01 / The topic shift
Reasoning & Agents grew 3× while LLMs & Language Models stayed flat
Each line is a topic’s share of that month’s papers. Hover or tab through the months.
3×share of all papers, Dec to Nov
02 / Momentum, all 21 technologies
Code generation is growing fastest
Change in each technology’s share of new papers, last 90 days against the 90 before. Click one to see its papers.
+32%code generation, the season’s fastest grower
- Rising
- Code generation+32%
- Robotics and embodied AI+26%
- Steady
- Speech and audio+18%
- State-space and linear attention+17%
- Mixture of experts+15%
- Tool use+15%
- AI agents+13%
- Retrieval (RAG)+11%
- AI for science+6%
- World models−3%
- 3D and Gaussian splatting−3%
- Small and efficient models−4%
- Reinforcement learning−11%
- Video generation−11%
- Cooling
- Computer-use and web agents−16%
- Hallucination−17%
- Safety and jailbreaks−20%
- Diffusion and flow matching−20%
- Long context and memory−22%
- Reasoning models−23%
- Vision-language models−29%
03 / Research against production
Companies build with tool use, retrieval, and safety and jailbreaks; research has moved on to vision-language models and world models
Share of new papers against share of the 444 company case studies that mention each technology.
18×more common in production: tool use
Research papersCompany case studiesGap
- Ahead in practice
- Tool use18×2.1% of papers · 37.8% of case studies
- Retrieval (RAG)11×1.1% of papers · 12.4% of case studies
- Safety and jailbreaks5×2.0% of papers · 10.1% of case studies
- AI agents1.6×32.9% of papers · 53.6% of case studies
- Ahead in research
- Vision-language models10×7.0% of papers · 0.7% of case studies
- World modelspapers only4.0% of papers · 0.0% of case studies
- Reasoning models5×13.3% of papers · 2.5% of case studies
- Robotics and embodied AI5×9.5% of papers · 1.8% of case studies
- Speech and audio5×4.6% of papers · 0.9% of case studies
- Diffusion and flow matching5×5.7% of papers · 1.1% of case studies
04 / Off the peak
Five technologies are well below their best month
The hatched bar is each one’s highest monthly share this year; the solid bar is Nov.
−71%vision-language models, peak to now
- Reasoning models11.8%peak 26.5%, Jan−55%
- Vision-language models4.3%peak 15.2%, Dec−71%
- Diffusion and flow matching6.6%peak 14.7%, Dec−55%
- Long context and memory2.9%peak 4.1%, May−29%
- Safety and jailbreaks1.7%peak 3.8%, Jan−55%
05 / Words on the rise
“self-improvement” is the word of the season
Title words used much more in the last three months than in the months before. Click one to see its papers.
5.6דself-improvement” in titles