SwarmEvolve: Latent-Space Agent Swarms Cut Search Time to 8 Mins

Most multi-agent AI systems communicate in much the same way humans do: one agent writes out text, the next reads it, processes it, and writes back. While functional, this approach carries a hidden cost. Every time an agent converts its internal reasoning into prose, it sheds crucial information—including probability distributions, uncertainty measures, and the nuanced context residing within the model's hidden states. The subsequent agent is then forced to reconstruct all of that context strictly from text.

A recent research paper authored by Jiang and Karniadakis at Brown University identifies this phenomenon as the "decoding tax" and introduces a framework named SwarmEvolve designed to eliminate it entirely. LLM-guided evolutionary search—where AI agents iteratively propose, evaluate, and refine programs or mathematical constructions—has emerged as a productive research domain. Frameworks such as OpenEvolve and AlphaEvolve have previously demonstrated that language models can discover novel scientific algorithms when given appropriate scaffolding.

The primary bottleneck in these setups is throughput. In a standard architecture, an agent generates a chain-of-thought in natural language, which is then tokenized and fed into the next agent to generate further text. For tasks where the fitness function can be checked entirely by a machine (such as verifying if code produces correct output or beats a packing record), prose acts largely as overhead. The agents are essentially writing for each other rather than for human readers.

Stock photo for illustration only, not from the actual event

SwarmEvolve addresses this by asking a fundamental question: what if agents bypassed text altogether? While latent-space reasoning has been explored previously in various forms, SwarmEvolve introduces a structured multi-agent architecture built specifically around this concept. Each round of evolution deploys 450 agents divided into 10 specialized swarms with distinct roles, such as refining existing programs, hybridizing parent programs, or exploring entirely new directions.

Within each swarm, a novelty filter calculates the mean hidden state of each agent and applies greedy farthest-point selection in cosine distance to retain the most distinct reasoning trajectories while discarding agents with overly similar internal states. The retained agents then pass their KV caches—the key-value attention matrices encoding their reasoning context—to a summarizer agent. This summarizer processes all concatenated caches simultaneously, synthesizing a new program proposal that integrates the diverse perspectives of the swarm without any agent writing a single word.