Anti-distillation classifiers flag GPT-5.6 grading of student code as suspected distillation
xuanalogue · x · 2026-09-09
The author suspects their most likely trigger was using GPT-5.6 to grade student code solutions against a rubric, feeding a small HMM that estimates student knowledge. Anti-distillation classifiers appear overzealous, treating any code grading as possible distillation — a notable false-positive problem in automated usage detection.
Related event: Teacher's GPT-5.6 code grading blocked by anti-distillation classifier(2 posts)→
More from Models
- Meta's Muse Spark 1.3 lands in Cursor, touted as cheaper than GPT 5.6 Sol — acbuller · 2026-09-09
- GPT-6 Astra via GitHub Copilot: faster than expected, priced 1x — same as Sonnet — revodavid · 2026-09-09
- Is GPT-6 Astra just autoregressive? Domingos and researcher spar over its architecture — eliebakouch · 2026-09-09
- Pedro Domingos: LLMs still can't add two numbers of arbitrary length — pmddomingos · 2026-09-09
- GPT-6 Astra is now available on AWS Bedrock — gabrielchua · 2026-09-09
- Dev dissects GPT-6 Astra's computer use loop, makes it 2.5x faster with Stagehand — TheMoonMidas · 2026-09-09