arXiv:2608. 14425v1 Announce Type: new Abstract: LLM evaluations often use fixed sampling budgets, testing every item the same number of times even after estimates are precise.
arXiv:2602. 05395v2 Announce Type: replace-cross Abstract: A simple strategy for improving LLM accuracy, especially in math and reasoning problems, is to sample multiple responses and submit the answer most consistently reached.
The paper investigates how to decide when to stop collecting posterior samples in classification‑oriented adaptive sensing. It shows that a simple threshold‑based plug‑in rule does not guarantee the desired confidence level, and proposes calibrated fixed‑sample and finite‑horizon sequential stopping rules that control the false‑declaration probability. Experiments on MNIST demonstrate that the sequential rule can reduce sensing cost the most, and that a curtailment strategy can save up to 62% of posterior samples while maintaining accuracy.
arXiv:2601.13885v2 Announce Type: replace-cross Abstract: Computerized Adaptive Testing (CAT) has proven effective for efficient LLM evaluation on multiple-choice benchmarks, but modern LLM evaluatio...
The paper introduces a risk‑controlled framework for using large language models (LLMs) as judges in tasks without reference answers. By calibrating uncertainty thresholds on a held‑out set, the method ensures that the false discovery rate of accepted verdicts stays below a user‑specified level α with high probability, using finite‑sample Clopper–Pearson intervals. When the parametric judge lacks confidence, the instance is routed to a retrieval‑augmented mode with a second calibrated threshold, preserving the error guarantee while achieving higher coverage than single‑mode baselines.
arXiv:2602. 13935v2 Announce Type: replace Abstract: While LLMs have seen substantial improvement in reasoning capabilities, they also sometimes overthink, generating unnecessary reasoning steps, particularly under uncertainty, given ill-posed or ambiguous queries.
arXiv:2606. 20820v2 Announce Type: replace Abstract: Can we trust evaluation scores to capture an LLM's true real-world performance?
arXiv:2607. 27023v1 Announce Type: new Abstract: Evaluating large generative models across benchmarks is time-consuming and computationally expensive.
arXiv:2607. 03991v1 Announce Type: new Abstract: Repeated LLM calls are the standard way to estimate how trustworthy a Text-to-SQL result is: run the pipeline multiple times, judge each SQL execution, and use the consistency of the verdicts as a confidence signal.
arXiv:2607. 08522v1 Announce Type: new Abstract: The inherent rigidity of fixed-size benchmarks makes them an inefficient tool for model evaluation.
arXiv:2604. 23099v2 Announce Type: replace-cross Abstract: Evaluating generative AI models is increasingly resource-intensive due to slow inference, expensive raters, and a rapidly growing landscape of models and benchmarks.
The paper introduces Speculative Evaluation, a method to reduce variance in evaluating stochastic large language models (LLMs) under a fixed rollout budget. It employs a Hierarchical Bayesian Neyman (HBN) policy that first runs a short uniform pilot, then pools task-level success counts via a hierarchical Bayesian model to compute posterior expectations of task-level sampling variances. Using these expectations, the method applies exact positive-integer Neyman allocation to allocate rollouts, and an asynchronous variant (HBN-async) speculatively executes continuations from partial pilot feedback to mitigate synchronization overhead. Across six checkpoints and 18 benchmark groups, Speculative Evaluation achieves 12.8%-33.6% lower variance compared to uniform allocation, outperforming empirical and independent Bayesian baselines, and demonstrates practical benefits in real-generation experiments.
LEAP (Likelihood Elicitation and Aggregation for Probabilistic forecasting) is a new approach that reorganizes how evidence is used in LLM-based forecasting systems. Instead of a monolithic prediction that aggregates all evidence at once, LEAP examines each evidence item separately, elicits likelihood parameters, and combines them with an explicit prior to produce a posterior distribution. The method supports continuous, single-choice, and multi-choice forecasts and has been shown to improve prediction and calibration metrics across models on a benchmark covering forecasting, information-seeking, and browsing tasks.