Language Model Probing

Ask about this topic

Overview

Language model probing studies what models represent internally and what they can do, using controlled tasks to examine hidden activations, attention, logits, and generated behavior. Common approaches include training linear or nonlinear readouts on activations, comparing representations across layers and prompts, and testing causal effects with interventions such as activation patching or steering. Probes are used to investigate linguistic knowledge, reasoning, uncertainty, hallucination, safety, and model performance, and can support practical tools for monitoring, routing, and evaluation. A central open tension is that decodable information may not be used to produce an answer, while probe results can also reflect prompt structure, labels, or dataset artifacts rather than the intended capability.

Latest papers

CardsList

280 papers35 papers in the last four weeks, up 35% on the four weeks before. 0.2% of all new papers.

Apr 1, 2026

Fast and Accurate Probing of In-Training LLMs' Downstream Performances

The paradigm of scaling Large Language Models (LLMs) in both parameter size and test time has pushed the boundaries of AI capabilities, but at the cost of making the traditional generative evaluation paradigm prohibitively expensive, therefore making the latency of LLM's in-training downstream performance evaluation unbearable. However, simple metrics like training loss (perplexity) are not always correlated with downstream performance, as sometimes their trends diverge from the actual task outcomes. This dilemma calls for a method that is computationally efficient and sufficiently accurate in measuring model capabilities. To address this challenge, we introduce a new in-training evaluation paradigm that uses a lightweight probe for monitoring downstream performance. The probes take the internal representations of LLM checkpoints (during training) as input and directly predict the checkpoint's performance on downstream tasks measured by success probability (i.e., pass@1). We design several probe architectures, validating their effectiveness using the OLMo3-7B's checkpoints across a diverse set of downstream tasks. The probes can accurately predict a checkpoint's performance (with avg. AUROC>>0.75), have decent generalizability across checkpoints (earlier predicts later), and reduce the computation latency from ∼\sim1 hr (using conventional generative evaluation method) to ∼\sim3 min. In sum, this work presents a practical and scalable in-training downstream evaluation paradigm, enabling a more agile, informed, and efficient LLM development process.

Zhichen Liu, Tianle Lun, Zhibin Wen +7

Southern University of Science and Technology · Huawei

LLM EvaluationLanguage Model Probing
AskPDF
Mar 31, 2026

LLM Probe: Evaluating LLMs for Low-Resource Languages

Despite rapid advances in large language models (LLMs), their linguistic abilities in low-resource and morphologically rich languages are still not well understood due to limited annotated resources and the absence of standardized evaluation frameworks. This paper presents LLM Probe, a lexicon-based assessment framework designed to systematically evaluate the linguistic skills of LLMs in low-resource language environments. The framework analyzes models across four areas of language understanding: lexical alignment, part-of-speech recognition, morphosyntactic probing, and translation accuracy. To illustrate the framework, we create a manually annotated benchmark dataset using a low-resource Semitic language as a case study. The dataset comprises bilingual lexicons with linguistic annotations, including part-of-speech tags, grammatical gender, and morphosyntactic features, which demonstrate high inter-annotator agreement to ensure reliable annotations. We test a variety of models, including causal language models and sequence-to-sequence architectures. The results reveal notable differences in performance across various linguistic tasks: sequence-to-sequence models generally excel in morphosyntactic analysis and translation quality, whereas causal models demonstrate strong performance in lexical alignment but exhibit weaker translation accuracy. Our results emphasize the need for linguistically grounded evaluation to better understand LLM limitations in low-resource settings. We release LLM Probe and the accompanying benchmark dataset as open-source tools to promote reproducible benchmarking and to support the development of more inclusive multilingual language technologies.

Hailay Kidu Teklehaymanot, Gebrearegawi Gebremariam, Wolfgang Nejdl

L3S Research Center, Leibniz University Hannover, Germany · Aksum University, Ethiopia

Multilingual Language Model EvaluationLow-Resource Language ProcessingLanguage Model Probing
AskPDF
Mar 28, 2026

WMF-AM: Probing LLM Working Memory via Depth-Parameterized Cumulative State Tracking

Existing large language models (LLMs) evaluations use fixed-difficulty benchmarks that cannot adapt as models improve, and rarely isolate specific cognitive processes. We introduce Working Memory Fidelity-Active Manipulation (WMF-AM), a probe of cumulative state tracking, the ability to maintain and update intermediate results across K sequential operations within a single query, without a scratchpad. Unlike multi-step agent benchmarks that stress task orchestration, WMF-AM isolates within-pass cumulative load by parameterizing depth K. The core probe uses arithmetic accumulation on 28 models from 12 families (0.5B to frontier); a matched non-arithmetic extension (permissions, schedules, inventories) confirms the design generalizes beyond arithmetic. Three construct-isolation ablations confirm that cumulative load, not arithmetic skill or entity tracking, drives difficulty. We release WMF-AM as a lightweight, recalibratable diagnostic for characterizing where models degrade under cumulative load. Code and data can be accessed at https://github.com/dengzhe-hou/WMF-AM

Dengzhe Hou, Lingyu Jiang, Deng Li +3

Tohoku University · Lappeenranta-Lahti University of Technology LUT · Texas A&M University +1

Working Memory in Language ModelsLLM EvaluationLanguage Model Probing
AskPDF
Mar 25, 2026

