A track about AI for science, on two coupled fronts:
- Using AI agents to do research — agents that plan, run, and interpret scientific computations (the harness itself is one working example).
- Building scientific knowledge bases for AI — turning the scientific
literature and method craft into structured, machine-consumable knowledge that
grounds those agents (the harness's
.knowledge/cards are one instance).
Unlike the method tracks (ED, MPS, PEPS, QMC, …), the deliverable here is not a reproduced physics figure but a more capable research agent or a better scientific knowledge base for AI: agent reasoning and tool use, retrieval and grounding faithfulness, knowledge-card design and coverage, or an evaluation that measures any of these.
This track suits participants from AI / systems / data backgrounds who want to advance how AI does science, rather than a single computational method.
Track leads: Kun Chen (陈锟), Jin-Guo Liu (刘金国).
Not yet fixed — to be decided at track start. This track is scaffolded without a fixed onboarding target; pick the concrete target interactively when you begin. Candidate framings (choose and pin one with the track lead at the help desk):
- Reproduce an AI-for-science / research-agent result — onboard by reproducing
a published autonomous-research-agent, tool-use, or retrieval-augmented-generation
benchmark, then go beyond it. Keeps the same reproduce → challenge shape as the
physics tracks; use
/download-refto bring the chosen paper into.knowledge/literature/. - Build a scientific knowledge base for AI — onboard by turning a slice of the scientific literature into structured, machine-consumable knowledge (cards + retrieval) and measuring how well an agent is grounded by it.
- Stand up a research-agent evaluation — onboard by building an eval that scores how faithfully an agent answers scientific questions grounded in its knowledge base (retrieval recall, citation faithfulness, task success), then improve the agent against it.
Once a target is chosen, record it here (and its tasks / success criteria) so the track has a fixed onboarding goal like the other tracks.
To be populated when the target is chosen. Foundational directions for this track:
retrieval-augmented generation, tool-use / reasoning-and-acting agents, agent
evaluation and grounding faithfulness. Use /download-ref <arXiv-id|DOI> to fetch
the chosen references into .knowledge/literature/ with correct metadata rather
than hand-entering citations here.