README.md

---
license: apache-2.0
tags:
  - llm-routing
  - model-selection
  - budget-optimization
  - knn
language:
  - en
library_name: sklearn
pipeline_tag: text-classification
---

# R2-Router: LLM Router with Joint Model-Budget Optimization

**R2-Router** intelligently routes each query to the optimal (LLM, token budget) pair, jointly optimizing accuracy and inference cost. Ranked **#1** on the [RouterArena](https://routerarena.github.io/) leaderboard.

**Paper**: [R2-Router (arxiv)](https://arxiv.org/abs/TODO)

## RouterArena Performance

Official leaderboard results on 8,400 queries:

| Metric | Value |
|--------|-------|
| Accuracy | 71.23% |
| Cost per 1K Queries | $0.061 |
| Arena Score (beta=0.1) | **71.60** |
| Robustness Score | 45.71% |
| Rank | **#1** |

## Quick Start

### Installation

```bash
pip install scikit-learn numpy joblib huggingface_hub sentence-transformers
```

### Complete Example

```python
from huggingface_hub import snapshot_download
from sentence_transformers import SentenceTransformer
import sys

# 1. Download router
path = snapshot_download("JiaqiXue/r2-router")
sys.path.insert(0, path)

from router import R2Router

# 2. Load pre-trained KNN checkpoints
router = R2Router.from_pretrained(path)

# 3. Embed your query with Qwen3-0.6B (1024-dim)
embedder = SentenceTransformer("Qwen/Qwen3-0.6B")
embedding = embedder.encode("What is the capital of France?")

# 4. Route!
result = router.route(embedding)
print(f"Model: {result['model_full_name']}")
print(f"Token Budget: {result['token_limit']}")
print(f"Predicted Quality: {result['predicted_quality']:.3f}")
```

### Train from Scratch

```python
from huggingface_hub import snapshot_download
import sys

path = snapshot_download("JiaqiXue/r2-router")
sys.path.insert(0, path)

from router import R2Router

# Train KNN from the provided sub_10 training data (custom hyperparameters)
router = R2Router.from_training_data(path, k=80)
```

### Alternative: vLLM Embeddings (Faster for Batches)

```python
from vllm import LLM
llm = LLM(model="Qwen/Qwen3-0.6B", runner="pooling")
outputs = llm.embed(["What is the capital of France?"])
embedding = outputs[0].outputs.embedding
```

Or with vLLM for faster batch inference:

```python
from vllm import LLM
llm = LLM(model="Qwen/Qwen3-0.6B", runner="pooling")
outputs = llm.embed(["What is the capital of France?"])
embedding = outputs[0].outputs.embedding
```

## Architecture

R2-Router jointly optimizes **which model** to use and **how many tokens** to allocate per query.

### Routing Formula

```
risk(M, b) = (1 - lambda) * predicted_quality(query, M, b) - lambda * predicted_tokens(query, M) * price_M / 1e6
(M*, b*) = argmax risk
```

### Pipeline

```
Input Query
    |
[1] Embed with Qwen3-0.6B -> 1024-dim vector
    |
[2] For each (model, budget) pair:
      - KNN predicts quality (accuracy)
      - KNN predicts output token count
      - Compute risk = (1-lambda) * quality - lambda * cost
    |
[3] Select (model, budget) with highest risk
    |
Output: (model_name, token_budget)
```

### Model Pool (6 LLMs)

| Model | Output $/M tokens |
|-------|------------------|
| Qwen3-235B-A22B | $0.463 |
| Qwen3-Next-80B-A3B | $1.10 |
| Qwen3-30B-A3B | $0.33 |
| Qwen3-Coder-Next | $0.30 |
| Gemini 2.5 Flash | $2.50 |
| Claude 3 Haiku | $1.25 |

### Token Budgets

4 output token limits: **100, 200, 400, 800** tokens.

### Key Parameters

| Parameter | Value |
|-----------|-------|
| KNN K | 80 |
| Lambda | 0.999 |
| Distance Metric | Cosine |
| KNN Weights | Distance-weighted |
| Embedding Dim | 1024 |

## Repository Contents

```
config.json             # Router configuration (models, budgets, prices, hyperparams)
router.py               # Self-contained inference code
training_data/
  embeddings.npy        # Sub_10 training embeddings (809 x 1024)
  labels.json           # Per-(model, budget) accuracy & token labels
checkpoints/
  quality_knn_*.joblib  # Pre-fitted KNN quality predictors (18 total)
  token_knn_*.joblib    # Pre-fitted KNN token predictors (6 total)
```

### Two Ways to Use

1. **Load checkpoints** (`from_pretrained`): Directly load pre-fitted KNN models. No training needed.
2. **Train from data** (`from_training_data`): Use the provided training embeddings and labels to fit your own KNN with custom hyperparameters (e.g., different K, distance metric).

## Training Details

- **Training Data**: RouterArena sub_10 split (809 queries, 10% of full 8,400)
- **Method**: KNeighborsRegressor with cosine distance, distance-weighted
- **Evaluation**: Full 8,400 RouterArena queries (no data leakage)
- **Training Time**: < 1 second (KNN fitting)

## Citation

```bibtex
@article{r2router2026,
  title={R2-Router: A New Paradigm for LLM Routing with Reasoning},
  author={TODO},
  year={2026},
  url={https://arxiv.org/abs/TODO}
}
```

## License

Apache 2.0