Research map / Large language models / Language, safety and interpretability
Multilingual LLM data and culture
Allocate multilingual data, select pretraining data, and adapt models to cultural differences.
69 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2022–2026) · an approach in Language, safety and interpretability, keyword cluster “natural language · speech · generation”
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.
18 papers in 2023–24 and 47 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic held steady, from 2.6% to 2.5%.
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
- 2022 · The BigScience ROOTS Corpus: A 1.6TB Composite Multilingual Dataset (NeurIPS 2022 · 61 citations)
- 2023 · MultiSpider: Towards Benchmarking Multilingual Text-to-SQL Semantic Parsing (AAAI 2023 · 22 citations)
- 2024 · CultureLLM: Incorporating Cultural Differences into Large Language Models (NeurIPS 2024 · 26 citations)
- 2025 · Language Ranker: A Metric for Quantifying LLM Performance Across High and Low-Resource Languages (AAAI 2025 · 21 citations)
- 2026 · MGAL: A Multilingual Granularity-Aware Long-Context Benchmark (ICML 2026 · 0 citations)
Earliest papers here
- The BigScience ROOTS Corpus: A 1.6TB Composite Multilingual Dataset (NeurIPS 2022 · 61 citations)
- Multi-LexSum: Real-world Summaries of Civil Rights Lawsuits at Multiple Granularities (NeurIPS 2022 · 20 citations)
- From Good to Best: Two-Stage Training for Cross-Lingual Machine Reading Comprehension (AAAI 2022 · 16 citations)
What has been tried
50 papers here report 512 measured changes (a component added, removed or swapped, compared with a baseline): helped 67%, no clear effect 8%, hurt 20%, mixed 5%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| Add semantic-template generation and synonym replacement to WVS seeds | Culture-related datasets · average performance | Helped | CultureLLM: Incorporating Cultural Differences into Large Language Models (NeurIPS 2024 · 26 citations) | |
| Replace specialized NTL augmenting model with vanilla PaLM2-XXS | FLORES-200 (XX-En) · languages improved over PaLM2-S | 175 -> 115 | Hurt | LLM Augmented LLMs: Expanding Capabilities through Composition (ICLR 2024 · 10 citations) |
| Add an Error Corrector before the LaTeX Translator | MathSpeech benchmark, Whisper-base · CER | 0.693 -> 0.336 | Helped | MathSpeech: Leveraging Small LMs for Accurate Conversion in Mathematical Speech-to-Formula (AAAI 2025 · 7 citations) |
| Add essential aspects (EA) | CNN/DailyMail · ROUGE-1 | 48.08 -> 48.54 | Helped | Distilling Structured Rationale from Large Language Models to Small Language Models for Abstractive Summarization (AAAI 2025 · 5 citations) |
| Add the commonalities enhancement module | Monolingual and multilingual scenarios · MRR | Helped | CONSIDER: Commonalities and Specialties Driven Multilingual Code Retrieval Framework (AAAI 2024 · 5 citations) | |
| deactivate language-specific neurons | XLSum average across French, Chinese, Spanish, and Russian · ROUGE-L | 26.6 -> 0.37 | Hurt | How do Large Language Models Handle Multilingualism? (NeurIPS 2024 · 5 citations) |
| Add the mapping stage | MGSM · low-resource average accuracy | 31.5 -> 51.7; 39.1 -> 53.5 | Helped | MindMerger: Efficiently Boosting LLM Reasoning in non-English Languages (NeurIPS 2024 · 4 citations) |
| Add chain-of-thought prompting to Zero-shot | Wikipedia · Continuous F1 | 0.455 -> 0.449 | No clear effect | End-to-End Ontology Learning with Large Language Models (NeurIPS 2024 · 4 citations) |
| Add the re-ranker | UN Chinese-to-English · BLUE-4 | 29.35 -> 29.07 | Hurt | Adaptive Few-shot Prompting for Machine Translation with Pre-trained Language Models (AAAI 2025 · 3 citations) |
| Use intermediate delexicalization levels | PtBrVid validation set · F1 | Helped | Enhancing Portuguese Variety Identification with Cross-Domain Approaches (AAAI 2025 · 2 citations) | |
| Increase X-InSTA demonstrations from 2-shot to 3-shot | French L1, L2, and L3 · F1 | 44.77 -> 46.62; 11.56 -> 11.92; 16.63 -> 17.74 | Helped | Multilingual LLMs Inherently Reward In-Language Time-Sensitive Semantic Alignment for Low-Resource Languages (AAAI 2025 · 2 citations) |
| Add IdiomKB meanings through KB-CoT prompting | Zh→En · GPT-4 | 1.66 -> 2.08 | Helped | Translate Meanings, Not Just Words: IdiomKB’s Role in Optimizing Idiomatic Translation with Language Models (AAAI 2024 · 2 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
- MMLU · 5 of 32 papers
- GSM8K · 5 of 32 papers
- OSCAR · 3 of 32 papers
- ARC · 3 of 32 papers
- NTREX · 3 of 32 papers
- FineWeb-2 · 3 of 32 papers
- DSL-TL · 2 of 32 papers
- FRMT · 2 of 32 papers
Baselines
- BART · 2 of 32 papers
- GPT-3.5 · 2 of 32 papers
- GPT-4o · 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 65 papers; from OpenAlex and the papers' affiliations).
- Google · 7 papers
- Harbin Institute of Technology · 4 papers
- Google DeepMind · 4 papers
- ByteDance · 4 papers
- Chinese Academy of Sciences · 3 papers
- Stanford University · 3 papers
- National University of Singapore · 3 papers
- EPFL · 3 papers
- Alibaba Group · 3 papers
- Tsinghua University · 3 papers
Use it
Download BibTeX of 28 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
- Text generation and evaluation metrics (79 papers)
- Speech and audio language models (111 papers)
- Sign language datasets and translation (23 papers)
- Sign and cross-lingual translation (96 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).