Bilinear Convolution Decomposition for Causal RL Interpretability

Understanding Addition and Subtraction in Transformers

Turing Complete Transformers: Two Transformers Are More Powerful Than One

TinySQL: A Progressive Text-to-SQL Dataset for Mechanistic Interpretability

Model Mapping: Can We Turn Transformers into Programs?

Method for Converting Models to Programs

Activation Space Interventions Can Be Transferred Between Large Language Models

Beyond Linear Steering: Unified Multi-Attribute Control for Language Models

Breaking the Mirror: Activation-Based Mitigation of Self-Preference in LLMs

Beyond Monoliths: Expert Orchestration for More Capable, Democratic, and Safe LLMs

Position: Require Frontier AI Labs To Release Small 'Analog' Models

Query Circuits: Explaining How Language Models Answer User Prompts

Spectral Superposition: A Theory of Feature Geometry

Towards Understanding Subliminal Learning: When and How Hidden Biases Transfer

Beyond Linear Probes: Dynamic Safety Monitoring for Language Models

Old Habits Die Hard: How Conversational History Geometrically Traps LLMs

Riemannian-Manifold Steering: Geometry-Aware Generative Autoencoders for Representation Control

STRIDE: Training Data Attribution via Sparse Recovery from Subset Perturbations

DreamReader: An Interpretability Toolkit for Text-to-Image Models

Ablation-Reversible Heads Don't Transfer: A Stress Test for Mechanistic Role Claims in Transformers

Understanding and Mitigating Dataset Corruption in LLM Steering

Geometry-Aware CLIP Retrieval via Local Cross-Modal Alignment and Steering

Curveball Steering: The Right Direction To Steer Isn't Always Linear

Make Mechanistic Interpretability Auditable: A Call to Develop Guidelines via Continuous Collaborative Reviewing

Debate, Deliberate, Decide (D3): A Cost-Aware Adversarial Framework for Reliable and Interpretable LLM Evaluation

Are LLM Evaluators Really Narcissists? Sanity Checking Self-Preference in LLMs

Fingerprinting All AI Cluster I/O Without Mutually Trusted Processors

The Capability Frontier: Benchmarks Miss 82% of Model Performance