Mo' Models, Mo' Problems: How to best select model pools when designing Multi-Agent Systems

AI summary

Overview

• Research area: Multi-Agent Systems (MAS) for LLM-based reasoning, specifically the understudied problem of model candidate selection.
• Technical level: Intermediate. The paper is empirically driven and its core claims are accessible, but some familiarity with ensembles, routing, majority voting, and LLM-as-a-judge architectures helps.
• Scope: A systematic empirical study of how the choice of which models to include in a multi-agent candidate pool affects downstream reasoning performance across three scientific benchmarks, three MAS architectures, and eight selection strategies.

What This Paper Is About

With millions of open-source models available, builders of multi-agent systems face a practical question: which models should actually be assembled into the pool? The paper tests whether intuitive heuristics — pick the largest models, the most accurate ones, or the most diverse ones — actually produce better systems than just using the single best model. It finds that they usually do not, and that adding more models to a heterogeneous pool typically harms rather than helps performance.

Key Contributions

1. First systematic evaluation of model candidate selection for MAS. The authors compare eight selection strategies — size, architectural family, LLM-chosen, accuracy, correct-answer diversity (IoU), error diversity, and two accuracy-diversity hybrids — spanning both pre-evaluation signals (available from model cards) and post-evaluation signals (requiring calibration data).

2. Separation of oracle potential from achieved performance. The paper quantifies how much improvement each model pool could theoretically yield (oracle pass@1 over the whole pool) and then measures how much of that potential the actual MAS architectures realize. The gap is large and consistent.

3. Evidence that heterogeneous MAS destabilize with scale. Across routing, majority voting, and LLM-as-a-judge systems, increasing candidate pool size degrades performance relative to the best single model in the pool, contradicting the monotonic-improvement assumption common in homogeneous-agent studies.

4. Finding that single-family selection is the most reliable strategy. The rare cases where a heterogeneous MAS outperforms its best base model occur when all candidates come from one architectural family (notably Gemma4), suggesting shared architecture is more useful than raw diversity.

Main Findings

• Accurate models agree on successes, not on failures. Models that score highly tend to get the same questions correct (Mantel r = 0.931, p < .01), but their errors are only weakly correlated (r = 0.384, p < .01). This means error diversity is available even among strong models, but correct-answer diversity is not.

• Size is a weak predictor. Parameter count correlates moderately with accuracy (Spearman r = 0.583) and only weakly with correct-answer similarity (r = 0.219), making size a poor sole selection criterion.

• Oracle potential rises with pool size, but achieved performance falls. Every selection strategy shows increasing oracle pass@1 as k grows, yet in actual MAS deployments larger pools almost always perform worse than the best single model in the pool, and often worse than a random baseline.

• Most intentional selection strategies underperform random. Accuracy- and diversity-based groupings frequently land below random subsets when run through majority vote or LLM-judge architectures, undermining the assumption that informed selection is straightforwardly beneficial.

• Routing benefits from correct-answer diversity; voting does not. The router performed best with IoU-diverse pools, consistent with Mixture-of-Experts theory. Majority vote and LLM-as-a-judge degraded sharply with increased response diversity, suggesting solution diversity pollutes the answer space on reasoning tasks.

• Homogeneous MAS improve with complexity; heterogeneous MAS do not. Adding agents, architectural complexity, and prompt variation helped when all agents shared a base model (best judge@5 = 36.5% versus best pass@1 = 29.4% on HLE), but the same scaling behavior did not transfer to mixed-model pools.

• Domain-specialized models do not honor their specialization. Llama-3.1-8B outperformed its own physics- and chemistry-fine-tuned derivatives (cosmosage-v3.1, Chem-R-8B) on physics and chemistry questions, and the specialists' gains were spread evenly across all domains rather than concentrated in their target fields.

• Architectural families cluster by error pattern, not accuracy pattern. DeepSeek-v4 and most Qwen3.5 models make highly similar errors (similarity > 0.8) distinct from Gemma, gpt-oss, OLMo3, Qwen3, and MiniMax.

• Families vary in reliability. Gemma4 yielded the most consistent relative MAS gains; OLMo3 and Qwen3 produced highly variable results.

