Measure the raw model
Benchmark reasoning, coding, memory, planning, tool use, uncertainty handling, persistence and action efficiency before adding architecture.
AX LABS · APPLIED INTELLIGENCE SYSTEMS
Model capability is not the same as agent capability.
We research whether automatically synthesized cognitive architecture—memory, perception, planning, verification, tool use and execution policy—can unlock materially more capability from fixed-weight models.
Research thesis
The model is only one component.
A model can fail because of weak memory, poor state representation, inefficient exploration, bad tool policy, lossy context handling or an unsuitable execution loop. AX treats those as compilable system components rather than fixed properties of the weights.
Generate the harness
Compile model-specific perception, memory, planner, verifier, context and execution policies from observed weaknesses and strengths.
Externally verify improvement
Candidate harnesses compete against the raw model and known-good genomes. Regressions are rejected; improvements earn promotion.
Measured evidence
Separate claims from ambition.
We publish evidence in layers: raw-model baseline, same model plus harness, same model plus tools, and—only when separately tested—learned adapters. We do not collapse those into one number.
Llama 3.2 1B: 0/4 → 4/4
On a four-case synthetic verification set, the same Cloudflare-hosted Llama 3.2 1B model moved from 0% raw baseline to 100% with the compiled AX harness. This is an internal controlled result, not an ARC-AGI-3 score.
Harness architecture can dominate outcome
ARC Prize reports GPT-6 Astra High at 54.82% under the Standard ARC-AGI-3 harness and 99.95% under its Provider Adapter harness—demonstrating that surrounding state and context architecture can materially alter effective agent performance.
Current programs
From harness uplift to general agent architecture.
Model-to-Agent Compiler
Given a fixed model M, synthesize HM: a model-specific architecture that maximizes generalizable interactive capability.
ARC-AGI-3 Meta-Harness
Build and recursively improve action-efficient agents without an arbitrary AX action budget, while respecting official environment scoring and hidden-set generalization.
Harness Distillation
Distill externally verified synthetic harness behavior into private LoRA/QLoRA adapters, keeping system-level and weight-level gains separately measurable.
Harness-Driven Capability Amplification in Fixed-Weight Models
We propose a model-to-agent compiler that characterizes a base model, synthesizes a fitted cognitive architecture, evaluates the resulting agent against immutable external tests, and recursively mutates the harness while keeping the evaluator outside the candidate.
Open the paper →