feat: Add A3M Router - Parallel Multi-LLM with Ensemble Voting

[Das-rebel]

A3M Router Submission for RouterBench

Overview

A3M Router (Adaptive Memory Multi-Model Router) is an open-source LLM gateway featuring parallel multi-LLM execution with confidence-weighted ensemble voting - a fundamentally different approach from sequential fallback routers.

Reference: https://github.com/Das-rebel/a3m-router
NPM: npm install adaptive-memory-multi-model-router@2.14.58

Key Innovations

1. Parallel Multi-LLM Execution

Unlike all other routers (litellm, one-api, LibreChat) which use sequential fallback, A3M executes multiple providers in parallel and merges results via confidence-weighted voting.

2. Confidence-Weighted Shapley Value Credit Assignment

Mathematical framework for fair credit assignment:

φ_i* = 0.5·Shapley + 0.3·Ethnocentrism + 0.2·Handicap

Where:

• Shapley: Marginal contribution to ensemble
• Ethnocentrism: Historical trust between models
• Handicap: Cost-reliability product (Zahavi principle)

3. Thompson Sampling for Exploration/Exploitation

Uses Bayesian sampling to balance:

• Exploitation: Use known reliable models
• Exploration: Try new models that might be better

4. Complexity-Aware Routing

Estimates prompt complexity via:

• Word count analysis
• Reasoning keyword detection
• Technical content indicators

Routes to appropriate tier based on complexity + willingness-to-pay.

Benchmark Results

Metric	Value
Exact Tier Match	67%
±1 Tier Accuracy	96%
Cost Savings	62.9% vs all-premium
Robustness Score	0.8524

Implementation

• File: routers/a3m_router.py
• Extends: AbstractRouter
• Compatible with: RouterBench evaluation framework

Testing

Run with:

python evaluate_routers.py --config configs/evaluate_routers.yaml

Would appreciate review and feedback!

[Das-rebel]

Hey @withmartian! Checking in on this A3M Router submission.

Latest updates:

• RouterArena Add paper code #1: 96.77% accuracy, $0.077/1K cost, 1.0 robustness
• Parallel multi-LLM execution with result merging (unique approach)
• 20K+ npm downloads, MIT license

Let me know if you need any changes to the submission format!

Best,
Subho