Chayan: Multi-Model LLM Router

This model is a high-performance LLM router presented in the paper RouterArena: An Open Platform for Comprehensive Comparison of LLM Routers.

• 📚 Paper (Hugging Face): RouterArena: An Open Platform for Comprehensive Comparison of LLM Routers
• 📚 Paper (arXiv): https://arxiv.org/abs/2510.00202
• 💻 Library Code: https://github.com/codelion/adaptive-classifier
• 🌐 RouterArena Project Page: https://routeworks.github.io/

Chayan intelligently routes between 4 models (gpt-4o-mini, gemini-2.5-flash-lite, gemini-2.5-flash, and gpt-4o) to optimize the accuracy-cost tradeoff.

🏆 RouterArena Performance

Official Leaderboard Results (8,400 queries):

• 🥇 #1 Optimal Accuracy Score: 88.7% - SOTA! (Best routing decision quality)
• 🥈 #2 Optimal Selection Score: 43.0% - Silver! (Second-best model selection)
• #7 Overall (#5 open-source): 64.9% accuracy, 63.8 arena score
• $0.60 per 1K queries - Cost-efficient routing

What do these metrics mean?

• Optimal Accuracy: When Chayan routes to a model, that model gives the correct answer 88.7% of the time
• Optimal Selection: Chayan selects the best available model 43% of the time

View full leaderboard: RouterArena | PR #24

Quick Start

pip install adaptive-classifier

from adaptive_classifier import AdaptiveClassifier

# Load router
router = AdaptiveClassifier.load("adaptive-classifier/chayan")

# Get routing decision
query = "What is the capital of France?"
predictions = router.predict(query, k=4)

# Route to top model
selected_model = predictions[0][0]  # e.g., "openai/gpt-4o-mini"

Recommended: Use with Calibration

# Apply calibration factors for best performance
calibration = {
    "openai/gpt-4o-mini": 0.9,
    "google/gemini-2.5-flash-lite": 1.5,
    "google/gemini-2.5-flash": 1.8,
    "openai/gpt-4o": 1.5
}

predictions = router.predict(query, k=4)
calibrated_scores = {model: score * calibration[model] for model, score in predictions}
selected_model = max(calibrated_scores.items(), key=lambda x: x[1])[0]

Architecture

Core Components:

• Base Model: BERT-base-uncased embeddings
• Classifier: Adaptive K-NN with prototype memory (FAISS-backed)
• Innovation: Calibrated confidence scores to correct training data imbalance

Supported Models:

How It Works

Training

• Dataset: RouterArena sub_10 (809 queries)
• Oracle Labels: 4-model cascade strategy (select cheapest successful model)
• Training Time: 19.2 minutes
• Method: K-NN classifier with 3000 prototypes, temperature 0.4

The Calibration Breakthrough

The uncalibrated router achieved 61.76% accuracy but was biased toward gpt-4o-mini (83% routing). This happened because the training data had class imbalance:

• 57% gpt-4o-mini examples
• 27% gpt-4o examples
• 12% gemini-flash-lite examples
• 4% gemini-flash examples

Solution: Apply post-training calibration factors to correct the bias without retraining.

Result: +7.29pp improvement (61.76% → 69.05% on sub_10 benchmark)

Performance Benchmarks

Sub_10 Benchmark (809 queries):

Key Insight: Chayan achieves 99% of perfect oracle performance at 57% lower cost.

Full Dataset (8,400 queries):

• Optimal Accuracy: 88.7% (🥇 #1)
• Optimal Selection: 43.0% (🥈 #2)
• Overall Accuracy: 64.9% (#7 overall, #5 open-source)
• Cost: $0.60/1K queries

Advanced Usage

Feature Augmentation

Chayan was trained with query features prepended as tokens:

from adaptive_classifier.complexity_features import augment_query_with_features

query = "What is 2+2?"
augmented = augment_query_with_features(query)
# Returns: "[LEN:12][WORDS:3][MATH:1][SENT:1][MC:0] What is 2+2?"

predictions = router.predict(augmented, k=4)

Limitations

• Calibration factors optimized on RouterArena sub_10; may require adjustment for other domains
• Requires the 4 specific models to be available via API
• Performance depends on query distribution similar to RouterArena benchmark
• Cost estimates assume ~500 tokens per query

Citation

@software{adaptive_classifier,
  title = {Adaptive Classifier: Dynamic Text Classification with Continuous Learning},
  author = {Sharma, Asankhaya},
  year = {2025},
  publisher = {GitHub},
  url = {https://github.com/codelion/adaptive-classifier}
}