Pairing two papers: subliminal distillation baggage and RL's pass@k ceiling
andrew_n_carr · x · 2026-11-08
The author pairs two seemingly unrelated papers: Subliminal Learning shows a teacher model's traits (even misalignment) transfer to students via datasets of pure number sequences, and filtering doesn't stop it — distillation carries invisible baggage. Does RL Really Incentivize Reasoning Capacity Beyond the Base Model? shows RLVR models beat base models at small k (e.g., pass@1), but base models score higher pass@k at large k — RL improves sampling of already-reachable answers without expanding capability. Together: training may be selecting existing behavior while smuggling in unseen side effects.
More from Research
- GPT-5.2 Pro solves COLT 2022 open problem on accelerated L1-regularized PageRank — kfountou · 2026-11-08
- LDLM opens up the frozen encoder: text latent diffusion that trains its own latent space at NeurIPS 2026 — kastnerkyle · 2026-11-08
- Professor slams ICLR reviews from non-experts, calls for qualification stage at top AI conferences — xbresson · 2026-11-08
- Google AI Detects Hypertension From a Regular Smartwatch, Catching 65.8% vs 55.3% for Cuff — emmanuelvivier · 2026-11-08
- Impagliazzo's five worlds table resurfaces: which complexity universe do we live in? — matthew_d_green · 2026-11-08
- Lilly and Novo use aging clocks as trial readouts, hearts up to 4 years younger — omarabudayyeh · 2026-11-08