Fine-tuning updates frequently stale activation monitors for language model safety while quantization does not, with degradation predictable and repairable via label-free realignment.
hub Mixed citations
Probing classifiers: Promises, shortcomings, and advances.Computational Linguistics, 48(1):207–219
Mixed citation behavior. Most common role is background (62%).
57
Pith papers citing it
130
external citations · Crossref
Background
62% of classified citations
abstract
Probing classifiers have emerged as one of the prominent methodologies for interpreting and analyzing deep neural network models of natural language processing. The basic idea is simple -- a classifier is trained to predict some linguistic property from a model's representations -- and has been used to examine a wide variety of models and properties. However, recent studies have demonstrated various methodological limitations of this approach. This article critically reviews the probing classifiers framework, highlighting their promises, shortcomings, and advances.
hub tools
citation-role summary
background 8
method 3
baseline 1
dataset 1
citation-polarity summary
representative citing papers
Adaptive scheduling of interventions in discrete diffusion language models, timed to attribute-specific commitment schedules discovered with sparse autoencoders, delivers precise multi-attribute steering up to 93% strength while preserving generation quality.
Factual associations in autoregressive transformers are localized to mid-layer feed-forward modules and can be edited via rank-one model editing while preserving both specificity and generalization on counterfactual tests.
LBR performs token-level test-time scaling via local branch routing on hidden states, enabling end-to-end RL training and improving Pass@1 and Pass@32 on math benchmarks over CoT and RLVR baselines.
Evaluation of two latent reasoning models against controls shows observable latent patterns appear without the proposed mechanisms, have graded causal effects on behavior, and concentrate in structured low-rank directions, arguing that patterns are insufficient evidence for reasoning.
PRISM is a new activation-conditioned model that recovers full sets of simultaneous instructions from LLM hidden states via judge-guided GRPO training and outperforms prior activation-to-language methods on security-relevant tasks.
Empirical study of five LVR variants finds cosine alignment negatively correlates with accuracy (r=-0.94), supervised latents are bypassed under corruption (max 4-point shift), and answers are decodable downstream but not at the latent.
Ablation of a stack-depth direction extracted via linear probes from transformer hidden states causes performance on counter languages to drop to near zero, showing causal necessity of the representation.
MLLMs exhibit spatial lexical bias on multiple-choice spatial questions, traced via mechanistic tools to language-side channels rather than vision, and largely mitigated by LLM-only DPO on synthetic data.
Chart information is encoded but not routed to predictions in VLMs for claim verification, unlike tables, revealed by layer-wise probing and attention analysis on three models.
A representation property is identifiable from the induced predictor iff it is constant on the fibers of the map from admissible (representation, head) pairs to the composite predictor.
Different scoring mechanisms cause encoder-based authorship attribution models to consolidate authorship signals at different layers, as shown by causal interventions and gradient analysis.
An 8B autoregressive LM implements a language-switching backdoor via a three-phase circuit with early trigger composition, orthogonal mid-layer propagation, and final-layer MLP conversion, routed through a single-position serial bottleneck.
QAOD projects away question-aligned directions from answer representations to isolate domain-agnostic factuality signals, enabling efficient hallucination detection with top in-domain AUROC and up to 21% better OOD transfer.
KamonBench is a grammar-based dataset of 20,000 synthetic Japanese crests with multi-format annotations that enables direct evaluation of factor recovery beyond caption accuracy in vision-language models.
Symmetry under affine reparameterizations of hidden coordinates selects a unique hierarchy of shallow coordinate-stable probes and a probe-visible quotient for cross-model transfer.
EEG foundation models encode 68.6% of a 63-feature clinical lexicon in a representation-causal way, with frequency-domain features dominant; these recover 79.3% of the models' advantage over random baselines on average.
Behavioral directions from one LLM family transfer to others via projection into a shared anchor coordinate space, yielding 0.83 ten-way detection accuracy and steering effects up to 0.46% on held-out models.
Tabular foundation models show substantial depthwise redundancy, so a looped single-layer version achieves comparable results with 20% of the original parameters.
Finite-answer projections of continuation probabilities stabilize before the answer is parseable, showing 17-31 token mean lead in delayed-verdict tasks with Qwen3-4B-Instruct.
Latent space probing on CogVideoX achieves 97.29% F1 for adult content detection on a new 11k-clip dataset with 4-6ms overhead.
