Paper argues mechanistic interpretability needs a logic of inference
The authors propose five principles as a recursive procedure for turning local findings into testable mechanistic claims.
Stories per day · 30 days
4 RSS
Relatedinterpretability3sparse-autoencoders3superposition3goodfire2inceptionv12qwen32ai-safety1autoresearch1causal-robustness1compositionality1data-efficiency1feature-reproducibility1gemma1gemma-21hallucinations1inference1large-language-models1llama1
34 stories, newest first · one card per story
Stories tagged #mechanistic-interpretability
Study tests whether causal neuron groups are unique or transferable
A few neurons often flip two-hop answers, but effects weaken on other chains.
Survey examines mechanistic interpretability for AI safety
The paper surveys methods for mechanistic interpretability and their role in AI safety, including knowledge editing and steering.
LLMs can read and write neural network weights
Research shows LLMs can directly read and edit neural network weights, enabling autonomous model improvement that cuts traditional autoresearch out of the loop.
1
Goodfire researcher claims neural nets are more interpretable than thought
Stanford's Chris Potts calls the 26,000-word post a hunch that does not make interpretability easier.
1
Is AI interpretability close to a breakthrough? A new essay has researchers divided
Nick Cammarata argues famous polysemantic neurons were misread, claiming most Qwen3-32B neurons have one job.
7
Neural Networks Are Already Clean
A Goodfire researcher argues neural networks are already interpretable and superposition was misread.
12
Causal robustness margin proposed to measure mechanistic LLM uncertainty
Incorrect or uncertain answers had smaller margins across three open-source models and five QA datasets.
Behavioral Compositionality of Sparse Autoencoder Representations
SAE features are more compositional than random bases, but faithful explanations still need pairwise interactions.
Diminishing Returns of Data in Sparse Autoencoders: Functional Behavior does not Guarantee Feature Reproducibility
SAE reconstruction converges on far less data, but one-to-one feature matches keep rising with more tokens.
Task Context Shapes Component Roles in Language Models
Task context changes which facts shared language-model heads use and how they influence later computation.
S-Space: Exploring Spatial Workspace in Multimodal Language Models
Researchers identify S-Space, a linear subspace where multimodal models encode object coordinates.
The Thinking Leakage in Reinforcement Learning for LLMs
Qwen3 NO-THINK RL gains largely reactivate existing THINK-mode reasoning rather than new skills.
When Forgetting is not Catastrophic: On the Mechanics of Spurious Forgetting
Spurious forgetting hides old facts through a reversible shared shift before slower fact-specific erosion.
U-Space: Uncovering When and Why Uncertainty Arises in Language Models
U-Space maps language-model residual states into a subspace that tracks token-level doubt and certainty without extra training.
Every Ablation Is a Dose: Counterweights and the Semblance of Self-Repair
Language-model self-repair is explained as pre-existing counterweights following an affine response to ablation dose.
Where Do LLMs Decide to Break the Rules? Mechanistic Localization of Prompt Injection Compliance
Causal patching localizes prompt-injection compliance to a late-layer bottleneck, reversible in 77-92% of cases.
Rethinking Circuit Evaluation: Do Circuits Explain Model Errors?
Circuits that match GPT-2's correct answers often fail to reproduce most of its errors.
Small transformers track Bayesian evidence for latent common causes via a context-invariant mechanism
Small transformers accumulate Bayesian evidence for latent common causes through a context-invariant, generalizing mechanism.
Tool Mediation Alters Refusal Mechanisms in Large Language Models
Tool access raises LLM refusal thresholds and makes those refusals easier to weaken than chat refusals.
A mechanistic study of language model introspection
Researchers identify attention-head circuits that let language models detect and localize hidden activation interventions.
Understanding On-Policy Distillation: A Mechanistic Interpretability Perspective via Sparse Crosscoders
Sparse crosscoders show on-policy distillation reweights features LLMs already share rather than creating new ones.
When Does Correction Become Repair? Mechanistic Auditing of Internal Interventions in Tool-Using LLMs
SAKIKO finds activation steering can raise tool-use scores while corrupting many previously correct LLM decisions.
See it, Say it, Sorted: Mechanistic Diagnosis and Parameter-Space Mitigation of Emergent Misalignment in LLMs
A geometric mitigation framework cuts emergent misalignment on Qwen2.5-14B-IT by up to 80%.