Models invent their own filler tokens during RL training
code_star · x · 2026-11-13
codestar notes an amusing RL training behavior: models spontaneously discover and use their own special filler tokens. A short observation pointing to unexpected emergent patterns in RL training.
More from Fun
- The Ringer's satire on Jensen Huang's hypothetical Nvidia exit goes viral — GaryMarcus · 2026-11-13
- AI-generated 'not canon' clip of Schwarzenegger driving with Michael Jordan — alanskimp · 2026-11-13
- Nvidia Earnings Nov 17, Anthropic IPO Nov 18, GTA 6 Nov 19: A Make-or-Break Week — noahsolomon · 2026-11-13
- Microsoft Teams 'refuses to die, like a cockroach with a corporate benefits package' — _jaydeepkarale · 2026-11-13
- Fable 5.2 refuses to swap a character's legs in game-editing fail — ezshine · 2026-11-13
- Joke: Jensen Huang Is Still CEO Because His Termination Email Was Sent via Outlook — Kyrannio · 2026-11-13