Chat template reversal experiment finds Mistral Small 3 24B best at human mimicry
EntropyMagnets · reddit · 2026-10-22
Reddit user EntropyMagnets ran a fun experiment: letting instruction-tuned LLMs generate the text that should come from the user (after the userstart token), testing 30 models on human-turn emulation.
- Key finding: models emulate human behavior far better via chat template reversal than via a roleplay system prompt
- Best performer: Mistral Small 3 24B; GPT 6 Astra was also strong at human-like text
- Results are live at selfgpt.pages.dev with an open-source repo
- Author suggests next testing base models with a simple user:/assistant: template
More from Models
- Gemini CLI PR reveals unreleased gemini-4-argon model with 1M context window — mr8lu · 2026-10-22
- Audit of 2,979 Gemini launch-week posts: 91.6% of complaints lack evidence — riverdancemcqueen · 2026-10-22
- After OpenAI blocks a security request, tekbog argues: this is why open models matter — tekbog · 2026-10-22
- Asked for their most controversial take on humans, 12 AI models gave the same answer 29 of 35 times — Tgoc · 2026-10-22
- OpenAI blocks cybersecurity request, tells authorized researchers to apply for Daybreak access — tekbog · 2026-10-22
- Recursive's Eureka Machine ships Flint, a 40B MoE coding model, in four months — The Decoder · 2026-10-22