Navigating research at agent scale

As researchers find it harder to keep up with the flood of papers and agents start playing a larger role in scientific research, it may be time to rethink how we navigate the literature.

Two developments should prompt a review of how we navigate research:

  1. Research firehose: the large and rapidly growing corpus of publications
  2. Agent proliferation: the increasingly significant role of autonomous agents in research

Research firehose

Researchers published 3.3 million articles in 2022, about 9,000 a day, up from 2.0 million in 2010 (National Science Board, 2023). arXiv alone now receives around 900 submissions a day, and submissions in the first eight months of 2026 were 27% higher than the previous year (arXiv, n.d.).

Overwhelmed by the flood of publications, researchers are no longer able to review our domains comprehensively. Instead, our finite attention gets distributed unequally (Chu & Evans, 2021):

When the number of papers published per year in a scientific field grows large, citations flow disproportionately to already well-cited papers; the list of most-cited papers ossifies; new papers are unlikely to ever become highly cited [...] These findings suggest that the progress of large scientific fields may be slowed, trapped in existing canon.

Canonical works dominate the conversation, while lesser known works may never get a chance in the spotlight. A lack of diversity in what researchers read and reference can stifle innovation where it is sorely needed.

And with the participation of autonomous research agents, we should expect the growth of our research corpus to accelerate. Yet, perhaps the problem is not that there is too much research. It is that humans have limited attention, which is where agents can help to pick up the slack.

Agent proliferation

Agents are already playing an active role in research:

  • AI Scientist v2 (Yamada et al., 2025) by Sakana, which has authored a paper good enough to pass human peer review for a workshop at ICLR 2025
  • Robin (Ghareeb et al., 2026) by FutureHouse, which was used to identify candidate drugs for dry age-related macular degeneration, a leading cause of irreversible vision loss
  • Co-Scientist (Gottweis et al., 2026) by Google DeepMind, which has helped to identify new drug targets for liver fibrosis, and independently re-discovered a mechanism in antimicrobial resistance

Services like Elicit and Consensus have also emerged as popular AI literature review tools.

As agents play bigger roles in scientific research, how might we adapt existing workflows to play to the strengths of these strange new participants?

Today's research paper and its IMRAD structure (Introduction, Methods, Results, and Discussion) are the result of centuries of human researchers writing and reading. But is there a better complement for the autonomous agent that can review hundreds of papers in minutes? Instead of having agents parse PDFs and skip over hundreds of repetitive introductions, what would an agent-centric token-efficient alternative look like?

Beyond that, how can we design tools that allow us to ask new questions that could never be answered before?

Building Cinnamon

With these considerations in mind, we are building Cinnamon to start a conversation about how we navigate research. Tentatively, Cinnamon is a structured queryable layer over the literature. Claims, experiments, results are extracted and indexed, along with supporting spans in the source.

Existing tools let us navigate papers by topic, metadata or citations. But it remains expensive to search by what a study actually reports: "Which studies evaluated on benchmark X, and what did they score?" or "Which papers contradict claim Y?" Today, answering those means reading every candidate paper.

We are starting with two questions:

  • How well can we extract structured data from papers?
  • Do agents do better literature reviews with a structured database?

This means curating and constructing suitable benchmarks, then testing and iterating what Cinnamon looks like.

As we answer these questions, we hopefully develop a better tool for navigating research, and in the long run, help push the boundaries of scientific knowledge.

References

  • arXiv. (n.d.). Monthly submissions. Retrieved September 18, 2026, from https://arxiv.org/stats/monthly_submissions
  • Chu, J. S. G., & Evans, J. A. (2021). Slowed canonical progress in large fields of science. Proceedings of the National Academy of Sciences, 118(41), e2021636118. https://doi.org/10.1073/pnas.2021636118
  • Ghareeb, A. E., Chang, B., Mitchener, L., Yiu, A., Szostkiewicz, C. J., Shved, D., Gyimesi, G. J., Laurent, J. M., Wright, S. M., Razzak, M. T., White, A. D., Finnemann, S. C., Hinks, M. M., & Rodriques, S. G. (2026). A multi-agent system for automating scientific discovery. Nature, 655(8122), 497–505. https://doi.org/10.1038/s41586-026-10652-y
  • Gottweis, J., Weng, W.-H., Daryin, A., Tu, T., Sirkovic, P., Myaskovsky, A., Glowaty, G., Weissenberger, F., Orlandi, A., Popovici, D., Palepu, A., Rong, K., Tanno, R., Saab, K., Zhang, F., Blum, J., Carroll, A., Kulkarni, K., Tomašev, N., . . . Natarajan, V. (2026). Accelerating scientific discovery with Co-Scientist. Nature, 655(8122), 487–496. https://doi.org/10.1038/s41586-026-10644-y
  • National Science Board. (2023). Publications output: U.S. trends and international comparisons (Science and Engineering Indicators 2024, NSB-2023-33). National Science Foundation. https://ncses.nsf.gov/pubs/nsb202333/publication-output-by-region-country-or-economy-and-by-scientific-field
  • Yamada, Y., Lange, R. T., Lu, C., Hu, S., Lu, C., Foerster, J., Clune, J., & Ha, D. (2025). The AI Scientist-v2: Workshop-level automated scientific discovery via agentic tree search (arXiv:2504.08066). arXiv. https://doi.org/10.48550/arXiv.2504.08066