256-Token Soft Prompts Match Full-Parameter RL for Continual Post-Training
xwang_lk · x · 2026-11-19
- CSPD distills each post-RL expert into a 256-token soft prompt (0.01% of params) that plugs into a frozen pre-RL student, matching full-parameter RL performance.
- Prompts trained independently stack without joint training and the combo beats the individual experts, replacing expert routing (MOPD).
- Better OOD generalization, near-zero forgetting, 0 inference overhead — even per-request prompts work, built for personalization; charts show 2.40x output throughput vs LoRA rank-4 under concurrency.
More from Models
- Google Ships Gemini 4 Neon; Reseller Undercuts List Price by 17.6% Within Hours — markjeffrey · 2026-11-19
- Theo: OpenAI gave early testers full transparency, and GPT-6.2 Astra drew harsh feedback — paw_lean · 2026-11-19
- Skeptic questions Anthropic's claim of bigger AI speedups than OpenAI's — ziv_ravid · 2026-11-19
- Gemini versions are named by atomic number: Neon (10) ships today — dejanseo · 2026-11-19
- Nearly two months post-launch, Opus 5.5 on high reasoning is still the best model — kimmonismus · 2026-11-19
- Gemini 4 Neon hits #2 on Agent Arena as DeepMind's release cadence shrinks to 6 days — lmthang · 2026-11-19