cs.CLNov 5, 2026

Decodability Is Not Validity: Evaluating Nuisance Sensitivity in Language Model Concept Probes

Authors: Ayal Klein

Abstract

Two concept probes can be equally accurate at predicting a concept from a language model's hidden states while their scores contain different mixtures of the target concept and a correlated nuisance attribute. In four target--nuisance settings (two aspect sentiments in CEBaB; occupation and gender in BiasBios; Amazon review sentiment with negation, and with first-person framing), two LLMs, and a population of concept extractors, we compare decodability with discriminant evidence: the score's dependence on the nuisance at fixed target, and its response to input edits intended to change only the nuisance, filtered by two LLM judges and compared against placebo edits. Under correlated training data, extractors that are statistically equivalent in held-out AUROC differ up to several-fold in nuisance sensitivity in three of the four settings, and decodability is not a reliable proxy for it: their rank correlation changes sign across settings. The effect follows the probe's training data, not the nuisance's natural association with the target: first-person framing, unrelated to sentiment in natural reviews, is absorbed at least as much as negation, which co-occurs with negative sentiment, and decorrelated training removes the systematic effect. Selecting extractors with nuisance-aware development diagnostics, which need nuisance labels but no counterfactuals, lowers counterfactual nuisance sensitivity wherever the default choice is sensitive, at a small accuracy cost. Concept measurements should be validated, not just decoded.

Figures & tables

Appendix figures & tables13 assets

Supplementary material from the paper’s appendix.

Appendix

Explore similar work

Nov 14, 2026

Confidence Without Cause: Predicting and Localizing Physics-Shortcut Sensitivity in Language Model Hidden States

A language model's confidence on a physics problem deserves trust only if it ignores details irrelevant to the answer. We introduce DECOY (Distractor-Embedded Certified cOunterfactuals for phYsics): 155 families of four prompts differing in one irrelevant detail. A symbolic solver certifies that each edit preserves the answer. Families are labelled by the largest drop in the model's log-probability of the certified answer, using a threshold fixed before probe training. Qwen2.5-7B-Instruct is sensitive on 50.0% of the 140 families yielding labels, and Llama-3.1-8B-Instruct on 22.9%. Effects are largest when distractor values lie numerically close to the answer. A linear probe on Qwen's layer-18 hidden state, averaged over four renderings, predicts sensitivity before generation at AUROC 0.760 with whole families held out. A separately tuned single-rendering probe reaches 0.747. A frozen layer-16 probe reaches 0.820 pooled AUROC on 15 families built after training, though within-cue-type AUROC is lower (0.667-0.6875). Under a paired family bootstrap, the probe performs on par with a fitted token-entropy baseline (AUROC 0.727) and achieves the lowest Brier score among tested predictors (0.204 versus 0.214). Knocking out four attention heads at layer 16 cancels sensitivity on a 20-family panel, localising it to a narrow mechanism. These results show that Qwen's confidence can track irrelevant details, that this failure is detectable before generation, and that targeted intervention can suppress it. DECOY provides certified ground truth and paired-tested baselines for evaluating hidden-state monitors.
Sep 4, 2026

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

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.
Sep 24, 2026

Encoded but Not Decoded: Layer-Localized Evidence for a Three-Level Gap in LLM Syntax

A language model can fail a syntactic test in two distinct ways: by not encoding the relevant structure, or by encoding it but failing to use it at the output. Behavioral evaluation alone cannot tell these apart. We propose a three-level evaluation framework (behavioral deployment, LM-head readout, and probe recoverability) measured on the same items under the same binary decision. Using a compact trilingual (English, Chinese, German) control-dependency benchmark, we find that probe recoverability exceeds or equals LM-head readout, which in turn exceeds or equals behavioral deployment, across seven models and all three languages in the aggregate. The recoverability surplus is never negative across all 14 (model, task) conditions. The disconnect concentrates in subject-control, where a nearest-noun heuristic gives the wrong answer. The single largest gap (0.653) appears on Qwen3-0.6B Instruct in question answering. The gap persists at Qwen3-14B Instruct. Instruction tuning degrades deployment more than encoding in percentage terms. We rule out option-position bias, late-layer erasure, output-formatting artifacts, and probe-training variance. The pattern is consistent with decoding that favors surface shortcuts, and the behavior-probe gap measures the strength of that preference. Activation patching shows the gap is layer-localized. Under instruction tuning, the LM-head-decoded layer shifts approximately ten layers later than the probe-decoded layer. These findings argue that behavioral evaluation understates what models encode, while probing alone overstates what they deploy.