TES hot path (int8, STM32H743-class)

Research release for the paper "Draft, Verify, Write" (Deployable AI workshop, AAAI-27, under review). The hot path is the draft stage of the paper. A larger model on the SD card verifies the answer. The verified answer is then written into a 32 KB episode. The hot path is a MobileNetV2-0.5 stem (160 px, SAME padding), a hashed text key and a fusion block (SAF), distilled without labels from SmolVLM-256M.

Files

  • int8/seed{0,1,2}-hotpath-int8.tflite: full int8 graph, 796,800 B each. TFLM arena 495,948 B.
  • int8/tes_stem_int8.tflite, int8/tes_fusion_int8.tflite: split graphs (seed 0).
  • fp32/seed{0,1,2}-hotpath_best.pt: PyTorch fp32 checkpoints.
  • episodes/episode_table_seed*_32KB.json: 32 KB episode tables (per context).

Results (3 seeds, measured)

ScienceQA test accuracy 0.671 (teacher 0.661). COCO-presence val 0.651. POPE test 0.605 (teacher 0.720). int8 export keeps ScienceQA accuracy (0.657 fp32, 0.657 int8) with 0.94 argmax agreement on the check set.

Limits

  • Instruction counts come from Renode and QEMU (164,073,092 per answer). No hardware measurement exists.
  • The int8 numbers come from the TFLite reference kernels on x86.
  • The int8 graphs in this release run with a near-null episode.
  • The hot path is near chance on topics unseen in training. A label-free 32 KB episode raises coverage at selective accuracy 0.90 from 0.000 to 0.232 on held-out topics (mean of 9 cells). A static fine-tuned student with 66,433 trainable parameters reaches 0.176 at best, so the episode gain over that baseline is small (+5.6 points, CI +2.5 to +10.2). No number comes from hardware.
  • Teacher: SmolVLM-256M.
Downloads last month
140
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support