CMBEvolve and CosmoEvolve
Cambridge cosmology pair of agentic systems for autonomous scientific discovery: CMBEvolve performs LLM-guided scientific-code evolution via typed tree search, and CosmoEvolve simulates a virtual research laboratory with a principal-investigator agent and student-scientist agents coordinated through a shared blackboard.
| Affiliation | University of Cambridge — Institute of Astronomy, Kavli Institute for Cosmology, Cavendish Astrophysics (paper) |
| First introduced | 2026-05 (arXiv:2605.14791, dated 2026-05-14) |
| Lifecycle stages | Multi-stage (CMBEvolve: hypothesis + experiment design via code evolution; CosmoEvolve: closed-loop hypothesis → exploration → analysis over a shared lab state) |
| Autonomy level | Semi-autonomous (human supplies the task and dataset; CMBEvolve iterates a tree search to a fixed budget; CosmoEvolve runs an open-ended virtual lab with no predefined scientific objective in the ACT DR6 demonstration) |
| Domain focus | Cosmology — weak-lensing out-of-distribution detection, CMB / ACT DR6 data analysis |
| Availability | Code on request — paper states both packages “Will be made available publicly” |
Approach
Two complementary systems for different scientific settings.
- CMBEvolve targets tasks with explicit quantitative objectives. The search is represented as a rooted tree where nodes have one of four types (task, idea generation and selection, code generation, code mutation). Each node stores search statistics (best score S*, visit count N) plus generated content and execution outputs; scores are backpropagated from leaves to the root after evaluation, updating ancestor statistics. The workflow cycles through idea generation, branch selection, targeted mutation, execution, scoring, and score backpropagation.
- CosmoEvolve targets open-ended scientific workflows. It simulates a virtual research laboratory consisting of one principal-investigator (PI) agent and a community of student-scientist agents. The PI observes a summary of the current lab state and selects from a discrete action space (group meeting, individual meeting, task assignment). Student agents carry out work independently and can dispatch subtasks to specialized sub-agents (data and file exploration, planning, code implementation). Agents share role-specific instructions, a compact skill index, persistent memory, and a shared lab-state blackboard; skills load on-demand and tool access is governed by per-agent allowlists.
Validation
Two cosmological demonstrations. (1) CMBEvolve is applied to the FAIR Universe Weak Lensing ML Uncertainty Challenge out-of-distribution-detection task on simulated Hyper Suprime-Cam-style weak-lensing maps. (2) CosmoEvolve is applied to autonomous analysis of public ACT DR6 data products with no predefined scientific objective, asked to explore the data and propose analysis paths.
Notable results
- CMBEvolve iteratively improves its benchmark score on weak-lensing OoD detection over successive code refinements during tree search.
- CosmoEvolve produced a beam-aware split-cross pseudo-Cℓ study of ACT DR6 temperature maps, finding percent-level within-channel stability, tighter same-band cross-array agreement at 90 GHz than at 150 GHz, and few-percent cross-frequency residuals consistent with effective-frequency and foreground-weighting differences.
- A related CosmoEvolve multi-frequency coherence analysis showed that no single multipole cut is appropriate for all ACT DR6 channel pairs, recommending pair- and scale-dependent stability windows — a non-trivial, analysis-grade diagnostic identified autonomously.
Primary paper
Xu & Borrett, “Beyond AI as Assistants: Toward Autonomous Discovery in Cosmology,” arXiv:2605.14791.
Other references
None yet.
Code
Not released at preprint time; both CMBEvolve and CosmoEvolve are stated as “Will be made available publicly.”