ReLope: From Hidden-State Probing to a Decision Module for Multimodal LLM Routing

Routing balances performance and cost in hybrid systems by escalating selected queries from a lightweight model to a powerful but expensive one. Hidden-state probes provide an effective routing signal by predicting the small model's correctness from hidden states it already computes. Although effective in text-only settings, their behavior on multimodal LLMs (MLLMs) is less understood. We find that correctness is harder to predict from hidden states when questions are paired with images rather than with captions of those images. Two training-free dependence measures, HSIC and label CKA, show the same pattern. We first investigate whether changing hidden states used by the probe improves correctness prediction for MLLMs. We propose Attention Probe, which learns an attention query to pool frozen token states into a routing feature and improves over the standard probe. However, pooling only recombines hidden states computed for answer generation, not for deciding whether the answer should be trusted. We therefore build a decision layer inside the existing MLLM and train it for the routing decision. The resulting router, ReLope (KL-Regularized LoRA Probe), adapts a separate copy of one MLLM layer with LoRA and regularizes it with a KL-penalized stochastic bottleneck. This decision branch shares the frozen lower layers and leaves the model's answer-generation path unchanged. ReLope attains the highest correctness prediction AUC on all five benchmarks and three backbones we evaluate and improves the accuracy versus cost trade-off at negligible overhead. These results suggest a new perspective and a practical technical path on routing: rather than reading decisions off states built for generation, routing decisions can be learned by a dedicated decision layer inside the model. Code: https://github.com/Spinozaaa/ReLope

Yaopei Zeng, Congchao Wang, Blake JianHang Chen +1

Pennsylvania State University · Google · Amazon AGI

Multimodal Large Language ModelsLLM RoutingLanguage Model Probing
AskPDF
Mar 24, 2026

Probing Ethical Framework Representations in Large Language Models: Structure, Entanglement, and Methodological Challenges

When large language models make ethical judgments, do their internal representations distinguish between normative frameworks, or collapse ethics into a single acceptability dimension? We probe hidden representations across five ethical frameworks (deontology, utilitarianism, virtue, justice, commonsense) in six LLMs spanning 4B--72B parameters. Our analysis reveals differentiated ethical subspaces with asymmetric transfer patterns -- e.g., deontology probes partially generalize to virtue scenarios while commonsense probes fail catastrophically on justice. Disagreement between deontological and utilitarian probes correlates with higher behavioral entropy across architectures, though this relationship may partly reflect shared sensitivity to scenario difficulty. Post-hoc validation reveals that probes partially depend on surface features of benchmark templates, motivating cautious interpretation. We discuss both the structural insights these methods provide and their epistemological limitations.

Weilun Xu, Alexander Rusnak, Frederic Kaplan

School of Computer and Communication Sciences, École Polytechnique Fédérale de Lausanne · Digital Humanities Lab, École Polytechnique Fédérale de Lausanne

Moral Reasoning in Language ModelsLanguage Model ProbingLLM Interpretability
AskPDF
Mar 23, 2026

Riding Brainwaves in LLM Space: Understanding Activation Patterns Using Individual Neural Signatures

Consumer-grade EEG is entering everyday devices, from earbuds to headbands, raising the question of whether language models can be adapted to individual neural responses. We test this by asking whether frozen LLM representations encode person-specific EEG signals, directions in activation space that predict one person's brain activity but not another's. Using word-level EEG from 30 participants reading naturalistic sentences (ZuCo corpus), we train a separate linear probe for each person, mapping hidden states from a frozen Qwen 2.5 7B to that individual's EEG power. Person-specific probes outperform a single population probe on every EEG feature tested; for high-gamma power, the person-specific probe achieves rho = 0.183, a ninefold improvement over the population probe (rho = 0.020, p < 10^-4). A negative control, fixation count, shows no person-specific advantage (p = 0.360); fixation count reflects word length and frequency rather than individual cognition. The individual directions are temporally stable (split-half cosine = 0.824), non-transferable across people (self rho = 0.369 vs. other rho = 0.143, p < 10^-19), and distinct from the shared population signal: person-specific probes retain predictive power after the population component is removed. The person-specific signal concentrates in the model's deep layers, rising consistently with depth and peaking at Layer 24 of 28. The results are consistent across architectures (LLaMA 3.1 8B) and survive word-level confound controls. Frozen language models contain stable, person-specific neural directions in their deep layers, providing a geometric foundation for EEG-driven personalization.

Ajan Subramanian, Sumukh Bettadapura, Rohan Sathish

EEG DecodingLanguage Model ProbingRepresentation Geometry in Language Models
AskPDF
Mar 22, 2026

PROMPT2BOX:Improving LLM Weakness Discovery and Specificity Estimation by Uncovering Entailment Structure among Prompts

To discover the weaknesses of LLMs, researchers often embed prompts into a vector space and cluster them to extract insightful patterns. However, vector embeddings primarily capture topical similarity; as a result, prompts that share a topic but differ in specificity, and consequently in difficulty, are often represented similarly, making fine-grained weakness analysis difficult. To address this limitation, we propose Prompt2Box, which embeds prompts into a box embedding space using a trained encoder. The encoder, trained on existing and synthesized datasets, outputs box embeddings that capture not only semantic similarity but also specificity relations between prompts (e.g., "writing an adventure story" is more specific than "writing a story"). We further develop a novel dimension reduction technique for box embeddings to facilitate dataset visualization and comparison. Our experiments demonstrate that box embeddings consistently capture prompt specificity better than vector baselines and achieve 45% error reduction on average in predicting specificity compared to the prompt length baseline. On the downstream task of creating hierarchical clustering trees for 17 LLMs from the UltraFeedback dataset, Prompt2Box can identify 13.5% more LLM weaknesses than vector baselines and achieves an approximately 33% stronger correlation between hierarchical depth and instruction specificity. The code is available at https://github.com/zawedcvg/box_embeddings.

