mlsanityData to 14 Oct 2026

Steering and interpreting LLM reasoning

102 papers1 this weekMedium confidence in this grouping: mostly one topic, with blurry edgesMedium confidence in this grouping: mostly one topic, with blurry edges · About

Keywords: reasoning, cot, steering, llms, language

Key facts. Steering and interpreting LLM reasoning: 102 papers on arXiv since 2023 (data through 14 Oct 2026). 9 of its 102 mapped papers (9%) cite “I Have Covered All the Bases Here: Interpreting Reasoning Features in Large Language Models via Sparse Autoencoders” (2025, arXiv 2503.18878). 5 new papers 24 Sep–14 Oct.

What’s new

24 Sep – 14 Oct 2026

5 new papers, 24 Sep – 14 Oct; 1 this week.

  1. This week · 1
  2. MAPLE: A Multi-Axis Parametric Knowledge Extraction Lens for LLM Evaluation also fits: Knowledge and representations inside LLMs
  3. Earlier · 4
  4. Learning to Decide, Not to Reason: Parameter-Efficient Decision Operators via Low-Rank Activation Steering also fits: Activation steering methods
  5. Concept Subspaces Compute Beyond the Logit Lens: A Weights-Only Test for Locating Representations Upstream of Readout
  6. RESCUE: Repairing Language Model Errors to Sparse Circuits via Reinforcement Learning also fits: Circuits and mechanistic analysis
  7. Understanding On-Policy Distillation: A Mechanistic Interpretability Perspective via Sparse Crosscoders

Foundations

Most cited by papers here

  1. I Have Covered All the Bases Here: Interpreting Reasoning Features in Large Language Models via Sparse Autoencoders2025cited by 9 here
  2. Understanding Reasoning in Thinking Language Models via Steering Vectors2025cited by 9 here
  3. SEAL: Steerable Reasoning Calibration of Large Language Models for Free2025cited by 8 here
  4. How does Chain of Thought Think? Mechanistic Interpretability of Chain-of-Thought Reasoning with Sparse Autoencoding2025cited by 7 here
  5. Unlocking General Long Chain-of-Thought Reasoning Capabilities of Large Language Models via Representation Engineering2025cited by 6 here

Lineage

The papers this direction builds on, oldest first.

Each paper on the line cites the one before.

  1. 2022Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 smallcited by 16 here
  2. 2023Progress measures for grokking via mechanistic interpretabilitycited by 8 here
  3. 2023Towards Automated Circuit Discovery for Mechanistic Interpretabilitycited by 9 here
  4. 2023Sparse Autoencoders Find Highly Interpretable Features in Language Modelscited by 31 here
  5. 2024Improving Dictionary Learning with Gated Sparse Autoencoderscited by 7 here
  6. 2024Scaling and evaluating sparse autoencoderscited by 14 here
  7. 2024Jumping Ahead: Improving Reconstruction Fidelity with JumpReLU Sparse Autoencoderscited by 10 here
  8. 2024Gemma Scope: Open Sparse Autoencoders Everywhere All At Once on Gemma 2cited by 13 here
  9. 2024Llama Scope: Extracting Millions of Features from Llama-3.1-8B with Sparse Autoencoderscited by 7 here
  10. 2025A Survey on Sparse Autoencoders: Interpreting the Internal Mechanisms of Large Language Modelscited by 6 here
  11. 2025I Have Covered All the Bases Here: Interpreting Reasoning Features in Large Language Models via Sparse Autoencoderscited by 9 here
13 more it builds on, not in this chain
  1. 2022Toy Models of Superpositioncited by 9 here
  2. 2023Steering Language Models With Activation Engineeringcited by 27 here
  3. 2023Representation Engineering: A Top-Down Approach to AI Transparencycited by 21 here
  4. 2023Attribution Patching Outperforms Automated Circuit Discoverycited by 5 here
  5. 2023Steering Llama 2 via Contrastive Activation Additioncited by 19 here
  6. 2024Personalized Steering of Large Language Models: Versatile Steering Vectors Through Bi-directional Preference Optimizationcited by 5 here
  7. 2024Programming Refusal with Conditional Activation Steeringcited by 5 here
  8. 2024Improving Instruction-Following in Language Models through Activation Steeringcited by 8 here
  9. 2024Automatically Interpreting Millions of Features in Large Language Modelscited by 5 here
  10. 2025Unlocking General Long Chain-of-Thought Reasoning Capabilities of Large Language Models via Representation Engineeringcited by 6 here
  11. 2025SEAL: Steerable Reasoning Calibration of Large Language Models for Freecited by 8 here
  12. 2025Understanding Reasoning in Thinking Language Models via Steering Vectorscited by 9 here
  13. 2025How does Chain of Thought Think? Mechanistic Interpretability of Chain-of-Thought Reasoning with Sparse Autoencodingcited by 7 here

