Adaptive eXecution with Intelligent Operations Memory
A sparse dynamic routing architecture for cost-efficient LLM inference.
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.
| 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) |
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
# Build
cargo build --release
# Run benchmark (trains + evaluates)
cargo run --release --bin axiom-bench
# Run tests
cargo test --workspaceaxiom-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)
The full paper is available as axiom_paper.docx and on Zenodo.