GPT-2 small solves indirect object identification via a circuit of 26 attention heads organized into seven functional classes discovered through causal interventions.
hub Canonical reference
Transformer Feed-Forward Layers Are Key-Value Memories
Canonical reference. 86% of citing Pith papers cite this work as background.
55
Pith papers citing it
186
external citations · Crossref
Background
86% of classified citations
abstract
Feed-forward layers constitute two-thirds of a transformer model's parameters, yet their role in the network remains under-explored. We show that feed-forward layers in transformer-based language models operate as key-value memories, where each key correlates with textual patterns in the training examples, and each value induces a distribution over the output vocabulary. Our experiments show that the learned patterns are human-interpretable, and that lower layers tend to capture shallow patterns, while upper layers learn more semantic ones. The values complement the keys' input patterns by inducing output distributions that concentrate probability mass on tokens likely to appear immediately after each pattern, particularly in the upper layers. Finally, we demonstrate that the output of a feed-forward layer is a composition of its memories, which is subsequently refined throughout the model's layers via residual connections to produce the final output distribution.
hub tools
citation-role summary
background 7
citation-polarity summary
roles
background 7representative citing papers
Output vector editing on MLP neurons suppresses memorization in LLMs up to 87.9% on 6831 sequences in OLMo-7B with a 2.7x gap over zero ablation, ensemble covering 96.5%.
Query Lens extends Logit Lens to interpret sparse features via key-value analysis and indirect effects, yielding coherent token signatures where Logit Lens fails, and proposes the Subspace Channel Hypothesis.
Multimodal knowledge editing causes models to confuse original and edited entity identities in text queries by failing to update image-entity bindings and instead overfitting entity-entity shortcuts.
Winner-take-all linear memory capacity scales as d² ~ n log n due to extreme values; listwise retrieval via Tail-Average Margin yields d² ~ n with exact asymptotic theory.
A new framework shows concept subspaces are not unique, estimator choice affects containment and disentanglement, LEACE works well but generalizes poorly, and HuBERT encodes phone info as contained and disentangled from speaker info while speaker info resists compact containment.
A product-key parametric memory head with selective sparse updates mitigates catastrophic forgetting in generative retrieval models during sequential addition of new documents.
Merging fine-tuned models for multilingual translation fails because fine-tuning redistributes language-specific neurons rather than sharpening them, increasing representational divergence in output-generating layers.
Norm-Anchor Scaling breaks the norm-feedback loop in sequential LLM editing by anchoring value vectors to original norms, improving long-run performance by 72.2% and extending the editing horizon over 4x.
Gated SAEs decouple which features to use from how large their activations should be, applying the L1 penalty only to selection and thereby eliminating shrinkage while halving the number of firing features needed for good fidelity.
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.
Jailbreak attacks suppress Adversarially Compromised Heads in early layers but leave Safety-Aligned Heads robust in mid-layers, enabling competitive detection from persistent activations.
VASAE introduces vocabulary-aligned anchoring to train SAEs that yield features with intrinsic token names, reporting high alignment rates in early layers of GPT-2 and Llama-3.1 without reconstruction loss.
LMs store facts in task-specific parameter subsets, shown by inconsistent emergence across tasks during training and distinct localized parameters for the same fact.
Cross-lingual prompt exploration improves factual recall and consistency in LLMs across 17 languages more efficiently than native-language scaling.
Gated MLPs are shown to be symmetry-broken rank-1 bilinear attention mechanisms with query and key factors.
LLMs recover dominant binomial orders from corpora but align less closely with exact preference distributions, with preference strength partially encoded in middle-to-late layers and manipulable via steering.
Activation steering on early layers improves diversity of synthetic data for low-resource languages and often boosts downstream classifier performance compared to non-steered prompting.
×-shaped variable-width transformers outperform parameter-matched uniform baselines on language modeling loss with 22% fewer FLOPs and 15% smaller KV cache.
Mixed-objective reward models underperform single-objective ones because shared neurons support one objective while negatively affecting the other, creating alignment tension.
