Research map / Large language models / Reasoning and chain of thought

Models Under SCOPE: Scalable and Controllable Routing via Pre-hoc Reasoning

ICML 2026 · Large language models › Reasoning and chain of thought · approach: Efficient and adaptive thinking · 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 SCOPE, a pre-hoc routing framework that uses retrieval-augmented anchor fingerprints and an RL-trained (GRPO) reasoning estimator to predict each model's correctness and token cost, enabling generalization to unseen models and user-controllable accuracy-cost trade-offs.

Problem

Existing model routers treat routing as classification over a fixed set of models, so they cannot handle newly released models without retraining and offer no explicit, adjustable accuracy-cost control.

Approach

Instead of choosing from fixed model names, SCOPE retrieves a model's behavioral fingerprint (correctness and token cost on similar anchor queries) and uses a reasoning LLM to predict correctness and token length for each candidate model; a utility function with a preference coefficient alpha, plus anchor-based calibration, then picks the model.

Key results

• On the test set, SCOPE (alpha=1.0) reaches 75.00% average accuracy versus 57.94% for xRouter at $1.01 cost, and Graph Router's 73.02%.
• With alpha near 1, SCOPE improves accuracy by +25.7% over Claude-Sonnet-4.5 while cutting cost by 74% ($3.08 vs $11.86) on the OOD set.
• Reducing predictor reasoning from 2354.9 to 238.7 tokens cuts token consumption by 90.8% versus test-time scaling with a 7-model pool (1.8k vs 19.2k tokens per query).

Surprising findings

• Untuned Qwen3-4B cannot predict performance via prompting alone, so the training pipeline is necessary.
• Supervised baselines re-trained on the anchor set drop to near-random accuracy on OOD unseen models.

Limitations

• Routing gains on OOD set rely on a 250-query anchor set and costs measured via API pricing.
• Supervised baselines are retrained on the anchor set, so comparisons depend on adaptation choices.
• Cost and accuracy estimates for the utility depend on hand-set hyperparameters (gamma_base, beta, w_base).

Experimental setup

Datasets: Scope-60K MMLU-Pro R-Bench GAOKAO GPQA AIME Humanity's Last Exam SimpleQA OlympiadBench

Benchmarks: Scope-60K test split OOD set (AIME, HLE, SimpleQA, OlympiadBench queries)

Metrics: PGR average accuracy total cost MAE for token length accuracy

Baselines: SVM Router MLP Router KNN Router Graph Router xRouter Qwen3-235B-A22B Claude-Sonnet-4.5

Builds on: GRPO Router-R1 Avengers-Pro Qwen3-Embedding

Closest papers by meaning

• Router-R1: Teaching LLMs Multi-Round Routing and Aggregation via Reinforcement Learning (NeurIPS 2025 · similarity 0.89)
• 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)
• R2-Router: A New Paradigm for LLM Routing with Reasoning (ICML 2026 · similarity 0.88)
• A Unified Approach to Routing and Cascading for LLMs (ICML 2025 · similarity 0.88)
• RACER: Risk-Aware Calibrated Efficient Routing for Large Language Models (ICML 2026 · similarity 0.87)