Sparse autoencoders applied to Neural Quantum States extract unsupervised features correlating with and causally steering physical observables such as order parameters while preserving variational energy.
hub
Locating and Editing Factual Associations in GPT
48 Pith papers cite this work, alongside 282 external citations. Polarity classification is still indexing.
48
Pith papers citing it
282
external citations · Pith
abstract
We analyze the storage and recall of factual associations in autoregressive transformer language models, finding evidence that these associations correspond to localized, directly-editable computations. We first develop a causal intervention for identifying neuron activations that are decisive in a model's factual predictions. This reveals a distinct set of steps in middle-layer feed-forward modules that mediate factual predictions while processing subject tokens. To test our hypothesis that these computations correspond to factual association recall, we modify feed-forward weights to update specific factual associations using Rank-One Model Editing (ROME). We find that ROME is effective on a standard zero-shot relation extraction (zsRE) model-editing task, comparable to existing methods. To perform a more sensitive evaluation, we also evaluate ROME on a new dataset of counterfactual assertions, on which it simultaneously maintains both specificity and generalization, whereas other methods sacrifice one or another. Our results confirm an important role for mid-layer feed-forward modules in storing factual associations and suggest that direct manipulation of computational mechanisms may be a feasible approach for model editing. The code, dataset, visualizations, and an interactive demo notebook are available at https://rome.baulab.info/
hub tools
citation-role summary
background 3
method 1
citation-polarity summary
representative citing papers
LLM activations encode current and prior entities in orthogonal slots, but models only use the current slot for explicit factual retrieval despite prior-slot information being linearly decodable.
Grokking arises from gradual amplification of a Fourier-based circuit in the weights followed by removal of memorizing components.
GPT-2 small solves indirect object identification via a circuit of 26 attention heads organized into seven functional classes discovered through causal interventions.
A causal audit of LMLMs finds near-zero parametric leakage after deletion, with surviving correctness coming from retrieval artifacts in the database.
Obliviate erases targeted concepts from autoregressive image generators via KL supervision on visual tokens over full trajectories, cutting nudity rates sharply on benchmarks while keeping general performance.
VLMs default to visual grounding but a sparse circuit of 2.5-4.8% attention heads in later layers mediates prior-knowledge overrides, identified causally via patching and ablation across three model families.
Auditability of subliminal learning is constrained by channel location, with initialization-dependent body channels allowing pre-training screens while vocabulary geometry and conditional body channels evade them.
SWITCH uses explicit <swi> and </swi> boundary tokens to make latent chain-of-thought compatible with on-policy RL (GRPO) and open to causal mechanistic probing, outperforming prior hidden-state recurrence methods.
Fragility, the activation noise level causing probe accuracy collapse, reveals evolving lexical-to-compositional moral encoding, layer robustness gradients, and fine-tuning differences invisible to saturated probing accuracy.
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.
Face-Feature Tuning is a label-free logit remapping method that reduces FPR/TPR gaps across groups in deepfake detection while preserving overall accuracy.
Locate-then-edit succeeds at the same early-to-mid MLP locations in masked diffusion models as in autoregressive models, but requires optimization over intermediate partial-mask states to handle multi-token targets.
CAS mitigates object hallucinations in MLLMs by extracting two context preference vectors from designed conflict samples and applying signed residual injection at mid-early MLP layers without retraining or added latency.
Counterfactual likelihood tests detect indirect influence through public channels in private reasoning models, validated on a 7B role-channel model showing asymmetric A-to-B influence and complete pathway identification via graph-separation controls.
Stealth Pretraining Seeding plants persistent unsafe behaviors in LLMs via diffuse poisoned web content that activates on precise triggers and evades standard evaluation.
DSCA turns concept isolation into an architectural property by dynamically creating orthogonal subspaces for non-interfering lifelong edits in vision-language models, sustaining over 95% success after 1000 sequential edits.
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.
Activation Addition steers language models by adding contrastive activation vectors from prompt pairs to control high-level properties like sentiment and toxicity at inference time without training.
KV-cache eviction, prompt compression, recurrent state bounding, and agent memory consolidation are unified as one rate-distortion problem with a shared lower bound, shared failure mode, and transferable mechanisms.
ConvMemory v3 introduces a dual-evidence gate for target-conditioned memory validity verification, reporting 90.12% accuracy on synthetic benchmarks, 98.8% transfer to real data, and H@1 improvement from 45.1% to 95.7% in deployment while preserving safety contracts.
Across four language models, the linear direction detecting hallucinated entities achieves perfect separability yet lies at cosine 0.12 from the direction that induces refusal, with the gap unchanged by instruction tuning.
Knowledge editing methods redistribute and suppress rather than overwrite facts in LLMs, creating narrow vulnerable regions in representation space that adversarial prompts can exploit.
