# EvoRoute: Experience-Driven Self-Routing LLM Agent Systems

Source: [EvoRoute: Experience-Driven Self-Routing LLM Agent Systems](<https://www.sota2.com/research/paper/695de87753fdd5c361743c7c>)

Collection index: [llms.txt](<https://www.sota2.com/research/llms.txt>)

Data updated: 2026-01-08T05:00:05.049Z

Authors: Guibin Zhang, Haiyang Yu, Kaiming Yang, Bingli Wu, Fei Huang, Yongbin Li, Shuicheng Yan

Published: 2026-01-06T04:06:46.000Z

## Abstract

Complex agentic AI systems, powered by a coordinated ensemble of Large Language Models (LLMs), tool and memory modules, have demonstrated remarkable capabilities on intricate, multi-turn tasks. However, this success is shadowed by prohibitive economic costs and severe latency, exposing a critical, yet underexplored, trade-off. We formalize this challenge as the \\textbf{Agent System Trilemma}: the inherent tension among achieving state-of-the-art performance, minimizing monetary cost, and ensuring rapid task completion. To dismantle this trilemma, we introduce EvoRoute, a self-evolving model routing paradigm that transcends static, pre-defined model assignments. Leveraging an ever-expanding knowledge base of prior experience, EvoRoute dynamically selects Pareto-optimal LLM backbones at each step, balancing accuracy, efficiency, and resource use, while continually refining its own selection policy through environment feedback. Experiments on challenging agentic benchmarks such as GAIA and BrowseComp+ demonstrate that EvoRoute, when integrated into off-the-shelf agentic systems, not only sustains or enhances system performance but also reduces execution cost by up to $80\\%$ and latency by over $70\\%$.

## Recorded benchmark results

A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage.

- [Multi-task Evaluation on Aggregate All Tasks](<https://www.sota2.com/research/sota/multi-task-evaluation-on-aggregate-all-tasks.md>): Multi-task Evaluation; Aggregate All tasks (summary); Score: 74.6.

- [General AI Assistant Tasks on GAIA All levels original (test)](<https://www.sota2.com/research/sota/general-ai-assistant-tasks-on-gaia-all-levels-original-test.md>): General AI Assistant Tasks; GAIA All levels original (test); Performance (%): 63.19.

- [General AI Assistant Tasks on GAIA Level 1 original (test)](<https://www.sota2.com/research/sota/general-ai-assistant-tasks-on-gaia-level-1-original-test.md>): General AI Assistant Tasks; GAIA Level 1 original (test); Performance (%): 83.02.

- [General AI Assistant Tasks on GAIA Level 2 original (test)](<https://www.sota2.com/research/sota/general-ai-assistant-tasks-on-gaia-level-2-original-test.md>): General AI Assistant Tasks; GAIA Level 2 original (test); Perf (%): 59.3.

- [Web Browsing and Tool Use on BrowseComp+ original (test)](<https://www.sota2.com/research/sota/web-browsing-and-tool-use-on-browsecomp-plus-original-test.md>): Web Browsing and Tool Use; BrowseComp+ original (test); Performance (%): 38.72.

- [Operational Metrics on DS-1000](<https://www.sota2.com/research/sota/operational-metrics-on-ds-1000.md>): Data Science; DS-1000; Performance Score: 56.5.

- [General AI Assistant Tasks on GAIA Level 3 original (test)](<https://www.sota2.com/research/sota/general-ai-assistant-tasks-on-gaia-level-3-original-test.md>): General AI Assistant Tasks; GAIA Level 3 original (test); Performance: 33.33.

- [Web Search on HotpotQA (Latency and Cost)](<https://www.sota2.com/research/sota/web-search-on-hotpotqa-latency-and-cost.md>): Web Search; HotpotQA; Performance Score: 87.8.

- [Medical Reasoning on DDXPlus](<https://www.sota2.com/research/sota/medical-reasoning-on-ddxplus.md>): Medical Reasoning; DDXPlus; Performance Score: 79.5.

## Original sources and implementation

[Original paper](<https://arxiv.org/abs/2601.02695>)

[PDF](<https://arxiv.org/pdf/2601.02695.pdf>)
