interpretability
41 posts tagged interpretability.
- Protein model steering still breaks at the decoder
- Does Looking Inside a Model Actually Make It Safer?
- MAEMs turn residual stream directions into testable text
- Same outputs are not enough for AI mechanism audits
- Forecasting an LLM’s answer before token one
- KLM Wants AI to Show Its Evidence. Its Own Negative Results Show Why That's Hard.
- TopKLoRA turns a poisoned fine-tune into a circuit you can cut out
- CausalCoT treats chain-of-thought as a dependency graph
- What Arabic Dialect Math Problems Reveal About Model Reasoning
- Diffusion Models Are Starting to Generate Robot Policies, Not Just Robot Videos
- Can vision-language models actually spot persuasion in memes?
- Hide and Seek tests black-box vision models by removing the thing you care about
- The Same Attention Head Does Different Jobs Depending on the Task
- Probes That Catch an Agent Lying Before It Says a Word
- Diffusion models may be sketching meaning before pixels
- Flow matching gives neural dissimilarity metrics one common frame
- Models may refuse based on who they think you are
- Audio models do not automatically share speech and text concepts
- When a model’s rejection reason changes its next choice
- Clifford-VAE puts pixels into symbolic memory space
- When a Model Gets Math Right but Reads the Same Problem Differently
- DiaVLo Turns Vision-Language Model Failures Into Named Behaviours
- Membership Inference Gets an Entropy Correction: Reading the ETD Paper
- Probabilistic Linear Explanations make interpretability more honest
- Can AI Agents Do Their Own Interpretability Research? A New Benchmark Says Not Yet
- CAST makes clinical model audits more concrete
- What a Moral Probe Finds Inside an LLM: Structure, Not a Single Switch
- Can Generated Text Prove Which Internal Path a Model Took?
- Model hypnosis turns harmless prompt quirks into a control surface
- What Six Years of TrustNLP Papers Say About Where AI Safety Research Actually Went
- CENDRe Brings Frequency-Domain Explanations to Time-Series CNNs
- The Gap Between Reading a Feature and Steering With It
- The Reconstruction Test Grades Vibes, Not Facts: Reading the RECAP Paper on Activation Explanations
- Judge Bias Lives in the Activations, Not Just the Prompt
- Transformer circuits may be lower-dimensional than they look
- Language model embeddings do not want to collapse
- SciReasoner Treats Molecular Structure as Evidence You Can Inspect
- C2R targets the hidden mess inside sparse autoencoder features
- Reasoning Traces as a Difficulty Sensor: What Epi2Diff Gets Right
- When Models Quietly Unlearn: The Natural Ungrokking Problem
- Reading Gradients to Catch Hallucinations Before They Ship