# When Do Internal Probes Beat Reading the Answer? Miscalibrated Readouts and Behavior-Concealed Knowledge in Language Models

arXiv: https://arxiv.org/abs/2609.04582 · HTML: https://mlsanity.com/paper/2609.04582/

## Key facts

- Cited by 1 papers in the map (data through 14 Oct 2026).
- Builds on 19 of its 29 references that are in the map.
- First posted on arXiv 4 Sep 2026.

## Builds on (most cited first, 19 of 19)

- [Language Models are Few-Shot Learners](https://mlsanity.com/paper/2005.14165/) (2020)
- [Qwen3 Technical Report](https://mlsanity.com/paper/2505.09388/) (2025)
- [On Calibration of Modern Neural Networks](https://mlsanity.com/paper/1706.04599/) (2017)
- [Language Models (Mostly) Know What They Know](https://mlsanity.com/paper/2207.05221/) (2022)
- [Calibrate Before Use: Improving Few-Shot Performance of Language Models](https://mlsanity.com/paper/2102.09690/) (2021)
- [Inference-Time Intervention: Eliciting Truthful Answers from a Language Model](https://mlsanity.com/paper/2306.03341/) (2023)
- [Towards Understanding Sycophancy in Language Models](https://mlsanity.com/paper/2310.13548/) (2023)
- [Discovering Latent Knowledge in Language Models Without Supervision](https://mlsanity.com/paper/2212.03827/) (2022)
- [The Geometry of Truth: Emergent Linear Structure in Large Language Model Representations of True/False Datasets](https://mlsanity.com/paper/2310.06824/) (2023)
- [Designing and Interpreting Probes with Control Tasks](https://mlsanity.com/paper/1909.03368/) (2019)
- [C-Pack: Packed Resources For General Chinese Embeddings](https://mlsanity.com/paper/2309.07597/) (2023)
- [Large Language Models Are Not Robust Multiple Choice Selectors](https://mlsanity.com/paper/2309.03882/) (2023)
- [SmolLM2: When Smol Goes Big -- Data-Centric Training of a Small Language Model](https://mlsanity.com/paper/2502.02737/) (2025)
- [Surface Form Competition: Why the Highest Probability Answer Isn't Always Right](https://mlsanity.com/paper/2104.08315/) (2021)
- [Semantic Entropy Probes: Robust and Cheap Hallucination Detection in LLMs](https://mlsanity.com/paper/2406.15927/) (2024)
- [LLMs Know More Than They Show: On the Intrinsic Representation of LLM Hallucinations](https://mlsanity.com/paper/2410.02707/) (2024)
- [A Theory of Usable Information Under Computational Constraints](https://mlsanity.com/paper/2002.10689/) (2020)
- [Beyond Performance: Quantifying and Mitigating Label Bias in LLMs](https://mlsanity.com/paper/2405.02743/) (2024)
- [The Geometry of Reasoning: Flowing Logics in Representation Space](https://mlsanity.com/paper/2510.09782/) (2025)

## Built on by (most cited first, 1 of 1)

- [FunDM: Cheap & Label-Free Decision Models from Frozen Language Models](https://mlsanity.com/paper/2610.17175/) (2026)

## Abstract (arXiv)

A 0.6B language model, asked to verify 1,200 logical conclusions (half valid, half corrupted by a single semantic edit), answers YES every time. Judged by behavior it discriminates nothing; linear probes on its hidden states read the correct verdict at 0.96 AUC, transferring to unseen logical structures and separating foils built from exactly the words of the true conclusion (0.90). We ask where the verdict is lost, and find the dominant failure is a single scalar. The verdict survives to the model's own output logits (margin AUC 0.89) along a well-aligned readout direction; a saturated decision threshold, offset by +4.6 sigma, erases it. The diagnosis generalizes: across 90 semantic-label configurations of a five-model, three-family factorial, behavioral accuracy collapses onto a single function of threshold offset (Spearman -0.93) while margin ranking moves far less. Across a 13x scale range, internal knowledge saturates while free-form behavior is non-monotone: an 8B model underperforms its 4B sibling through an answer-channel failure rather than the threshold; forced-choice accuracy is monotone. The diagnosis is actionable: a one-parameter correction, never fit on evaluated structures, repairs behavior from 50% to 81% (0.6B); calibrated margin decoding recovers 94% at 8B; few-shot prompting works the same way, recentering the threshold (+4.6 sigma to 0.0 sigma) while preserving ranking. Comparing probe to margin separates three regimes: concealed, miscalibrated, and undetected. On a maze task built so foils carry no surface cues, the audit correctly reports the third. In the standard generation setting, answer-surface features and heuristic labels reproduce published probing results without any internal access.