The paper introduces an evidence‑referenced audit for large language models that assesses how much predictive information proxy fields contribute, comparing model reliance to a known outcome rule and a Bayesian linear regression. Applied to a simulated patient‑ranking task, the study finds that models rely on non‑informative fields, grow less reliance as proxies become predictive, are sensitive to display format, and that ranking accuracy does not reliably indicate proxy reliance. The authors argue that audits should report reliance against explicit predictive references alongside accuracy to properly evaluate model behavior.
Large language models (LLMs) are entering decisions in triage and lending, where task-relevant inference must be distinguished from impermissible proxy use. Current audits ask whether decisions change...
arXiv:2608.22887v1 Announce Type: new Abstract: Large language models (LLMs) are entering decisions in triage and lending, where task-relevant inference must be distinguished from impermissible proxy...
The paper investigates how large language models (LLMs) from OpenAI and Anthropic exhibit probability bias when deciding whether to trust information from reporters of different racial descriptions. Using a two‑urn Bayesian game and 6,600 sessions, the authors find that models often refuse to assess a reporter’s credibility—especially for Black‑described reporters—yet when they do assess, Black reporters receive slightly higher trust. The bias is subtle, growing with reasoning effort, and is not detected by standard audits that compare average trust or only assessed cases.
Large Language Model judges are commonly used to rank texts via pairwise comparison, with reliability traditionally measured by position bias, transitivity, and pairwise agreement. This paper argues that these proxies are misleading because they are dominated by close‑rank‑gap pairs, which contribute little to the overall ranking, while far‑gap pairs carry the true ranking signal. Experiments on simulations and human‑rated corpora show weak correlation between the proxies and actual ranking accuracy, suggesting judges should be evaluated using rank‑gap‑conditional metrics against human rankings.
The study evaluates ten bias audit instruments across ten advanced language models on occupational gender, age, and socioeconomic status. While each tool reliably detects bias, their rankings of model performance are essentially random, indicating that different audits measure distinct constructs. The findings show that a single audit can identify bias direction within its own framework, but no audit can consistently rank models against one another.
arXiv:2608. 14399v1 Announce Type: cross Abstract: Patients increasingly ask large language model (LLM) assistants which doctor to see, making these systems AI infomediaries: algorithms that intermediate one person's choice among other people and thereby decide, silently and at scale, which physicians become visible.
The paper investigates how benchmark contamination—leakage of test items into training data—affects large language model (LLM) leaderboards. By comparing original test items with semantically equivalent paraphrases, the authors measure contamination as a violation of anchor-item invariance and find that it inflates absolute scores but rarely changes model rankings. Across 47 public models and 74 finetuned models on four benchmarks, the rank correlation between standard and paraphrase-controlled leaderboards is 0.997, with only a handful of cases showing differential contamination that could alter rankings.
arXiv:2605.01048v2 Announce Type: replace-cross Abstract: Counterfactual prompting (i.e., perturbing a single factor and measuring output change) is widely used to evaluate things like LLM bias and C...
The paper discusses how large language models (LLMs) can be fine‑tuned with observational data to improve alignment with human preferences and business goals. It highlights that directly using such data can cause models to learn spurious correlations, and introduces DeconfoundLM, a method that removes known confounders from reward signals. Experiments show that DeconfoundLM better recovers causal relationships and outperforms baseline methods by over 16% in objective score when confounding is present.
arXiv:2609.37616v1 Announce Type: new Abstract: Language models tend to agree with whatever a user asserts, and post-training increasingly targets this sycophancy so that models evaluate claims on th...
The paper "Medical Causal Hypothesis Verification with Large Language Models" reports a small-scale study evaluating eight LLMs on 17 medical causal hypotheses. The authors introduce an evaluation framework and annotate 1,067 evidence points across six criteria, using nine metrics to assess performance. Results show that while LLMs have strong recall, they frequently fail to provide valid scientific articles, evidence, or reject unsupported hypotheses, revealing a critical limitation for their use in healthcare.
arXiv:2606. 03650v1 Announce Type: cross Abstract: Choosing or ranking language models for a specific application is hardest when no task-specific labeled data exists, and standard public benchmarks cannot be trusted, their items having likely leaked into pretraining, so scores reflect memorization rather than fitness.