The paper introduces ReVeR, a pointwise rubric reward model that derives hard rules and graded principles directly from instructions, avoiding response-induced bias. It employs tool-augmented judges and label‑blind distillation to create ReVeR‑Traces, a large dataset of rubric reasoning and judging trajectories. Experiments show ReVeR surpasses pairwise rubric models in preference accuracy, especially for instruction following, and reveal that higher preference accuracy does not always lead to better downstream policies.
arXiv:2602.13576v2 Announce Type: replace-cross
Abstract: Evaluation and alignment pipelines for large language models increasingly rely on LLM-based judges, whose behavior is guided by natural-langu...
By Ruomeng Ding, Yifei Pang, He Sun, Yizhong Wang, Zhiwei Steven Wu, Zhun Deng
The paper introduces Graph Direct Preference Optimization (GraphDPO), a generalization of Direct Preference Optimization that uses directed acyclic graphs to capture richer preference structures from multiple rollouts per prompt. GraphDPO encodes dominance relations as edges and optimizes a Plackett–Luce-inspired objective that aggregates supervision over dominated neighborhoods, preserving transitivity and handling ties or incomparable responses via an equivalence-class construction. Experiments on reasoning and program synthesis tasks show that this graph-structured approach outperforms traditional pairwise and listwise alignment objectives while maintaining linear per‑prompt complexity.
By Ning Liu, Chuanneng Sun, Kristina Klinkner, Shervin Malmasi
The paper introduces CoCA, a closed‑loop framework that jointly evolves a policy and its rubric by diagnosing each rubric criterion against the current policy’s capabilities. CoCA contrasts query‑only and rubric‑guided rollouts to identify mismatched difficulty, recalibrates criteria, and uses frontier‑based learning or self‑distillation to align the rubric with the policy’s evolving strengths. Experiments on various benchmarks and large language models show that CoCA consistently outperforms existing rubric‑based reinforcement learning methods and helps the policy internalize guidance‑revealed capabilities while preventing degradation.
By Chengsheng Zhang, Wenshen Xu, Yuchuan Wu, Zhiyuan Yu, Junfeng Fang, Tianlong Wang, Bingxue Qiu, Yongbin Li, Xiang Wang
The paper investigates how the quality of instructions used to generate response pairs affects preference learning for language models. It shows that low‑quality or ambiguous instructions limit the range of response quality, weakening preference signals, and introduces an instruction‑refinement pipeline that improves data quality without discarding examples. Experiments across models and benchmarks demonstrate that refining instructions leads to better alignment and complements other data‑improvement methods.
By Seohyeong Lee, Hwaran Lee, Buru Chang
The paper investigates what makes rubric design effective for evaluating language models, analyzing criterion organization, operationalization, and score aggregation. It finds that a well-structured organization supports comprehensive criteria, semantic specification alone is insufficient for accurate judgments, and complex aggregation yields limited gains. Based on these insights, the authors propose ARTS‑Hybrid, an adaptive tree search with hybrid refinement that improves both evaluation performance and rubric quality, and shows promise for guiding model optimization.
By Feifei Qian, Xuan Chen, Lixin Cui, Chaoyang Mei, Fanlin Meng, Shuyao Xiao, Ming Cui, Junxi Yin, Lu Bai