Deceptive forward passes show 2.1-2.3x higher residual rank than naive-liar passes on identical wrong answers, enabling label-free lie identification at 100% accuracy across GPT-2, Qwen, and Phi models with cross-family and cross-language transfer.
hub Canonical reference
Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 small
Canonical reference. 70% of citing Pith papers cite this work as background.
67
Pith papers citing it
50
external citations · Pith
Background
70% of classified citations
abstract
Research in mechanistic interpretability seeks to explain behaviors of machine learning models in terms of their internal components. However, most previous work either focuses on simple behaviors in small models, or describes complicated behaviors in larger models with broad strokes. In this work, we bridge this gap by presenting an explanation for how GPT-2 small performs a natural language task called indirect object identification (IOI). Our explanation encompasses 26 attention heads grouped into 7 main classes, which we discovered using a combination of interpretability approaches relying on causal interventions. To our knowledge, this investigation is the largest end-to-end attempt at reverse-engineering a natural behavior "in the wild" in a language model. We evaluate the reliability of our explanation using three quantitative criteria--faithfulness, completeness and minimality. Though these criteria support our explanation, they also point to remaining gaps in our understanding. Our work provides evidence that a mechanistic understanding of large ML models is feasible, opening opportunities to scale our understanding to both larger models and more complex tasks.
hub tools
citation-role summary
background 8
method 2
citation-polarity summary
representative citing papers
A Particle Transformer jet tagger contains a sparse six-head circuit whose source-relay-readout structure recovers most performance and whose residual stream preferentially encodes 2-prong energy correlators.
Grokking arises from gradual amplification of a Fourier-based circuit in the weights followed by removal of memorizing components.
Positional schemes set the default spectral algebra of attention heads: previous-token heads are rotational under RoPE and content-like under absolute/ALiBi, as a post-function fingerprint rather than a hard constraint.
Conditional Co-Ablation recovers self-repair backup heads in transformers by scoring conditional ablation growth, raising ROC-AUC from 0.33 to 0.91 on the IOI circuit and transferring to induction across models.
Activation patching reveals that citation decisions in Llama-3.1-8B RAG are implemented by a distributed attributional ensemble of heads and layers; targeted interventions fix most missed and spurious citations on PopQA.
SASA replaces single-vector decoders in SAEs with learned subspaces plus block sparsity and nuclear-norm regularization, proving that a single group becomes the global minimizer once block size meets intrinsic dimension and yielding polynomial rather than exponential sample complexity.
The same composed tasks are realized by different attention-head patterns in different models when the same selectivity-plus-ablation protocol is applied.
STRIDE formulates TDA as sparse recovery using steering operators that mimic subset training effects in activation space, claiming SOTA LLM pre-training attribution at 13x prior speed.
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.
Transformer Field Theory frames the residual stream as a field, models patching as source insertion, and uses first-order sensitivities plus Green functions to predict and describe responses, with empirical tests on GPT-2 autoregressive models.
KAN-SAE applies nonlinear per-feature B-spline activations in sparse autoencoders to discover 72% more alive climate features and interpretable patterns such as European heatwaves and Pacific typhoons in deep learning weather models.
Gender bias and factual gender knowledge are severely entangled in language model circuits and neurons, making neuron ablation an unreliable method for debiasing.
Temporarily reducing the learning rate on upper-layer query and key projections during early GPT pretraining prevents premature attention specialization and improves model performance.
Graph Memory Transformer replaces FFN sublayers with a graph memory cell using 128 centroids and transition matrices per block, yielding stable training at 82.2M parameters but higher validation loss than a 103M dense baseline.
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.
Diffusion models show grokking on modular addition by composing periodic operand representations in simple data regimes or by separating arithmetic computation from visual denoising across timesteps in varied regimes.
CURE disentangles LLM recommendation circuits into forget-specific, retain-specific, and task-shared modules with tailored update rules to achieve more effective unlearning than weighted baselines.
Massive activations are constant large values in LLMs that function as indispensable bias terms and concentrate attention probabilities on specific tokens.
Training per-layer affine probes on frozen transformers yields more reliable latent predictions than the logit lens and enables detection of malicious inputs from prediction trajectories.
Latent reasoning faithfulness is a property of training stage and answer format, not of architecture or the final checkpoint alone.
On Llama-2-7B, path-rerouting magnitude in paired transcoder attribution graphs correlates with jailbreak success (r=0.461), while static node metrics and top-feature ablations do not.
Prediction agreement between open and closed LLMs substantially overstates agreement on attributions and causal reasons.
ScAle learns scalar coefficients to modulate last-token attention and MLP activations in frozen VLMs, achieving up to 134.1% relative accuracy gains on spatial benchmarks with only 1K parameters.
citing papers explorer
-
Rift: A Conflict Signature for Deception in Language Models
Deceptive forward passes show 2.1-2.3x higher residual rank than naive-liar passes on identical wrong answers, enabling label-free lie identification at 100% accuracy across GPT-2, Qwen, and Phi models with cross-family and cross-language transfer.
