REVIEW 8 cited by
Theoretical guarantees on the best-of-n alignment policy
Not yet reviewed by Pith; the record is open.
This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.
SPECIMEN: schema-true, not a live event
T0 review · schema-true
One-sentence machine reading of the paper's core claim.
pith:XXXXXXXX · record.json · timestamp
abstract
A simple and effective method for the inference-time alignment and scaling test-time compute of generative models is best-of-$n$ sampling, where $n$ samples are drawn from a reference policy, ranked based on a reward function, and the highest ranking one is selected. A commonly used analytical expression in the literature claims that the KL divergence between the best-of-$n$ policy and the reference policy is equal to $\log (n) - (n-1)/n.$ We disprove the validity of this claim, and show that it is an upper bound on the actual KL divergence. We also explore the tightness of this upper bound in different regimes, and propose a new estimator for the KL divergence and empirically show that it provides a tight approximation. We also show that the win rate of the best-of-$n$ policy against the reference policy is upper bounded by $n/(n+1)$ and derive bounds on the tightness of this characterization. We conclude with analyzing the tradeoffs between win rate and KL divergence of the best-of-$n$ alignment policy, which demonstrate that very good tradeoffs are achievable with $n < 1000$.
Forward citations
Cited by 8 Pith papers
-
Multi-Turn On-Policy Distillation with Prefix Replay
ReOPD offline-distills multi-turn agentic LLMs via teacher-prefix replay plus step-decay sampling, matching online OPD accuracy at ≥4× speed with zero tool calls.
-
Improving Large Vision and Language Models by Learning from a Panel of Peers
A panel of LVLMs that generate, evaluate, and learn from each other's outputs improves average benchmark scores by 9 points across 15 tasks.
-
Does More Inference-Time Compute Really Help Robustness?
With exposed reasoning chains, increasing inference-time compute consistently decreases measured robustness across 12 open-source reasoning models, while hidden chains show improvements.
-
Saffron-1: Safety Inference Scaling
A multifurcation reward model that scores all next-token candidates in one call makes inference-time safety scaling with tree search far more compute-efficient than best-of-N sampling.
-
Safe Inference-Time Alignment via Lagrangian Reward Augmentation
Dualizing Safe RLHF yields a one-dimensional convex calibration of λ that defines a drop-in safety-aware reward for Best-of-N and token-level inference-time decoders.
-
Test-time reward-guided alignment of language models by importance sampling on pre-logit space
AISP is a training-free decode-time alignment method: Gaussian-perturb LLM pre-logits, score sampled responses with a reward model, and iteratively shift the perturbation mean by reward-weighted importance sampling.
-
Position: Machine Learning Conferences Should Establish a "Refutations and Critiques" Track
ML conferences should create an official peer-reviewed track dedicated to refuting and critiquing previously published work.
-
DynScaling: Efficient Verifier-free Inference Scaling via Dynamic and Integrated Sampling
DynScaling improves verifier-free inference-time scaling by merging parallel and sequential sampling and allocating budget across queries with a UCB-based uncertainty rule.
Discussion (0). Sign in to comment.