The Big Picture
An iterative, trustless voting loop lets ensembles of small models (<20B parameters) match or exceed single large models (70–100B) while tightening errors and reducing conformity.
ON THIS PAGE
The Evidence
Recurrent rounds of deliberation—where a shared consensus fed back into the group—enable small, diverse models to arrive at high-quality answers without scaling a single model’s size. A trustless voting design prevents agents from voting for their own proposals, reducing groupthink and improving robustness. Runtime selection of which agents to use is framed as an optimization problem (trade-offs between cost, speed, and quality). Empirical tests across math, coding, and safety benchmarks show ensembles of small open-weight models can match or beat much larger single models and that an analytic efficiency model predicts when to stop deliberating. consensus-based decision pattern
Not sure where to start?Get personalized recommendations
Data Highlights
1Ensembles of small models (<20B parameters) matched or exceeded performance of 70–100B single models on the evaluated benchmarks.
2Thermodynamic efficiency model of the deliberation process fit observed behavior with R² ≈ 0.99, supporting predictable optimal stopping.
3Math benchmark standard error tightened from ≈ ±4.2% in early rounds to ≈ ±2.7% at peak convergence (p ≥ 0.90) across 120 total trials.
What This Means
Engineers building AI agents can use this to get near-state-of-the-art results without running huge models, lowering cost and hardware demands. Technical leads and platform teams should consider the trustless voting and runtime agent selection for better auditability, reduced bias, and clearer failure signals. Safety and evaluation teams can use the approach to curb sycophancy (blind agreement) while keeping deliberation transparent. Supervisor Pattern
Ready to evaluate your AI agents?
Learn how ReputAgent helps teams build trustworthy AI through systematic evaluation.
Learn MoreKeep in Mind
Experiments used a fixed, pre-selected agent roster rather than solving the runtime selection optimization for every prompt, so real-time brokerage gains are still unproven. Benchmarks were run in a text-only regime (no code execution or external tool feedback), so tasks that rely on execution or external verification may change results. Datasets were modest in size and used open-weight models, so results may not fully generalize to proprietary or very large-scale production settings. Memory Poisoning
Methodology & More
N-Way Self-Evaluating Deliberation (NSED) replaces one-pass, feed-forward agent pipelines with an iterative loop where a shared consensus acts like memory and is fed back into the group for further refinement. Agents propose solutions, a trustless voting step prevents self-influence, and a dynamic broker conceptually selects which agents to engage based on cost, latency, and expected quality. The topology is designed so deliberation depth grows by time (more rounds) rather than by stacking more models, keeping memory costs low while enabling prolonged, “deep” reflection. Validation used ensembles of open-weight models (all under ~20B parameters) across three benchmarks: a math set with exact answers, a hard coding benchmark, and a safety benchmark for manipulative or sycophantic replies. Runs totaled 120 trials using bootstrap sampling. Results showed small-model ensembles matching or exceeding 70–100B single-model baselines, a tight fit (R² ≈ 0.99) for a derived efficiency-fatigue model that predicts when to stop deliberating, and measurable error reduction on the math tasks. Practical implications include cheaper inference with verifiable, less centralized decision-making and straightforward human-in-the-loop integration; practical gaps to fill include deploying the runtime broker, adding tool-based verification (compilers, tests), and validating broadly on proprietary models and larger datasets. Chain of Thought Pattern Tree of Thoughts Pattern
Avoid common pitfallsLearn what failures to watch for
Credibility Assessment:
No recognizable affiliations or high h-index authors; arXiv preprint.