Research map / Large language models / Reasoning and chain of thought
RADAR: Reasoning–Ability and Difficulty-Aware Routing in Language Models
Nigel Fernandez, Branislav Kveton, Ryan A. Rossi, Andrew Lan, Zichao Wang
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 RADAR, a difficulty-aware router that chooses both model size and reasoning budget for reasoning language models to trade off performance against cost.
Problem
Choosing a reasoning language model for deployment involves a performance-cost tradeoff across model size and reasoning budget, where larger models and larger budgets improve accuracy but increase cost and latency.
Approach
Routes each query jointly over model size and reasoning budget using reasoning-ability and difficulty awareness, rather than choosing a single model or fixed budget. Details beyond the abstract are not stated.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| route with RADAR instead of IRT-Router | IRT-Router · GPQA-Diamond in-distribution | hypervolume: 0.6942 -> 0.7513 | Improved |
| route with RADAR instead of IRT-Router | IRT-Router · FRAMES out-of-distribution | hypervolume: 0.8548 -> 0.8865 | Improved |
| route with RADAR instead of IRT-Router | IRT-Router · AIME out-of-distribution | hypervolume: 0.7915 -> 0.7566 | Hurt |
| use RADAR to match 90% of o4-mini-high performance | o4-mini high budget · MATH-500 | cost fraction: 1.31% of cost | Improved |
| add Qwen3-14B configurations via adaptive testing (RADAR++) | RADAR · FRAMES in-distribution | hypervolume: 0.8762 -> 0.8830 | Improved |
| replace linear difficulty transform with two-layer MLP (ablation) | RADAR · eight benchmarks | hypervolume: GPQA-Diamond 0.7513 -> 0.7308; FRAMES 0.8762 -> 0.8777 | No clear effect |
| use Chebyshev instead of linear scalarization (ablation) | RADAR · GPQA-Diamond out-of-distribution | CPT(90%): 29.94% -> 17.60% | Improved |
| use Chebyshev instead of linear scalarization (ablation) | RADAR · in-distribution benchmarks | hypervolume | No clear effect |
| train on only 20% of subsampled training queries (ablation) | RADAR · eight benchmarks | hypervolume | No clear effect |
| expose 5% of AIME and MATH queries in training (ablation) | RADAR · AIME out-of-distribution | hypervolume: 0.7566 -> 0.7665 (30%: 0.7787) | Improved |
Closest papers by meaning
- RADAR: Benchmarking Language Models on Imperfect Tabular Data (NeurIPS 2025 · similarity 0.90)
- Confidence-Guided Stepwise Model Routing for Cost-Efficient Reasoning (AAAI 2026 · similarity 0.89)
- Plan and Budget: Effective and Efficient Test-Time Scaling on Reasoning Large Language Models (ICLR 2026 · similarity 0.88)
- Reasoning Is Not Free: Robust Adaptive Cost-Efficient Routing for LLM-as-a-Judge (ICML 2026 · similarity 0.88)
- Training Language Models to Reason Efficiently (NeurIPS 2025 · similarity 0.87)
- Reasoning-as-Logic-Units: Scaling Test-Time Reasoning in Large Language Models Through Logic Unit Alignment (ICML 2025 · similarity 0.86)
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.
