Research map / Large language models / Language, safety and interpretability
Controlled text generation
Steer generation with plug-in attribute models, instructions, and token-level input refinement.
220 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2017–2026) · an approach in Language, safety and interpretability, keyword cluster “prompt · uncertainty · shot”
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.
51 papers in 2023–24 and 86 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic shrank, from 7.3% to 4.7%.
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 · Neural Models for Sequence Chunking (AAAI 2017 · 104 citations)
- 2018 · Empower Sequence Labeling with Task-Aware Neural Language Model (AAAI 2018 · 300 citations)
- 2019 · GLUE: A Multi-Task Benchmark and Analysis Platform for Natural Language Understanding (ICLR 2019 · 4,101 citations)
- 2020 · Do Not Have Enough Data? Deep Learning to the Rescue! (AAAI 2020 · 308 citations)
- 2021 · Learning Contextual Representations for Semantic Parsing with Generation-Augmented Pre-Training (AAAI 2021 · 76 citations)
- 2022 · COLD Decoding: Energy-based Constrained Text Generation with Langevin Dynamics (NeurIPS 2022 · 43 citations)
- 2023 · Structure-informed Language Models Are Protein Designers (ICML 2023 · 56 citations)
- 2024 · MedAlign: A Clinician-Generated Dataset for Instruction Following with Electronic Medical Records (AAAI 2024 · 49 citations)
- 2025 · Spider 2.0: Evaluating Language Models on Real-World Enterprise Text-to-SQL Workflows (ICLR 2025 · 7 citations)
- 2026 · CharBench: Evaluating the Role of Tokenization in Character-Level Tasks (AAAI 2026 · 3 citations)
Earliest papers here
- Neural Models for Sequence Chunking (AAAI 2017 · 104 citations)
- Joint Copying and Restricted Generation for Paraphrase (AAAI 2017 · 60 citations)
- Empower Sequence Labeling with Task-Aware Neural Language Model (AAAI 2018 · 300 citations)
What has been tried
167 papers here report 1,774 measured changes (a component added, removed or swapped, compared with a baseline): helped 70%, no clear effect 7%, hurt 20%, mixed 3%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| increase GPT-4 context length from 2k to 32k | MedAlign · correctness | 51.8% -> 60.1%; 8.3% | Helped | MedAlign: A Clinician-Generated Dataset for Instruction Following with Electronic Medical Records (AAAI 2024 · 49 citations) |
| Train on non-Python data (32K) | HumanEval+ · pass@1 | 34.1 -> 44.5 | Helped | Magicoder: Empowering Code Generation with OSS-Instruct (ICML 2024 · 26 citations) |
| Increase chunk size from 2048 to 88000 | Newly-published books · BooookScore | 78.6 -> 90.9 | Helped | BooookScore: A systematic exploration of book-length summarization in the era of LLMs (ICLR 2024 · 14 citations) |
| Add self-curated augmented data to seed data | Instruction-following evaluation · win rate against text-davinci-003 | Helped | Self-Alignment with Instruction Backtranslation (ICLR 2024 · 13 citations) | |
| Increase ContextCite ablations from 32 to 64, 128, or 256 | TyDi QA, Hotpot QA, and CNN DailyMail · attribution quality | Helped | ContextCite: Attributing Model Generation to Context (NeurIPS 2024 · 13 citations) | |
| Use learned LLM base distributions for lattice parameters and fractional coordinates instead of simple base distributions | MP-20 · Stability Rate (%) | 8.95 -> 17.82 | Helped | FlowLLM: Flow Matching for Material Generation with Large Language Models as Base Distributions (NeurIPS 2024 · 8 citations) |
| Add oracle SQL function documentation | Spider 2.0-Lite · execution accuracy | 5.68% -> 5.85% | Helped | Spider 2.0: Evaluating Language Models on Real-World Enterprise Text-to-SQL Workflows (ICLR 2025 · 7 citations) |
| add IUPAC name to the prompt | BBBP and TOX21 classification, 5% few-shot · Avg ROC-AUC | 63.5 -> 64.9 | Helped | Bridging Molecular Graphs and Large Language Models (AAAI 2025 · 6 citations) |
| add segmentation-mask supervision for cross-attention (Lreg) | FlintStonesSV · Char-Acc | 88.86 -> 90.36 | Helped | StoryGPT-V: Large Language Models as Consistent Story Visualizers (CVPR 2025 · 6 citations) |
| Use top-k selection instead of τ=2.0 sampling | SlimPajama validation · perplexity | 8.91 -> 10.59 | Hurt | QuRating: Selecting High-Quality Data for Training Language Models (ICML 2024 · 6 citations) |
| Select the sentence with the highest global distribution score instead of using all generated sentences | SemEval test · Micro-F1 | 86.1±0.9 -> 88.4±0.4 | Helped | Reliable Data Generation and Selection for Low-Resource Relation Extraction (AAAI 2024 · 4 citations) |
| Add RRHF alignment after SFT | DABench, COCO Caption, and Parti Prompts (SD XL) · Unified Metric | 55.8 -> 57.5; 55.6 -> 57.4; 56.5 -> 58.8 | Helped | DiffAgent: Fast and Accurate Text-to-Image API Selection with Large Language Model (CVPR 2024 · 4 citations) |
What reviewers will expect
Datasets and baselines named in the 57 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- CNN/DailyMail · 2 of 57 papers
- SQuAD · 2 of 57 papers
- The Pile · 2 of 57 papers
- GSM8K · 2 of 57 papers
- BIRD · 2 of 57 papers
- ARC-C · 2 of 57 papers
- ARC-E · 2 of 57 papers
- Winogrande · 2 of 57 papers
Baselines
- GPT-4o · 6 of 57 papers
- GPT-4o-mini · 3 of 57 papers
- Llama3.1-8B-Instruct · 3 of 57 papers
- DIN-SQL · 2 of 57 papers
- CHESS · 2 of 57 papers
- QwQ 32B · 2 of 57 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 137 papers; from OpenAlex and the papers' affiliations).
- Tsinghua University · 6 papers
- Meta · 6 papers
- Microsoft · 6 papers
- Chinese Academy of Sciences · 5 papers
- IBM · 5 papers
- Stanford University · 5 papers
- Google · 5 papers
- Princeton University · 5 papers
- Tencent · 4 papers
- Amazon · 4 papers
Use it
Download BibTeX of 40 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
- LLM unlearning and benchmarks (298 papers)
- LLM pruning and scaling data (285 papers)
- Faithfulness of LLM explanations (169 papers)
- Data and evaluation methodology (136 papers)
- In-context learning analysis (95 papers)
- Parallel and early-exit decoding (100 papers)
Elsewhere in the topic
- Uncertainty quantification for LLMs (136 papers)
- Pruning and circuits in LLMs (135 papers)
- Statistical frameworks for LLM evaluation (124 papers)
- Safety alignment of LLMs (123 papers)
- Linguistic encoding in LLM representations (123 papers)
- Reasoning mechanisms and failure modes (114 papers)
- Speech and audio language models (111 papers)
- Sign and cross-lingual translation (96 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).