Tag
This paper measures the per-token 'sufficient compute' of LLMs using a Mixture-of-Agents panel of fifteen models across Qwen, OLMo, and R1-distilled families, finding a 0.5B agent reproduces 92-95% of reference tokens while the costliest 10% of tokens account for 64-80% of FLOPs. The resulting per-token compute map improves model routing latency from 7.59s to 5.12s and enables speculative drafting with 32.6% fewer draft tokens than fixed-window baselines.
The paper proposes SMART, a self-evolving multi-agent system for long-form subtitle translation that uses test-time training with a dynamic router, Mixture-of-Agents layer, and judge-refiner loop to build series-level memory and adapt agent prompts without retraining the underlying LLMs. It also introduces the SubtitleArena benchmark (14 genres, 15 target locales) and the SubMQM evaluation framework, achieving a 6.9% average MQM penalty reduction over competing agent systems and top results on the MuSC benchmark.
CERA-MoA introduces a co-evolving framework for mixture-of-agents systems that uses reinforcement learning to dynamically route queries and adapt agent capabilities, enhancing task performance and efficiency.
A user highlights a finding from Hugh Madden's writeup on multi-agent systems: even a strong arbiter (GPT-5.5) can be biased by seeing weaker agents' outputs first, collapsing from ~98% solo accuracy to 7/9.
A detailed technical report on Hermes Mixture-of-Agents (MoA) findings, including benchmark results, caching economics, GPU topology studies, and a roadmap for future development.
The author shares findings from Hermes Mixture-of-Agents experiments, including voter upgrades, GPU topology, and caching economics, showing that local prefix caching can make long agent sessions nearly free and that two independent GPU instances outperform a single partitioned one.
Hermes Agent by Nous Research introduces Mixture of Agents (MoA), allowing users to define presets that combine multiple models for consultation and a final answer model, improving performance by covering blind spots. The feature integrates seamlessly into the existing agent loop, maintaining tools, memory, and context.
Hermes Agent ships 'Mixture of Agents', a feature that runs multiple frontier AI models in parallel and aggregates their outputs, reportedly outperforming individual models like Opus 4.8 and GPT 5.5 on internal benchmarks.
A mixture-of-agents paper (arxiv 2406.04692) shows that a committee of cheap open models can outperform GPT-4o on AlpacaEval 2.0 by leveraging decorrelated errors, and the author shares similar real-world findings where multiple cheap models catch more bugs than a single expensive model.
This paper identifies a fundamental constraint on multi-model LLM systems: accuracy is capped by the rate at which all models fail on the same query. Across 67 frontier models, the all-wrong rate is significantly underestimated by common metrics, limiting gains from voting, routing, and ensemble strategies.
ReM-MoA introduces a memory-augmented Mixture-of-Agents framework that sustains scaling through ranked reasoning memory and curated diversified memory routing, outperforming prior MoA variants across five reasoning benchmarks.
This paper reveals that aggregating complete reasoning traces from multiple LLM agents, rather than just their final answers, can correct errors even when agents unanimously agree, introducing the 'aggregation paradox' and the Self-Consistent Mixture of Agents method.
Proposes MMoA, a novel AI-agent framework that incorporates recurrence mechanisms for a memoried mixture-of-agent architecture. The paper introduces a method to improve agent collaboration and memory in multi-agent systems.