Research map / Large language models / Language, safety and interpretability
Task vectors and steering
Locate task representations in attention heads and steer models via representation interventions.
84 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2019–2026) · an approach in Language, safety and interpretability, keyword cluster “languages · pretraining · multilingual”
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.
9 papers in 2023–24 and 73 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic grew, from 1.3% to 4.0%.
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 · Violence Rating Prediction from Movie Scripts (AAAI 2019 · 29 citations)
- 2022 · Few-Shot Cross-Lingual Stance Detection with Sentiment-Based Pre-training (AAAI 2022 · 52 citations)
- 2024 · A Theoretical Understanding of Self-Correction through In-context Alignment (NeurIPS 2024 · 6 citations)
- 2025 · POI-Enhancer: An LLM-based Semantic Enhancement Framework for POI Representation Learning (AAAI 2025 · 22 citations)
- 2026 · The Assistant Axis: Situating and Stabilizing the Default Persona of Language Models (ICML 2026 · 4 citations)
Earliest papers here
- Violence Rating Prediction from Movie Scripts (AAAI 2019 · 29 citations)
- Few-Shot Cross-Lingual Stance Detection with Sentiment-Based Pre-training (AAAI 2022 · 52 citations)
- A Theoretical Understanding of Self-Correction through In-context Alignment (NeurIPS 2024 · 6 citations)
What has been tried
27 papers here report 303 measured changes (a component added, removed or swapped, compared with a baseline): helped 59%, no clear effect 13%, hurt 22%, mixed 7%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| Add LLM semantic features | POI Recommendation, NY · Hit@1 | 7.029 -> 8.009 | Helped | POI-Enhancer: An LLM-based Semantic Enhancement Framework for POI Representation Learning (AAAI 2025 · 22 citations) |
| Replace multi-round checking with diverse checking context | AdvBench, AutoDAN · ASR | 88% -> 80% | Helped | A Theoretical Understanding of Self-Correction through In-context Alignment (NeurIPS 2024 · 6 citations) |
| Use probability instead of logit difference | GPT-2 XL factual recall, STR, window size 5 · last-to-middle subject-token effect ratio | 4.33× vs 1.22× | Mixed | Towards Best Practices of Activation Patching in Language Models: Metrics and Methods (ICLR 2024 · 5 citations) |
| Train CAVs from filtered outputs rather than concept-related input prompts | toxicity_toxic · toxicity | 0.1231 -> 0.0879 | Helped | Controlling Large Language Models Through Concept Activation Vectors (AAAI 2025 · 4 citations) |
| Probe the last exact answer token instead of the last generated token | TriviaQA · AUC | 0.71 ± 0.006 -> 0.85 ± 0.004 | Helped | LLMs Know More Than They Show: On the Intrinsic Representation of LLM Hallucinations (ICLR 2025 · 4 citations) |
| Use 5-BFT rather than 3-BFT | YouTube replay values, Past · RMSE | 8.12 -> 11.53 | Hurt | Large Content And Behavior Models To Understand, Simulate, And Optimize Content And Behavior (ICLR 2024 · 1 citation) |
| Reduce attention at middle-input sinks and redistribute it to other tokens rather than using temperature scaling | MMLU · accuracy | 44.89 -> 46.82 | Helped | Unveiling and Harnessing Hidden Attention Sinks: Enhancing Large Language Models without Training through Attention Calibration (ICML 2024 · 1 citation) |
| Preserve FV heads rather than ablate them, excluding top induction heads | Few-shot ICL accuracy on 19 natural language tasks · accuracy | Helped | Which Attention Heads Matter for In-Context Learning? (ICML 2025 · 0 citations) | |
| Replace SAE decoder-feature addition with clamping | Concept500 model steering · overall score | 0.165 -> 0.088 | Hurt | AxBench: Steering LLMs? Even Simple Baselines Outperform Sparse Autoencoders (ICML 2025 · 0 citations) |
| Add TSV to the latent representations | TruthfulQA · AUROC | 52.2 -> 80.9 | Helped | Steer LLM Latents for Hallucination Detection (ICML 2025 · 0 citations) |
| Increase the fraction of memorization neurons activated per sequence to p=0.7 instead of p=0.3 | TinyStories repeated training sequences · loss on memorization examples | p = 0.3; p=0.7 | Hurt | Memorization Sinks: Isolating Memorization during LLM Training (ICML 2025 · 0 citations) |
| Use an activation-vector dataset with matching temporal variation | VidHalluc BQA, MCQ, STH, and TSH · accuracy or score | 75.77 -> 79.45; 83.55 -> 83.77; 56.55 -> 66.37; 58.17 -> 58.50 | Helped | Mitigating Hallucination in VideoLLMs via Temporal-Aware Activation Engineering (NeurIPS 2025 · 0 citations) |
What reviewers will expect
Datasets and baselines named in the 32 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- TruthfulQA · 5 of 32 papers
- RealToxicityPrompts · 2 of 32 papers
- Alpaca · 2 of 32 papers
- Alpaca-Eval · 2 of 32 papers
- TriviaQA · 2 of 32 papers
- WinoGrande · 2 of 32 papers
- GSM8K · 2 of 32 papers
- POPE · 2 of 32 papers
Baselines
- LoRA · 3 of 32 papers
- LoReFT · 3 of 32 papers
- ITI · 3 of 32 papers
- SAE · 2 of 32 papers
- DiffMean · 2 of 32 papers
- Semantic Entropy · 2 of 32 papers
- HaloScope · 2 of 32 papers
- Prompt · 2 of 32 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 82 papers; from OpenAlex and the papers' affiliations).
- University of California, Berkeley · 4 papers
- Stanford University · 4 papers
- Google · 3 papers
- Beihang University · 2 papers
- City University of Hong Kong · 2 papers
- Google DeepMind · 2 papers
- Georgia Institute of Technology · 2 papers
- Shanghai Jiao Tong University · 2 papers
- Massachusetts Institute of Technology · 2 papers
- University of Toronto · 2 papers
Use it
Download BibTeX of 22 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
- Pruning and circuits in LLMs (135 papers)
- Reasoning mechanisms and failure modes (114 papers)
- Linguistic encoding in LLM representations (123 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)
- Statistical frameworks for LLM evaluation (124 papers)
- Safety alignment of LLMs (123 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).