The paper investigates how the order of candidate documents in a prompt affects the decisions made by large‑language‑model (LLM) scorers, even when their ranking quality is similar. It shows that five scorers with only a 0.010 nDCG@10 difference can produce retained‑set overlaps as low as 0.66–0.84, and that existing rerankers still exhibit significant order dependence. The authors propose Order‑Consistency SFT (OC‑SFT), a training method that reduces this dependence, maintaining ranking quality while improving decision stability across multiple tasks.
arXiv:2608.21382v1 Announce Type: new Abstract: Multiple-choice benchmarks fix the questions and the correct answers, but not the harness: the order of the options, the wording of the prompt, and whe...
arXiv:2609.24128v1 Announce Type: cross Abstract: Judge-specific sensitivity is useful for aggregating pairwise LLM evaluations, but its interpretation depends on which systematic presentation effect...
The study investigates whether small differences in leaderboard rankings between large language models (LLMs) are robust to changes in benchmark composition. Using item‑level responses from five benchmarks and a spectral approximation to multidimensional item‑response theory, the authors find that while overall rankings remain highly correlated, a significant portion (30.9–47.1%) of near‑tie pairs reverse order when benchmark items are recomposed based on low differential item functioning. This suggests that sub‑one‑percentage‑point leaderboard gaps may not reliably reflect true model superiority.
arXiv:2606. 13685v1 Announce Type: cross Abstract: LLM-as-a-Judge is now widely used to rank model outputs, train reward models, and populate public leaderboards, but its run-to-run reliability remains under-characterized.
arXiv:2608. 15428v1 Announce Type: cross Abstract: Multiple-choice benchmarks are graded on whether a model picks the right option, not on whether it needed the question.
The paper introduces a pluralistic agreement index, Gamma, to quantify how often wrong runs of large language models (LLMs) agree with the majority consensus. By decomposing Gamma into a mechanical component and a preference‑unexplained residual, the authors show that on GPT‑4.1 the mechanical part explains most of the agreement on multiple‑choice benchmarks but only about half on open‑domain tasks, revealing a residual bias that can cause self‑consistency to backfire on hard questions. The study provides a quantitative framework for understanding when majority voting over LLM samples improves or harms accuracy, without proposing new voting methods.
The paper investigates whether accuracy and order sensitivity in large language model (LLM) multiple‑choice question (MCQ) evaluation move together. Using proprietary and open‑weight models on MMLU and ARC‑Challenge, the authors measure order sensitivity with flip rate and recall standard deviation (RStd) and find that these metrics often diverge from accuracy and from each other. They conclude that accuracy alone is insufficient and recommend reporting both flip rate and RStd alongside accuracy.
The paper investigates how temporal aggregation and ranking preservation affect aggressive decoding-time KV cache eviction. It shows that exponential‑moving‑average (EMA) aggregation makes many scorer modifications indistinguishable at the eviction‑set level, while certain variants (KeyDiff, key norm, recency, learned scorer) alter rankings and degrade performance. Building on this, the authors propose InertiaKV and its lazy variant, achieving 1.34–1.46× faster decoding, and also explore a score‑free decoding approach that slightly improves quality while eliminating further scoring.
Judge-specific sensitivity is useful for aggregating pairwise LLM evaluations, but its interpretation depends on which systematic presentation effects the ranking model includes. We introduce OSCAR, a...
arXiv:2607.27422v2 Announce Type: replace Abstract: Goal-conditioned action generators can produce many plausible actions. Best-of-\(K\) selection then uses a critic to rank those actions for the cur...
Given a state and a set of options, such as intents, tools or interface elements, System One decision models return a distribution over the options in one forward pass, without generated reasoning. We...
The paper investigates when it is better to return an existing draft answer or revise it using retrieved evidence in retrieval‑augmented QA systems. By grading both the draft and its candidate revision with the same correctness judge, the authors define a paired effect called recoverability and train policies to predict it before revision. Experiments on 25,870 open‑domain questions show that a recoverability‑based scorer outperforms a draft‑correctness scorer across multiple Llama setups, improving accuracy–revision trade‑offs and closing a significant portion of the oracle gap, though it still applies harmful revisions in a substantial fraction of cases.