SwarmTRM: Coordination Laws for Tiny Recursive Agent Swarms

Large reasoning systems have mostly improved by scaling parameter count, context length, and generated deliberation. This paradigm has produced powerful general models, but it couples reasoning depth to memory traffic, energy cost, and latency. Recent tiny recursive models suggest a complementary scaling axis: a compact neural core can refine latent states over multiple recurrent steps, increasing effective computation without increasing the number of stored parameters. However, isolated recursion follows a single latent trajectory, making it vulnerable to premature commitment, weak self-verification, and limited decomposition. We introduce SwarmTRM, a collective reasoning architecture in which many tiny recursive agents share one compact reasoning core, maintain separate latent trajectories, exchange bottlenecked messages over an adaptive communication graph, and converge through verifier-gated consensus. The central claim is bounded and falsifiable: for structured reasoning tasks with deterministic or high-precision verifiers, coordination among tiny recursive agents may improve accuracy-energy-latency tradeoffs relative to isolated recursion, independent ensembling, non-recursive swarms, or larger monolithic baselines under matched budgets. This version tightens the technical specification by removing absolute correctness claims, making the Intelligence Density Index optional and measurable, defining all loss components with implementation-safe forms, integrating citations into the related-work prose, and presenting the ARC-style example as an illustrative role configuration rather than a guaranteed emergent behavior. SwarmTRM is framed as a verifier-bounded, empirically testable architecture for efficient collective reasoning, not as a universal replacement for large pretrained models.