Skip to content

olliverc1985/AXIOM

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

43 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

AXIOM

Adaptive eXecution with Intelligent Operations Memory

A sparse dynamic routing architecture for cost-efficient LLM inference.

DOI

AXIOM routes incoming queries across three model tiers — Surface, Reasoning, and Deep — using a 128-dimensional structural encoder and a hierarchical resolver with dynamic coalition formation and non-local graph communication. No preference data, no GPU, no ML frameworks. Pure Rust.

Results

Metric Value
Validation accuracy 89.5% (105 queries)
Simple routing accuracy 95.0%
Adversarial accuracy 65.0% (40 queries)
Cost savings vs all-Opus 58.1%
Routing latency 1,311 μs
Parameters 1,205,376
Training time 3.4 minutes (Apple Silicon, no GPU)

Architecture

AXIOM's sparse computation graph supports four traversal directions — forward, lateral, feedback, and temporal — enabling non-local communication between routing nodes. No existing LLM router provides this.

RouteLLM:  Input → [BERT] → score → model selection
FrugalGPT: Input → [Model1] → score → maybe [Model2]
AXIOM:     Input → [Surface1] ←lateral→ [Surface2] → (conditional edge) →
           [Reasoning3] ←coalition→ [Deep6] → (feedback signal upward)
           with temporal_buffer blending throughout

Quick start

# Build
cargo build --release

# Run benchmark (trains + evaluates)
cargo run --release --bin axiom-bench

# Run tests
cargo test --workspace

Project structure

axiom-core/     Core library — encoder, graph, resolver, cache
axiom-bench/    Benchmark binary — training, validation, adversarial eval
axiom-tuner/    Auto-tuner — adjusts thresholds between runs
axiom-datasets/ Corpus data (simple/moderate/complex sentences)

Paper

The full paper is available as axiom_paper.docx and on Zenodo.

License

MIT

About

No description, website, or topics provided.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors