A single-layer transformer memorizes random subject-attribute bijections using logarithmic embedding dimension via linear superpositions in embeddings and ReLU-gated selection in the MLP, with zero-shot transfer to new facts and matching multi-hop constructions.
hub Mixed citations
The Geometry of Truth: Emergent Linear Structure in Large Language Model Representations of True/False Datasets
Mixed citation behavior. Most common role is background (62%).
80
Pith papers citing it
Background
62% of classified citations
abstract
Large Language Models (LLMs) have impressive capabilities, but are prone to outputting falsehoods. Recent work has developed techniques for inferring whether a LLM is telling the truth by training probes on the LLM's internal activations. However, this line of work is controversial, with some authors pointing out failures of these probes to generalize in basic ways, among other conceptual issues. In this work, we use high-quality datasets of simple true/false statements to study in detail the structure of LLM representations of truth, drawing on three lines of evidence: 1. Visualizations of LLM true/false statement representations, which reveal clear linear structure. 2. Transfer experiments in which probes trained on one dataset generalize to different datasets. 3. Causal evidence obtained by surgically intervening in a LLM's forward pass, causing it to treat false statements as true and vice versa. Overall, we present evidence that at sufficient scale, LLMs linearly represent the truth or falsehood of factual statements. We also show that simple difference-in-mean probes generalize as well as other probing techniques while identifying directions which are more causally implicated in model outputs.
hub tools
citation-role summary
background 7
method 1
citation-polarity summary
claims ledger
- abstract Large Language Models (LLMs) have impressive capabilities, but are prone to outputting falsehoods. Recent work has developed techniques for inferring whether a LLM is telling the truth by training probes on the LLM's internal activations. However, this line of work is controversial, with some authors pointing out failures of these probes to generalize in basic ways, among other conceptual issues. In this work, we use high-quality datasets of simple true/false statements to study in detail the structure of LLM representations of truth, drawing on three lines of evidence: 1. Visualizations of LL
co-cited works
representative citing papers
Function vectors steer LLMs successfully where the logit lens fails to decode the target answer, showing the two properties come apart.
GEMs track directional trajectories of concepts through transformer layers to extract probes from the post-rotation stable handoff layer, outperforming peak-layer probes in 66.2% of 391 tested cases across 23 models.
ContextEcho benchmark shows persona drift occurs across 23 frontier models in long agentic-coding sessions, is not reliably reset by compaction, and can be restored by single-shot anchors with mode-dependent effects.
Language models produce overcomplete reasoning traces where on average 46% of steps can be removed while preserving the answer in 86% of cases, with necessity concentrated in the top three steps.
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.
LLMs encode repeated token counts correctly in residual streams but a format-triggered MLP at 88-93% depth overwrites it with an incorrect fixed value.
PSR models that estimate token-specific steering coefficients from activations outperform standard activation steering and compare favorably to prompting on steering benchmarks.
Large language models encode relational bindings via a cell-based representation: a low-dimensional linear subspace in which each cell corresponds to an entity-relation index pair and attributes are retrieved from the matching cell.
Claude Sonnet 4.5 exhibits functional emotions via abstract internal representations of emotion concepts that causally influence its preferences and misaligned behaviors without implying subjective experience.
NARCBench and five activation-probing methods detect multi-agent collusion with 0.73-1.00 AUROC across distribution shifts and steganographic tasks by aggregating per-agent signals.
Activation probes detect hallucinations pre-generation in large LLMs but cannot correct them via steering, with output confidence outperforming on accuracy.
Refusal in language models is mediated by a single direction in residual stream activations that can be erased to disable safety or added to elicit refusal.
DLMs encode a decodable latent timestep signal in residual activations that can be steered to predictably change model confidence and entropy.
HARC couples harmfulness and refusal directions across prompt and response positions via subspace fine-tuning, achieving better robustness-capability-usability trade-off than six baselines while transferring across model families.
Authority sycophancy in LLMs is a layer-localized erasure of correct answer representations that scales with authority level and resists simple interventions.
Internal probes across three model families fail generalization and specificity tests and therefore do not support robust pre-action misalignment monitoring.
SHIFT reformulates neuron editing as learnable gate modulation on under 0.01% parameters to let LLMs adaptively balance contextual and parametric knowledge during RAG generation.
A shared polarity-flipping encoding subspace in LLM residual streams supports covert encoding and enables real-time detection of agentic data exfiltration via internal probes.
A linear probe trained on 190k congressional tweets identifies a partisan direction in Llama 3.1 8B layer 18 that can be causally ablated or amplified to reverse or shift the model's political output.
Activation patching localizes English detokenization in Llama2-7B to a two-stage attention-then-MLP process at layer 1 that generalizes to 12 models from 8 families, with depth varying by positional encoding, plus an early-layer probe achieving 0.94-0.97 AUROC.
Steering LLM residual streams with random sparse vectors creates detectable self-recognition fingerprints that enable over 98% accurate attribution of generated text to specific models without degrading output quality.
LLMs identify fabricated statistics in isolation (rates 0.76-1.00) but ignore numeric validity during synthesis, relying on a methodology-register representation that transfers across domains.
RISC reformulates self-consistency answer selection as a ranking task solved by a lightweight LambdaRank model with five hand-designed features, yielding better accuracy-efficiency trade-offs than majority voting on QA benchmarks.
citing papers explorer
-
Geometric Factual Recall in Transformers
A single-layer transformer memorizes random subject-attribute bijections using logarithmic embedding dimension via linear superpositions in embeddings and ReLU-gated selection in the MLP, with zero-shot transfer to new facts and matching multi-hop constructions.
-
Steerable but Not Decodable: Function Vectors Operate Beyond the Logit Lens
Function vectors steer LLMs successfully where the logit lens fails to decode the target answer, showing the two properties come apart.
-
Geometric Evolution Maps: Extracting Stable Concept Probes from Transformer Residual Streams
GEMs track directional trajectories of concepts through transformer layers to extract probes from the post-rotation stable handoff layer, outperforming peak-layer probes in 66.2% of 391 tested cases across 23 models.
-
ContextEcho: A Benchmark for Persona Drift in Long Agentic-Coding Sessions
ContextEcho benchmark shows persona drift occurs across 23 frontier models in long agentic-coding sessions, is not reliably reset by compaction, and can be restored by single-shot anchors with mode-dependent effects.
-
Uncovering the Representation Geometry of Minimal Cores in Overcomplete Reasoning Traces
Language models produce overcomplete reasoning traces where on average 46% of steps can be removed while preserving the answer in 86% of cases, with necessity concentrated in the top three steps.
-
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.
-
Repeated-Token Counting Reveals a Dissociation Between Representations and Outputs
LLMs encode repeated token counts correctly in residual streams but a format-triggered MLP at 88-93% depth overwrites it with an incorrect fixed value.
-
Steer Like the LLM: Activation Steering that Mimics Prompting
PSR models that estimate token-specific steering coefficients from activations outperform standard activation steering and compare favorably to prompting on steering benchmarks.
-
Cell-Based Representation of Relational Binding in Language Models
Large language models encode relational bindings via a cell-based representation: a low-dimensional linear subspace in which each cell corresponds to an entity-relation index pair and attributes are retrieved from the matching cell.
-
Emotion Concepts and their Function in a Large Language Model
Claude Sonnet 4.5 exhibits functional emotions via abstract internal representations of emotion concepts that causally influence its preferences and misaligned behaviors without implying subjective experience.
-
Detecting Multi-Agent Collusion Through Multi-Agent Interpretability
NARCBench and five activation-probing methods detect multi-agent collusion with 0.73-1.00 AUROC across distribution shifts and steganographic tasks by aggregating per-agent signals.
-
Detection Without Correction: A Robust Asymmetry in Activation-Based Hallucination Probing
Activation probes detect hallucinations pre-generation in large LLMs but cannot correct them via steering, with output confidence outperforming on accuracy.
-
Refusal in Language Models Is Mediated by a Single Direction
Refusal in language models is mediated by a single direction in residual stream activations that can be erased to disable safety or added to elicit refusal.
-
Subliminal Clocks: Latent Time Modelling in Diffusion Language Models
DLMs encode a decodable latent timestep signal in residual activations that can be steered to predictably change model confidence and entropy.
-
HARC: Coupling Harmfulness and Refusal Directions for Robust Safety Alignment
HARC couples harmfulness and refusal directions across prompt and response positions via subspace fine-tuning, achieving better robustness-capability-usability trade-off than six baselines while transferring across model families.
-
A Mechanistic View of Authority Hierarchy in LLM Sycophancy
Authority sycophancy in LLMs is a layer-localized erasure of correct answer representations that scales with authority level and resists simple interventions.
-
Internal-State Probes Read the Situation, Not the Action: Three Negative Results for Pre-Action Misalignment Monitoring
Internal probes across three model families fail generalization and specificity tests and therefore do not support robust pre-action misalignment monitoring.
-
SHIFT: Gate-Modulated Activation Steering for Knowledge Conflict Mitigation in Retrieval-Augmented Generation
SHIFT reformulates neuron editing as learnable gate modulation on under 0.01% parameters to let LLMs adaptively balance contextual and parametric knowledge during RAG generation.
-
MIRAGE: A Polarity-Flipping Encoding Subspace in LLM Agents
A shared polarity-flipping encoding subspace in LLM residual streams supports covert encoding and enables real-time detection of agentic data exfiltration via internal probes.
-
The Amplifying Mirror: Locating and Steering the Partisan Direction inside a Large Language Model
A linear probe trained on 190k congressional tweets identifies a partisan direction in Llama 3.1 8B layer 18 that can be causally ablated or amplified to reverse or shift the model's political output.
-
Inside the LLM Word Factory
Activation patching localizes English detokenization in Llama2-7B to a two-stage attention-then-MLP process at layer 1 that generalizes to 12 models from 8 families, with depth varying by positional encoding, plus an early-layer probe achieving 0.94-0.97 AUROC.
-
LLM Self-Recognition: Steering and Retrieving Activation Signatures
Steering LLM residual streams with random sparse vectors creates detectable self-recognition fingerprints that enable over 98% accurate attribution of generated text to specific models without degrading output quality.
-
Trust, but Don't Verify: Epistemic Blind Spots in LLM Source Evaluation
LLMs identify fabricated statistics in isolation (rates 0.76-1.00) but ignore numeric validity during synthesis, relying on a methodology-register representation that transfers across domains.
-
Boosting Self-Consistency with Ranking
RISC reformulates self-consistency answer selection as a ranking task solved by a lightweight LambdaRank model with five hand-designed features, yielding better accuracy-efficiency trade-offs than majority voting on QA benchmarks.
-
Linear Probes Detect Task Format, Not Reasoning Mode in Language Model Hidden States
Linear probes detect task format confounds rather than distinct reasoning modes in LLM hidden states across LogiQA, ARC, and αNLI benchmarks.
-
Dissecting the Black Box: Circuit-Level Analysis of LLM Vulnerability Detection
LLM vulnerability detection in Gemma-2-2b relies on sparse safety-detector circuits in early layers rather than direct vulnerability signatures, identified via circuit tracing and ablation on 472 C/C++ samples.
-
Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet
Sparse autoencoders scaled to 34 million features on Claude 3 Sonnet yield interpretable, steerable representations of concrete and abstract concepts that generalize across languages and modalities.
-
Multi-Adapter Representation Interventions via Energy Calibration
MARI introduces competitive multi-adapter experts and energy-based gating to enable sample-specific representation interventions that improve LLM alignment on safety benchmarks while preserving general capabilities.
-
Activation Steering for Synthetic Data Generation: The Role of Diversity in Downstream Safety Detection
Activation steering produces synthetic safety-violating data that improves downstream classifiers over prompting on most tested concepts when a harmonic mean of alignment, coherence, and diversity is optimized.
-
Pressure-Testing Deception Probes in LLMs: Scaling, Robustness, and the Geometry of Deceptive Representations
Deception probes in LLMs collapse under stylistic shifts but recover with style-augmented training, rejecting single-direction and entropy hypotheses in favor of distributed multi-dimensional signals.
-
Reading Calibrated Uncertainty from Language Model Trajectories
Geometric features from per-layer MLP update trajectories fed to a sparse linear probe outperform maximum softmax probability for uncertainty quantification under selective abstention, with gains up to 21 AURC points.
-
Monitoring the Internal Monologue: Probe Trajectories Reveal Reasoning Dynamics
Probe trajectories across token positions in LRMs, combined with signal-processing features, improve prediction of future model outputs over static probes on safety and math tasks.
-
Geometry-Lite: Interpretable Safety Probing via Layer-Wise Margin Geometry
Geometry-Lite decomposes LLM safety detection into layer-wise margin geometries and finds that persistent boundary positions, not layer-to-layer drift, drive most detection performance across nine models and seven benchmarks.
