mlsanityData to 14 Oct 2026

Mechanistic interpretability

13 directions · 1,591 papers · 19 this week

Busiest this week

Most new since Mon 12 Oct

Activation steering methods

239 papers5 this weekMedium confidence in this grouping: mostly one topic, with blurry edgesMedium confidence in this grouping: mostly one topic, with blurry edges · About

  1. Causal Syntax Circuits in CodeLMs: Budget and Probe Decide What Looks Localised
  2. Genesis Steering: Better Activation Steering Through Training Data
  3. Perturbation, Not Provenance: Why Activation Steering Does Not Make a Robust Text Watermark
  4. Beyond Bias Scores: Answerability Confounds in Activation Steering with Gradient-Constrained Direction Design (GCD)
  5. The Sharp Directions Are Against You: Curvature Analysis of Activation Steering
Open →

All directions

Largest first · week from Mon 12 Oct

  1. Circuits and mechanistic analysiscircuit, mechanistic, interpretability, circuitsHigh confidence in this grouping: the papers form one clear topic · About2721High confidence in this grouping: the papers form one clear topic
  2. Activation steering methodssteering, activation, activation steering, languageMedium confidence in this grouping: mostly one topic, with blurry edges · About2395Medium confidence in this grouping: mostly one topic, with blurry edges
  3. SAE features and their applicationsfeatures, sparse, saes, saeMedium confidence in this grouping: mostly one topic, with blurry edges · About2314Medium confidence in this grouping: mostly one topic, with blurry edges
  4. Sparse autoencoder training and evaluationsaes, sparse, features, featureHigh confidence in this grouping: the papers form one clear topic · About1754High confidence in this grouping: the papers form one clear topic
  5. Knowledge and representations inside LLMslanguage, interpretability, llms, largeMedium confidence in this grouping: mostly one topic, with blurry edges · About1360Medium confidence in this grouping: mostly one topic, with blurry edges
  6. Steering and probing for LLM safetysafety, refusal, steering, harmfulMedium confidence in this grouping: mostly one topic, with blurry edges · About1190Medium confidence in this grouping: mostly one topic, with blurry edges
  7. Interpreting vision and multimodal modelsvisual, vision, multimodal, vision-languageHigh confidence in this grouping: the papers form one clear topic · About1050High confidence in this grouping: the papers form one clear topic
  8. Steering and interpreting LLM reasoningreasoning, cot, steering, llmsMedium confidence in this grouping: mostly one topic, with blurry edges · About1021Medium confidence in this grouping: mostly one topic, with blurry edges
  9. Probing deception, bias and beliefsbias, probes, llms, linearLow confidence in this grouping: the papers may belong to more than one topic, or the group may change · About882Low confidence in this grouping: the papers may belong to more than one topic, or the group may change
  10. Interpreting biomedical and science modelsprotein, medical, features, interpretabilityMedium confidence in this grouping: mostly one topic, with blurry edges · About600Medium confidence in this grouping: mostly one topic, with blurry edges
  11. Interpreting and steering diffusion modelsdiffusion, concept, erasure, sparseMedium confidence in this grouping: mostly one topic, with blurry edges · About270Medium confidence in this grouping: mostly one topic, with blurry edges
  12. Hallucination detection from model internalshallucination, hallucinations, steering, detectionHigh confidence in this grouping: the papers form one clear topic · About231High confidence in this grouping: the papers form one clear topic
  13. Interpreting time-series and weather modelstime series, series, time, interpretabilityHigh confidence in this grouping: the papers form one clear topic · About140High confidence in this grouping: the papers form one clear topic
1 new this week isn’t in a direction yet
  1. Prompt-Injection Probes Learn the Conversation, Not Just the Injection

Weekly activity

Papers per week, 29 Jun – 12 Oct (last week partial)