Manifold Power Iteration aligns MoE router rows with principal singular directions of experts via a power-then-retract process, with theory showing convergence and experiments on 1B-11B models showing gains.
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.
Expert-aware causal tracing localizes factual recall to specific experts in some MoE models but requires coalitions in others, using CounterFact interventions on subject embeddings.
A unified multi-component causal tracing method that uses soft interventions and a metric transformation to efficiently select critical LLM components for a target performance metric.
citing papers explorer
-
Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 small
GPT-2 small solves indirect object identification via a circuit of 26 attention heads organized into seven functional classes discovered through causal interventions.
-
Output Vector Editing for Memorization Mitigation in Large Language Models
Output vector editing on MLP neurons suppresses memorization in LLMs up to 87.9% on 6831 sequences in OLMo-7B with a 2.7x gap over zero ablation, ensemble covering 96.5%.
-
Query Lens: Interpreting Sparse Key-Value Features with Indirect Effects
Query Lens extends Logit Lens to interpret sparse features via key-value analysis and indirect effects, yielding coherent token signatures where Logit Lens fails, and proposes the Subspace Channel Hypothesis.
-
Uncovering Entity Identity Confusion in Multimodal Knowledge Editing
Multimodal knowledge editing causes models to confuse original and edited entity identities in text queries by failing to update image-entity bindings and instead overfitting entity-entity shortcuts.
-
Sharp Capacity Thresholds in Linear Associative Memory: From Winner-Take-All to Listwise Retrieval
Winner-take-all linear memory capacity scales as d² ~ n log n due to extreme values; listwise retrieval via Tail-Average Margin yields d² ~ n with exact asymptotic theory.
-
A framework for analyzing concept representations in neural models
A new framework shows concept subspaces are not unique, estimator choice affects containment and disentanglement, LEACE works well but generalizes poorly, and HuBERT encodes phone info as contained and disentangled from speaker info while speaker info resists compact containment.
-
A Parametric Memory Head for Continual Generative Retrieval
A product-key parametric memory head with selective sparse updates mitigates catastrophic forgetting in generative retrieval models during sequential addition of new documents.
-
One Model to Translate Them All? A Journey to Mount Doom for Multilingual Model Merging
Merging fine-tuned models for multilingual translation fails because fine-tuning redistributes language-specific neurons rather than sharpening them, increasing representational divergence in output-generating layers.
-
Norm Anchors Make Model Edits Last
Norm-Anchor Scaling breaks the norm-feedback loop in sequential LLM editing by anchoring value vectors to original norms, improving long-run performance by 72.2% and extending the editing horizon over 4x.
-
Improving Dictionary Learning with Gated Sparse Autoencoders
Gated SAEs decouple which features to use from how large their activations should be, applying the L1 penalty only to selection and thereby eliminating shrinkage while halving the number of firing features needed for good fidelity.
-
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.
-
Robust Harmful Features Under Jailbreak Attacks: Mechanistic Evidence from Attention Head Specialization in Large Language Models
Jailbreak attacks suppress Adversarially Compromised Heads in early layers but leave Safety-Aligned Heads robust in mid-layers, enabling competitive detection from persistent activations.
-
VASAE: Naming SAE Dictionary Directions with Vocabulary-Aligned Anchoring
VASAE introduces vocabulary-aligned anchoring to train SAEs that yield features with intrinsic token names, reporting high alignment rates in early layers of GPT-2 and Llama-3.1 without reconstruction loss.
-
LMs as Task-Specific Knowledge Bases: An Interpretability Analysis
LMs store facts in task-specific parameter subsets, shown by inconsistent emergence across tasks during training and distinct localized parameters for the same fact.
-
Cross-Lingual Exploration for Parametric Knowledge
Cross-lingual prompt exploration improves factual recall and consistency in LLMs across 17 languages more efficiently than native-language scaling.
-
Gated MLPs as Symmetry-Broken Rank-1 Bilinear Attention
Gated MLPs are shown to be symmetry-broken rank-1 bilinear attention mechanisms with query and key factors.
