Back to AI News
arxiv

SciMARS: Recursive Self-Improvement of Multi-Agent Harnesses for Closed-Loop Scientific Discovery

SciMARS is a framework for recursive, task-local adaptation of multi-agent harnesses used in closed-loop scientific discovery, where agents coordinate analysis, experimental design and verification as evidence evolves. It introduces a shared evidence graph plus separate component editors that adapt agent roles and coordination, with an added memory editor controlling how agents record, retrieve and receive evidence; candidate probes guide selection of the next harness. Evaluation used experimental-data replay and a digital twin: SciMARS + Memory resolved 12/18 GPT-6-Astra and 2/18 GLM-5.3 attempts on Autonomous Science Bench versus 8/18 and 1/18 for the base team, and solved 14/38 and 11/20 tasks on ScienceAgentBench and AppWorld versus 11/38 and 5/20. The authors conclude that combining shared memory with harness adaptation helps across scientific and interactive tasks.

AI Brief

It offers a concrete architectural recipe and comparative benchmark numbers for teams building multi-agent scientific-discovery systems, especially on memory and self-adaptation of agent roles.

arXiv:2611.12552v1 Announce Type: new Abstract: We propose SciMARS, a framework for recursive task-local adaptation of multi-agent harnesses in closed-loop scientific discovery. Multi-agent systems coordinate analysis, experimental design, and verification as scientific evidence evolves. However, two problems limit their effectiveness: (i) without shared memory, agents lose observations and repeat work; (ii) fixed role procedures, handoffs, and memory policies perpetuate mistakes despite retained evidence. We address these problems with a shared evidence graph and separate component editors. SciMARS adapts roles and coordination; SciMARS + Memory adds a third editor for how agents record, retrieve, and receive evidence. Candidate probes guide composition and selection of the next harness. We evaluate scientific discovery through experimental-data replay and a digital twin. SciMARS + Memory resolves 12/18 GPT-6-Astra and 2/18 GLM-5.3 attempts on Autonomous Science Bench, compared with 8/18 and 1/18 for the base multi-agent team. On ScienceAgentBench and AppWorld, it solves 14/38 and 11/20 tasks, compared with 11/38 and 5/20 for the base team. These results support combining shared memory with harness adaptation across scientific and interactive tasks.
Be the first to rate this article.

Comments & Questions (0)

Captcha
Please be respectful — let's keep the conversation friendly.

No comments yet

Be the first to comment!