Trends

Oct 5 – Nov 1 vs. the four weeks before · Updated Nov 4

Memory Management for LLM Agents gained a 521% larger share, while Cross-Cultural Language Model Evaluation gained a 242% larger share; papers on long-horizon agent tasks and contextualized social media translation suggest areas of focus. Policy Learning’s share fell 72%, and Speech and Audio AI’s fell 18%.AI-written summary of the figures below

Papers in the last four weeks
15,218+1% on the 15,128 before
Topics gaining share
3clearly ahead of the overall rate
Topics losing share
17clearly behind it
Strongest rise
Memory Management for LLM Agents+521% share of papers
Last four weeksFour weeks before

The clearest changes in each topic’s share of all new papers, strongest first. Bars show its share week by week.

Gaining

Change in share
  1. Memory Management for LLM Agents50 papers, 8 the four weeks before+521%
  2. Cross-Cultural Language Model Evaluation31 papers, 9 the four weeks before+242%
  3. Efficient Language Model Reasoning74 papers, 41 the four weeks before+79%

Losing

Change in share
  1. Policy Learning26 papers, 92 the four weeks before−72%
  2. Vision-Language-Action Models131 papers, 221 the four weeks before−41%
  3. Underwater Robotics7 papers, 36 the four weeks before−81%
  4. Speech Processing30 papers, 70 the four weeks before−57%
  5. Robotic Control65 papers, 115 the four weeks before−44%
  6. LLM Inference Efficiency27 papers, 62 the four weeks before−57%
  7. AI Governance20 papers, 50 the four weeks before−60%
  8. LLM Uncertainty Estimation20 papers, 48 the four weeks before−59%
  9. Language-Conditioned Robot Manipulation23 papers, 52 the four weeks before−56%
  10. Dexterous Manipulation8 papers, 28 the four weeks before−72%

New topics already drawing papers, and topics whose share jumped in the last two weeks.

Breakout topics

Change in share
  1. Embodied Agent Benchmarks30 papers in four weeks · emerged Sep 22+326%
  2. Typed Decision Models18 papers in four weeks · emerged Oct 6+156%
  3. Chemical Reasoning in LLMs10 papers in four weeks · emerged Oct 1+397%
  4. Memory-Augmented RL9 papers in four weeks · emerged Sep 30+795%
  5. Storage-Efficient Continual Learning8 papers in four weeks · emerged Oct 8+695%
  6. Procedural Memory for LLM Agents8 papers in four weeks · emerged Oct 19+298%

Speeding up

Change in share, last two weeks
  1. Language Model Reasoning Evaluation52 papers in the last two weeks, 14 the two before+350%
  2. LLM Agent Efficiency41 papers in the last two weeks, 10 the two before+397%
  3. Visual Evidence Sensitivity in VLMs36 papers in the last two weeks, 8 the two before+445%
  4. Process-Based Agent Evaluation29 papers in the last two weeks, 5 the two before+603%
  5. Reasoning for VLMs27 papers in the last two weeks, 4 the two before+718%
  6. Behavioral Consistency of LLM Agents25 papers in the last two weeks, 4 the two before+657%

The map regions gaining and losing the most share of papers, week by week. Hover a field to pick out its line.

0%5%10%15%Aug 10Aug 24Sep 7Sep 21Oct 5Oct 19
View as table
Week ofLLM Agents and EvaluationAI for Molecular and Materials DesignAI for Software EngineeringSpeech and Audio AITime Series ForecastingAutonomous Robotics and Navigation
Aug 108.2%1.1%3.5%4.6%4.3%6.7%
Aug 178.5%1.4%3.3%3.6%4.8%7.4%
Aug 248.8%1.5%3.2%4%4.4%5.1%
Aug 318.6%1.9%2.6%5.1%3.9%5.4%
Sep 77.4%1.3%2.5%6.7%4.6%7.2%
Sep 146.9%1%2.8%5.2%4.5%10.6%
Sep 219.3%1.3%2.6%6.7%5.3%7.5%
Sep 2810.3%1.6%2.9%4.6%4.4%6%
Oct 59.4%1.8%3.3%5.1%4.1%7.7%
Oct 1211.3%1.7%3.2%4.6%3.6%6%
Oct 1910.3%1.6%3.5%4.3%3.3%6.4%
Oct 269.9%1.8%2.9%4.2%3.9%6.7%

The newest papers from the topics gaining share fastest.

Universities and labs on a growing share of new papers, by author affiliation. Badges show the change in share.

  1. University of California78 papers, 26 the four weeks before+198%
  2. UC Berkeley77 papers, 43 the four weeks before+78%
  3. Shanghai Jiao Tong University427 papers, 342 the four weeks before+24%
  4. Meituan44 papers, 21 the four weeks before+108%
  5. Shanghai AI Laboratory47 papers, 24 the four weeks before+95%
  6. Google90 papers, 57 the four weeks before+57%
  7. University of Amsterdam54 papers, 30 the four weeks before+79%
  8. Princeton University111 papers, 79 the four weeks before+40%
  9. Michigan State University44 papers, 25 the four weeks before+75%
  10. MIT107 papers, 76 the four weeks before+40%

Affiliations are matched by name, so a lab listed under several spellings, or by department, is counted separately.

Topics from different areas of the map that share more papers than they used to, with their newest shared paper. New subfields often start this way.

Research labels turning up in more papers than before: papers using each in the four weeks before → the last four.

visual evidence sensitivity in vision-language models0 → 45reasoning for vision-language models1 → 24large language model safety monitoring0 → 21large language model context compression0 → 16step-level credit assignment in reinforcement learning10 → 37large language models for evolutionary optimization0 → 15temporal modeling for vision-language-action models0 → 15temporal reasoning in vision-language models0 → 14representation geometry in vision-language models1 → 16reinforcement learning for gui agents2 → 18world models for vision-language-action models2 → 18flow-based vision-language-action models1 → 15energy-aware large language model inference0 → 12instruction generalization in vision-language-action models0 → 12multimodal large language model evaluation0 → 12prompt injection attacks on large language models1 → 143d gaussian splatting training0 → 11evidence-grounded question answering0 → 11