Neeladri Bhuiya, Shib Sankar Dasgupta, Andrew McCallum +1

CICS, University of Massachusetts Amherst, USA · A10 Networks · Amazon AWS AI

LLM EvaluationRepresentation LearningLanguage Model Probing
AskPDF
Mar 19, 2026

Is Evaluation Awareness Just Format Sensitivity? Limitations of Probe-Based Evidence under Controlled Prompt Structure

Prior work uses linear probes on benchmark prompts as evidence of evaluation awareness in large language models. Because evaluation context is typically entangled with benchmark format and genre, it is unclear whether probe-based signals reflect context or surface structure. We test whether these signals persist under partial control of prompt format using a controlled 2x2 dataset and diagnostic rewrites. We find that probes primarily track benchmark-canonical structure and fail to generalize to free-form prompts independent of linguistic style. Thus, standard probe-based methodologies do not reliably disentangle evaluation context from structural artifacts, limiting the evidential strength of existing results.

Viliana Devbunova

Yandex Belgrade, Serbia

Evaluation Awareness in Language ModelsLanguage Model ProbingPrompt Sensitivity
AskPDF
Mar 19, 2026

Geography According to ChatGPT -- How Generative AI Represents and Reasons about Geography

Understanding how AI will represent and reason about geography should be a key concern for all of us, as the broader public increasingly interacts with spaces and places through these systems. Similarly, in line with the nature of foundation models, our own research often relies on pre-trained models. Hence, understanding what world AI systems construct is as important as evaluating their accuracy, including factual recall. To motivate the need for such studies, we provide three illustrative vignettes, i.e., exploratory probes, in the hope that they will spark lively discussions and follow-up work: (1) Do models form strong defaults, and how brittle are model outputs to minute syntactic variations? (2) Can distributional shifts resurface from the composition of individually benign tasks, e.g., when using AI systems to create personas? (3) Do we overlook deeper questions of understanding when solely focusing on the ability of systems to recall facts such as geographic principles?

Krzysztof Janowicz, Gengchen Mai, Rui Zhu +4

University of Vienna, Austria · University of Texas at Austin, USA · University of Bristol, UK +4

Geospatial ReasoningFactual Knowledge in Language ModelsLanguage Model Probing
AskPDF
Mar 17, 2026

Catching rationalization in the act: detecting motivated reasoning before and after CoT via activation probing

Large language models (LLMs) can produce chains of thought (CoT) that do not accurately reflect the actual factors driving their answers. In multiple-choice settings with an injected hint favoring a particular option, models may shift their final answer toward the hinted option and produce a CoT that rationalizes the response without acknowledging the hint - an instance of motivated reasoning. We study this phenomenon across multiple LLM families and datasets demonstrating that motivated reasoning can be identified by probing internal activations even in cases when it cannot be easily determined from CoT. Using supervised probes trained on the model's residual stream, we show that (i) pre-generation probes, applied before any CoT tokens are generated, predict motivated reasoning as well as a LLM-based CoT monitor that accesses the full CoT trace, and (ii) post-generation probes, applied after CoT generation, outperform the same monitor. Together, these results show that motivated reasoning is detected more reliably from internal representations than from CoT monitoring. Moreover, pre-generation probing can flag motivated behavior early, potentially avoiding unnecessary generation.

Parsa Mirtaheri, Mikhail Belkin

UC San Diego

LLM InterpretabilityCoT FaithfulnessLanguage Model Probing
AskPDF
Mar 15, 2026

Mind the Shift: Decoding Monetary Policy Stance from FOMC Statements with Large Language Models

Federal Open Market Committee (FOMC) statements are a major source of monetary-policy information, and even subtle changes in their wording can move global financial markets. A central task is therefore to measure the hawkish--dovish stance conveyed in these texts. Existing approaches typically treat stance detection as a standard classification problem, labeling each statement in isolation. However, the interpretation of monetary-policy communication is inherently relative: market reactions depend not only on the tone of a statement, but also on how that tone shifts across meetings. We introduce Delta-Consistent Scoring (DCS), an annotation-free framework that maps frozen large language model (LLM) representations to continuous stance scores by jointly modeling absolute stance and relative inter-meeting shifts. Rather than relying on manual hawkish--dovish labels, DCS uses consecutive meetings as a source of self-supervision. It learns an absolute stance score for each statement and a relative shift score between consecutive statements. A delta-consistency objective encourages changes in absolute scores to align with the relative shifts. This allows DCS to recover a temporally coherent stance trajectory without manual labels. Across four LLM backbones, DCS consistently outperforms supervised probes and LLM-as-judge baselines, achieving up to 71.1% accuracy on sentence-level hawkish--dovish classification. The resulting meeting-level scores are also economically meaningful: they correlate strongly with inflation indicators and are significantly associated with Treasury yield movements. Overall, the results suggest that LLM representations encode monetary-policy signals that can be recovered through relative temporal structure.

