SciHorizon (Computer Network Information Center, Chinese Academy of Sciences) · 2026-10-26
SciHorizon Self-Evolution Leaderboard, Edition 1
“Each agent first evolves on evolution tasks, then its state is frozen and it is tested on test tasks that do not overlap the evolution tasks; testing never writes back to the agent's state.”
- Who
- SciHorizon (Computer Network Information Center, Chinese Academy of Sciences)
- Date
- 2026-10-26 Release date of Edition 1 in the leaderboard page's own code (publishedAt "2026-10-26"), which the page shows as its release date. SciHorizon announced it on X on 2026-10-28.
- Evidence
- Not rated. The quoted text describes the protocol. The results are leaderboard data, and the technical report is not out yet. how we rate evidence
- Replication
- Nothing to rerun. all claims
- Checked
- 2026-10-29, against the source page
- Links
- Source
Also on the page
- The quotes here are from the site's own English interface strings. Edition 1 "covers data-driven scientific research tasks (data analysis and scientific modeling)". The main track tests "20 held-out modeling targets", and its overall score "is the equal-weight mean of four domains": physics, life sciences, earth & space and chemistry. SciHorizon describes itself as "developed by the Computer Network Information Center, Chinese Academy of Sciences (CNIC), in collaboration with leading institutions".
- Three agent systems are listed, each on "a common evaluation harness and base model"; the pages we saved do not name the base model. On the main track, in the leaderboard data as fetched on 2026-10-29: EvoScientist (memory and skill library) gains 1.42%, with 9 wins and 11 losses on 20 tasks; HyperAgents (meta-agent rewrites the task agent) -1.64%; SIA-H (harness self-improvement) -3.34%.
- Each run is scored against a reference solution on a 0 to 1 scale, where 0.5 matches the reference, and a task's score is "the mean of the 3 qualified runs". Gain is the "Mean of each task's percentage improvement over the initial state."
- The data give 38 evolution tasks for the main track and 38 for the cross-type track. The page text names 40 source tasks for each and shows source tasks as completed of planned, so these read as runs cut before the planned 40. The page marks the cross-domain track as coming soon, and a technical report as "Stay tuned".
Benchmarks it cites
SciHorizon Self-Evolution Leaderboard: what each measures, in its makers' words.