Pith. sign in

REVIEW 16 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

arxiv 2401.01879 v3 pith:GRI5B4Q7 submitted 2024-01-03 cs.LG cs.CLcs.ITmath.IT

classification cs.LGcs.CLcs.ITmath.IT
keywords policybest-of-divergencealignmentreferenceupperboundrate
verification ladder T0 review T1 audit T2 compute T3 formal

Signed reviews

No signed human review yet.

0 comments
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 $n < 1000$.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 16 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. When May a Model Replace the Experiment? Audits, Licenses, and the Price of Trust in Surrogate-Driven Design

    cs.LG 2026-08 conditional novelty 7.0 of 10

    Safety in surrogate-driven design requires oracle-only certification; rank preservation is the exact criterion for oracle use, and selection-aware auditing is the cheapest way to certify it.

  2. ThinkRetrieve: Retrieval-Augmented Reasoning Traces for Test-Time Scaling

    cs.AI 2026-08 conditional novelty 6.0 of 10

    Per-step retrieval of solved exemplars injected into the reasoning trace improves test-time scaling accuracy, with up to 13.4 absolute points gained on AIME 2025.

  3. Multi-Turn On-Policy Distillation with Prefix Replay

    cs.LG 2026-07 conditional novelty 6.0 of 10

    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.

  4. Improving Large Vision and Language Models by Learning from a Panel of Peers

    cs.CV 2025-09 conditional novelty 6.0 of 10

    A panel of LVLMs that generate, evaluate, and learn from each other's outputs improves average benchmark scores by 9 points across 15 tasks.

  5. Does More Inference-Time Compute Really Help Robustness?

    cs.AI 2025-07 conditional novelty 6.0 of 10

    With exposed reasoning chains, increasing inference-time compute consistently decreases measured robustness across 12 open-source reasoning models, while hidden chains show improvements.

  6. Saffron-1: Safety Inference Scaling

    cs.LG 2025-06 conditional novelty 6.0 of 10

    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.

  7. Soft Best-of-n Sampling for Model Alignment

    cs.IT 2025-05 conditional novelty 6.0 of 10

    Soft Best-of-n sampling provably approaches the optimal tilted reward distribution at O(1/n) KL divergence and relative reward error, with sample complexity that grows exponentially in sequence length for blockwise sampling.

  8. InfAlign: Inference-aware language model alignment

    cs.LG 2024-12 conditional novelty 6.0 of 10

    Reward calibration plus a procedure-specific reward transformation lets RLHF optimize inference-time (best-of-N / worst-of-N) win rates better than standard RLHF, IPO, BoND, and BoNBoN.

  9. Safeguarding Text-to-Image Generation via Inference-Time Prompt-Noise Optimization

    cs.CV 2024-12 conditional novelty 6.0 of 10

    Prompt-Noise Optimization jointly tunes the prompt embedding and diffusion noise at inference time to suppress unsafe images while keeping outputs close to the prompt.

  10. Safe Inference-Time Alignment via Lagrangian Reward Augmentation

    cs.LG 2026-07 conditional novelty 5.5 of 10

    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.

  11. Refining Over Resampling: Test-Time Self-Correction for LLM Reasoning

    cs.AI 2026-08 conditional novelty 5.0 of 10

    Sampling multiple reasoning paths, refining each with self-critique and self-correction, then majority voting improves math reasoning accuracy over width-only or verifier-based test-time scaling on several open-weight LLMs.

  12. Test-time reward-guided alignment of language models by importance sampling on pre-logit space

    cs.LG 2025-10 conditional novelty 5.0 of 10

    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.

  13. Position: Machine Learning Conferences Should Establish a "Refutations and Critiques" Track

    cs.LG 2025-06 conditional novelty 5.0 of 10

    ML conferences should create an official peer-reviewed track dedicated to refuting and critiquing previously published work.

  14. DynScaling: Efficient Verifier-free Inference Scaling via Dynamic and Integrated Sampling

    cs.CL 2025-06 conditional novelty 5.0 of 10

    DynScaling improves verifier-free inference-time scaling by merging parallel and sequential sampling and allocating budget across queries with a UCB-based uncertainty rule.

  15. CoDe: Blockwise Control for Denoising Diffusion Models

    cs.CV 2025-02 conditional novelty 5.0 of 10

    CoDe applies blockwise best-of-N sampling during diffusion denoising, with Tweedie-based reward estimates, to align generated images to differentiable or non-differentiable rewards.

  16. LIAR: Leveraging Inference Time Alignment (Best-of-N) to Jailbreak LLMs in Seconds

    cs.CL 2024-12 reject novelty 4.0 of 10

    LIAR shows that best-of-N sampling of suffixes from a GPT-2 model jailbreaks several aligned LLMs with low-perplexity prompts and far faster time-to-attack than training-based attacks.

Pith tools