Tag
The paper proposes PACE (Proxy Alignment and Curvature Expansion), a pre-optimization diagnostic that uses repeated human ratings to predict the human-gain curve induced by a fixed proxy, modeling proxy optimization as a KL-regularized exponential tilt with first-order alignment and second-order correction terms, and shows reduced held-out gain-prediction error across multiple benchmarks.
论文提出 Preference-Behavior Model (PBM)，通过逆博弈问题联合推断 LLM 的偏好与不完美决策行为参数，并在公共物品博弈等三种社会困境中评估，发现传统行为经济学偏好测量工具无法一致预测 LLM 在战略环境中的行为，而跨博弈拟合的模型则具有预测价值。
RLCD is explained as a schema-conditioned Plackett–Luce objective that advances reward modeling from scalar rewards to pairwise preferences to multiway calibrated decisions, simplifying the understanding of Jev.
The paper proposes that intransitive preferences in language models arise from multiple latent orderings and introduces a mixture Bradley–Terry model to analyze them, showing better explanatory power across various models and tasks.
HypReflect is a framework for continual personalization in LLM assistants that infers explicit preference hypotheses from user signals and uses self-distillation to adapt, outperforming baselines and generalizing to unseen users and cross-domain settings.
This paper proposes a unified framework for memory access and selection in long-context dialogue systems, using Bayes factors to quantify the utility of historical turns for modeling changing user preferences. Experiments show it outperforms embedding-based retrieval on preference-intensive tasks.
This paper presents a system for constrained humor generation that uses a generate-many select-best strategy with a preference model learned from human comparisons. It achieved top ranks in English and Chinese subtasks and second in Spanish at SemEval-2026 Task 1.