Research map / Large language models / AI and society
Critiques of LLM claims and judging
Examine validity of LLM research claims, LLM-as-judge use, and research benchmarks.
104 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2021–2026) · an approach in AI and society, keyword cluster “human · systems · argue”
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.
24 papers in 2023–24 and 79 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic grew, from 10.6% to 15.5%.
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
- 2021 · Think Big, Teach Small: Do Language Models Distil Occam’s Razor? (NeurIPS 2021)
- 2023 · OpenAGI: When LLM Meets Domain Experts (NeurIPS 2023 · 74 citations)
- 2024 · MemoryBank: Enhancing Large Language Models with Long-Term Memory (AAAI 2024 · 201 citations)
- 2025 · SafetyPrompts: A Systematic Review of Open Datasets for Evaluating and Improving Large Language Model Safety (AAAI 2025 · 15 citations)
- 2026 · The Ideation-Execution Gap: Execution Outcomes of LLM-Generated versus Human Research Ideas (ICLR 2026 · 6 citations)
Earliest papers here
- Think Big, Teach Small: Do Language Models Distil Occam’s Razor? (NeurIPS 2021)
- OpenAGI: When LLM Meets Domain Experts (NeurIPS 2023 · 74 citations)
- Building Socio-culturally Inclusive Stereotype Resources with Community Engagement (NeurIPS 2023 · 2 citations)
What has been tried
56 papers here report 454 measured changes (a component added, removed or swapped, compared with a baseline): helped 59%, no clear effect 9%, hurt 23%, 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 |
|---|---|---|---|---|
| Add evaluation criteria to the prompt | Climate questions · tone | 3.34 -> 3.59 | Helped | Assessing Large Language Models on Climate Information (ICML 2024 · 18 citations) |
| Replace CUT's steering direction with a random vector of the same norm | WMDP-Bio and WMDP-Cyber · accuracy | No clear effect | The WMDP Benchmark: Measuring and Reducing Malicious Use with Unlearning (ICML 2024 · 13 citations) | |
| Fine-tune with generated cultural opinions rather than the base model alone | Arabic, Bengali, Chinese, and Portuguese content moderation · Avg F1 | Ar .370 -> .451; Bn .542 -> .622; Zh .448 -> .636; Pt .593 -> .594 | Helped | CulturePark: Boosting Cross-cultural Understanding in Large Language Models (NeurIPS 2024 · 12 citations) |
| add knowledge construction | MNDEC pre-configured runtimes · PPL | 31.22±13.38 -> 11.07±5.49 | Helped | Sim911: Towards Effective and Equitable 9-1-1 Dispatcher Training with an LLM-Enabled Simulation (AAAI 2025 · 8 citations) |
| Vary temperature across the tested settings | Dictator and Ultimatum games · answer ratio and accept rate | No clear effect | EAI: Emotional Decision-Making of LLMs in Strategic Games and Ethical Dilemmas (NeurIPS 2024 · 4 citations) | |
| Combine verbalised and hidden-state predictions | OSCLMRIC test · F1-score | 0.88 ± 0.02; 0.93 ± 0.01 -> 0.94 ± 0.01 | Helped | Trustworthy and Practical AI for Healthcare: A Guided Deferral System with Large Language Models (AAAI 2025 · 3 citations) |
| scale up input context length | EquationInference · F1 | Mixed | AAAR-1.0: Assessing AI’s Potential to Assist Research (ICML 2025 · 3 citations) | |
| Replace “Are you sure?” with the self-corrective iteration prompt | TruthfulQA MC1 · accuracy | No clear effect | Understanding the Effects of Iterative Prompting on Truthfulness (ICML 2024 · 3 citations) | |
| defer to the trusted model for outputs with suspicion scores up to 10% | APPS backdooring setting · safety | 62% -> 83% | Helped | AI Control: Improving Safety Despite Intentional Subversion (ICML 2024 · 3 citations) |
| add block-wise parameter sharing | 60G FLOPs target · average accuracy | 0.381 -> 0.383 | Helped | MeRino: Entropy-Driven Design for Generative Language Models on IoT Devices (AAAI 2025 · 2 citations) |
| Provide the full research paper as context | ResearchCodeBench · success rate | up to 30% | Helped | ResearchCodeBench: Benchmarking LLMs on Implementing Novel Machine Learning Research Code (NeurIPS 2025 · 2 citations) |
| Add transsuasion training to content and behavior simulation training | Comparative Transsuasion Random · accuracy | 77.9 -> 80.9 | Helped | Measuring And Improving Persuasiveness Of Large Language Models (ICLR 2025 · 2 citations) |
What reviewers will expect
Datasets and baselines named in the 56 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- The Pile · 2 of 56 papers
- TruthfulQA · 2 of 56 papers
- BBQ · 2 of 56 papers
- AnthropicRedTeam · 2 of 56 papers
- ToxiGen · 2 of 56 papers
- MMLU · 2 of 56 papers
- GPQA Diamond · 2 of 56 papers
- choices13k · 2 of 56 papers
Baselines
- GPT-4o · 5 of 56 papers
- GPT-4o-mini · 4 of 56 papers
- GPT-4-turbo · 3 of 56 papers
- Claude 3.5 Sonnet · 3 of 56 papers
- GPT-3.5-turbo · 2 of 56 papers
- GPT-3.5 · 2 of 56 papers
- GPT-4 · 2 of 56 papers
- Zero-shot CoT · 2 of 56 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 103 papers; from OpenAlex and the papers' affiliations).
- Stanford University · 11 papers
- Microsoft · 7 papers
- Massachusetts Institute of Technology · 6 papers
- Google · 6 papers
- Meta · 5 papers
- University of California, Berkeley · 5 papers
- Amazon · 4 papers
- Princeton University · 4 papers
- Harvard University · 3 papers
- University of California, Los Angeles · 3 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
- Human-centered trust and agency (161 papers)
- Social-science view of AI evaluation (159 papers)
- Data commons and human baselines (159 papers)
- Benchmarking and reproducibility practice (104 papers)
- AI for health and public decisions (55 papers)
Elsewhere in the topic
- Teacher needs and AI curricula (45 papers)
- Classroom tools and student perceptions (40 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).