Skip to content

ResearchTown

external · status: active · focus: ideation · discipline: general · started: 2024

Project page: https://github.com/ulab-uiuc/research-town

Source: projects/landscape/research-town.yml

Positioning

An ICML 2025 multi-agent platform for community-level automatic research simulation. ResearchTown models a research community as agents (Researchers), environments (collaboration rooms), and engines (state-machine controllers that route agents between tasks like idea discussion, rebuttal writing, paper writing, and reviewing). Sits in the ideation + lit-synthesis + drafting + review block.

Distinctive contribution

Studies community dynamics rather than single-pipeline output: how groups of agents interact, divide labor, and shape each other's work. The simulator-vs-pipeline framing makes it a natural vehicle for studying field-level RISE questions (cf. 1, 2).

Evaluation scores

Dimension Score (0–3) Note
Lifecycle coverage 2 Five stages spanning ideation through review (in simulation).
Autonomy level 3 Community runs autonomously; user configures the simulation.
Architectural transparency 3 Open under Apache-2.0; ICML 2025 publication; researcher/environment/engine abstractions documented.
Inputs supported 2 Community/paper-seed inputs; configurable agent skill sets.
Outputs / reproducibility 2 PyPI-installable; trajectories persisted; LLM nondeterminism limits exact reproduction.
Internal evaluation 2 ICML paper presents systematic evaluation of community-level metrics.
Openness 3 Apache-2.0; pip-installable; active community channels.
Maturity / traction 2 204 stars; ICML 2025 acceptance; active development through 2026-05.
Cross-family policy 0 Single-family agent population in published runs.
Runtime assurance 1 Community-level metrics + state-machine engines provide some structural gating.
Cross-platform portability 1 Pip-installable; OpenAI-tied in default config.

Scored on 2026-05-18. See the evaluation rubric.

Tags

Pipeline stages: literature-synthesis rq-formulation hypothesis-generation paper-drafting referee-simulation

Architectural features: multi-agent dag-orchestration persistent-memory artifact-versioning

Inputs: research-community-spec paper-seed

Outputs: agent-trajectories simulated-papers simulated-reviews

Knowledge sources: paper-corpus

Limitations

  • Community simulation, not a deployable RISE pipeline — outputs are research about RISE, not research output.
  • OpenAI API + database required to run end-to-end.

Papers describing this project

  • ResearchTown: Simulator of Human Research Community — ULab UIUC team (2025). ICML 2025. link

  1. Gartenberg, C., Murray, F., Hasan, S., & Pierce, L. (2026). More versus better: Artificial intelligence, incentives, and the emerging crisis in peer review. Organization Science, 37(3). https://doi.org/10.1287/orsc.2026.ed.v37.n3 ↩

  2. Filimonovic, D., Rutzer, C., & Wunsch, C. (2025). Can GenAI improve academic performance? Evidence from the social and behavioral sciences. https://arxiv.org/abs/2510.02408 ↩