How LSP Training Makes R2T2 Never Rewrite: Mechanism-to-Behavior Mapping
dr_cintas · x · 2026-11-06
drcintas maps R2T2's LSP (Longest Stable Prefix) training target to observable behavior: "never rewrites" comes from training the model to emit only text matching the full-context output's prefix; low latency at 160ms chunks comes from an explicit WAIT token that defers uncertain continuations; and the paper shows R2T2 narrows the stability-latency gap plaguing other streaming models. Figure 2's "I have two cats" example illustrates it: "I have" stabilizes early while "two" waits through the /tuː/ vs /tu/ ambiguity before committing.
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