LAND: Learning an Adaptive Number of Draws for Bayesian Best-of-N Selection
Inference-side entry: per-input sample counts instead of a fixed N.
AI Storyline
3 items · 2 sources · 3 days
Operational story trace
Learning Adaptive
Follow in this browser to see new updates on your Live feed.
Latest change
A Nov 9 theory paper shows adaptive quantization keeps TD learning's fast convergence rate even when an agent's updates cross a rate-limited channel.
Earlier contextThe story so far
This is not one developing event: three unrelated arXiv papers share the word "adaptive" and one idea, spending scarce resources where they help most. LAND (Nov 2) varies Best-of-N draws per input, and ASP (Nov 6) schedules RL rollouts to cut idle GPU time.
editor-curated · source-linked
Arc
Nov 2Nov 9 · now
INFERENCE · Nov 2LAND learns how many Best-of-N draws each input deserves1 source · show source ▾
TRAINING · Nov 6ASP schedules asynchronous RL rollouts to reduce underutilization1 source · show source ▾
Adaptive Scheduling Pipeline for Multi-Instance Asynchronous Reinforcement Learning
Training-side entry: rollout scheduling for asynchronous RL of LLMs.
THEORY · Nov 9Adaptive quantization preserves TD-learning rates over rate-limited links1 source · show source ▾
Adaptive Quantization Preserves Fast Rates for TD Learning over Rate-Limited Channels
Theory entry: convergence guarantee under a communication budget.
Day 1 Monday, Nov 2, 2026
LAND: Learning an Adaptive Number of Draws for Bayesian Best-of-N Selection
Inference-side entry: per-input sample counts instead of a fixed N.
Day 2 Friday, Nov 6, 2026
Adaptive Scheduling Pipeline for Multi-Instance Asynchronous Reinforcement Learning
Training-side entry: rollout scheduling for asynchronous RL of LLMs.
Day 3 Monday, Nov 9, 2026
Adaptive Quantization Preserves Fast Rates for TD Learning over Rate-Limited Channels
Theory entry: convergence guarantee under a communication budget.
What to watch — open questions
- Does LAND publish compute savings against fixed-N Best-of-N on a standard reward-model benchmark?
- Is ASP's scheduler released as code, or integrated into an open RL framework?
- Does the adaptive-quantization result extend beyond linear function approximation?
How this thread was built
editor wrote the arc · 3 beats
Storylines are threaded mechanically from the feed: stories that share a distinctive anchor across multiple days and sources. Each item links to its original source. The evidence trace, current state, and open questions are written by the editor routine and refreshed whenever a new beat lands.