Validation Summary
Last updated: May 2026
This page lists public validation results for the current Qriton HLM documentation release. It is intentionally limited to benchmark-level evidence; reproduction material and partner-specific technical packages are handled through Early Access or commercial engagements.
Public Results
| Area | Evaluation | Result | Status |
|---|---|---|---|
| HLM3-Mix K=16 | WikiText-103 language modelling | val PPL 10.66 | validated |
| HLM3-Mix K=4 | WikiText-103 language modelling | val PPL 11.44 | validated |
| HLM3 Base + FFN | WikiText-103 non-mixer baseline | val PPL 48.3 | baseline |
| HLM-Vision-Mix Large | CIFAR-10 image classification | 83.91% val accuracy | demo-ready |
| HLM-Vision-Mix Small | CIFAR-10 image classification | 78.63% val accuracy | demo-ready |
| HLM-TTS-Mix | LJSpeech text-to-speech | val_loss 0.3845 | demo-ready |
| HLM-Micro HAR v1 | UCI HAR human activity recognition | 93.52% test accuracy | early access |
| HLM-Micro Anomaly v1 | Multimodal industrial anomaly | 99.00% val accuracy (synthetic) | early access |
| HLM-Micro ECG v1 | ECG anomaly research artifact | 83.62% val accuracy (synthetic) | research |
| HLM-Micro Gesture v1 | IMU gesture recognition | 100% (synthetic ceiling) | early access |
Architecture Principle
The public validation pattern is:
modality frontend + mixer + readout + shared Hopfield coreLanguage uses causal token mixing. Vision uses a convolutional frontend with a 2D spatial mixer and pooled readout. Audio and sensor time-series use temporal mixing, with task-appropriate augmentation when the deployment data calls for it. The shared piece is the polynomial Hopfield core and its energy-landscape semantics, not a single universal input frontend.
The figure uses a shared template, but the frontend remains modality-shaped: language, vision, audio, spatial, time-series, and robotics inputs do not enter the Hopfield core through the same raw representation.
In Progress
Speech recognition, larger language checkpoints, code modelling, time-series benchmarks, and multimodal fusion are active workstreams. They are not presented as public headline results until matched validation runs are complete.
Access
Validated demo checkpoints, model cards, and reproduction material are available through the Early Access program or a commercial engagement.