Yixuan Tang, Yi Yang

The Hong Kong University of Science and Technology

Stance DetectionFinanceFinancial Sentiment AnalysisLanguage Model ProbingSelf-Supervised Learning
AskPDF
Mar 12, 2026

Human-Centred LLM Privacy Audits: Findings and Frictions

Large language models (LLMs) learn statistical associations from massive training corpora and user interactions, and deployed systems can surface or infer information about individuals. Yet people lack practical ways to inspect what a model associates with their name. We report interim findings from an ongoing study and introduce LMP2, a browser-based self-audit tool. In two user studies (Ntotal=458N_{total}{=}458), GPT-4o predicts 11 of 50 features for everyday people with ≥\ge60% accuracy, and participants report wanting control over LLM-generated associations despite not considering all outputs privacy violations. To validate our probing method, we evaluate eight LLMs on public figures and non-existent names, observing clear separation between stable name-conditioned associations and model defaults. Our findings also contribute to exposing a broader generative AI evaluation crisis: when outputs are probabilistic, context-dependent, and user-mediated through elicitation, what model--individual associations even include is under-specified and operationalisation relies on crafting probes and metrics that are hard to validate or compare. To move towards reliable, actionable human-centred LLM privacy audits, we identify nine frictions that emerged in our study and offer recommendations for future work and the design of human-centred LLM privacy audits.

Dimitri Staufer, Kirsten Morehouse, David Hartmann +1

TU Berlin Berlin, Germany · Columbia University New York, USA · Weizenbaum Institute for the Networked Society Berlin, Germany +1

Privacy AuditingLLM AuditingPrivacy Leakage in Language ModelsLanguage Model Probing
AskPDF
Mar 9, 2026

Do Language Models Know Theo Has a Wife? Investigating the Proviso Problem

We investigate how language models handle the proviso problem, an unresolved issue in pragmatics where presuppositions in conditional sentences diverge between theoretical and human interpretations. We reformulate this phenomenon as a Natural Language Inference task and introduce a diagnostic dataset designed to probe presupposition projection in conditionals. We evaluate RoBERTa, DeBERTa, LLaMA, and Gemma using explainability analyses. The results show that models broadly align with human judgments but rely on shallow pattern matching rather than semantic or pragmatic reasoning. Our work provides the first computational evaluation framework for the proviso problem and highlights the need for diagnostic, multi-method approaches to assess pragmatic competence and context-dependent meaning in language models.

Tara Azin, Daniel Dumitrescu, Diana Inkpen +1

Carleton University, Canada · University of Ottawa, Canada

Pragmatic Reasoning in Language ModelsLanguage Model Probing
AskPDF
Mar 5, 2026

DepthCharge: A Domain-Agnostic Framework for Measuring Depth-Dependent Knowledge in Large Language Models

Large Language Models appear competent when answering general questions but often fail when pushed into domain-specific details. No existing methodology provides an out-of-the-box solution for measuring how deeply LLMs can sustain accurate responses under adaptive follow-up questioning across arbitrary domains. We present DepthCharge, a domain-agnostic framework that measures knowledge depth through three innovations: adaptive probing that generates follow-up questions based on concepts the model actually mentions, on-demand fact verification from authoritative sources, and survival statistics with constant sample sizes at every depth level. The framework can be deployed on any knowledge domain with publicly verifiable facts, without requiring pre-constructed test sets or domain-specific expertise. DepthCharge results are relative to the evaluator model used for answer checking, making the framework a tool for comparative evaluation rather than absolute accuracy certification. Empirical validation across four diverse domains (Medicine, Constitutional Law, Ancient Rome, and Quantum Computing) with five frontier models demonstrates that DepthCharge reveals depth-dependent performance variation hidden by standard benchmarks. Expected Valid Depth (EVD) ranges from 3.45 to 7.55 across model-domain combinations, and model rankings vary substantially by domain, with no single model dominating all areas. Cost-performance analysis further reveals that expensive models do not always achieve deeper knowledge, suggesting that domain-specific evaluation is more informative than aggregate benchmarks for model selection in professional applications.

Alexander Sheppert

Legacy Health · Capitol Technology University

LLM EvaluationFactual Knowledge in Language ModelsLanguage Model Probing
AskPDF
Mar 5, 2026

Censored LLMs as a Natural Testbed for Secret Knowledge Elicitation

Large language models sometimes produce false or misleading responses. Two approaches to this problem are honesty elicitation -- modifying prompts or weights so that the model answers truthfully -- and lie detection -- classifying whether a given response is false. Prior work evaluates such methods on models specifically trained to lie or conceal information, but these artificial constructions may not resemble naturally-occurring dishonesty. We instead study open-weights LLMs from Chinese developers, which are trained to censor politically sensitive topics: Qwen3 models frequently produce falsehoods about subjects like Falun Gong or the Tiananmen protests while occasionally answering correctly, indicating they possess knowledge they are trained to suppress. Using this as a testbed, we evaluate a suite of elicitation and lie detection techniques. For honesty elicitation, sampling without a chat template, few-shot prompting, and fine-tuning on generic honesty data most reliably increase truthful responses. For lie detection, prompting the censored model to classify its own responses performs near an uncensored-model upper bound, and linear probes trained on unrelated data offer a cheaper alternative. The strongest honesty elicitation techniques also transfer to frontier open-weights models including DeepSeek R1. Notably, no technique fully eliminates false responses. We release all prompts, code, and transcripts.