On 2,520 programming tasks, matched Qwen general and coder models reliably raise Bloom cognitive demand but fail to lower it, so execution skill does not imply educational control.
Linear probes on intermediate LLM activations produce better-calibrated confidence than verbalized probabilities, detect hidden evidence influence, and reveal that forecasts are largely pre-committed before reasoning begins.
Empirical evaluation on the PLUME benchmark shows steering vectors vary widely in trait expressibility, degrade on task transfer, and lose effectiveness when multiple vectors are composed.
citing papers explorer
-
Do Activation Monitors Survive Model Updates? Benchmarking, Predicting, and Repairing Activation-Monitor Staleness
Fine-tuning updates frequently stale activation monitors for language model safety while quantization does not, with degradation predictable and repairable via label-free realignment.
-
Steering Without Breaking: Mechanistically Informed Interventions for Discrete Diffusion Language Models
Adaptive scheduling of interventions in discrete diffusion language models, timed to attribute-specific commitment schedules discovered with sparse autoencoders, delivers precise multi-attribute steering up to 93% strength while preserving generation quality.
-
Locating and Editing Factual Associations in GPT
Factual associations in autoregressive transformers are localized to mid-layer feed-forward modules and can be edited via rank-one model editing while preserving both specificity and generalization on counterfactual tests.
-
Efficient and Trainable Language Model Test-Time Scaling via Local Branch Routing
LBR performs token-level test-time scaling via local branch routing on hidden states, enabling end-to-end RL training and improving Pass@1 and Pass@32 on math benchmarks over CoT and RLVR baselines.
-
Observable Patterns Are Not Explanations: A Causal-Geometric Analysis of Latent Reasoning Models
Evaluation of two latent reasoning models against controls shows observable latent patterns appear without the proposed mechanisms, have graded causal effects on behavior, and concentrate in structured low-rank directions, arguing that patterns are insufficient evidence for reasoning.
-
PRISM: Recovering Instruction Sets from Language Model Activations
PRISM is a new activation-conditioned model that recovers full sets of simultaneous instructions from LLM hidden states via judge-guided GRPO training and outperforms prior activation-to-language methods on security-relevant tasks.
-
Cosine Misleads: Auxiliary Losses Reshape Vision Language Models, Not Their Latents
Empirical study of five LVR variants finds cosine alignment negatively correlates with accuracy (r=-0.94), supervised latents are bypassed under corruption (max 4-point shift), and answers are decodable downstream but not at the latent.
-
Causal Evidence of Stack Representations in Modeling Counter Languages Using Transformers
Ablation of a stack-depth direction extracted via linear probes from transformer hidden states causes performance on counter languages to drop to near zero, showing causal necessity of the representation.
-
Mechanistic Diagnostics of Spatial Lexical Bias in Multimodal Large Language Model Spatial Reasoning
MLLMs exhibit spatial lexical bias on multiple-choice spatial questions, traced via mechanistic tools to language-side channels rather than vision, and largely mitigated by LLM-only DPO on synthetic data.
-
Encoded but Not Routed: Explaining the Table-Chart Gap in Scientific Claim Verification
Chart information is encoded but not routed to predictions in VLMs for claim verification, unlike tables, revealed by layer-wise probing and attention analysis on three models.
-
A Fiber Criterion for Representation Identifiability in Supervised Learning
A representation property is identifiable from the induced predictor iff it is constant on the fibers of the map from admissible (representation, head) pairs to the composite predictor.
-
Where Does Authorship Signal Emerge in Encoder-Based Language Models?
Different scoring mechanisms cause encoder-based authorship attribution models to consolidate authorship signals at different layers, as shown by causal interventions and gradient analysis.
-
Language-Switching Triggers Take a Latent Detour Through Language Models
An 8B autoregressive LM implements a language-switching backdoor via a three-phase circuit with early trigger composition, orthogonal mid-layer propagation, and final-layer MLP conversion, routed through a single-position serial bottleneck.
-
When Answers Stray from Questions: Hallucination Detection via Question-Answer Orthogonal Decomposition
QAOD projects away question-aligned directions from answer representations to isolate domain-agnostic factuality signals, enabling efficient hallucination detection with top in-domain AUROC and up to 21% better OOD transfer.
-
KamonBench: A Grammar-Based Dataset for Evaluating Compositional Factor Recovery in Vision-Language Models
KamonBench is a grammar-based dataset of 20,000 synthetic Japanese crests with multi-format annotations that enables direct evaluation of factor recovery beyond caption accuracy in vision-language models.