A 2x2 ablation shows repeated shared access enables grokking while addressable memory (not recurrence) enables edit propagation in transformer variants on synthetic KG QA.
citing papers explorer
-
Mechanistic Interpretability and Causal Feature Steering of Neural Quantum States via Sparse Autoencoders
Sparse autoencoders applied to Neural Quantum States extract unsupervised features correlating with and causally steering physical observables such as order parameters while preserving variational energy.
-
Slot Machines: How LLMs Keep Track of Multiple Entities
LLM activations encode current and prior entities in orthogonal slots, but models only use the current slot for explicit factual retrieval despite prior-slot information being linearly decodable.
-
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.
-
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.
-
Auditing Forgetting in Limited Memory Language Models
A causal audit of LMLMs finds near-zero parametric leakage after deletion, with surviving correctness coming from retrieval artifacts in the database.
-
Obliviate: Erasing Concepts from Autoregressive Image Generation Models
Obliviate erases targeted concepts from autoregressive image generators via KL supervision on visual tokens over full trajectories, cutting nudity rates sharply on benchmarks while keeping general performance.
-
Vision-Default, Prior-Override: Causal Mechanisms of Perception-Knowledge Conflict in Vision-Language Models
VLMs default to visual grounding but a sparse circuit of 2.5-4.8% attention heads in later layers mediates prior-knowledge overrides, identified causally via patching and ablation across three model families.
-
Channel Location Constrains the Auditability of Subliminal Learning
Auditability of subliminal learning is constrained by channel location, with initialization-dependent body channels allowing pre-training screens while vocabulary geometry and conditional body channels evade them.
-
Demystifying Hidden-State Recurrence: Switchable Latent Reasoning with On-Policy Reinforcement Learning
SWITCH uses explicit <swi> and </swi> boundary tokens to make latent chain-of-thought compatible with on-policy RL (GRPO) and open to causal mechanistic probing, outperforming prior hidden-state recurrence methods.
-
When Probing Accuracy Saturates, Fragility Resolves: A Complementary Metric for LLM Pre-Training Analysis
Fragility, the activation noise level causing probe accuracy collapse, reveals evolving lexical-to-compositional moral encoding, layer robustness gradients, and fine-tuning differences invisible to saturated probing accuracy.
-
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.
-
Toward Calibrated, Fair, and accurate Deepfake Detection
Face-Feature Tuning is a label-free logit remapping method that reduces FPR/TPR gaps across groups in deepfake detection while preserving overall accuracy.
-
Knowledge Editing in Masked Diffusion Language Models
Locate-then-edit succeeds at the same early-to-mid MLP locations in masked diffusion models as in autoregressive models, but requires optimization over intermediate partial-mask states to handle multi-token targets.
-
Rethinking Visual Neglect: Steering via Context-Preference for MLLM Hallucination Mitigation
CAS mitigates object hallucinations in MLLMs by extracting two context preference vectors from designed conflict samples and applying signed residual injection at mid-early MLP layers without retraining or added latency.
-
Counterfactual Likelihood Tests for Indirect Influence in Private Reasoning Channels
Counterfactual likelihood tests detect indirect influence through public channels in private reasoning models, validated on a 7B role-channel model showing asymmetric A-to-B influence and complete pathway identification via graph-separation controls.
-
PermaFrost-Attack: Stealth Pretraining Seeding(SPS) for planting Logic Landmines During LLM Training
Stealth Pretraining Seeding plants persistent unsafe behaviors in LLMs via diffuse poisoned web content that activates on precise triggers and evades standard evaluation.
-
DSCA: Dynamic Subspace Concept Alignment for Lifelong VLM Editing
DSCA turns concept isolation into an architectural property by dynamically creating orthogonal subspaces for non-interfering lifelong edits in vision-language models, sustaining over 95% success after 1000 sequential edits.
-
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.
-
Steering Language Models With Activation Engineering
Activation Addition steers language models by adding contrastive activation vectors from prompt pairs to control high-level properties like sentiment and toxicity at inference time without training.
-
What to Keep, What to Forget: A Rate--Distortion View of Memory Compaction in LLMs and Agents
KV-cache eviction, prompt compression, recurrent state bounding, and agent memory consolidation are unified as one rate-distortion problem with a shared lower bound, shared failure mode, and transferable mechanisms.
-
ConvMemory v3: A Validity Context Layer for Conversational Memory via Target-Conditioned Relation Verification
ConvMemory v3 introduces a dual-evidence gate for target-conditioned memory validity verification, reporting 90.12% accuracy on synthetic benchmarks, 98.8% transfer to real data, and H@1 improvement from 45.1% to 95.7% in deployment while preserving safety contracts.
