Research map / Multi-agent systems and games / LLM agents

DiSRouter: Distributed Self-Routing for LLM Selections

ICLR 2026 · Multi-agent systems and games › LLM agents · approach: LLM multi-agent decision making · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes DiSRouter, a distributed self-routing system in which a cascade of Qwen2.5-Instruct agents each decides whether to answer or reject (route onward) using self-awareness trained by SFT then RL with a scenario-conditioned reward.

Problem

Centralized external query routers for LLM pools are rigid to changes in the agent pool and struggle to judge each LLM's knowledge boundaries, limiting utility under cost-performance trade-offs.

Approach

Replace the central router with a network of LLM agents that each use self-awareness (answer vs. reject with "I don't know") as a local routing policy, trained with a localized, preference-conditioned reward so agents train independently and in parallel and adapt to performance-first vs cost-first scenarios via prompt instructions.

Key results

• In the Balance scenario (alpha=0.5) on in-domain tasks, DiSRouter+RL reaches utility 0.61 versus 0.53 for GraphRouter and 0.57 for DiSRouter SFT, and reaches at least 74.29% of the Oracle topline.
• On out-of-domain tasks (Balance), DiSRouter+RL achieves utility 0.48 versus 0.39 for GraphRouter and 0.42 for FrugalGPT.
• Self-assessment classifier of the 7B DiSRouter model reaches 0.80 accuracy versus 0.71 for an 8B LLM-based external classifier and 0.56 for a BERT-based one.

Surprising findings

• Pass@10 change after training is negligible, suggesting gains come from self-awareness rather than improved task capability.
• Automix and FrugalGPT incur latency exceeding the LLM inference time itself in their reported measurements.

Limitations

• Only a simple cascade topology is evaluated; tree or mesh structures are discussed but not tested.
• The self-awareness training pipeline is basic; smaller models show weaker self-awareness.
• Routing uses a fixed, predefined cost per model and only Qwen2.5 family agents in main experiments.

Experimental setup

Datasets: GSM8K ARC MMLU RACE_HIGH OpenbookQA DROP CosmosQA

Benchmarks: GSM8K ARC MMLU RACE_HIGH OpenbookQA DROP CosmosQA SQuAD HellaSwag HeadQA

Metrics: utility accuracy cost F1 Pass@10

Baselines: Oracle Smallest LLM Largest LLM Random RouteLLM FrugalGPT Automix FORC GraphRouter

Builds on: FrugalGPT RouteLLM AutoMix GraphRouter FORC Reinforce++

Closest papers by meaning

• A Unified Approach to Routing and Cascading for LLMs (ICML 2025 · similarity 0.88)
• Router-R1: Teaching LLMs Multi-Round Routing and Aggregation via Reinforcement Learning (NeurIPS 2025 · similarity 0.88)
• Augmented Runtime Collaboration for Self-Organizing Multi-Agent Systems: A Hybrid Bi-Criteria Routing Approach (AAAI 2026 · similarity 0.87)
• Distilling LLM Agent into Small Models with Retrieval and Code Tools (NeurIPS 2025 · similarity 0.87)
• Graph-of-Agents: A Graph-based Framework for Multi-Agent LLM Collaboration (ICLR 2026 · similarity 0.87)
• MAS-GPT: Training LLMs to Build LLM-based Multi-Agent Systems (ICML 2025 · similarity 0.87)