Research map / Large language models / Reasoning and chain of thought
Causal Dependency-Aware Unsupervised Routing for Large Reasoning Models
Jiacheng Liu, Hao Liu, Xiaofeng Hou, Wei Xue, Yike Guo
Read the paperCheck your idea against 137,217 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 an unsupervised, causal dependency-aware router that selects among a pool of LLMs for reasoning queries without human-annotated preference data.
Problem
Routing queries to the most suitable LLM in a diverse pool normally requires expensive human-annotated preference data for supervised routers, which are brittle under distribution shift.
Approach
Unsupervised routing that models causal dependencies to select models without supervised preference labels; full text unavailable so details beyond the abstract are not confirmed.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use ReasoningRouter instead of RandomRouter | routing over reasoning-model pool · 9-benchmark average (BGE-M3) | routing accuracy: 84.99 -> 88.20 | Improved |
| use ReasoningRouter instead of RandomRouter | routing over reasoning-model pool · 9-benchmark average (Embeddinggemma-300m) | routing accuracy: 84.99 -> 88.23 | Improved |
| use thinking-and-answer scoring instead of monolithic OutputRouter | ReasoningRouter · 9-benchmark average (BGE-M3) | routing accuracy: 87.81 -> 88.20 | Improved |
| use thinking-and-answer scoring instead of monolithic OutputRouter | ReasoningRouter · 9-benchmark average (Embeddinggemma-300m) | routing accuracy: 88.16 -> 88.23 | No clear effect |
| route on thinking and answer jointly instead of thinking only | ReasoningRouter · TruthfulQA (BGE-M3) | routing accuracy: 80.48 -> 81.98 | Improved |
| route on thinking and answer jointly instead of answer only | ReasoningRouter · TruthfulQA (BGE-M3) | routing accuracy: 76.58 -> 81.98 | Improved |
| use ReasoningRouter instead of monolithic OutputRouter | ReasoningRouter · MathQA (Embeddinggemma-300m) | routing accuracy: 93.27 -> 93.98 | Improved |
| vary proportion of thinking weight w_t versus answer weight w_a (ablation) | ReasoningRouter · nine benchmarks | routing accuracy | No clear effect |
| increase the number of models in the candidate pool (ablation) | ReasoningRouter · nine benchmarks (BGE-M3) | average accuracy | Improved |
Closest papers by meaning
- Confidence-Guided Stepwise Model Routing for Cost-Efficient Reasoning (AAAI 2026 · similarity 0.89)
- Lookahead Routing for Large Language Models (NeurIPS 2025 · similarity 0.88)
- CP-Router: An Uncertainty-Aware Router Between LLM and LRM (AAAI 2026 · similarity 0.88)
- Router-R1: Teaching LLMs Multi-Round Routing and Aggregation via Reinforcement Learning (NeurIPS 2025 · similarity 0.87)
- R2R: Efficiently Navigating Divergent Reasoning Paths with Small-Large Model Token Routing (NeurIPS 2025 · similarity 0.87)
- RouteLLM: Learning to Route LLMs from Preference Data (ICLR 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