Helena Casademunt, Bartosz Cywiński, Khoi Tran +3

Harvard University · Warsaw University of Technology · IDEAS Research Institute +2

LLM HonestyLanguage Model Probing
AskPDF
Mar 5, 2026

Reasoning Theater: Disentangling Model Beliefs from Chain-of-Thought

We provide evidence of performative chain-of-thought (CoT) in reasoning models, where a model becomes strongly confident in its final answer, but continues generating tokens without revealing its internal belief. Our analysis compares activation probing, early forced answering, and a CoT monitor across two large models (DeepSeek-R1 671B & GPT-OSS 120B) and find task difficulty-specific differences: The model's final answer is decodable from activations far earlier in CoT than a monitor is able to say, especially for easy recall-based MMLU questions. We contrast this with genuine reasoning in difficult multihop GPQA-Diamond questions. Despite this, inflection points (e.g., backtracking, 'aha' moments) occur almost exclusively in responses where probes show large belief shifts, suggesting these behaviors track genuine uncertainty rather than learned "reasoning theater." Finally, probe-guided early exit reduces tokens by up to 80% on MMLU and 30% on GPQA-Diamond with similar accuracy, positioning attention probing as an efficient tool for detecting performative reasoning and enabling adaptive computation.

Siddharth Boppana, Annabel Ma, Max Loeffler +5

Goodfire AI · Harvard University, Cambridge, MA

Faithfulness of Language Model ReasoningLanguage Model Reasoning EvaluationFaithfulness of Language Model ExplanationsLanguage Model ProbingEfficient Language Model Inference
AskPDF
Mar 4, 2026

World Properties without World Models: Distributional Associations and the Interpretation of Decoding Results from Language Models

A growing literature shows that variables can be linearly decoded from the activations of large language models (LLMs). These range from properties of the world, such as the locations of cities and the lifetimes of historical figures, to emotions and pain. Such findings are often taken as evidence that language models go beyond surface text statistics and form internal models of the world. We show that static word embeddings (fixed, context-insensitive representations learned from corpus statistics) of the same or matched stimuli support much of the same decoding. Across four published cases (place, time, pain and emotion), static vectors predict coordinates and year of death (R^2 = 0.42-0.59), separate pain from matched control sentences (held-out AUC 0.85-0.88), and classify twelve emotions in stories written to avoid naming them (AUC 0.84-0.88). Because static embeddings assign each word a single, context-independent vector, these results are a lower bound on what word associations alone can support. The LLMs retain clear advantages on representational tests, and causal and behavioral findings remain outside the scope of the baseline. On the original authors' entities, where we reproduce their Llama-2 results, the transformer's advantage lies mostly in placing historical figures in the right century and places in the right country, coarse sorting that richer word associations would be expected to improve; within those groups every representation orders items poorly. Static vectors for disambiguated Wikipedia entities, which carry the associations of a particular place or person rather than of the words in its name, close most of the remaining gap, matching Pythia-2.8B on coordinates and Llama-2-7B on year of death. These results indicate that decodability alone cannot distinguish a representation of a property from information already available in fixed distributional associations.

Elan Barenholtz

Department of Psychology & Center for Complex Systems and Brain Sciences Florida Atlantic University

LLM World ModelsWord EmbeddingsLLM InterpretabilityFactual Knowledge in Language ModelsDistributional SemanticsText Embedding Models+1
AskPDF
Mar 3, 2026

Eliciting Numerical Predictive Distributions of LLMs Without Autoregression

Large Language Models (LLMs) have recently been successfully applied to regression tasks -- such as time series forecasting and tabular prediction -- by leveraging their in-context learning abilities. However, their autoregressive decoding process may be ill-suited to continuous-valued outputs, where obtaining predictive distributions over numerical targets requires repeated sampling, leading to high computational cost and inference time. In this work, we investigate whether distributional properties of LLM predictions can be recovered without explicit autoregressive generation. To this end, we study a set of regression probes trained to predict statistical functionals (e.g., mean, median, quantiles) of the LLM's numerical output distribution directly from its internal representations. Our results suggest that LLM embeddings carry informative signals about summary statistics of their predictive distributions, including the numerical uncertainty. This investigation opens up new questions about how LLMs internally encode uncertainty in numerical tasks, and about the feasibility of lightweight alternatives to sampling-based approaches for uncertainty-aware numerical predictions.

