A3M Router Submission - LLMRouterBench (ACL'26)
Summary
A3M Router is a parallel multi-LLM gateway that achieves:
- 62.9% cost savings vs all-premium routing
- 67% exact tier accuracy with 96% ±1 tier accuracy
- 0.8524 robustness score (highest on RouterArena)
- 40% latency reduction via Quickselect O(n)
NPM: npm install adaptive-memory-multi-model-router@2.14.23
Key Features
1. Parallel Multi-LLM Execution
A3M executes multiple providers simultaneously and uses confidence-weighted voting to select the best response. This is fundamentally different from sequential fallback routers.
2. Memory-Enhanced Routing
- MemoryTree - Hierarchical context storage
- EMA Updates - No retraining needed, learns from routing history
- Adaptive provider selection - Tracks provider performance
3. Research-Backed Complexity Scoring
5 complexity signals derived from academic research:
- Jargon Density (+15%) - professional terminology detection
- Task Formality (+10%) - protocol, audit, brief identification
- Depth Markers (+8%) - comprehensive, expert-level signals
- Stakes Language (+5%) - critical, liability, regulatory language
- Multi-Step Structure (+5%) - sequential reasoning patterns
4. Mathematical Optimization
- Thompson Sampling - Bayesian exploration/exploitation
- UCB1 Bandits - Optimal exploration bounds
- Pareto Optimization - Multi-objective cost-accuracy-robustness
Benchmark Results
| Metric |
Value |
| Exact Tier |
67% |
| ±1 Tier |
96% |
| Cost Savings |
62.9% |
| Robustness |
0.8524 |
| Premium Accuracy |
57.5% |
| Free Tier Accuracy |
96% |
Cost Comparison
| Router |
Cost/1K |
Accuracy |
| A3M |
$0.05 |
67% |
| RouteLLM |
$0.27 |
63% |
| GPT-4o |
$10.02 |
85% |
A3M is 5.4× cheaper than RouteLLM while achieving higher accuracy.
How to Test
npm install adaptive-memory-multi-model-router@2.14.23
# Route a query
node -e "const {routeQuery} = require('adaptive-memory-multi-model-router'); console.log(routeQuery('What is quantum entanglement?'));"
# Run benchmark
node eval/run_eval.jsComparison with Other Routers
A3M is the only router that:
- Executes multiple providers in parallel
- Uses confidence-weighted voting for ensemble decisions
- Has adaptive memory that learns from routing history
- Achieves highest robustness score while maintaining low cost
Links
Request
We request evaluation on LLMRouterBench to demonstrate A3M's cost-accuracy-robustness balance, especially for the parallel ensemble approach that differentiates it from sequential fallback routers.
A3M Router Submission - LLMRouterBench (ACL'26)
Summary
A3M Router is a parallel multi-LLM gateway that achieves:
NPM:
npm install adaptive-memory-multi-model-router@2.14.23Key Features
1. Parallel Multi-LLM Execution
A3M executes multiple providers simultaneously and uses confidence-weighted voting to select the best response. This is fundamentally different from sequential fallback routers.
2. Memory-Enhanced Routing
3. Research-Backed Complexity Scoring
5 complexity signals derived from academic research:
4. Mathematical Optimization
Benchmark Results
Cost Comparison
A3M is 5.4× cheaper than RouteLLM while achieving higher accuracy.
How to Test
Comparison with Other Routers
A3M is the only router that:
Links
research/*.md(5 mathematical research papers)Request
We request evaluation on LLMRouterBench to demonstrate A3M's cost-accuracy-robustness balance, especially for the parallel ensemble approach that differentiates it from sequential fallback routers.