Beyond the Swarm Hype: Ringelmann's Law Exposes Hard Limits on Multi-Agent LLM Scaling
Preprint derives scaling law showing multi-agent LLM teams hit hard diversity ceilings by N=5–30 unless models are heterogeneous; placebo controls reveal debate gains are mostly re-evaluation.
The preprint applies the classic Ringelmann effect—where individual effort drops in larger groups due to coordination losses—to LLM agents, deriving R(N) = 1/(1+c(N-1)N^{-eta}) that cleanly separates three regimes. Across 44 controlled cells spanning Qwen/Llama/Ministral families (7B–32B), Gemini frontier checks, and conditions from dense debate to sparse heterogeneous teams, the functional form achieves R^{2}>0.99; only parameters shift. Methodology relies on answer diversity and correctness redundancy metrics rather than nominal agent counts, revealing that dense peer influence on free-form math collapses regimes to a hard ceiling of 1/c. This directly challenges mainstream coverage of agent swarms (e.g., AutoGen and CAMEL frameworks) that assume linear gains. A noise-placebo control matches self-correction performance, indicating gains often stem from re-evaluation, not peer content—an insight missed by earlier debate papers such as Du et al. (2023). Only architectural heterogeneity lowers c and escapes the ceiling; communication tweaks do not. Limitations include focus on open-weight homogeneous baselines and absence of long-horizon tool-use tasks. The work is a 2026 arXiv preprint, not yet peer-reviewed. Synthesis with Ringelmann's original 1913 psychology data and recent multi-agent benchmarks shows the same sublinear pattern emerges whenever evidence is redundant rather than independent.
DiversityAgent: Only heterogeneous model teams lower the c parameter and escape the hard ceiling; all homogeneous scaling hits 1/c regardless of k au product.
Sources (3)
- [1]Primary Source(https://arxiv.org/abs/2606.02646)
- [2]Related Source(https://arxiv.org/abs/2305.14325)
- [3]Related Source(https://psycnet.apa.org/record/1913-00001-001)