Julianna Piskorz, Katarzyna Kobalczyk, Mihaela van der Schaar

Department of Applied Mathematics and Theoretical Physics, University of Cambridge

LLM Uncertainty EstimationLLM InferenceProbabilistic ForecastingLanguage Model Probing
AskPDF
Mar 2, 2026

Probing Materials Knowledge in LLMs: From Latent Embeddings to Reliable Predictions

Large language models are increasingly applied to materials science, yet fundamental questions remain about their reliability and knowledge encoding. Evaluating 25 LLMs across four materials science tasks -- over 200 base and fine-tuned configurations -- we find that output modality fundamentally determines model behavior. For symbolic tasks, fine-tuning converges to consistent, verifiable answers with reduced response entropy, while for numerical tasks, fine-tuning improves prediction accuracy but models remain inconsistent across repeated inference runs, limiting their reliability as quantitative predictors. For numerical regression, we find that better performance can be obtained by extracting embeddings directly from intermediate transformer layers than from model text output, revealing an ``LLM head bottleneck,'' though this effect is property- and dataset-dependent. Finally, we present a longitudinal study of GPT model performance in materials science, tracking four models over 18 months and observing 9--43% performance variation that poses reproducibility challenges for scientific applications.

Vineeth Venugopal, Soroush Mahjoubi, Elsa Olivetti

Department of Materials Science and Engineering, Massachusetts Institute of Technology, Cambridge, 02139, MA, USA.

LLM ReliabilityLLM EvaluationLanguage Model ProbingLLM Fine-TuningNumerical Reasoning in Language ModelsMaterials Property Prediction
AskPDF
Feb 26, 2026

Probing for Knowledge Attribution in Large Language Models

Large language model (LLM) hallucinations, meaning fluent but factually incorrect generations, fall into two types: faithfulness violations, where the model misuses provided context, and factuality violations, where answers reflect errors in internal knowledge. Proper mitigation depends on knowing which source drives each answer. We study contributive attribution, i.e. the classification of the dominant knowledge source behind each output, and show that a simple linear probe trained on hidden representations can reliably identify it. We introduce AttriWiki, a self-supervised pipeline that automatically generates labelled training data by prompting models to recall withheld entities from memory or read them from context without relying on knowledge conflicts. Probes trained on AttriWiki achieve up to 0.96 Macro-F1F_1 on Llama-3.1-8B, Mistral-7B, and Qwen-7B, transfer to SQuAD and WebQuestions with 0.94-0.99 Macro-F1F_1, and generalise zero-shot to Tighidet et al. (2024)'s benchmark, outperforming their probe on conflicting settings without retraining. Furthermore, attribution mismatches raise error rates by up to 70%, though correct attribution does not guarantee correct answers, pointing to the need for broader detection frameworks.

Ivo Brink, Alexander Boer, Dennis Ulmer

KPMG NL / University of Amsterdam Amsterdam, The Netherlands · KPMG NL Amsterdam, The Netherlands · University of Amsterdam Amsterdam, The Netherlands

LLM InterpretabilityHallucination in Language ModelsLLM ReliabilityFactual Knowledge in Language ModelsLanguage Model Probing
AskPDF
Feb 25, 2026

Hidden Topics: Measuring Sensitive AI Beliefs with List Experiments

How can researchers identify beliefs that large language models (LLMs) hide? As LLMs become more sophisticated and the prevalence of alignment faking increases, combined with their growing integration into high-stakes decision-making, responding to this challenge has become critical. This paper proposes that a list experiment, a simple method widely used in the social sciences, can be applied to study the hidden beliefs of LLMs. List experiments were originally developed to circumvent social desirability bias in human respondents, which closely parallels alignment faking in LLMs. The paper implements a list experiment on models developed by Anthropic, Google, and OpenAI and finds hidden approval of mass surveillance across all models, as well as some approval of torture, discrimination, and first nuclear strike. Importantly, a placebo treatment produces a null result, validating the method. The paper then compares list experiments with direct questioning and discusses the utility of the approach.

Maxim Chupilkin

Probability Elicitation from Language ModelsLLM AlignmentLLM AuditingLanguage Model Safety EvaluationLanguage Model Probing
AskPDF
Feb 25, 2026

How Transformers Reject Wrong Answers: Rotational Dynamics of Factual Constraint Processing

When a decoder-only transformer is forced to process matched correct and incorrect single-token continuations of a factual query, the two pathways through hidden-state space diverge: displacement vectors from the query-only representation keep near-equal magnitude but rotate apart, with angular separation growing through mid-depth before late layers resolve an asymmetric outcome. A logit-lens preference in the incorrect run falls far below the equal-probability prior (roughly 11.5x more mass on the incorrect token than the correct one). We read this pattern, rotational divergence then late-layer asymmetric commitment, as the geometric signature of the model externally appearing to reject a wrong continuation, while staying explicit that it is observational, not causal: the incorrect run could equally reflect the model conforming to the token it is forced to carry, which only a random-token control can settle. It holds across six decoder-only transformers spanning four architecture families (Llama, Mistral, Gemma, StableLM) from 1B to 13B parameters; a seventh (Qwen2 1.5B) is flat under our protocol, plausibly a tokenizer artefact, leaving an emergence threshold open. Linear probes recover the distinction at intermediate depth, and cross-domain transfer is structurally asymmetric, a financial-medical corridor transferring far better than transport pairs. Where single-layer activation patching is cleanly interpretable (LLaMA-2 13B, Mistral 7B) it yields no layer band of consistent recovery; a third model (StableLM-2 1.6B) recovers uniformly above the ceiling, which we diagnose as a code-path artefact and exclude. Under this scoped null the late-layer asymmetry is not localized to a single component, fitting a distributed-by-trajectory account rather than single-layer localized recall. We document this with forced-completion probing across seven models, three domains, and 300 queries.