-
Perfect Detection, Failed Control: The Geometry of Knowing vs. Steering in Language Models
Across four language models, the linear direction detecting hallucinated entities achieves perfect separability yet lies at cosine 0.12 from the direction that induces refusal, with the gap unchanged by instruction tuning.
-
Exposing the Illusion of Erasure in Knowledge Editing for LLMs
Knowledge editing methods redistribute and suppress rather than overwrite facts in LLMs, creating narrow vulnerable regions in representation space that adversarial prompts can exploit.
-
Repeated Shared Access Enables Grokking, but Edit Propagation Depends on an Addressable Memory
A 2x2 ablation shows repeated shared access enables grokking while addressable memory (not recurrence) enables edit propagation in transformer variants on synthetic KG QA.
-
Benchmarking Knowledge Editing using Logical Rules
Introduces a benchmark using logical rules from knowledge graphs to generate multi-hop questions that evaluate whether knowledge edits in LLMs propagate to entailed facts, finding up to 24% performance gaps for methods like ROME and FT.
-
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.
-
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.
-
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.
-
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.
-
Dissociating Decodability and Causal Use in Bracket-Sequence Transformers
In Dyck-language transformers, depth, distance, and top-of-stack signals are decodable from both residual stream and attention, but only attention-based top-of-stack signals are causally used for task performance.
-
Co-Located Tests, Better AI Code: How Test Syntax Structure Affects Foundation Model Code Generation
Co-locating tests with implementation code yields substantially higher preservation and correctness in foundation-model-generated programs than separated test syntax.
-
State Transfer Reveals Reuse in Controlled Routing
Fixed-interface state transfer provides stronger evidence of internal reuse in controlled routing than prompt retraining success alone.
-
Meltdown: Circuits and Bifurcations in Point-Cloud-Conditioned 3D Diffusion Transformers
Tiny on-surface point perturbations trigger a bifurcation in the reverse diffusion process of 3D transformers, localized to a low-rank cross-attention write that can be reshaped at test time to suppress the failure.
-
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.
-
Painless Activation Steering: An Automated, Lightweight Approach for Post-Training Large Language Models
PAS automates activation steering for LLMs using labeled data to improve behavior control on tasks like bias and alignment, with gains over ICL and SFT but limited effect on intelligence tasks.
-
Sparse Feature Circuits: Discovering and Editing Interpretable Causal Graphs in Language Models
Sparse feature circuits are introduced as interpretable causal subnetworks in language models, supporting unsupervised discovery of thousands of circuits and a method called SHIFT to improve classifier generalization by ablating irrelevant features.
-
GPT-NeoX-20B: An Open-Source Autoregressive Language Model
GPT-NeoX-20B is a publicly released 20B parameter autoregressive language model trained on the Pile that shows strong gains in five-shot reasoning over similarly sized prior models.
-
Rare Events, Real Signals: Functional Ensembles as Units of Computation in Deep Spiking Networks
In spiking ResNets, 1FC ensembles defined by pairwise correlations show ReLU-like cofiring-to-response mapping whose gain scales with ensemble size, with reliable class encoding restricted to infrequent high-cofiring events.
-
From Correlation to Cause: A Five-Stage Methodology for Feature Analysis in Transformer Language Models
A five-stage causal feature analysis methodology is proposed and tested on GPT-2 for IOI, showing partial causality of SAE features, robustness differences under shifts, and deployment cost benefits.
-
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.
-
Towards Scalable Lifelong Knowledge Editing with Selective Knowledge Suppression
LightEdit enables scalable lifelong knowledge editing in LLMs via selective knowledge retrieval and probability suppression during decoding, outperforming prior methods on ZSRE, Counterfact, and RIPE while reducing training costs.
-
Attention, May I Have Your Decision? Localizing Generative Choices in Diffusion Models
Implicit generative choices in diffusion models concentrate in self-attention layers; targeted ICM interventions there outperform broader debiasing methods with fewer artifacts.
-
MemOS: A Memory OS for AI System
MemOS introduces a unified memory management framework for LLMs using MemCubes to handle and evolve different memory types for improved controllability and evolvability.
-
How to use and interpret activation patching
Activation patching provides evidence about neural network circuits when the choice of metric is aligned with the hypothesis and common interpretation errors are avoided.
-
A Negative Result on Cross-Model Activation Transfer in a Pythia Multi-Hop Setting
A learned linear activation bridge achieves high alignment (cosine ~0.97) between Pythia-160M and Pythia-410M states but produces no improvement in downstream multi-hop answering when injected into the receiver.
- Decodable but Not Faithful: Coupling Natural-Language Rationales to Programmatic Verifiers
- RepSelect: Robust LLM Unlearning via Representation Selectivity
- Not Just RLHF: Why Alignment Alone Won't Fix Multi-Agent Sycophancy