NeurIPS · ICML · ICLR · CVPR · ICCV · AAAI · ACL · KDD  ·  main-conference papers, 2022–2026

Automating research: the paper list

Accepted papers on auto-research agents, AutoML, ML-engineering agents, AI for science and self-improving LLMs — curated from the conferences' own accepted-paper listings across eight venues. Title-only coverage for ICML 2026, CVPR 2022/2026 and ICCV 2023; KDD 2026 supplemented from the ACM DL; NeurIPS 2026 not out yet.

1628papers
compiled 2026-08-26
1628 shown

A Auto-Research / AI-Scientist agents 255

LLM agents and studies that automate parts of the research process itself: ideation, hypothesis generation, literature work, experiment execution, paper writing and peer review.

A1 End-to-end research agents & discovery systems 19

A2 Research idea & hypothesis generation, evaluation, validation 41

A3 Literature, papers & scientific writing 89

A4 Peer review & meta-science 49

A5 Benchmarks, evaluation & position papers 30

A6 Adjacent: web "deep research" agents (report-writing over the web, not scientific research) 27

B Auto-MLE / data-science agents 101

Agents that do machine-learning engineering, Kaggle-style modelling, data analysis and AI R&D, plus the benchmarks and environments that evaluate them.

B1 Agents & systems 38

B2 Benchmarks & environments 47

B3 Paper reproduction & research-code generation 16

C Self-improving LLMs 458

Models and agents that improve from their own outputs: self-training, self-play, self-rewarding, self-correction, iterative preference optimisation, self-evolving agents, and the theory and failure modes of these loops.

C1 Self-training & bootstrapped reasoning (STaR family, self-generated data) 105

C2 Self-play & zero-data RL 54

C3 Self-rewarding, self-verification & unsupervised / test-time RL 70

C4 Self-correction, self-refinement & self-debugging 69

C5 Iterative preference optimisation & self-alignment 30

C6 Self-evolving agents & systems 83

C7 Theory, limits & synthetic-data loops / model collapse 47

D AutoML 499

Neural architecture search, hyperparameter optimisation, AutoML systems, automated feature engineering, prior-data fitted networks, and the newer LLM-driven AutoML and agent/workflow search.

D1 Neural architecture search (methods, proxies, benchmarks) 196

D2 Hyperparameter optimisation (multi-fidelity, meta / in-context BO, transfer) 89

D3 AutoML systems, CASH, algorithm selection & configuration, model selection 66

D4 Automated feature engineering & data augmentation 27

D5 Prior-data fitted networks & tabular foundation models 39

D6 LLM-driven AutoML & automated algorithm discovery 54

D7 AutoML for LLM systems: agent, workflow & pipeline search 28

E AI for Science 315

Scoped to LLM/agent-driven science: agents for scientific domains and lab automation, scientific reasoning benchmarks, LLM-driven equation/law/program discovery, and a selective set of scientific LLMs and foundation models. The long tail of domain-specific modelling papers (individual property predictors, PDE solvers, molecule generators) is deliberately excluded.

E1 LLM / agent systems for scientific domains & lab automation 92

E2 Scientific reasoning & knowledge benchmarks 119

E3 LLM-driven discovery: equations, laws, programs, mathematics 29

E4 Scientific LLMs & landmark scientific foundation models (selective) 75

142 Made with Syncric