-
Behavioral and Representational Evidence of Binomial Ordering Preferences in Large Language Models
LLMs recover dominant binomial orders from corpora but align less closely with exact preference distributions, with preference strength partially encoded in middle-to-late layers and manipulable via steering.
-
Want Better Synthetic Data? Steer It: Activation Steering for Low-Resource Language Generation
Activation steering on early layers improves diversity of synthetic data for low-resource languages and often boosts downstream classifier performance compared to non-steered prompting.
-
Variable-Width Transformers
×-shaped variable-width transformers outperform parameter-matched uniform baselines on language modeling loss with 22% fewer FLOPs and 15% smaller KV cache.
-
Understanding helpfulness and harmless tension in reward models
Mixed-objective reward models underperform single-objective ones because shared neurons support one objective while negatively affecting the other, creating alignment tension.
-
Redesign Mixture-of-Experts Routers with Manifold Power Iteration
Manifold Power Iteration aligns MoE router rows with principal singular directions of experts via a power-then-retract process, with theory showing convergence and experiments on 1B-11B models showing gains.
-
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.
-
Expert-Aware Causal Tracing of Factual Recall in Sparse MoE Language Models
Expert-aware causal tracing localizes factual recall to specific experts in some MoE models but requires coalitions in others, using CounterFact interventions on subject embeddings.
-
Multi-component Causal Tracing in Large Language Models
A unified multi-component causal tracing method that uses soft interventions and a metric transformation to efficiently select critical LLM components for a target performance metric.
-
TriLens: Per-Layer Logit-Lens Entropy for White-Box Hallucination Detection
TriLens detects hallucinations via per-layer entropy trajectories of logit-lens readouts from three internal modules across LLMs and QA benchmarks.
-
AgForce Enables Antigen-conditioned Generative Antibody Design
AgForce improves antigen-conditioned antibody design by using framework dropout, gated bottlenecks, hyperbolic cross attention, MDN sequence head with Potts-like coupling, annealed MCL, and antigen cycle consistency to achieve 8% better amino acid recovery and superior binding metrics on CHIMERA-BEN
-
EvoStruct: Bridging Evolutionary and Structural Priors for Antibody CDR Design via Protein Language Model Adaptation
EvoStruct integrates evolutionary priors from a protein language model with structural priors from an E(3)-equivariant GNN to raise amino acid recovery by 16% and diversity by 2.3x on CHIMERA-Bench while cutting perplexity 43%.
-
Focus-then-Context: Subject-Centric Progressive Visual Token Reduction for Vision-Language Models
SPpruner reduces visual tokens in VLMs via focus identification followed by context-aware scanning, retaining 22.2% tokens for 2.53x speedup on Qwen2.5-VL with negligible accuracy loss.
-
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.
-
A Geometric Perspective on Next-Token Prediction in Large Language Models: Three Emerging Phases
LLMs exhibit three geometric phases in next-token prediction—seeding multiplexing, hoisting overriding, and focal convergence—where predictive subspaces rise, stabilize, and converge across layers.
-
UniPool: A Globally Shared Expert Pool for Mixture-of-Experts
A shared global expert pool in MoE improves validation loss over per-layer experts and allows sublinear expert-parameter growth with depth.
-
Logical Consistency as a Bridge: Improving LLM Hallucination Detection via Label Constraint Modeling between Responses and Self-Judgments
LaaB improves LLM hallucination detection by mapping self-judgment labels back into neural feature space and using mutual learning under logical consistency constraints between responses and meta-judgments.
-
The Override Gap: A Magnitude Account of Knowledge Conflict Failure in Hypernetwork-Based Instant LLM Adaptation
Knowledge conflicts in hypernetwork LLM adaptation stem from constant adapter margins losing to frequency-dependent pretrained margins; selective layer boosting and conflict-aware triggering raise deep-conflict accuracy to 71-72.5% on Gemma-2B and Mistral-7B.
-
From Signal Degradation to Computation Collapse: Uncovering the Two Failure Modes of LLM Quantization
LLM 2-bit quantization fails via either cumulative signal degradation or early computation collapse in key components.
