SwarmEvolve: 450 latent-space agents hit in 3-8 min what rivals need 60 min for
burny_tech · x · 2026-11-09
Brown University researchers (Qile Jiang & George Em Karniadakis) introduce SwarmEvolve, tackling the bottleneck that LLM-guided evolutionary search wastes time generating textual chains of thought instead of exploring solutions.
- Latent reasoning: hundreds of cheap agents per round reason internally without decoding to text, sharing information by passing KV caches
- Novelty filter: measures distances between agents' hidden-state vectors to keep the most distinct ideas; a summarizer agent merges all members' KV caches for latent-space crossover
- Results: across 21 mathematical/scientific problems it achieves the highest average score, reaches in a median of 3–8 minutes what SOTA frameworks need 60 minutes for, and explores 5–13x more programs per hour
- Low-cost wins: using only Qwen3-8B, it improves solutions AlphaEvolve originally found with frontier LLMs
Takeaway: agent swarms communicating in latent space offer a fast, cheap route to program search for scientific problems.
More from coding & agent
- Publishing an MCP server to the official registry: the 403 gotcha and immutable versions — srikanth_builds · 2026-11-10
- Sonner creator ships /gestures skill packing rubber-banding and momentum drags for apps — floguo · 2026-11-10
- NotiBuddy: Open-Source MCP Tool Lets Coding Agents Ping Your iPhone via Lockscreen — dbenc · 2026-11-10
- Skyrim modder uses AI to reverse-engineer the 15-year-old engine for real-time terrain editing — 141_1337 · 2026-11-10
- MCP for Thunderbird finds a calendar clash hiding in another email account in 20 seconds — SaCreFizZ · 2026-11-10
- Circle: Over 99% of x402 agent payments settle in USDC, ~2,800 endpoints listed — kleffew94 · 2026-11-10