEW 6 cited by

Inside-Out: Hidden Factual Knowledge in LLMs

Not yet reviewed by Pith; the record is open.

This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.

SPECIMEN: schema-true, not a live event

T0 review · schema-true

One-sentence machine reading of the paper's core claim.

pith:XXXXXXXX · record.json · timestamp

arxiv 2503.15299 v4 pith:76FT3KBH submitted 2025-03-19 cs.CL

Inside-Out: Hidden Factual Knowledge in LLMs

classification cs.CL
keywords knowledgellmsanswerfactualhiddentheyanswersclosed-book
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved
0 comments
Share X Bluesky LinkedIn Reddit HN
read the original abstract

This work presents a framework for assessing whether large language models (LLMs) encode more factual knowledge in their parameters than what they express in their outputs. While a few studies hint at this possibility, none has clearly defined or demonstrated this phenomenon. We first propose a formal definition of knowledge, quantifying it for a given question as the fraction of correct-incorrect answer pairs where the correct one is ranked higher. This gives rise to external and internal knowledge, depending on the information used to score individual answer candidates: either the model's observable token-level probabilities or its intermediate computations. Hidden knowledge arises when internal knowledge exceeds external knowledge. We then present a case study, applying this framework to three popular open-weights LLMs in a closed-book QA setup. Our results indicate that: (1) LLMs consistently encode more factual knowledge internally than what they express externally, with an average relative gap of 40%. (2) Surprisingly, some knowledge is so deeply hidden that a model can internally know an answer perfectly, yet fail to generate it even once, despite large-scale repeated sampling of 1,000 answers. This reveals fundamental limitations in the generation capabilities of LLMs, which (3) put a practical constraint on scaling test-time compute via repeated answer sampling in closed-book QA: significant performance improvements remain inaccessible because some answers are practically never sampled, yet if they were, we would be guaranteed to rank them first.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 6 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Improving LLMs via Validator-to-Generator Alignment

    cs.CL 2026-07 conditional novelty 7.0

    Frequency-corrected rank alignment of an LLM generator to its own validator improves generator AUROC and G-V Pearson correlation by up to 27 points while preserving validator quality.

  2. Cross-Lingual Exploration for Parametric Knowledge

    cs.CL 2026-06 unverdicted novelty 6.0

    Cross-lingual prompt exploration improves factual recall and consistency in LLMs across 17 languages more efficiently than native-language scaling.

  3. DECOR: Auditing LLM Deception via Information Manipulation Theory

    cs.CL 2026-05 unverdicted novelty 6.0

    DECOR introduces a theory-grounded multi-agent system that decomposes contexts into atomic units, scores four manipulation dimensions per unit, and aggregates profiles into a global deception index, reporting SOTA res...

  4. Hallucinations Undermine Trust; Metacognition is a Way Forward

    cs.CL 2026-05 unverdicted novelty 6.0

    LLMs need metacognition to align expressed uncertainty with their actual knowledge boundaries, moving beyond knowledge expansion to reduce confident errors.

  5. Input-Time Scaling: Adding Noise and Irrelevance into Less-Is-More Drastically Improves Reasoning Performance and Efficiency

    cs.LG 2025-08 unverdicted novelty 6.0

    Adding controlled noise and irrelevant persona contexts across training and testing stages for strong LLMs yields better reasoning and efficiency than high-quality data alone, reaching 76.7% on AIME24/25 with Qwen2.5-32B.

  6. Improving Factuality in LLMs via Inference-Time Knowledge Graph Construction

    cs.CL 2025-08 unverdicted novelty 5.0

    A framework for inference-time knowledge graph construction and expansion improves factual accuracy in LLMs on three QA benchmarks by combining internal LLM knowledge with selective external retrieval.