The 24 most cited of 153.

Browse

30 of 102, ranked 6–35

  1. 6Feature Extraction and Steering for Enhanced Chain-of-Thought Reasoning in Language Models20255
  2. 7Improving Reasoning Performance in Large Language Models via Representation Engineering20254
  3. 8Controllable LLM Reasoning via Sparse Autoencoder-Based Steering20263
  4. 9Fantastic Reasoning Behaviors and Where to Find Them: Unsupervised Discovery of the Reasoning Process20253
  5. 10Steering LLM Reasoning Through Bias-Only Adaptation20252
  6. 11SAFER: Probing Safety in Reward Models with Sparse Autoencoder20252
  7. 12Activation Steering for Chain-of-Thought Compression20252
  8. 13Interpretable Reward Model via Sparse Autoencoder20252
  9. 14Reasoning Circuits in Language Models: A Mechanistic Interpretation of Syllogistic Inference20242
  10. 15Do Sparse Autoencoders Identify Reasoning Features in Language Models?20261
All 30
  1. 16What's the plan? Metrics for implicit planning in LLMs and their application to rhyme generation and question answering20261
  2. 17Interaction Locality in Hierarchical Recursive Reasoning20261
  3. 18Mechanistic Interpretability of Chain-of-Thought Reasoning via Sequential Activation Patching20261
  4. 19Mitigating Content Effects on Reasoning in Language Models through Fine-Grained Activation Steering20251
  5. 20Improving LLM Reasoning through Interpretable Role-Playing Steering20251
  6. 21Fractional Reasoning via Latent Steering Vectors Improves Inference Time Compute20251
  7. 22Small Vectors, Big Effects: A Mechanistic Study of RL-Induced Reasoning via Steering Vectors20251
  8. 23Learning to Ponder: Adaptive Reasoning in Latent Space20251
  9. 24SparseRM: A Lightweight Preference Modeling with Sparse Autoencoder20251
  10. 25Constrain Alignment with Sparse Autoencoders20241
  11. 26Mechanistic Interpretability of Large-Scale Counting in LLMs through a System-2 Strategy20260
  12. 27ATLAS: Verifier-Guided Adaptive Latent Activation Steering for Efficient LLM Reasoning20260
  13. 28Triggering Chain-of-Thought via Latent Feature Interventions in Large Language Models20260
  14. 29Is Grokking Worthwhile? Functional Analysis and Transferability of Generalization Circuits in Transformers20260
  15. 30RISER: Orchestrating Latent Reasoning Skills for Adaptive Activation Steering20260
  16. 31Identifying and Transferring Reasoning-Critical Neurons: Improving LLM Inference Reliability via Activation Steering20260
  17. 32Do Latent-CoT Models Think Step-by-Step? A Mechanistic Study on Sequential Reasoning Tasks20260
  18. 33Beyond Dense States: Sparse Transcoders as Causally Testable Operators for LLM Latent Reasoning20260
  19. 34Data-Centric Interpretability for LLM-based Multi-Agent Reinforcement Learning20260
  20. 35DistillLens: Symmetric Knowledge Distillation Through Logit Lens20260