-
Deep Minds and Shallow Probes
Symmetry under affine reparameterizations of hidden coordinates selects a unique hierarchy of shallow coordinate-stable probes and a probe-visible quotient for cross-model transfer.
-
What Do EEG Foundation Models Capture from Human Brain Signals?
EEG foundation models encode 68.6% of a 63-feature clinical lexicon in a representation-causal way, with frequency-domain features dominant; these recover 79.3% of the models' advantage over random baselines on average.
-
Cross-Family Universality of Behavioral Axes via Anchor-Projected Representations
Behavioral directions from one LLM family transfer to others via projection into a shared anchor coordinate space, yielding 0.83 ten-way detection accuracy and steering effects up to 0.46% on held-out models.
-
Is One Layer Enough? Understanding Inference Dynamics in Tabular Foundation Models
Tabular foundation models show substantial depthwise redundancy, so a looped single-layer version achieves comparable results with 20% of the original parameters.
-
When Does a Language Model Commit? A Finite-Answer Theory of Pre-Verbalization Commitment
Finite-answer projections of continuation probabilities stabilize before the answer is parseable, showing 17-31 token mean lead in delayed-verdict tasks with Qwen3-4B-Instruct.
-
Latent Space Probing for Adult Content Detection in Video Generative Models
Latent space probing on CogVideoX achieves 97.29% F1 for adult content detection on a new 11k-clip dataset with 4-6ms overhead.
-
From Execution to Education: A Bloom-Aligned Framework for Measuring Educational Control in LLMs
On 2,520 programming tasks, matched Qwen general and coder models reliably raise Bloom cognitive demand but fail to lower it, so execution skill does not imply educational control.
-
What LLM Forecasters Know but Don't Say: Probing Internal Representations for Calibration and Faithfulness
Linear probes on intermediate LLM activations produce better-calibrated confidence than verbalized probabilities, detect hidden evidence influence, and reveal that forecasts are largely pre-committed before reasoning begins.
-
On the Limits of Steering Vectors for Preference-Aligned Generation
Empirical evaluation on the PLUME benchmark shows steering vectors vary widely in trait expressibility, degrade on task transfer, and lose effectiveness when multiple vectors are composed.
-
Aionoscope: Debugging Latent-State Accessibility in Time-Series Representations
Aionoscope shows that time-series representations recover coarse signal types reliably but expose dense latent states like phase and amplitude much less reliably, with best dense-probe R² at 0.689 versus oracle 0.999.
-
Validating Causal Abstraction Metrics on Simulated Complex Systems
Authors create a benchmark across discrete/continuous and static/dynamical systems and introduce the Causal Abstraction Error (CAE) metric that reliably distinguishes valid from invalid causal abstractions when it includes faithfulness testing.
-
The strength of clinical evidence is recoverable from language model representations but not from their stated grades
Linear probes recover evidence grades from LLM activations (median AUROC 71.8) across 22 models but the models' stated grades perform at chance level and the signal is largely lexical.
-
ToxiREX: A Dataset on Toxic REasoning in ConteXt
ToxiREX is a new dataset of 128k Reddit comments in six languages with hierarchical annotations for implicit toxicity in conversational context based on an existing reasoning schema.
-
From Texts to Scores: Tracing the Emergence of Essay Quality Representations in Large Language Models
LLM representations encode essay quality in a linearly decodable form that emerges across layers and includes identifiable scoring neurons whose distribution shifts with essay length.
-
Where Computation Lives Inside TabPFN: Causal Localisation of Attention Head Function
TabPFN shows temporal specialization where one attention head dominates causal necessity at shifting peak layers depending on task complexity, while contrastive activation steering fails to transfer across samples due to context-dependent attention.
-
Rethinking Molecular Text Representations for LLMs: An Empirical Study
Structured text representations like CML and MolJSON outperform SMILES variants on structural tasks while IUPAC dominates semantic tasks such as molecule retrieval across all tested LLMs.
-
Child-directed speech facilitates production, not comprehension, in BabyLMs
CDS-trained BabyLMs show earlier and more appropriate production in a new frame-completion task while FineWeb-edu models lead on comprehension benchmarks, indicating current tests underestimate CDS benefits.
