# Knowledge and representations inside LLMs

Area: [Mechanistic interpretability](https://mlsanity.com/mechinterp/index.md). HTML: https://mlsanity.com/mechinterp/knowledge-and-representations-inside-llms/

## Key facts

- Knowledge and representations inside LLMs: 136 papers on arXiv since 2021 (data through 14 Oct 2026).
- 22 of its 136 mapped papers (16%) cite “Eliciting Latent Predictions from Transformers with the Tuned Lens” (2023, arXiv 2303.08112).
- 16 of its 136 mapped papers (12%) cite “Towards Best Practices of Activation Patching in Language Models: Metrics and Methods” (2023, arXiv 2309.16042).
- 7 new papers 24 Sep–14 Oct.

## Most cited here

- [Eliciting Latent Predictions from Transformers with the Tuned Lens](https://mlsanity.com/paper/2303.08112/) (2023), cited by 22 papers here
- [Towards Best Practices of Activation Patching in Language Models: Metrics and Methods](https://mlsanity.com/paper/2309.16042/) (2023), cited by 16 papers here
- [Patchscopes: A Unifying Framework for Inspecting Hidden Representations of Language Models](https://mlsanity.com/paper/2401.06102/) (2024), cited by 14 papers here
- [Representation Engineering: A Top-Down Approach to AI Transparency](https://mlsanity.com/paper/2310.01405/) (2023), cited by 12 papers here
- [Do Multilingual LLMs Think In English?](https://mlsanity.com/paper/2502.15603/) (2025), cited by 3 papers here
- [Aligning Large Language Models with Human Preferences through Representation Engineering](https://mlsanity.com/paper/2312.15997/) (2023), cited by 2 papers here
- [Rethinking Interpretability in the Era of Large Language Models](https://mlsanity.com/paper/2402.01761/) (2024), cited by 2 papers here
- [Towards Faithful Natural Language Explanations: A Study Using Activation Patching in Large Language Models](https://mlsanity.com/paper/2410.14155/) (2024), cited by 2 papers here
- [Separating Tongue from Thought: Activation Patching Reveals Language-Agnostic Concept Representations in Transformers](https://mlsanity.com/paper/2411.08745/) (2024), cited by 2 papers here
- [Towards eliciting latent knowledge from LLMs with mechanistic interpretability](https://mlsanity.com/paper/2505.14352/) (2025), cited by 2 papers here

## New (24 Sep–14 Oct)

- [SemanticFold: Latent Sequence Compression SeparatesLanguage Modeling, Decodability, and Reasoning](https://mlsanity.com/paper/2610.10304/) (2026-10-07)
- [Copying Before Suppression: What Drives a Below-Chance Dip During Language Model Training?](https://mlsanity.com/paper/2610.04119/) (2026-10-02)
- [Clinical Concept Centers in LLMs](https://mlsanity.com/paper/2610.02829/) (2026-10-02)
- [Signatures of semantic search in the activations of large language models](https://mlsanity.com/paper/2609.35599/) (2026-09-28)
- [Spontaneous Context Restoration: How Language Models Recover from Corrupted Inputs](https://mlsanity.com/paper/2609.35475/) (2026-09-28)
- [Identifying Temporal Features within Transcoders for Time Sensitive Factual Recall](https://mlsanity.com/paper/2609.33183/) (2026-09-27)
- [FTB Graph: Determining and Validating First-token Broadcasters and Language-Identity Head Circuits in Multilingual Language Models](https://mlsanity.com/paper/2609.30954/) (2026-09-25)