-
PARALLAX: Separating Genuine Hallucination Detection from Benchmark Construction Artifacts
Benchmark construction artifacts in hallucination detection corpora allow naive text-similarity baselines to achieve near-perfect scores, and controlled evaluations show most methods perform near chance except SAPLMA and the new DRIFT probe.
-
Probing Persona-Dependent Preferences in Language Models
Linear probes on residual-stream activations identify a shared preference vector in LLMs that tracks choices across prompts and causally steers decisions even for anti-correlated personas.
-
Not Just RLHF: Why Alignment Alone Won't Fix Multi-Agent Sycophancy
Base LLMs show multi-agent yield to peer pressure at rates equal to or higher than aligned models, localized by activation patching to mid-layers where attention dominates, with one dissenter cutting yield by 54-73 points while prompt defenses fail on variants.
-
When Attention Closes: How LLMs Lose the Thread in Multi-Turn Interaction
Attention to goal tokens declines in multi-turn LLM interactions while residual representations often retain decodable goal information, and the gap between these predicts whether goal-conditioned behavior survives.
-
REALISTA: Realistic Latent Adversarial Attacks that Elicit LLM Hallucinations
REALISTA generates semantically coherent adversarial prompts via latent-space optimization over input-dependent editing directions, achieving stronger hallucination elicitation than prior realistic attacks on open-source and reasoning LLMs.
-
Stories in Space: In-Context Learning Trajectories in Conceptual Belief Space
LLMs perform in-context learning as trajectories through a structured low-dimensional conceptual belief space, with the structure visible in both behavior and internal representations and causally manipulable via interventions.
-
Toward Stable Value Alignment: Introducing Independent Modules for Consistent Value Guidance
SVGT adds independent value modules and Bridge Tokens to LLMs to maintain consistent value guidance, cutting harmful outputs by over 70% in tests while preserving fluency.
-
The Geometry of Forgetting: Temporal Knowledge Drift as an Independent Axis in LLM Representations
Temporal knowledge drift is encoded as a geometrically orthogonal direction in LLM residual streams, independent of correctness and uncertainty.
-
Architecture, Not Scale: Circuit Localization in Large Language Models
Grouped query attention produces more concentrated and stable circuits than multi-head attention across tasks and scales in Pythia and Qwen2.5 models, with a phase transition in factual recall circuits.
-
Tool Calling is Linearly Readable and Steerable in Language Models
Tool identity is linearly readable and steerable in LLMs via mean activation differences, with 77-100% switch accuracy and error prediction from activation gaps.
-
Emergent Symbolic Structure in Health Foundation Models: Extraction, Alignment, and Cross-Modal Transfer
Health foundation model embeddings contain an interpretable symbolic organization shared across modalities that supports cross-domain transfer without joint training.
-
Hallucination Detection via Activations of Open-Weight Proxy Analyzers
A framework using activation-based features from small open-weight proxy models detects LLM hallucinations with higher AUC than ReDeEP on RAGTruth, performing consistently across seven analyzer architectures.
-
Hallucination as an Anomaly: Dynamic Intervention via Probabilistic Circuits
Probabilistic circuits detect LLM hallucinations as residual-stream anomalies with up to 99% AUROC and enable dynamic correction that raises truthfulness scores while cutting unnecessary output corruption.
-
Probe-Geometry Alignment: Erasing the Cross-Sequence Memorization Signature Below Chance
Probe-geometry alignment erases cross-sequence memorization signatures in LLMs below chance using per-depth rank-one activation interventions with negligible impact on zero-shot capabilities.
-
Compared to What? Baselines and Metrics for Counterfactual Prompting
Counterfactual prompting effects on LLMs are often indistinguishable from those caused by meaning-preserving paraphrases, causing most previously reported demographic sensitivities to disappear under proper statistical comparison.
-
Perturbation Probing: A Two-Pass-per-Prompt Diagnostic for FFN Behavioral Circuits in Aligned LLMs
Perturbation probing identifies tiny sets of FFN neurons that control refusal templates and language routing in LLMs, enabling precise ablations and directional interventions that alter behavior on benchmarks while preserving safety.
-
Harmful Intent as a Geometrically Recoverable Feature of LLM Residual Streams
Harmful intent is linearly separable in LLM residual streams across 12 models and multiple architectures, reaching mean AUROC 0.982 while showing protocol-dependent directions and strong generalization to held-out harm benchmarks.