-
Dissecting Jet-Tagger Through Mechanistic Interpretability
A Particle Transformer jet tagger contains a sparse six-head circuit whose source-relay-readout structure recovers most performance and whose residual stream preferentially encodes 2-prong energy correlators.
-
Progress measures for grokking via mechanistic interpretability
Grokking arises from gradual amplification of a Fourier-based circuit in the weights followed by removal of memorizing components.
-
Fingerprint, Not Blueprint: How Positional Schemes Set the Default Spectral Algebra of Attention
Positional schemes set the default spectral algebra of attention heads: previous-token heads are rotational under RoPE and content-like under absolute/ALiBi, as a post-function fingerprint rather than a hard constraint.
-
Conditional Co-Ablation: Recovering Self-Repair Backups in Transformer Circuits
Conditional Co-Ablation recovers self-repair backup heads in transformers by scoring conditional ablation growth, raising ROC-AUC from 0.33 to 0.91 on the IOI circuit and transferring to induction across models.
-
How Do LLMs Cite? A Mechanistic Interpretation of Attribution in Retrieval-Augmented Generation
Activation patching reveals that citation decisions in Llama-3.1-8B RAG are implemented by a distributed attributional ensemble of heads and layers; targeted interventions fix most missed and spurious citations on PopQA.
-
Subspace-Aware Sparse Autoencoders for Effective Mechanistic Interpretability
SASA replaces single-vector decoders in SAEs with learned subspaces plus block sparsity and nuclear-norm regularization, proving that a single group becomes the global minimizer once block size meets intrinsic dimension and yielding polynomial rather than exponential sample complexity.
-
Pattern Selectivity is Not Task-Causal Structure: A Cross-Architecture Mechanistic Study of Composed-Task Circuits in 1B-Class Language Models
The same composed tasks are realized by different attention-head patterns in different models when the same selectivity-plus-ablation protocol is applied.
-
STRIDE: Training Data Attribution via Sparse Recovery from Subset Perturbations
STRIDE formulates TDA as sparse recovery using steering operators that mimic subset training effects in activation space, claiming SOTA LLM pre-training attribution at 13x prior speed.
-
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.
-
Transformer Field Theory: A Response-Theoretic Approach to Mechanistic Interpretability
Transformer Field Theory frames the residual stream as a field, models patching as source insertion, and uses first-order sensitivities plus Green functions to predict and describe responses, with empirical tests on GPT-2 autoregressive models.
-
Beyond Linear Superposition: Discovering Climate Features in AI Weather Models with KAN-SAE
KAN-SAE applies nonlinear per-feature B-spline activations in sparse autoencoders to discover 72% more alive climate features and interpretable patterns such as European heatwaves and Pacific typhoons in deep learning weather models.
-
GKnow: Measuring the Entanglement of Gender Bias and Factual Gender
Gender bias and factual gender knowledge are severely entangled in language model circuits and neurons, making neuron ablation an unreliable method for debiasing.
-
Learning Less Is More: Premature Upper-Layer Attention Specialization Hurts Language Model Pretraining
Temporarily reducing the learning rate on upper-layer query and key projections during early GPT pretraining prevents premature attention specialization and improves model performance.
-
Graph Memory Transformer (GMT)
Graph Memory Transformer replaces FFN sublayers with a graph memory cell using 128 centroids and transition matrices per block, yielding stable training at 82.2M parameters but higher validation loss than a 103M dense baseline.
-
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.
-
Grokking of Diffusion Models: Case Study on Modular Addition
Diffusion models show grokking on modular addition by composing periodic operand representations in simple data regimes or by separating arithmetic computation from visual denoising across timesteps in varied regimes.
-
CURE:Circuit-Aware Unlearning for LLM-based Recommendation
CURE disentangles LLM recommendation circuits into forget-specific, retain-specific, and task-shared modules with tailored update rules to achieve more effective unlearning than weighted baselines.
-
Massive Activations in Large Language Models
Massive activations are constant large values in LLMs that function as indispensable bias terms and concentrate attention probabilities on specific tokens.
-
Eliciting Latent Predictions from Transformers with the Tuned Lens
Training per-layer affine probes on frozen transformers yields more reliable latent predictions than the logit lens and enables detection of malicious inputs from prediction trajectories.
-
Final Checkpoints Are Not Enough: Analyzing Latent Reasoning Faithfulness Along Training Trajectories
Latent reasoning faithfulness is a property of training stage and answer format, not of architecture or the final checkpoint alone.
-
Mechanistic Interpretability of LLM Jailbreaks via Internal Attribution Graphs
On Llama-2-7B, path-rerouting magnitude in paired transcoder attribution graphs correlates with jailbreak success (r=0.461), while static node metrics and top-feature ablations do not.
