LLMs and multimodal AI

Publishing authors

Base scenario: 9.6k publishing authors. Paper count × 5.14 known authors per paper ÷ 3 papers per author per year. The range assumes 1–5 papers per author; the base and both ends of the range are at least the 9,561 distinct Author IDs already observed in the sample.

The publication rate is an assumption, not a measured rate for this field. The range is a scenario, not a confidence interval.

Random sample: 2,000 of 4,069 papers (49.2%). 13.9% of authorship records lack an Author ID; 9.4% of retrieved works lack an abstract. Retrieved 2026-09-26.

Mean known authors per paper: 5.14; 95% bootstrap interval 4.84–5.48. This describes sampling variability, not search accuracy or uncertainty in the number of researchers.

Applied coauthors are included. Missing or incorrectly linked author records affect these figures. They do not measure research jobs or everyone working in a field.

Private investment

An assumed allocation of broader investment segments. The range is 0.5–2× the base case. Shares are editorial assumptions, not measured deals or research spending.

Base case: $57.3B.

AI infrastructure/models/research/governance: $143.2B × 40.00%

Company investment, not revenue or research spending. Ranges are scenarios, not confidence intervals.

ACL 2026 — research topics ICML 2026 — research topics

Scope and related research

Using a pre-trained model in an application does not necessarily involve pre-training research.

Organizations

Selected examples

OpenAI Research
Research on frontier models, reasoning and multimodal systems.
Anthropic Models
Claude language and multimodal models; system cards document their training.
Google DeepMind Research
Models, world models and robotics.
Meta Models
Llama research includes pretrained multimodal language models.
Alibaba Cloud / Qwen Models
Alibaba Cloud's Qwen team publishes dense and mixture-of-experts language models with reasoning variants.
DeepSeek Models
Open language and reasoning models.

More organizations (17)

AI21 Labs Models
Language models and enterprise agents.
Aleph Alpha Models
Enterprise and sovereign AI.
ByteDance Seed Models
Multimodal model research.
Cohere Models
Enterprise language models and retrieval.
Huawei Models
Industry foundation models.
IBM Product
Enterprise models.
LG AI Research Models
EXAONE and scientific AI.
Magic Research
Magic describes model architecture and pretraining research for code and long-context tasks.
MiniMax Models
Multimodal foundation models.
Mistral AI Models
Open and enterprise models.
NAVER Cloud / HyperCLOVA Models
HyperCLOVA language models and enterprise APIs.
Poolside Models
Foundation models for enterprise work.
Reka AI Models
Multimodal models.
Sakana AI Models
Foundation-model research.
SpaceXAI (xAI) Models
Grok language and reasoning models.
Part of SpaceX
Upstage Models
Document AI and language models.
Z.ai / GLM Models
GLM language models and coding assistants.

LLM post-training and fine-tuning

Supervised fine-tuning, instruction tuning, RLHF, RLAIF, DPO, RLVR, PEFT, and LoRA.

RLHF DPO Fine-tuning

3,757 papers

4.7× vs. 2023

Authors · scenario: 4.8k–19k
Investment · scenario: $11.5B–$45.8B

Sources and methodology

Publication activity

2019: 1
2020: 3
2021: 18
2022: 50
2023: 793
2024: 2,433
2025: 3,757

Matches include papers applying these methods. Search precision and recall have not been systematically measured.

Title and abstract matches for articles, preprints and reviews; retracted work is excluded. Versions and overlapping areas may be counted more than once.

("instruction tuning" OR "supervised fine tuning" OR "preference optimization" OR "reinforcement learning from human feedback" OR "RLHF" OR "RLVR" OR "parameter efficient fine tuning" OR "low rank adaptation") AND (("language model" OR "large language model" OR "vision language model"))

Retrieved 2026-09-26.

Publishing authors

Base scenario: 6.2k publishing authors. Paper count × 4.98 known authors per paper ÷ 3 papers per author per year. The range assumes 1–5 papers per author; the base and both ends of the range are at least the 4,829 distinct Author IDs already observed in the sample.

The publication rate is an assumption, not a measured rate for this field. The range is a scenario, not a confidence interval.

Random sample: 1,000 of 3,757 papers (26.6%). 14.2% of authorship records lack an Author ID; 7.5% of retrieved works lack an abstract. Retrieved 2026-09-26.

Mean known authors per paper: 4.98; 95% bootstrap interval 4.64–5.34. This describes sampling variability, not search accuracy or uncertainty in the number of researchers.

Applied coauthors are included. Missing or incorrectly linked author records affect these figures. They do not measure research jobs or everyone working in a field.

Private investment

An assumed allocation of broader investment segments. The range is 0.5–2× the base case. Shares are editorial assumptions, not measured deals or research spending.

Base case: $22.9B.

AI infrastructure/models/research/governance: $143.2B × 16.00%

Company investment, not revenue or research spending. Ranges are scenarios, not confidence intervals.

ACL 2026 — research topics ICML 2026 — research topics

Scope and related research

Includes parameter-efficient adaptation. Training infrastructure is covered separately.

Organizations

Selected examples

Anthropic Models
Claude post-training uses reinforcement learning with human and AI feedback.
OpenAI Research
InstructGPT research established an instruction-tuning and RLHF pipeline.
Google DeepMind Models
Gemini post-training includes instruction, preference and tool-use data.