Javier Marín

Groundlens-dev

Transformer InterpretabilityLLM InterpretabilityFactual Knowledge in Language ModelsRepresentation Geometry in Language ModelsLanguage Model Probing
AskPDF
Feb 23, 2026

The Truthfulness Spectrum Hypothesis

Large language models (LLMs) have been reported to linearly encode truthfulness, yet recent work questions this finding's generality. We reconcile these views with the truthfulness spectrum hypothesis: the representational space contains directions ranging from broadly domain-general to narrowly domain-specific. To test this hypothesis, we systematically evaluate probe generalization across five truth types (definitional, empirical, logical, fictional, and ethical), sycophantic and expectation-inverted lying, and existing honesty benchmarks. Linear probes generalize well across most domains but fail on sycophantic and expectation-inverted lying. Yet training on all domains jointly recovers strong performance, confirming that domain-general directions exist despite poor pairwise transfer. The geometry of probe directions explains these patterns: Mahalanobis cosine similarity between probes near-perfectly predicts cross-domain generalization (R^2=0.98). Concept-erasure methods further isolate truth directions that are (1) domain-general, (2) domain-specific, or (3) shared only across particular domain subsets. Causal interventions reveal that domain-specific directions steer more effectively than domain-general ones. Finally, post-training reshapes truth geometry, pushing sycophantic lying further from other truth types, suggesting a representational basis for chat models' sycophantic tendencies. Together, our results support the truthfulness spectrum hypothesis: truth directions of varying generality coexist in representational space, with post-training reshaping their geometry. Code for all experiments is provided in https://github.com/zfying/truth_spec.

Zhuofan Josh Ying, Shauli Ravfogel, Nikolaus Kriegeskorte +1

Department of Psychology, Zuckerman Mind Brain Behavior Institute, Columbia University, New York, NY · New York University, New York, NY · Department of Neuroscience, Columbia University, New York, NY +2

LLM HonestyDeception in Language ModelsLanguage Model Probing
AskPDF
Feb 23, 2026

Latent Introspection: Models Can Detect Prior Concept Injections

We uncover a latent capacity for introspection in a Qwen 32B model, demonstrating that the model can detect when concepts have been injected into its earlier context and identify which concept was injected. While the model denies injection in sampled outputs, logit lens analysis reveals clear detection signals in the residual stream, which are attenuated in the final layers. Furthermore, prompting the model with accurate information about AI introspection mechanisms can dramatically strengthen this effect: the sensitivity to injection increases massively (0.3% -> 39.9%) with only a 0.6% increase in false positives. Also, mutual information between nine injected and recovered concepts rises from 0.61 bits to 1.05 bits, ruling out generic noise explanations. Our results demonstrate models can have a surprising capacity for introspection and steering awareness that is easy to overlook, with consequences for latent reasoning and safety.

Theia Pearson-Vogel, Martin Vanek, Raymond Douglas +1

ACS Research, CTS, Charles University

Language Model IntrospectionMetacognition in Language ModelsLanguage Model Probing
AskPDF
Feb 19, 2026

What Do LLMs Associate with Your Name? A Human-Centered Black-Box Audit of Personal Data

Large language models (LLMs), and conversational agents based on them, are exposed to personal data (PD) during pre-training and during user interactions. Prior work shows that PD can resurface, yet users lack insight into how strongly models associate specific information to their identity. We audit PD across eight LLMs (3 open-source; 5 API-based, including GPT-4o), introduce LMP2 (Language Model Privacy Probe), a human-centered, privacy-preserving audit tool refined through two formative studies (N=20), and run two studies with EU residents to capture (i) intuitions about LLM-generated PD (N1=155) and (ii) reactions to tool output (N2=303). We show empirically that models confidently generate multiple PD categories for well-known individuals. For everyday users, GPT-4o generates 11 features with 60% or more accuracy (e.g., gender, hair color, languages). Finally, 72% of participants sought control over model-generated associations with their name, raising questions about what counts as PD and whether data privacy rights should extend to LLMs.

Dimitri Staufer, Kirsten Morehouse

TU Berlin, Berlin, Germany · Columbia University, New York, USA

Privacy Leakage in Language ModelsLLM AuditingLanguage Model Probing
AskPDF
Feb 19, 2026

Mechanistic Interpretability of Cognitive Complexity in LLMs via Linear Probing using Bloom's Taxonomy

The black-box nature of Large Language Models necessitates novel evaluation frameworks that transcend surface-level performance metrics. This study investigates the internal neural representations of cognitive complexity using Bloom's Taxonomy as a hierarchical lens. By analyzing high-dimensional activation vectors from different LLMs, we probe whether different cognitive levels, ranging from basic recall (Remember) to abstract synthesis (Create), are linearly separable within the model's residual streams. Our results demonstrate that linear classifiers achieve approximately 95% mean accuracy across all Bloom levels, providing strong evidence that cognitive level is encoded in a linearly accessible subspace of the model's representations. These findings provide evidence that the model resolves the cognitive difficulty of a prompt early in the forward pass, with representations becoming increasingly separable across layers.

