Research map / Large language models / Language, safety and interpretability
Pruning and circuits in LLMs
Prune or extract task-specific circuits and structure from large language models.
135 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2017–2026) · 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.
34 papers in 2023–24 and 94 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic held steady, from 4.9% to 5.1%.
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
- 2017 · On the Transitivity of Hypernym-Hyponym Relations in Data-Driven Lexical Taxonomies (AAAI 2017 · 16 citations)
- 2020 · Mixout: Effective Regularization to Finetune Large-scale Pretrained Language Models (ICLR 2020 · 97 citations)
- 2021 · XL-WSD: An Extra-Large and Cross-Lingual Evaluation Framework for Word Sense Disambiguation (AAAI 2021 · 45 citations)
- 2022 · Memorization Without Overfitting: Analyzing the Training Dynamics of Large Language Models (NeurIPS 2022 · 44 citations)
- 2023 · UniMax: Fairer and More Effective Language Sampling for Large-Scale Multilingual Pretraining (ICLR 2023 · 6 citations)
- 2024 · LLMs are Good Sign Language Translators (CVPR 2024 · 70 citations)
- 2025 · Prompt Compression with Context-Aware Sentence Encoding for Fast and Improved LLM Inference (AAAI 2025 · 10 citations)
- 2026 · mmBERT: A Modern Multilingual Encoder with Annealed Language Learning (ICML 2026 · 4 citations)
Earliest papers here
- On the Transitivity of Hypernym-Hyponym Relations in Data-Driven Lexical Taxonomies (AAAI 2017 · 16 citations)
- Mixout: Effective Regularization to Finetune Large-scale Pretrained Language Models (ICLR 2020 · 97 citations)
- XL-WSD: An Extra-Large and Cross-Lingual Evaluation Framework for Word Sense Disambiguation (AAAI 2021 · 45 citations)
What has been tried
83 papers here report 903 measured changes (a component added, removed or swapped, compared with a baseline): helped 60%, no clear effect 10%, hurt 23%, mixed 6%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| Feed visual encoder outputs directly to the frozen LLM | Phoenix-2014T test · BLEU-4 | 23.40 -> 8.42 | Hurt | LLMs are Good Sign Language Translators (CVPR 2024 · 70 citations) |
| Use adaptive layers and modules (AL-AM) instead of uniform layers and modules (UL-UM) | WikiText2 validation, LLaMA-7B at 50% pruning · perplexity | 57.57 -> 31.80 | Helped | Fluctuation-Based Adaptive Structured Pruning for Large Language Models (AAAI 2024 · 43 citations) |
| increase gamma to retain more performance-critical heads | WikiText-2 · perplexity | Helped | Fairness-Aware Structured Pruning in Transformers (AAAI 2024 · 19 citations) | |
| Use memory lookup during both encoding and decoding instead of decoding only | ∞-Bench Retrieve.KV · R.KV | 85.2 -> 96.8 | Helped | InfLLM: Training-Free Long-Context Extrapolation for LLMs with an Efficient Context Memory (NeurIPS 2024 · 11 citations) |
| remove the positional vector from initial tokens 1–4 | RedPajama · PPL | 7.56 -> >1000 | Hurt | Exploring Context Window of Large Language Models via Decomposed Positional Vectors (NeurIPS 2024 · 11 citations) |
| add L_MNTP to the training loss | LongBench, SingleDoc, 2,000 tokens · accuracy | 35.46 -> 42.38 | Helped | Prompt Compression with Context-Aware Sentence Encoding for Fast and Improved LLM Inference (AAAI 2025 · 10 citations) |
| train with two stages rather than a single stage with in-batch negatives | MTEB (56 tasks) · Avg. | 70.83 -> 72.31 | Helped | NV-Embed: Improved Techniques for Training LLMs as Generalist Embedding Models (ICLR 2025 · 10 citations) |
| Remove the layer completely instead of applying low-rank approximation | language understanding tasks · performance | Hurt | The Truth is in There: Improving Reasoning in Language Models with Layer-Selective Rank Reduction (ICLR 2024 · 7 citations) | |
| Use 64 or 32 memory slots instead of k=128 | Long contexts, including length 500 · BLEU | Hurt | In-context Autoencoder for Context Compression in a Large Language Model (ICLR 2024 · 7 citations) | |
| Use single-nucleotide tokenization instead of K-mer, BPE, or non-overlapping tokenization | RNA benchmark tasks · task wins | best performance on 11 out of 13 tasks | Helped | BEACON: Benchmark for Comprehensive RNA Tasks and Language Models (NeurIPS 2024 · 7 citations) |
| add instrument metadata | MID-FiLD test set · RMSE | 29.4925(±18.45) -> 21.6794(±17.79) | Helped | MID-FiLD: MIDI Dataset for Fine-Level Dynamics (AAAI 2024 · 5 citations) |
| Add Opposing Operation Simplification (OOS) | WikiText2, 50% unstructured sparsity · perplexity | 6.8395 -> 6.7079 | Helped | Pruner-Zero: Evolving Symbolic Pruning Metric From Scratch for Large Language Models (ICML 2024 · 4 citations) |
What reviewers will expect
Datasets and baselines named in the 58 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- C4 · 5 of 58 papers
- WikiText2 · 4 of 58 papers
- MMLU · 4 of 58 papers
- FineWeb-Edu · 4 of 58 papers
- OpenWebText · 3 of 58 papers
- HellaSwag · 3 of 58 papers
- GSM8K · 3 of 58 papers
- ImageNet-1K · 2 of 58 papers
Baselines
- Uniform · 4 of 58 papers
- BM25 · 3 of 58 papers
- ITI · 2 of 58 papers
- zero ablation · 2 of 58 papers
- LLMLingua · 2 of 58 papers
- LLMLingua-2 · 2 of 58 papers
- LongLLMLingua · 2 of 58 papers
- Selective-Context · 2 of 58 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 128 papers; from OpenAlex and the papers' affiliations).
- Google · 9 papers
- Tsinghua University · 8 papers
- Microsoft · 7 papers
- University of California, Berkeley · 5 papers
- Allen Institute for Artificial Intelligence · 5 papers
- Meta · 4 papers
- University of Washington · 4 papers
- Hong Kong University of Science and Technology · 4 papers
- University of Chinese Academy of Sciences · 3 papers
- Renmin University of China · 3 papers
Use it
Download BibTeX of 27 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
- Reasoning mechanisms and failure modes (114 papers)
- Linguistic encoding in LLM representations (123 papers)
- Task vectors and steering (84 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).