Methodology in Plain English

The researchers assembled a pool of 23 language models released between 2024 and 2026, spanning six architecture families and ranging from 2 billion to 1.6 trillion parameters, including dense and mixture-of-experts designs, reasoning and non-reasoning models, and four science-specialized fine-tunes. They drew five generations per question from each model on three difficult science benchmarks: Humanity's Last Exam, GPQA-Diamond, and Frontier Science–Olympiad. Because these are open-form answer datasets, correctness was judged by gpt-oss-120b, which agreed with a human annotator on 93% of a 100-question subsample (Cohen's kappa 0.63). Training and calibration used 15.5k questions drawn from three closed-source STEM datasets, validated as highly correlated with test performance (r > 0.9).

Selection strategies were then applied to produce ordered model lists, and the top-k models under each ordering formed the candidate pool. Pools were tested in three MAS configurations: a trained clustering router (before-generation), embedding-based majority vote (after-generation), and a tournament-style LLM judge using GenSelect (after-generation). Performance was reported both as absolute accuracy and as "MAS Gain" — the change relative to the best single model already inside that pool — to isolate the effect of the multi-agent structure itself.

Why This Matters

• Research impact. The paper challenges a foundational assumption in the multi-agent literature — that more agents and more diversity monotonically improve reasoning. It shows that published gains in homogeneous systems do not transfer to heterogeneous ones, and that oracle metrics alone are insufficient evidence of a good architecture.

• Real-world applications:

  • Enterprise AI orchestration platforms that route queries across internal model catalogs for cost and latency reasons
  • Scientific and medical question-answering systems where domain-specialized fine-tunes are commonly assumed to outperform generalists
  • Model-hosting and inference providers building ensemble endpoints or agent-swarm offerings
  • Benchmark developers and evaluation labs designing tests for multi-model systems rather than single models

• Industry relevance. NVIDIA-affiliated research directly informs how model hubs, routing layers, and ensemble APIs should be designed. The finding that specialist fine-tunes frequently underperform their own base model in-domain suggests that the financial case for training narrow specialists is weaker than commonly assumed, and that candidate pool curation deserves dedicated engineering effort rather than being treated as a configuration detail.

Future Directions

• Develop architectures purpose-built for heterogeneous MAS. The authors note that routing, voting, and judging designs were largely optimized for homogeneous agents; stronger orchestrators or more sophisticated judges may be required to absorb model-pool diversity.

• Evaluate additional diversity criteria. The paper only tested correct-answer and error diversity. Similar accuracy, similar size, or similar architecture constraints might restrict pools usefully, though they also shrink the available candidate set.

• Test tool use and retrieval. Tool use was deliberately disabled to equalize model comparison, but it is common in deployed MAS and could alter response diversity and voting dynamics substantially.

• Expand coverage of MAS paradigms. During-generation systems, more routing backbones, more elaborate voting mechanisms, and other reasoning domains (math, coding) remain unexplored, as do larger architectural families that would permit testing family-based pooling at higher k.

Target Audience

This paper is most valuable to MAS architects and engineers choosing which models to deploy, ML platform teams building routing or ensemble infrastructure over large model catalogs, and researchers studying ensemble methods or multi-agent coordination. It is also relevant to evaluation scientists designing benchmarks for multi-model systems, and to anyone considering investment in narrow specialist fine-tunes based on the assumption that domain specialization translates into domain performance.

Authors’ abstract

Multi-agent Systems (MAS) combine multiple model outputs to solve complex reasoning tasks. However, despite rapid growth of available open-source models, there is limited research on how to select optimal model candidates out of this massive pool. We systematically evaluate 8 model selection strategies (including model size, accuracy and answer diversity) across before-generation (routing) and after-generation (majority-voting, LLM-as-a-judge) MAS architectures on challenging scientific benchmarks. Our findings show a significant gap between theoretical oracle potential and actual performance: Expanding candidate pool sizes often degrades performance below that of the top performing base-model. We find that candidate selection within a single model family is the strategy that yields the best relative performance over a standalone model. These results demonstrate that adding arbitrary models to a heterogeneous MAS can introduce system instability, highlighting model selection as a critical design choice for multi-agent systems.