#interpretability
- A Global Workspace in Language Models Anthropic research
- How Transparent is DiffusionGemma? Interpretability Study Closes the Gap to Autoregressive Models Google DeepMind research
- The Mask Is Not the Model: Auditing Prefix Invariance in Attention, State-Space, and Hybrid Sequence Models research
- Natural Language Autoencoders: Turning Claude's Thoughts into Text Anthropic research
- Anthropic Introduces Natural Language Autoencoders for Scalable LLM Interpretability Anthropic research
- Judge Circuits: Mechanistic Explanation of LLM-as-Judge Format Inconsistency research
- CiteVQA: Benchmarking Evidence Attribution for Trustworthy Document Intelligence (178 HF upvotes) Peking University / Shanghai Artificial Intelligence Laboratory research
- SAE Interventions Are Unreliable: Suppressed Behaviors Recover Post-Intervention Hong Kong Polytechnic University research
- BadWAM: When World-Action Models Dream Right but Act Wrong research
- Super Weights in LLMs and the Failure of Selective Training Amazon Web Services research
- Quantifying Faithful Confidence Expression in Large Reasoning Models Yale NLP research
- Anatomy of Post-Training: Using Interpretability to Audit and Fix Preference Data research
- Formalizing Latent Thoughts: Axiomatic Framework for Evaluating LLM Reasoning Representations University of British Columbia research
- GradCuit: Credit-Assigned Gradient Flow Enables Robust and Interpretable Test-Time Latent Reasoning research
- Anthropic loosens Claude Fable 5's biology safety classifiers, cutting false-positive blocks 85% Anthropic research
- Stealing Reasoning Traces from Proprietary LLM APIs research
- LLM Safety From Within (SIREN) University of Toronto CSSLab / McGill / LMU Munich research