ARA Cognitives builds AI that solves hard problems through iterative self-modification — not brute-force scale. A 10M-parameter model that rivals billion-parameter baselines on mathematical reasoning.
ARA partitions every weight matrix into two structurally distinct zones. The 40% reasoning core is permanently frozen after pre-training — it encodes the invariant logic of mathematical thought. The 60% plastic workspace updates via gradient descent at every inference step, building temporary circuits tuned to the specific problem at hand.
4M parameters. Pre-trained on symbolic solution traces, then permanently frozen. Encodes 25+ mathematical primitives: distribute, substitute, differentiate, factorise, and more. Proposes operations and runs the coherence checker.
6M parameters. Updated by gradient descent on L_coherence at every micro-iteration — during inference, not training. Holds intermediate state, absorbs retrieved knowledge, and self-organises into a three-tier resistance hierarchy.
Per-neuron resistance scores R_i track which plastic weights have proven useful. Successful neurons are consolidated; neurons responsible for failed steps soften, enabling backtracking without forgetting hard-won progress.
When the core detects a knowledge gap, it issues structured library queries at increasing depth — from domain identification through edge-case retrieval. Retrieved content is absorbed via K_absorb=3–5 gradient steps into the live workspace.
The iterative architecture gives ARA structural advantages that raw parameter counts cannot buy. More iterations on harder problems. No depth limit. Knowledge decoupled from memory.
Within 7 points of billion-parameter baselines, at a fraction of the compute cost.
Competitive across all MATH categories. Projected to exceed 1B baselines on Level 5 hard problems.
570B FLOPs for 1,000 micro-iterations vs 1,030B FLOPs for a single 1B model forward pass.
Fits on any modern GPU or Apple Silicon. No datacenter required for inference.
Standard transformers are capped at L layers. ARA iterates for as long as the problem requires.
N_eff = N × T^0.5 × (1 − f_facts)⁻¹. 10M parameters punch at 775M equivalent capacity.
ARA solves problems through a structured loop — propose, check, update, adapt — until the solution converges or the budget is exhausted. Every step is mathematically grounded.
The frozen reasoning core proposes the next logical operation based on the current problem state and its pre-trained operation library.
theta_R · frozenThe proposed operation is applied to the current state, producing a candidate new state with updated symbolic assertions and variable bindings.
Symbolic engineThe frozen coherence checker validates consistency and measures progress toward the goal. Failed checks trigger backtracking with resistance softening.
Cons(s_t) · Prog(t)The plastic weights update via gradient descent on L_coherence. Active neurons gain resistance; failed neurons soften. The workspace self-organises.
theta_P · EWCWhen the core detects a knowledge gap, it queries an external library and absorbs the retrieved content through mini inference-time fine-tuning.
Progressive deepeningThe loop terminates when all subgoals are satisfied, the budget is exhausted, or confidence falls below threshold. Result is returned with full step trace.
h(s_t, s_goal) = 0ARA is not a metaphor. Every architectural decision — the 40/60 split, the coherence loss, the adaptive resistance — is derived from first principles and rendered as an equation.
ARA Cognitives was founded to answer a single question: can rigorous mathematical architecture make AI genuinely more capable — without requiring more power?
We're inviting builders who believe AI efficiency matters as much as capability.
Sign in to the Hub →The Shadow Demo is in active development. We're working with a small group of research partners and enterprise pilots. Apply to join the early access programme.