June 2026 · Proprietary Technology

TernaT

First learned neural reasoner on Vector-Symbolic Architecture with ternary weights
90% exact multi-hop QA. No LLM. No GPU. No FPU. 16 KB ternary resonator.
90% exact 100% 1-hop 100% 2-hop 16 KB model 0.15ms inference TernaT DOI
Read More GitHub
90%
Exact Multi-Hop
16 KB
Ternary Resonator
100%
1 & 2-Hop Accuracy
<100 KB
Total Pipeline

Abstract

We present TernaT, the first learned neural reasoner over Vector-Symbolic Architecture (VSA) using discrete ternary weights {−1, 0, +1}. Our system combines a 16 KB ternary GraphKANResonator, a <1 KB FastController, and a <1 KB ChainScorer to achieve 90% exact multi-hop QA on a synthetic benchmark of 96 facts and 30 queries — without any LLM, GPU, or floating-point hardware.

The complete pipeline fits in <100 KB, deployable on $0.50 microcontrollers. Key insight: VSA superposition noise (~80% ceiling) is resolved via a hybrid exact+VSA memory architecture, while negative training eliminates false positives in the learned resonator.

Pipeline Architecture

Question NL Parser (12 patterns) entity + predicate Hybrid Memory: exact dict + VSA fallback PredicateShardedStore VSA memory, predicate-sharded reduces superposition noise GraphKANResonator (16 KB ternary) FastController (<1 KB) + ChainScorer Beam search (width 1-3) Answer (entity)
TernaT Pipeline: hybrid exact+VSA memory with learned ternary components

Multi-Hop QA Results

Method Overall 1-hop 2-hop 3-hop VSA direct query 30% 90% 0% 0% Resonator only 30% 90% 0% 0% Controller + Resonator 73% 80% 60% 80% TernaT (full pipeline) 90% 100% 100% 70% Benchmark: 96 facts, 30 queries, D=1024. 3 failures all from controller confusion. 30% 30% 73% 90% VSA direct Resonator Ctrl+Res TernaT ✓
Full system achieves 90% exact match — 100% 1-hop, 100% 2-hop, 70% 3-hop

Component Sizes

ComponentParametersSizeDeployment
GraphKANResonator65,536 ternary16 KBAny MCU
FastController37,384 float<1 KBAny MCU
ChainScorer~50,000 float<1 KBAny MCU
VSA Memory (96 facts)D=1024~44 KBCortex-M4+
Total pipeline<100 KB$0.50 MCUs

Key Breakthroughs

1. Learned Neural VSA Resonator First learned cleanup for VSA queries Ternary weights: 16 KB, 1.1ms inference replaces O(n) brute-force (77s → 1.1ms) 2. Negative Training Eliminates FPs Training with contrastive examples FP rate: 1.3% → 0% eliminates false positives entirely 3. Multi-hop Without LLM Controller + ChainScorer + beam search 90% exact — first VSA + learned components 1-hop 100%, 2-hop 100%, 3-hop 70% 4. VSA Superposition Noise Solved ~80% ceiling proven independent of dim Hybrid exact+VSA store: 80% → 90% verified across VSA dimensions
Four fundamental contributions — all fitting in <100 KB

Comparison with Prior Art

Method Learned? Ternary? Year Kanerva (VSA) ✗ ✗ 1988 Plate (HRR) ✗ ✗ 2003 Frady (Resonators) ✗ ✗ 2021 TernaT (ours) ✓ ✓ 2026 First learned neural VSA reasoner — all prior work is algorithmic, not learned

Hardware Efficiency

Total Pipeline: <100 KB Resonator 16 KB Controller <1 KB ChainScorer <1 KB VSA Memory 44 KB Deployment: Cortex-M0+ ($0.50) · ESP32-S3 · RISC-V GD32V · Smartwatch DSP · Arduino RP2040
Combined with GraphKAN for perception: 38 KB total for a complete AI pipeline

Author

YV

Yuri Venediktov (Fakeonomics)

Independent researcher, 17 years old. Invented ternary KAN and VSA multi-hop reasoning.

github.com/Fakeonomics

June 2026. Proprietary technology — All Rights Reserved.