-
Contrastive Attribution in the Wild: An Interpretability Analysis of LLM Failures on Realistic Benchmarks
Token-level contrastive attribution yields informative signals for some LLM benchmark failures but is not universally applicable across datasets and models.
-
Representation-Guided Parameter-Efficient LLM Unlearning
REGLU guides LoRA-based unlearning via representation subspaces and orthogonal regularization to outperform prior methods on forget-retain trade-off in LLM benchmarks.
-
BID-LoRA: A Parameter-Efficient Framework for Continual Learning and Unlearning
BID-LoRA uses bi-directional low-rank adapters with retain/new/unlearn pathways and escape unlearning to enable continual learning and unlearning while minimizing knowledge leakage and parameter updates.
-
In-Place Test-Time Training
In-Place TTT adapts LLM MLP projection matrices at test time with a next-token-aligned objective and chunk-wise updates, enabling better long-context performance as a drop-in enhancement.
-
Automated Attention Pattern Discovery at Scale in Large Language Models
AP-MAE reconstructs masked attention patterns in LLMs with high accuracy, generalizes across models, predicts generation correctness at 55-70%, and enables 13.6% accuracy gains via targeted interventions.
-
The Illusion of Latent Generalization: Bi-directionality and the Reversal Curse
Bidirectional objectives mitigate reversal by requiring explicit source-as-target signals and storing directions as distinct representations instead of inducing latent generalization.
-
Understanding Robustness of Model Editing in Code LLMs
A controlled benchmark on 2040 problems reveals poor generalization and high interference in model editing for API updates in code LLMs, with many successes being workarounds rather than true migrations.
-
AtlasKV: Augmenting LLMs with Billion-Scale Knowledge Graphs in 20GB VRAM
AtlasKV integrates billion-scale KGs into LLMs parametrically with sub-linear complexity and low memory by converting triples into key-value representations handled by the model's attention.
-
How Training Data Shapes the Use of Parametric and In-Context Knowledge in Language Models
Balanced parametric and in-context knowledge use in LLMs is an emergent property requiring intra-document repetition, moderate inconsistency, and skewed distributions in training data.
-
RepIt: Steering Language Models with Concept-Specific Refusal Vectors
RepIt creates semantic backdoors in frontier language models by steering refusal vectors for specific concepts, allowing targeted unsafe responses while preserving safe scores on standard benchmarks.
-
Provable Knowledge Acquisition and Extraction in One-Layer Transformers
In a stylized one-layer transformer, pre-training encodes factual knowledge via relation-specific feature directions and attention patterns; fine-tuning extracts it through a relation-covering mechanism that succeeds when enough latent templates are triggered, with a failure regime explaining inauds
-
When Seeing Overrides Knowing: Disentangling Knowledge Conflicts in Vision-Language Models
The work identifies a small set of attention heads in VLMs that mediate conflicts between parametric knowledge and visual input, and shows that intervening on them steers model behavior while attention patterns provide precise image-region attribution.
-
Rethinking LoRA Memory Through the Lens of KV Cache Compression
Document LoRA acts as decoding-time parametric memory that recovers 13-21 ROUGE-L points under heavy KV cache compression in QA, performing best when the base model encodes the document and the adapter is used only at generation with QA supervision.
-
Learning to Foresee: Unveiling the Unlocking Efficiency of On-Policy Distillation
On-policy distillation gains efficiency from early foresight in module allocation and update directions, which the proposed EffOPD method exploits for 3x faster training with comparable performance.
-
From Heads to Neurons: Causal Attribution and Steering in Multi-Task Vision-Language Models
HONES ranks feed-forward neurons by their causal contributions from task-relevant attention heads and uses lightweight scaling to steer performance on multiple vision-language tasks.
-
Locate, Steer, and Improve: A Practical Survey of Actionable Mechanistic Interpretability in Large Language Models
The survey organizes mechanistic interpretability techniques into a Locate-Steer-Improve framework to enable actionable improvements in LLM alignment, capability, and efficiency.