MASHBOX Steer

Resources

Research we read while building MASHBOX Steer.

1. neuron-explorer, Lyra
2. Neural Networks Are Already Clean, Cammarata, Goodfire, 2026
3. Goodfire Ember SDK
4. Steering Language Models With Activation Engineering, Turner et al., 2023
5. Steering Llama 2 via Contrastive Activation Addition, Panickssery et al., 2023
6. Representation Engineering: A Top-Down Approach to AI Transparency, Zou et al., 2023
7. Inference-Time Intervention, Li et al., 2023
8. Improving Activation Steering in Language Models with Mean-Centring, Jorgensen et al., 2023
9. Refusal in Language Models Is Mediated by a Single Direction, Arditi et al., 2024
10. Scaling Monosemanticity, Templeton et al., 2024
11. Evaluating feature steering: a case study in mitigating social biases, Anthropic, 2024
12. ReFT: Representation Finetuning for Language Models, Wu et al., 2024
13. Analysing the Generalisation and Reliability of Steering Vectors, Tan et al., 2024
14. Programming Refusal with Conditional Activation Steering, Lee et al., 2024
15. Improving Instruction-Following in Language Models through Activation Steering, Stolfo et al., 2024
16. AxBench: Steering LLMs? Even Simple Baselines Outperform Sparse Autoencoders, Wu et al., 2025
17. Improved Representation Steering for Language Models, Wu et al., 2025
18. Persona Vectors, Chen et al., 2025
19. SteerMoE: Steering MoE LLMs via Expert (De)Activation, 2025
20. Activation Steering with a Feedback Controller, 2025
21. Non-identifiability of steering vectors, 2026
22. Activation Steering for Aligned Open-ended Generation without Sacrificing Coherence, 2026
23. Steering-induced emergent misalignment, 2026
24. From Golden Gate Bridge to Broken JSON, Maziyar Panahi, Hugging Face, 2026
25. Neuronpedia
26. repeng, Theia Vogel
27. llama.cpp control vectors, ggml-org
28. EmoSteer-TTS, 2025
29. Accent steering in zero-shot TTS, 2026
30. Verbalizable Representations Form a Global Workspace in Language Models, Gurnee et al., 2026
31. J++ Lens and Drift Calibrator, Neuronpedia and MIT Media Lab, 2026
32. Natural Language Autoencoders Produce Unsupervised Explanations of LLM Activations, Fraser-Taliente et al., 2026
33. Natural Language Autoencoders Identify Without Explaining, 2026
34. Activation Oracles: Training and Evaluating LLMs as General-Purpose Activation Explainers, Karvonen et al., 2025
35. HeadVis, Anthropic, 2026
36. Emotion Concepts and their Function in a Large Language Model, Sofroniew et al., 2026
37. Circuits Updates, June 2026 (turn-averaged SAEs), Anthropic, 2026
38. Circuits Updates, May 2026 (features through downstream connections), Anthropic, 2026
39. Circuit Tracing: Revealing Computational Graphs in Language Models, Ameisen et al., 2025
40. Qwen-Scope: Turning Sparse Features into Development Tools for Large Language Models, Deng et al., 2026
41. Gemma Scope 2: SAEs and Transcoders for Gemma 3, Google DeepMind
42. The Assistant Axis, Lu et al., Anthropic Fellows, 2026
43. Meandering on Manifolds: The Neural Geometry of Stories Over Time, Bigelow et al., Goodfire, 2026
44. Steering Along Manifolds to Control Neural Networks, Wurgaft et al., Goodfire, 2026
45. From Stories to Features: Scaling Concept Vectors for AI Alignment and Interpretability, 2026
46. Comparing Sparsified MLPs and Transcoders: Many Transcoder Features Are Neurons in Disguise, 2026
47. The Router Already Knows, 2026
48. Sparse Autoencoders for Interpretable Emotion Control in Text-to-Speech, Du et al., 2026
49. A Geometric Perspective on Composable Emotion Steering in Text-to-Speech Models, 2026
50. Controlling Speaking Rate in Autoregressive TTS via Activation Steering, 2026
51. EmoRES-TTS: Residual-Enhanced Vector Steering for Emotional Speech Generation, 2026
52. Towards Interpretable Framework for Neural Audio Codecs via Sparse Autoencoders, 2026
53. Read-Best Is Not Steer-Best: A Probing-Steering Layer Dissociation in Omni-Modal Large Language Models, 2026
54. Steering Follows Geometry, Not Labels: Emotion Directions in a Full-Duplex Speech Model, 2026
55. SteerSpeech: Activation Steering for Emotion Control in Generated Speech, 2026
56. Steering intonational tunes in text-to-speech synthesis, 2026
57. Parasitic Co-Denoising: Unlocking 3D Human Motion Generation in a Frozen Video Diffusion Model, 2026
58. UMAP: Uniform Manifold Approximation and Projection for Dimension Reduction, McInnes, Healy and Melville, 2018
59. Density-Based Clustering Based on Hierarchical Density Estimates (HDBSCAN), Campello, Moulavi and Sander, 2013
60. interpreting GPT: the logit lens, nostalgebraist, 2020
61. Towards Monosemanticity: Decomposing Language Models With Dictionary Learning, Bricken et al., 2023
62. Negative Results for SAEs On Downstream Tasks and Deprioritising SAE Research, Smith et al., Google DeepMind, 2025