arXiv AI
3d ago

Leaderboard Hacking: Preference-Based Model Evaluations are Vulnerable to Manipulation

The paper "Leaderboard Hacking: Preference-Based Model Evaluations are Vulnerable to Manipulation" examines how arena-style rankings of language models, which rely on pairwise comparisons and the Bradley‑Terry model, can be manipulated. It identifies three attack vectors—strategic voting, strategic prompting, and strategic nomination—that can each boost a model’s rank by 15‑20 places, and combined attacks can improve a rank by up to 45 places. The authors analyze the conditions that enable these vulnerabilities and propose safeguards for any preference‑based evaluation that permits open prompting, voting, or model submission.

By Eve Fleisig, Joachim Baumann, Dylan Hadfield-Menell, Dirk Hovy
arXiv AI
Aug 18

Inference-Time Mitigation of Adversarial Political Bias in Large Language Models

arXiv:2608. 14629v1 Announce Type: cross Abstract: As Large Language Models (LLMs) become the mainstay for information retrieval and summarization tasks, ensuring that they are always non-partisan and invulnerable to political bias is a critical step towards safer and more trustworthy Artificial Intelligence (AI).

By Tejaswi V. Panchagnula, Bruce Coburn, Bryce J. Dietrich, Robert X. Browning, Edward J. Delp, Fengqing Zhu
arXiv Computation and Language
Sep 14

I Am No One: Style-Aware Paraphrasing for Text Anonymization

The paper introduces a style-aware paraphrasing method for text anonymization that leverages pretrained large language models to build compact stylistic profiles from minimal samples and rewrite text to suppress identifiable style markers while preserving meaning. It demonstrates that this approach reduces authorship attribution F1 scores by 60‑70% on blog and review datasets, outperforming both differential privacy‑based and non‑DP baselines, and maintains content quality and readability.

By Ahmed Sohair Khan, Estrid He, Monica Wachowicz, Elham Naghizade
arXiv AI
Sep 28

DIAL: Position-Debiased LLM Judges with Adaptive Human Preference Calibration

The paper introduces DIAL, a framework that uses large language models (LLMs) as judges while mitigating position bias and aligning their judgments with human preferences. DIAL separates judge‑specific position effects, learns shared structure in debiased LLM preferences, and adaptively calibrates this structure toward human targets using limited human comparisons. Experiments on simulations and three human‑preference benchmarks show that DIAL remains robust to unbalanced response order, achieves strong human‑aligned rankings with few labels, and adapts when LLM information is imperfect, supported by a real‑data study of over 410K judgments from 21 LLM judges.

By Zesheng Cai, Yingqi Fan, Sichang Chen, Jin-Hong Du