More organizations (7)

DeepSeek Research
DeepSeek-R1 combines cold-start data and reinforcement learning.
Fireworks AI Tools & infrastructure
Inference and model tuning.
Labelbox Data & labeling
Data labeling and model training workflows.
Mercor Data & labeling
Expert training data.
Microsoft Research
LoRA adapts pretrained models using low-rank trainable updates.
Together AI Tools & infrastructure
Training and inference cloud.
Turing Data & labeling
Expert data for model training.

LLM reasoning and test-time scaling

Chain-of-thought, verifiers, answer search, and scaling inference-time computation.

Reasoning models Test-time compute Chain of thought

3,135 papers

7.4× vs. 2023

Authors · scenario: 9.4k–16k
Investment · scenario: $6.44B–$25.8B

Sources and methodology

Publication activity

2019: 1
2020: 1
2021: 6
2022: 56
2023: 426
2024: 1,138
2025: 3,135

Matches include papers applying these methods. Search precision and recall have not been systematically measured.

Title and abstract matches for articles, preprints and reviews; retracted work is excluded. Versions and overlapping areas may be counted more than once.

("chain of thought" OR "test time compute" OR "test time scaling" OR "reasoning model" OR "language model reasoning") AND (("language model" OR "large language model" OR "vision language model"))

Retrieved 2026-09-26.

Publishing authors

Base scenario: 9.4k publishing authors. Paper count × 5.09 known authors per paper ÷ 3 papers per author per year. The range assumes 1–5 papers per author; the base and both ends of the range are at least the 9,404 distinct Author IDs already observed in the sample.

The publication rate is an assumption, not a measured rate for this field. The range is a scenario, not a confidence interval.

Random sample: 2,000 of 3,135 papers (63.8%). 15.2% of authorship records lack an Author ID; 8.9% of retrieved works lack an abstract. Retrieved 2026-09-26.

Mean known authors per paper: 5.09; 95% bootstrap interval 4.83–5.34. This describes sampling variability, not search accuracy or uncertainty in the number of researchers.

Applied coauthors are included. Missing or incorrectly linked author records affect these figures. They do not measure research jobs or everyone working in a field.

Private investment

An assumed allocation of broader investment segments. The range is 0.5–2× the base case. Shares are editorial assumptions, not measured deals or research spending.

Base case: $12.9B.

AI infrastructure/models/research/governance: $143.2B × 9.00%

Company investment, not revenue or research spending. Ranges are scenarios, not confidence intervals.

ACL 2026 — research topics AAAI 2026 — AI taxonomy

Scope and related research

Related to planning and symbolic reasoning, but focused on model reasoning at inference time.

Organizations

Selected examples

OpenAI Research
Research on frontier models, reasoning and multimodal systems.
Anthropic Models
Claude models support extended thinking for multi-step reasoning.
Google DeepMind Models
Gemini models use thinking and inference-time computation.

More organizations (4)

Alibaba Cloud / Qwen Models
Alibaba Cloud's Qwen team publishes dense and mixture-of-experts language models with reasoning variants.
DeepSeek Models
Open language and reasoning models.
Harmonic Models
Mathematical reasoning.
SpaceXAI (xAI) Models
Grok language and reasoning models.
Part of SpaceX

Model architectures and memory

Attention, state-space models, hybrids, mixture-of-experts, long context, and architectural memory.

Transformers State-space models Memory

1,886 papers

5.3× vs. 2023

Authors · scenario: 4.6k–9.3k
Investment · scenario: $4.30B–$17.2B

Sources and methodology

Publication activity

2019: 20
2020: 46
2021: 98
2022: 120
2023: 354
2024: 1,222
2025: 1,886

Matches include papers applying these methods. Search precision and recall have not been systematically measured.

Title and abstract matches for articles, preprints and reviews; retracted work is excluded. Versions and overlapping areas may be counted more than once.

("state space model" OR "linear attention" OR "mixture of experts" OR "long context" OR "memory augmented" OR "transformer architecture") AND (("language model" OR "large language model" OR "vision language model"))

Retrieved 2026-09-26.

Publishing authors

Base scenario: 4.6k publishing authors. Paper count × 4.93 known authors per paper ÷ 3 papers per author per year. The range assumes 1–5 papers per author; the base and both ends of the range are at least the 4,648 distinct Author IDs already observed in the sample.

The publication rate is an assumption, not a measured rate for this field. The range is a scenario, not a confidence interval.

Random sample: 1,000 of 1,886 papers (53.0%). 15.2% of authorship records lack an Author ID; 8.4% of retrieved works lack an abstract. Retrieved 2026-09-26.

Mean known authors per paper: 4.93; 95% bootstrap interval 4.42–5.48. This describes sampling variability, not search accuracy or uncertainty in the number of researchers.

Applied coauthors are included. Missing or incorrectly linked author records affect these figures. They do not measure research jobs or everyone working in a field.

Private investment

An assumed allocation of broader investment segments. The range is 0.5–2× the base case. Shares are editorial assumptions, not measured deals or research spending.

Base case: $8.59B.

AI infrastructure/models/research/governance: $143.2B × 6.00%

Company investment, not revenue or research spending. Ranges are scenarios, not confidence intervals.