Replaying encrypted reasoning state into another model recovers the original task answer
Tophant_ · reddit · 2026-09-10
A Reddit experiment tested whether reasoning-state information can carry across models:
- The source model only output "READY," but the API response contained 13,958 reasoning tokens with encrypted reasoning state
- Replaying the intact state into a compatible model—without the original number—still recovered the same factorization: 8,154,241 = 839 × 9,719
- Removing the state stopped recovery; changing one character in the encrypted content broke the request
The authors are interested not in reconstructing chain-of-thought token by token, but in whether task information persists through replayed reasoning state, and ask how to design stronger controls.
More from Models
- DeepSeek's answer to surging demand: make its model cheaper and faster — yacineMTB · 2026-09-10
- Prelim assessment: GLM-4.1 behaves similar to V4, Zhipu's post-training called more advanced — menhguin · 2026-09-10
- Huge share of post-2022 web data is AI content mislabeled as human-written — menhguin · 2026-09-10
- DeepSeek V4.1 Hailed as the 'First Gamer Model' After Blowing Away an FPS-Generation Test — teortaxesTex · 2026-09-10
- Bug Hunt Bench grades frontier models on 105 real bugs; DeepSeek-V4.1-Flash lands 24/105 for $1.80 — PawelHuryn · 2026-09-10
- RSI is here, just disaggregated: DeepSeek using LLMs to design algorithms — teortaxesTex · 2026-09-10