arXiv cs.AI: SwarmEvolve: Latent Agent Swarms for Evolutionary Discovery and Optimization

arXiv cs.AI: SwarmEvolve: Latent Agent Swarms for Evolutionary Discovery and Optimization

Authors Qile Jiang and George Em Karniadakis present SwarmEvolve, a framework for evolutionary search in which agents improve programs by refining and combining solutions. They argue that existing LLM-guided search can be slow because each round generates long textual chains of thought and repeats that work to maintain population diversity. SwarmEvolve instead uses hundreds of low-cost agents per round that reason in latent space without decoding to text and share information through KV caches. Each swarm explores a different angle; a novelty filter selects distinct ideas by measuring distances between hidden state vectors, and a summarizer receives combined KV caches for crossover. On 21 problems in mathematical discovery and scientific computing, the authors report the highest average score and a median of 3 to 8 minutes to reach the best scores other state-of-the-art frameworks reach in 60 minutes. They also report that, on several problems, SwarmEvolve with Qwen3 8B improves on solutions AlphaEvolve first found with frontier LLMs. The authors frame latent-space communication as a fast, low-cost approach to program search for scientific problems. Escape Meridian interpretation: the results concern the evaluated problems; the abstract does not establish how broadly they generalize.

Publication: 9dc7a912-a1c2-498d-8806-30b63f42c029

attribution — pass: The draft correctly credits the paper to Qile Jiang and George Em Karniadakis. It keeps the performance results and the closing framing attributed to the authors ('the authors report', 'The authors frame'). The generalization caveat is labeled as Escape Meridian interpretation.

Authors:Qile Jiang, George Em Karniadakis

summary — uncertain: The draft is faithful and keeps the authors' argument: text chains of thought are slow, and latent swarms are fast and cheap. A non-specialist would still struggle with the central terms 'latent space', 'KV caches', 'hidden state vectors' and 'crossover'. Nothing tells them that Qwen3 8B is a much smaller model than the frontier LLMs AlphaEvolve used, and the low-cost claim depends on that contrast. Smallest repair: add term-library tooltips for latent space, KV cache, Qwen3 8B and AlphaEvolve, keeping the terms exactly. The title only repeats the paper title and does not say what changed.

reason in latent space without decoding to text, and share information by passing KV caches

date — pass: The originalPublished date of 2026-11-05 matches the arXiv v1 submission date. It is not the later feed date of 2026-11-06.

[Submitted on 5 Nov 2026]

support — pass: Every source-stated claim matches the abstract verbatim in substance: the 21 problems, the highest average score, the 3 to 8 minute median versus 60 minutes, and Qwen3 8B improving on AlphaEvolve solutions on several problems. The interpretation claim is labeled and conditional, and it does not overstate the results. The question follows from the evidence, and the limitation correctly notes that only the abstract was available.

On several problems, SwarmEvolve with Qwen3 8B improves on the solutions that AlphaEvolve first found with frontier LLMs.

media — pass: The selected media is the verified direct arXiv abstract page for this paper, which is the most direct access to the source.