-
Represented Is Not Computed: A Causal Test of Candidate Algorithmic Intermediates in a Transformer
Transformer represents but does not causally transmit staged algorithmic intermediates for base-digit extraction, diverging from probe predictions.
-
Mechanistic Interpretability of EEG Foundation Models via Sparse Autoencoders
Sparse autoencoders on EEG transformers extract clinical features, identify three steering regimes, expose age-pathology entanglements and wrecking-ball failures, and map interventions to frequency spectra.
-
Correcting Influence: Unboxing LLM Outputs with Orthogonal Latent Spaces
A latent mediation framework with sparse autoencoders enables non-additive token-level influence attribution in LLMs by learning orthogonal features and back-propagating attributions.
-
Predicting Decisions of AI Agents from Limited Interaction through Text-Tabular Modeling
A tabular foundation model with LLM-as-Observer features predicts AI agent decisions in controlled games, outperforming baselines by 4 AUC points and 14% lower error at K=16 interactions.
-
Instructions Shape Production of Language, not Processing
Instructions trigger a production-centered mechanism in language models, with task-specific information stable in input tokens but varying strongly in output tokens and correlating with behavior.
-
Molecules Meet Language: Confound-Aware Representation Learning and Chemical Property Steering in Transformer-VAE Latent Spaces
Chemically meaningful steering for properties like cLogP and TPSA emerges in entangled Transformer-VAE latent spaces only after controlling for SELFIES representation confounds through residualization and decoded traversals.
-
Conceptors for Semantic Steering
Conceptors as soft projection matrices from bipolar activations offer a multidimensional, compositional, and geometrically principled method for semantic steering in LLMs that outperforms single-vector baselines in multi-dimensional subspaces.
-
Beyond Decodability: Reconstructing Language Model Representations with an Encoding Probe
An encoding probe reconstructs transformer representations from acoustic, phonetic, syntactic, lexical and speaker features, showing independent syntactic/lexical contributions and training-dependent speaker effects.
-
Why Do LLMs Struggle in Strategic Play? Broken Links Between Observations, Beliefs, and Actions
LLMs encode accurate but brittle internal beliefs about latent game states and convert them poorly into actions, creating systematic gaps that explain strategic failures.
-
Architecture Determines Observability of Transformers
Architecture and training determine whether transformers retain a readable internal signal that lets activation monitors catch errors missed by output confidence.
-
Prophecy: Inferring Formal Properties from Neuron Activations
Prophecy infers formal properties of feed-forward neural networks by extracting rules from neuron activation patterns that imply desirable output behaviors.
-
The Geometry of Truth: Emergent Linear Structure in Large Language Model Representations of True/False Datasets
At sufficient scale, LLMs linearly represent the truth value of factual statements, as shown by visualizations, cross-dataset generalization, and causal interventions that flip truth judgments.
-
BLOOM: A 176B-Parameter Open-Access Multilingual Language Model
BLOOM is a 176B-parameter open-access multilingual language model trained on the ROOTS corpus that achieves competitive performance on benchmarks, with improved results after multitask prompted finetuning.
-
When Context Misleads: Surprisal, Energy and Attention Entropy as Metrics of Coherence Illusions in LLMs
Dutch LLMs display coherence illusions tracked by surprisal, with attention entropy identifying affected heads and a new energy metric quantifying discourse coherence.
-
Characterize Then Distill: Mechanistic Reasoning in Large Output Spaces
Reasoning in large output spaces proceeds via shortlisting then fine-grained reasoning; this characterization enables a mechanistic distillation strategy that outperforms standard distillation.
-
Linguistic Productivity in Large Language Models: Models Coerce, but do not Preempt
Larger LLMs reproduce constructional productivity via entrenchment in coercion cases with nonce words but fail to use statistical preemption to avoid overgeneralizing semantically plausible but unobserved patterns.
-
Trait-space Monitoring for Emergent Misalignment During Supervised Finetuning
Trait-space drift monitoring detects emergent misalignment checkpoints in 7-9B LLMs with 2.2% FNR, 2.9% FPR and 0.99 AUROC, outperforming PCA and SAE baselines.
-
Prefix-Safe Bayesian Belief Tracking for LLM Reasoning Reliability:Separating Calibration from Ranking
SBBT separates Brier-score calibration gains from AUROC ranking gains in prefix-conditioned success estimation for LLM math reasoning, with structure-aware signals yielding up to +0.110 AUROC over baselines.