-
Surrogate Fidelity: When Can Open LLMs Explain Closed Ones?
Prediction agreement between open and closed LLMs substantially overstates agreement on attributions and causal reasons.
-
ScAle: Attention Head Scaling as a Minimal Adapter for Spatial Reasoning in Vision Language Models
ScAle learns scalar coefficients to modulate last-token attention and MLP activations in frozen VLMs, achieving up to 134.1% relative accuracy gains on spatial benchmarks with only 1K parameters.
-
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.
-
Sentence-Level Contextual Entrainment in Large Language Models
Sentence-level contextual entrainment exists across LLMs, weakens with scale, and is localized to 2-4% of attention heads whose deactivation removes the effect without performance loss.
-
Factual Retrieval in LLMs Is a Redundant, Distributed and Non-Contiguous Process
Attribute retrieval in LLMs follows non-contiguous, redundant layer paths identified via iterative patching, implying highly distributed knowledge storage.
-
Learning Dynamics of Chain-of-Thought State Tracking in a Solvable Transformer Model
Mean-field equations for attention retrieval, teacher alignment, and logic overlap quantitatively match simulations and predict a sharp accuracy transition in a solvable transformer for permutation state tracking.
-
Sparsely gated tiny linear experts
Sgatlin replaces transformer FF layers with sparse single linear neurons, improving perplexity across compute budgets and enabling direct interpretation of semantically clustered circuits for factual recall.
-
Beyond Encoder Accumulation: Measuring Encoder Roles in Multi-Encoder VLMs
Retraining all 31 subsets of five vision encoders shows Capacity and Necessity are distinct, pre-projector effective rank predicts residual performance at fixed parameter count, and high-Capacity plus adaptive complement pairs match the full five-encoder model.
-
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.
-
Tracing Computation Density in LLMs
LLM computation follows a consistent two-phase pattern: a sparse early-layer core reconstructs the head of the output distribution, with later layers and attention heads providing incremental refinements that correlate with model uncertainty.
-
Binding Visual Features Point by Point
Training VLMs to point via text induces serial processing that eliminates binding errors and enables compositional generalization on multi-object tasks.
-
Manifold-Guided Attention Steering
MAGS learns low-dimensional subspaces from correct versus incorrect reasoning traces and applies targeted projection corrections to attention heads when they deviate from the correctness manifold during inference.
-
How to Interpret Agent Behavior
ACT*ONOMY is a Grounded-Theory-derived hierarchical taxonomy and open repository that enables systematic comparison and characterization of autonomous agent behavior across trajectories.
-
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.
-
Not How Many, But Which: Parameter Placement in Low-Rank Adaptation
Gradient-informed placement of LoRA parameters recovers full performance under GRPO while random placement does not, due to differences in gradient rank and stability across training regimes.
-
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.
-
Where's the Plan? Locating Latent Planning in Language Models with Lightweight Mechanistic Interventions
Future-rhyme information is linearly decodable at line boundaries across model families and strengthens with scale, yet only Gemma-3-27B causally depends on it, with the driver migrating to the boundary around layer 30 and localizing to five attention heads.
-
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.
-
The Position Curse: LLMs Struggle to Locate the Last Few Items in a List
LLMs exhibit the Position Curse, with backward position retrieval in lists lagging far behind forward retrieval, showing only partial gains from PosBench fine-tuning.
-
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.
-
The Illusion of Equivalence: Systematic FP16 Divergence in KV-Cached Autoregressive Inference
FP16 KV caching in transformers causes deterministic token divergence versus cache-free inference due to non-associative floating-point accumulation orderings.
-
Inside-Out: Measuring Generalization in Vision Transformers Through Inner Workings
Circuit-based metrics from Vision Transformer internals provide better label-free proxies for generalization under distribution shift than existing methods like model confidence.
-
PhiNet: Speaker Verification with Phonetic Interpretability
PhiNet adds phonetic interpretability to speaker verification while matching the accuracy of standard black-box models on VoxCeleb, SITW, and LibriSpeech.
-
Beyond I'm Sorry, I Can't: Dissecting Large Language Model Refusal
Sparse autoencoders plus greedy filtering and factorization-machine interaction modeling identify minimal sets of features in Gemma-2-2B-IT and LLaMA-3.1-8B-IT whose ablation produces jailbreaks by flipping refusal to compliance.
-
Sparse Autoencoders Find Highly Interpretable Features in Language Models
Sparse autoencoders applied to language model activations yield more interpretable and monosemantic features than alternative approaches, enabling finer causal analysis on the indirect object identification task.
-
Inherited Circuits, Learned Semantics: How Fine-Tuning Creates Evasion Vulnerabilities Invisible to Standard Evaluation
Fine-tuning security LLMs specializes inherited classification circuits into token-level indicators that preserve canonical accuracy but fail under behavior-preserving transformations like aliasing and case mutation.