Research map / Large language models / Language, safety and interpretability
Mechanistic analysis of attention and representations
Trace attention heads and representation editing to explain in-context learning and model behavior.
43 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2019–2026) · an approach in Language, safety and interpretability, keyword cluster “representations · interpretability · heads”
State History Tried Checklist Preprints Who Export
Where it stands
Not enough benchmark data Fewer than four papers report the same benchmark, so there is no record to follow.
4 papers in 2023–24 and 38 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic grew, from 0.6% to 2.1%.
How this is decided
A benchmark is tracked when at least four papers of this approach report their own method on the same dataset and metric in at least two different years. Its record is the best number reported up to 2024 compared with the best reported in 2025–26: still improving means at least 1% better, flat means at most 0.2% better with at least two papers from 2025–26 trying. Numbers on a different scale from the rest (a fraction among percentages) and far outliers are left out. Results come from different papers with different setups, so they are signals, not head-to-head comparisons. The share of papers is counted at ICML, NeurIPS, ICLR, CVPR and AAAI.
How it got here
Most cited paper of each year
- 2019 · Anti-efficient encoding in emergent communication (NeurIPS 2019 · 68 citations)
- 2023 · Joint processing of linguistic properties in brains and language models (NeurIPS 2023)
- 2024 · UniBias: Unveiling and Mitigating LLM Bias through Internal Attention and FFN Manipulation (NeurIPS 2024 · 3 citations)
- 2025 · Why and How LLMs Hallucinate: Connecting the Dots with Subsequence Associations (NeurIPS 2025 · 3 citations)
- 2026 · Mechanistic Data Attribution: Tracing the Training Origins of Interpretable LLM Units (ICML 2026 · 7 citations)
Earliest papers here
- Anti-efficient encoding in emergent communication (NeurIPS 2019 · 68 citations)
- Joint processing of linguistic properties in brains and language models (NeurIPS 2023)
- UniBias: Unveiling and Mitigating LLM Bias through Internal Attention and FFN Manipulation (NeurIPS 2024 · 3 citations)
What has been tried
20 papers here report 202 measured changes (a component added, removed or swapped, compared with a baseline): helped 66%, no clear effect 6%, hurt 18%, mixed 9%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| Remove biased attention heads (FFN-only) | Across 12 datasets · accuracy | Mixed | UniBias: Unveiling and Mitigating LLM Bias through Internal Attention and FFN Manipulation (NeurIPS 2024 · 3 citations) | |
| Mask Retrieval heads instead of leaving the model unmodified | Extractive_QA · accuracy (%) | 57.14 -> 0 | Hurt | Cognitive Mirrors: Exploring the Diverse Functional Roles of Attention Heads in LLM Reasoning (NeurIPS 2025 · 1 citation) |
| Add tuning of o_proj to Emb + Head-only tuning | AIME 2024 · accuracy | 0.200 -> 0.367 | Helped | Who Reasons in the Large Language Models? (NeurIPS 2025 · 1 citation) |
| Add a harmful-prompt condition to refusal activation steering | Sorry-Bench harmful prompts · refusal rate | 100.00% -> 90.67% | Mixed | Programming Refusal with Conditional Activation Steering (ICLR 2025 · 1 citation) |
| Knock out True/False heads and intervene on the opposite-class heads | GAR classification examples sampled as correct · accuracy | 100% to 60% | Hurt | Benchmarking and Understanding Compositional Relational Reasoning of LLMs (AAAI 2025 · 0 citations) |
| Increase the fixed RMU coefficient c | WMDP · drop-in-accuracy | Helped | On Effects of Steering Latent Representation for Large Language Model Unlearning (AAAI 2025 · 0 citations) | |
| Use two-stage DPO rather than one-stage DPO | In-distribution rubric elicitation · Verifier Score | -1.13 -> -0.70 | Helped | Eliciting Language Model Behaviors with Investigator Agents (ICML 2025 · 0 citations) |
| Increase the Jacobian loss coefficient above the useful range | Pythia MLPs · Reconstruction quality and model performance preservation | Hurt | Jacobian Sparse Autoencoders: Sparsify Computations, Not Just Activations (ICML 2025 · 0 citations) | |
| remove MLP sublayers (use attention-only) | Tanh counterfactual regression · MSE | 0.0132 vs 0.0058 | Hurt | Counterfactual reasoning: an analysis of in-context emergence (NeurIPS 2025 · 0 citations) |
| apply functional homotopy to GR | AdvBench + HarmBench, Llama-2 · ASR | 37.5 → 99.5 | Helped | Functional Homotopy: Smoothing Discrete Optimization via Continuous Parameters for LLM Jailbreak Attacks (ICLR 2025 · 0 citations) |
| add learned positional embeddings instead of using sinusoidal embeddings | Single-pattern ICL, block-noisy, d_E = 100 · accuracy | 0.55 -> 1.00 | Helped | From Unstructured Data to In-Context Learning: Exploring What Tasks Can Be Learned and When (NeurIPS 2024 · 0 citations) |
What reviewers will expect
Datasets and baselines named in the 20 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- AdvBench · 2 of 20 papers
- TruthfulQA · 2 of 20 papers
- GSM8K · 2 of 20 papers
- Alpaca · 2 of 20 papers
- Sorry-Bench · 2 of 20 papers
- HarmBench · 2 of 20 papers
Baselines
- GCG · 2 of 20 papers
- AutoDAN · 2 of 20 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 42 papers; from OpenAlex and the papers' affiliations).
- University of California, Berkeley · 4 papers
- RIKEN · 2 papers
- Stanford University · 2 papers
- University of California, San Diego · 2 papers
- Microsoft · 2 papers
- University of Cambridge · 2 papers
Use it
Download BibTeX of 21 key papersOpen the topic in the interactive map
Get alerts. Describe your idea in scime atlas and save it (free scime account): you get a weekly email when new preprints land close to it.
Related approaches
Same keyword cluster
- Safety alignment of LLMs (123 papers)
- Jailbreak attacks and defenses (57 papers)
Elsewhere in the topic
- LLM unlearning and benchmarks (298 papers)
- LLM pruning and scaling data (285 papers)
- Controlled text generation (220 papers)
- Faithfulness of LLM explanations (169 papers)
- Data and evaluation methodology (136 papers)
- Uncertainty quantification for LLMs (136 papers)
- Pruning and circuits in LLMs (135 papers)
- Statistical frameworks for LLM evaluation (124 papers)
Benchmark numbers, ablations and baseline numbers are extracted by AI from each paper's full text and can be wrong; the paper is the authority. Data: results index of 2026-11-02, citation counts from OpenAlex (2026-10-09).