  1. DirectionConfidencepapers16 wks
  2. Activation steering methodsMedium confidence in this grouping: mostly one topic, with blurry edges · AboutMedium confidence in this grouping: mostly one topic, with blurry edgesPer week: 29 Jun: 5; 6 Jul: 3; 13 Jul: 0; 20 Jul: 2; 27 Jul: 4; 3 Aug: 7; 10 Aug: 2; 17 Aug: 1; 24 Aug: 6; 31 Aug: 6; 7 Sep: 5; 14 Sep: 4; 21 Sep: 6; 28 Sep: 13; 5 Oct: 6; 12 Oct: 523975
  3. Sparse autoencoder training and evaluationHigh confidence in this grouping: the papers form one clear topic · AboutHigh confidence in this grouping: the papers form one clear topicPer week: 29 Jun: 2; 6 Jul: 2; 13 Jul: 3; 20 Jul: 1; 27 Jul: 1; 3 Aug: 0; 10 Aug: 0; 17 Aug: 1; 24 Aug: 3; 31 Aug: 3; 7 Sep: 5; 14 Sep: 0; 21 Sep: 2; 28 Sep: 6; 5 Oct: 3; 12 Oct: 417536
  4. SAE features and their applicationsMedium confidence in this grouping: mostly one topic, with blurry edges · AboutMedium confidence in this grouping: mostly one topic, with blurry edgesPer week: 29 Jun: 2; 6 Jul: 1; 13 Jul: 3; 20 Jul: 2; 27 Jul: 0; 3 Aug: 3; 10 Aug: 4; 17 Aug: 0; 24 Aug: 3; 31 Aug: 4; 7 Sep: 2; 14 Sep: 3; 21 Sep: 3; 28 Sep: 5; 5 Oct: 2; 12 Oct: 423141
  5. Circuits and mechanistic analysisHigh confidence in this grouping: the papers form one clear topic · AboutHigh confidence in this grouping: the papers form one clear topicPer week: 29 Jun: 4; 6 Jul: 3; 13 Jul: 3; 20 Jul: 1; 27 Jul: 0; 3 Aug: 6; 10 Aug: 3; 17 Aug: 5; 24 Aug: 4; 31 Aug: 3; 7 Sep: 2; 14 Sep: 3; 21 Sep: 4; 28 Sep: 5; 5 Oct: 2; 12 Oct: 127249
  6. Steering and probing for LLM safetyMedium confidence in this grouping: mostly one topic, with blurry edges · AboutMedium confidence in this grouping: mostly one topic, with blurry edgesPer week: 29 Jun: 0; 6 Jul: 2; 13 Jul: 2; 20 Jul: 2; 27 Jul: 1; 3 Aug: 1; 10 Aug: 0; 17 Aug: 2; 24 Aug: 4; 31 Aug: 2; 7 Sep: 1; 14 Sep: 1; 21 Sep: 1; 28 Sep: 6; 5 Oct: 3; 12 Oct: 011928
  7. Knowledge and representations inside LLMsMedium confidence in this grouping: mostly one topic, with blurry edges · AboutMedium confidence in this grouping: mostly one topic, with blurry edgesPer week: 29 Jun: 2; 6 Jul: 0; 13 Jul: 1; 20 Jul: 0; 27 Jul: 2; 3 Aug: 1; 10 Aug: 1; 17 Aug: 1; 24 Aug: 3; 31 Aug: 3; 7 Sep: 1; 14 Sep: 2; 21 Sep: 3; 28 Sep: 4; 5 Oct: 1; 12 Oct: 013625
  8. Steering and interpreting LLM reasoningMedium confidence in this grouping: mostly one topic, with blurry edges · AboutMedium confidence in this grouping: mostly one topic, with blurry edgesPer week: 29 Jun: 0; 6 Jul: 0; 13 Jul: 0; 20 Jul: 2; 27 Jul: 2; 3 Aug: 1; 10 Aug: 0; 17 Aug: 1; 24 Aug: 3; 31 Aug: 3; 7 Sep: 1; 14 Sep: 4; 21 Sep: 3; 28 Sep: 4; 5 Oct: 0; 12 Oct: 110225
  9. Probing deception, bias and beliefsLow confidence in this grouping: the papers may belong to more than one topic, or the group may change · AboutLow confidence in this grouping: the papers may belong to more than one topic, or the group may changePer week: 29 Jun: 2; 6 Jul: 2; 13 Jul: 2; 20 Jul: 1; 27 Jul: 1; 3 Aug: 1; 10 Aug: 0; 17 Aug: 0; 24 Aug: 2; 31 Aug: 0; 7 Sep: 2; 14 Sep: 1; 21 Sep: 1; 28 Sep: 2; 5 Oct: 3; 12 Oct: 28822
  10. Interpreting vision and multimodal modelsHigh confidence in this grouping: the papers form one clear topic · AboutHigh confidence in this grouping: the papers form one clear topicPer week: 29 Jun: 4; 6 Jul: 2; 13 Jul: 0; 20 Jul: 1; 27 Jul: 2; 3 Aug: 1; 10 Aug: 2; 17 Aug: 2; 24 Aug: 1; 31 Aug: 2; 7 Sep: 2; 14 Sep: 1; 21 Sep: 2; 28 Sep: 3; 5 Oct: 0; 12 Oct: 010525
  11. Interpreting and steering diffusion modelsMedium confidence in this grouping: mostly one topic, with blurry edges · AboutMedium confidence in this grouping: mostly one topic, with blurry edgesPer week: 29 Jun: 1; 6 Jul: 0; 13 Jul: 0; 20 Jul: 0; 27 Jul: 0; 3 Aug: 0; 10 Aug: 1; 17 Aug: 0; 24 Aug: 0; 31 Aug: 1; 7 Sep: 0; 14 Sep: 0; 21 Sep: 1; 28 Sep: 3; 5 Oct: 0; 12 Oct: 0277
  12. Hallucination detection from model internalsHigh confidence in this grouping: the papers form one clear topic · AboutHigh confidence in this grouping: the papers form one clear topicPer week: 29 Jun: 0; 6 Jul: 0; 13 Jul: 0; 20 Jul: 0; 27 Jul: 0; 3 Aug: 1; 10 Aug: 0; 17 Aug: 0; 24 Aug: 1; 31 Aug: 0; 7 Sep: 0; 14 Sep: 0; 21 Sep: 1; 28 Sep: 0; 5 Oct: 0; 12 Oct: 1234
  13. Interpreting time-series and weather modelsHigh confidence in this grouping: the papers form one clear topic · AboutHigh confidence in this grouping: the papers form one clear topicPer week: 29 Jun: 0; 6 Jul: 0; 13 Jul: 0; 20 Jul: 0; 27 Jul: 1; 3 Aug: 0; 10 Aug: 0; 17 Aug: 1; 24 Aug: 0; 31 Aug: 0; 7 Sep: 0; 14 Sep: 0; 21 Sep: 0; 28 Sep: 1; 5 Oct: 1; 12 Oct: 0144
  14. Interpreting biomedical and science modelsMedium confidence in this grouping: mostly one topic, with blurry edges · AboutMedium confidence in this grouping: mostly one topic, with blurry edgesPer week: 29 Jun: 0; 6 Jul: 0; 13 Jul: 0; 20 Jul: 0; 27 Jul: 0; 3 Aug: 1; 10 Aug: 2; 17 Aug: 0; 24 Aug: 4; 31 Aug: 2; 7 Sep: 1; 14 Sep: 0; 21 Sep: 0; 28 Sep: 1; 5 Oct: 0; 12 Oct: 06011