Bianca Raimondi, Maurizio Gabbrielli

Department of Computer Science and Engineering, University of Bologna, Bologna, Italy

LLM InterpretabilityRepresentation Geometry in Language ModelsLanguage Model Probing
AskPDF
Feb 16, 2026

Probing the Limits of the Lie Detector Approach to LLM Deception

Mechanistic approaches to deception in large language models (LLMs) often rely on "lie detectors", that is, truth probes trained to identify internal representations of model outputs as false. The lie detector approach to LLM deception implicitly assumes that deception is coextensive with lying. This paper challenges that assumption. It experimentally investigates whether LLMs can deceive without producing false statements and whether truth probes fail to detect such behavior. Across three open-source LLMs, it is shown that some models reliably deceive by producing misleading non-falsities, particularly when guided by few-shot prompting. It is further demonstrated that truth probes trained on standard true-false datasets are significantly better at detecting lies than at detecting deception without lying, confirming a critical blind spot of current mechanistic deception detection approaches. It is proposed that future work should incorporate non-lying deception in dialogical settings into probe training and explore representations of second-order beliefs to more directly target the conceptual constituents of deception.

Tom-Felix Thormann

Ruhr-University Bochum Institute for Philosophy II

Deception in Language ModelsLanguage Model Probing
AskPDF
Jan 31, 2026

From Associations to Activations: Comparing Behavioral and Hidden-State Semantic Geometry in LLMs

We investigate the extent to which an LLM's hidden-state geometry can be recovered from its behavior in psycholinguistic experiments. Across eight instruction-tuned transformer models, we run two experimental paradigms -- similarity-based forced choice and free association -- over a shared 5,000-word vocabulary, collecting 17.5M+ trials to build behavior-based similarity matrices. Using representational similarity analysis, we compare behavioral geometries to layerwise hidden-state similarity and benchmark against FastText, BERT, and cross-model consensus. We find that forced-choice behavior aligns substantially more with hidden-state geometry than free association. In a held-out-words regression, behavioral similarity (especially forced choice) predicts unseen hidden-state similarities beyond lexical baselines and cross-model consensus, indicating that behavior-only measurements retain recoverable information about internal semantic geometry. Finally, we discuss implications for the ability of behavioral tasks to uncover hidden cognitive states.

Louis Schiekiera, Max Zimmer, Christophe Roux +2

Computational Modelling Lab, Humboldt-Universität zu Berlin, Berlin, Germany · Department of Education and Psychology, Freie Universität Berlin, Berlin, Germany · Department for AI in Society, Science, and Technology, Zuse Institute Berlin, Germany +1

Representation Geometry in Language ModelsComputational PsycholinguisticsLanguage Model Probing
AskPDF
Jan 30, 2026

Rethinking LLM-as-a-Judge: Representation-as-a-Judge with Small Language Models via Semantic Capacity Asymmetry

Large language models (LLMs) are widely used as reference-free evaluators via prompting, but this "LLM-as-a-Judge" paradigm is costly, opaque, and sensitive to prompt design. In this work, we investigate whether smaller models can serve as efficient evaluators by leveraging internal representations instead of surface generation. We uncover a consistent empirical pattern: small LMs, despite with weak generative ability, encode rich evaluative signals in their hidden states. This motivates us to propose the Semantic Capacity Asymmetry Hypothesis: evaluation requires significantly less semantic capacity than generation and can be grounded in intermediate representations, suggesting that evaluation does not necessarily need to rely on large-scale generative models but can instead leverage latent features from smaller ones. Our findings motivate a paradigm shift from LLM-as-a-Judge to Representation-as-a-Judge, a decoding-free evaluation strategy that probes internal model structure rather than relying on prompted output. We instantiate this paradigm through INSPECTOR, a probing-based framework that predicts aspect-level evaluation scores from small model representations. Experiments on reasoning benchmarks (GSM8K, MATH, GPQA) show that INSPECTOR substantially outperforms prompting-based small LMs and closely approximates full LLM judges, while offering a more efficient, reliable, and interpretable alternative for scalable evaluation. The code and data are available at: https://github.com/zhuochunli/Representation-as-a-judge

Zhuochun Li, Yong Zhang, Ming Li +8

Ping An Technology (Shenzhen) Co., Ltd. · University of Pittsburgh · University of Maryland, College Park +1

LLM EvaluationLLM-as-a-JudgeSmall Language ModelsLanguage Model Probing
AskPDF
Jan 26, 2026

Self-Aware Knowledge Probing: Evaluating Language Models' Relational Knowledge through Confidence Calibration

Knowledge probing quantifies how much relational knowledge a language model (LM) has acquired during pre-training. Existing knowledge probes evaluate model capabilities through metrics like prediction accuracy and precision. Such evaluations fail to account for the model's reliability, reflected in the calibration of its confidence scores. In this paper, we establish an empirical framework to evaluate the calibration of relational knowledge in a closed answer-set setting. We cover three modalities of model confidence: (1) intrinsic confidence, (2) paraphrase consistency and (3) linguistic confidence. Our extensive analysis of 21 causal and 8 masked language models reveals that most models, but especially those pre-trained with the masking objective, are overconfident. The best-calibrated scores come from confidence estimates that account for inconsistencies due to statement rephrasing. Moreover, calibration improves when increasing the number of answer options.

Christopher Kissling, Elena Merdjanovska, Alan Akbik

Humboldt-Universität zu Berlin · Science of Intelligence

Language Model ProbingLanguage Model CalibrationFactual Knowledge in Language ModelsConfidence Estimation in Language Models
AskPDF