R2T2's Real Contribution Is a Learning Target: Formalizing Longest Stable Prefix
dr_cintas · x · 2026-11-06
drcintas argues R2T2's real contribution isn't architecture but a learning target: LSP (Longest Stable Prefix) — at each time step t, find the longest text prefix the full-context model has committed to, used as the training target, with an explicit WAIT action teaching the model to defer uncertain continuations. Key theoretical result: LSP is monotone (never shrinks), consistent with full-audio output, and maximal among valid labels. As the author puts it, this is a formal definition of what "stable" even means, not just "train with wait tokens".
More from Research
- Unverified claims: frontier labs allegedly gaming ICLR reviews with malicious reviews — examachine · 2026-11-07
- SSF teaches robots to reason about likely hiding spots when searching for hidden objects — _krishna_murthy · 2026-11-07
- 'Die getting Schmidhubered or live long enough to become one': scaling law priority drama — teortaxesTex · 2026-11-07
- Stanford and Together AI propose Next-Concept Modeling to replace next-token prediction — bendee983 · 2026-11-07
- New benchmark shows some LLMs refuse over 50% of legitimate biology queries — nlarusstone · 2026-11-07
- From AI-designed phage DNA to patient care: the ecosystem behind the model — exnx · 2026-11-07