arXiv AI

Just Ask Jev: Reinforcement Learning for Calibrated Decisions as a Zero-Shot Detector of AI Alignment Failures

The paper introduces Jev, a reinforcement‑learning‑trained model that provides calibrated probability answers to typed questions about a single input in one call. Jev is evaluated on RLCDAlignBench, a benchmark covering ten alignment failures across 44 tests and five target models, achieving a median AUROC of 0.886 zero‑shot and outperforming supervised baselines on most tasks. The study shows that question wording has little impact, while contextual fields that encode labels are more influential, and that Jev matches human‑label agreement while being 63× cheaper than LLM‑judge scorers.

arXiv AI
Sep 28

JevAdvBench: A Benchmark and Black-Box Attacks for Reinforcement Learning for Calibrated Decisions Models

JevAdvBench introduces the first adversarial benchmark for reinforcement‑learning‑based calibrated decision (RLCD) models, providing 812 typed questions across 66 scenarios and a black‑box attack suite of 9,744 single‑edit variants. The benchmark evaluates attacks by comparing each perturbed decision to the model’s own clean decision and to an identical re‑run, revealing that rewording changes decisions by only 1.2 percentage points while certain injected opinions can flip 12.1% of decisions and lower confidence below 0.8 in 38% of cases. These findings demonstrate that RLCD models can be significantly misled by seemingly innocuous input edits, underscoring the need to treat the state as untrusted in applications.

By Jianyi Hu, Hangtao Zhang, Yi Liu, Yeqi Zeng, Li Zeng, Xianlong Wang, Rui Wang, Leo Yu Zhang
arXiv AI
Oct 5

Test-time Calibration Learning for Large Language Model Reasoning

The paper introduces Test-Time Calibration Learning (TTCL), a label‑free framework that adapts both reasoning accuracy and verbalized confidence of large language models directly on unlabeled target‑task data. TTCL generates self‑supervision signals from multiple model responses, enabling calibration without ground‑truth labels and proving theoretically as a bounded surrogate for the ideal calibration objective. Experiments on mathematical reasoning and factual question answering show consistent improvements, with base models gaining an average 40.13% in accuracy and 70.80% in ECE reduction across eight benchmarks, and further gains under domain shift.

By Zizhuo Zhang, Xiong Peng, Jingwei Sun, Rong Yao, Shixiong Kai, Mingxuan Yuan, Bo Han
arXiv Machine Learning
Nov 2

Step-wise Rubric Rewards for LLM Reasoning

The paper introduces Step-wise Rubrics as Rewards (SRaR), a reinforcement learning framework that assigns rubric-based scores to individual reasoning steps in large language models. By normalizing step scores and combining them with outcome rewards through a decoupled advantage estimator, SRaR preserves criterion-level structure and mitigates issues like rewarding flawed steps or encouraging endless self‑correction. Experiments on 16K problems and six math benchmarks show that SRaR outperforms prior rubric methods, improving accuracy, increasing faithful reasoning rates, and reducing self‑correction loops.

By Weichu Xie, Haozhe Zhao, Wenpu Liu, Yongfu Zhu, Liang Chen, Minghao Ye, Zirong Chen, Yuqi Xu, Shuai Dong, Ziyue Wang, Xinbo Xu, Kean Shi, Ruoyu Wu, Xiaoying Zhang, Wenqi Shao, Baobao Chang, Nan Duan, Jiaqi Wang
arXiv Machine Learning
Aug 19

Thinking in a Low-Resource Language: What SFT Builds, What RL Fixes, What Accuracy Cannot See

The paper evaluates how three large mixture‑of‑experts models (Alibaba, OpenAI, NVIDIA) can be fine‑tuned to reason in a low‑resource language, specifically Greek. Accuracy metrics show little change, but the authors uncover significant qualitative improvements: after supervised fine‑tuning, models reason in Greek on ~98% of items, with better grammaticality and retained general ability. Reinforcement learning with pre‑registered rewards further eliminates reasoning‑channel leaks and format skips, while the Greek‑reasoning habit remains robust to an accuracy‑only gradient.

By Ayoub Kirouane, Christos Petrocheilos
arXiv Computation and Language
Oct 27

SearchJev: A Calibrated System-1 Model for Search Agents

arXiv:2610.05107v2 Announce Type: replace-cross Abstract: Search agents repeatedly make short decisions about relevance, evidence sufficiency, and search actions. Using generative language models for...

By Congfeng Cao, Lipeng Zuo, Konstantinos Papakostas, Qiwei Xu, Songwei Xu, Qianru Meng, Lun Zhou, Zhaochun Ren, Yougang Lyu, Xiaohui Yan
arXiv Computation and Language
Nov 2

EDU-Jev Training System-One Decision Models for Education

EDU-Jev is a family of single‑pass decision models (0.8B, 2B, and 4B parameters) designed for educational tasks that require selecting from a fixed set of options. The models output a probability distribution over predefined choices without generating text, using a pointer head to score each option. In twelve education benchmarks, EDU‑Jev 4B achieved 67.6% accuracy, outperforming five commercial decision APIs by 9.4 percentage points and the best open 4B model by 18.7 points, while the 2B version can process about 150 decisions per second on a single H100 GPU.

By Jiazheng Li, Yulan He
arXiv AI
Nov 5

Rubric Curriculum RL: Exploiting the Generation-Verification Gap in Non-Verifiable Domains

Rubric Curriculum RL (RcRL) is a reinforcement learning approach that leverages the generation‑verification gap in non‑verifiable domains by using pairwise preferences over rollouts and a curriculum over rubric criteria. Unlike traditional methods that rely on absolute scoring and stationary judges, RcRL provides a more discriminative, non‑stationary signal, allowing policies to maintain output entropy and continue improving beyond 1,000 training steps. In experiments on creative writing and HealthBench, RcRL outperforms baseline models, achieving a 70.5% human win rate and significant accuracy gains over the base model.

By Sumeet Ramesh Motwani, Tejas Krishnan, Alistair Letcher, Charles London, Suhaas Bhat, Huitian Jiao, Philip Torr, Riashat Islam, Christopher Summerfield, Jakob Foerster, Christian Schroeder de Witt, Qilong Gu, Shital Shah
arXiv Computation and Language
Oct 1

Listening to the Wise Few: Query-Key Alignment Unlocks Latent Correct Answers in Large Language Models

arXiv:2410.02343v2 Announce Type: replace Abstract: Large language models (LLMs) routinely fail to output the correct option in multiple-choice question answering (MCQA) while encoding the answer int...

By Eduard Tulchinskii, Kristian Kuznetsov, Laida Kushnareva, Anastasia Voznyuk, Andrei Andriiainen, Irina Piontkovskaya, Evgeny Burnaev, Serguei Barannikov