Skip to content
TOPOVOID

Whitepaper

Customizing models with Reflecho

Results on three classification tasks: email filing, scanned documents and satellite images.

Base model
Qwen3.8-27B-FP8
Updated

Test-set accuracy

  1. Email filing

    Enron · 611 test emails · 8 folders

    Base model, zero-shot
    65.8%
    Reflecho
    98.5%

  2. Scanned documents

    RVL-CDIP · 960 test images · 16 classes

    Base model, zero-shot
    64.2%
    Reflecho
    97.6%

  3. Satellite images

    EuroSAT · 1,000 test images · 10 classes

    Base model, zero-shot
    64.7%
    Reflecho
    98.3%

Results

Task Base model, zero-shot Prompt examples Reflecho LoRA Other baseline
Email filing 65.8% 80.7% 98.5% 90.5% TF-IDF + logistic regression: 89.7%
Scanned documents 64.2% 71.6% 97.6% 82.5% Linear probe on vision features: 73.0%
Satellite images 64.7% 70.4% 98.3% 90.7% Linear probe on vision features: 94.5%

With 10 labeled images per class, Reflecho already reaches 75.2% on scanned documents and 81.8% on satellite images.

How it works

Reflecho is our post-training representation engineering method for adapting an existing foundation model to your task.

  1. You provide labeled examples from your own work, such as filed emails or sorted documents.
  2. Reflecho measures the model’s hidden-state geometry and fits a lightweight representation intervention. The base model’s weights stay unchanged.
  3. The adapted model answers with your labels.

About these numbers

  • Every method uses the same base model, Qwen3.8-27B-FP8, and the same held-out test set for each task.
  • Zero-shot: the model receives the task instruction and the label names. Prompt examples: labeled examples are added to the prompt, 24 emails or 1 image per class.
  • Reflecho and LoRA are trained on 1,583 emails, or 100 images per class.
  • All results come from offline evaluation on public datasets.

From post-training representation engineering to the TopoVoid model

Reflecho adapts existing foundation models after pretraining without modifying their weights. To remove the entanglement between syntax, tone, and factual memory present inside existing residual streams, we also designed and pretrained the TopoVoid model from random initialization. Its architecture partitions the hidden space into mutually orthogonal regions for general reasoning operators, mid-layer concept bottlenecks that control tone and reasoning depth without system prompts, and a null-space memory chamber that writes new facts and long-conversation memory into weights in seconds via closed-form algebra.

Try Reflecho on your task

We can run the same comparison on your labels and data. Email hello@topovoid.com.