Research map / Large language models / Reasoning and chain of thought
Anytime-Valid Inference for Online Ranking of Large Language Models
Runzhe Gu, Wenguang Sun, Bowen Gang, Xintao Xia
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes SERPANT, an anytime-valid sequential procedure for online pairwise-preference ranking of LLMs that remains statistically valid under adaptive sampling and continuous monitoring.
Problem
Online LLM evaluation from sequentially collected pairwise preferences: adaptive sampling and continuous monitoring invalidate classical fixed-sample inference, so existing ranking procedures are largely heuristic.
Approach
Applies anytime-valid sequential inference to online pairwise-preference ranking so that model rankings can be stopped and monitored at any time with valid error guarantees. Details beyond the abstract are not available.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add transitivity-based pruning (ablation) | SERPANT (random sampling) · Databricks, t=2000 | partial-order pairs discovered: 35 (78%) -> 41 (91%) | Improved |
| add transitivity-based pruning (ablation) | SERPANT (random sampling) · Databricks, t=1000 | partial-order pairs discovered: 24 (53%) -> 28 (62%) | Improved |
| add transitivity-based pruning (ablation) | SERPANT (random sampling) · Databricks, t=3000 | partial-order pairs discovered: 38 (84%) -> 43 (96%) | Improved |
| replace random pair sampling with tournament sampling | SERPANT · Databricks, t=500 | partial-order pairs discovered: 5 (11%) -> 12 (27%) | Improved |
| replace random pair sampling with tournament sampling | SERPANT · TriviaQA, t=2000 | partial-order pairs discovered: 23 (51%) -> 32 (71%) | Improved |
| replace random pair sampling with tournament sampling | SERPANT · SQuAD2.0, t=2000 | partial-order pairs discovered: 29 (64%) -> 30 (67%) | Improved |
| replace random pair sampling with tournament sampling | SERPANT · MMLU, t=2000 | partial-order pairs discovered: 33 (73%) -> 32 (71%) | No clear effect |
| replace random pair sampling with tournament sampling | SERPANT · Databricks, t=2000 | partial-order pairs discovered: 41 (91%) -> 41 (91%) | No clear effect |
| replace random pair sampling with tournament sampling | SERPANT · simulation, m=20, w1=0.2, w2=0.8 | power | Improved |
| replace random pair sampling with tournament sampling | SERPANT top-10 sets · simulation, m=20 | top-10 set size | Improved |
| replace empirical win-rate ranking with SERPANT e-process testing | ranking of 10 LLMs · Databricks, t=4000 | false discoveries: 4 (9%) -> 1 (2%) | Improved |
| replace empirical win-rate ranking with SERPANT e-process testing | ranking of 10 LLMs · Databricks, t=4000 | true discoveries: 41 (91%) -> 32 (71%) | Hurt |
Closest papers by meaning
- Prediction-Powered Ranking of Large Language Models (NeurIPS 2024 · similarity 0.88)
- Language Ranker: A Lightweight Ranking framework for LLM Decoding (NeurIPS 2025 · similarity 0.87)
- Peering Through Preferences: Unraveling Feedback Acquisition for Aligning Large Language Models (ICLR 2024 · similarity 0.86)
- From Uncertain Judgments to Calibrated Rankings: Conformal Elo Estimation for LLM Evaluation (NeurIPS 2026 · similarity 0.86)
- A Statistical Framework for Ranking LLM-based Chatbots (ICLR 2025 · similarity 0.86)
- Language Ranker: A Metric for Quantifying LLM Performance Across High and Low-Resource Languages (AAAI 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
