Reinforcement learning has powered many of the recent breakthroughs in large language models, especially for tasks where rewards can be computed automatically, such as code generation. However, these methods deteriorate in open-ended domains like medical consultation, where feedback is inherently ambiguous, highly context-dependent, and cannot be reduced to a reliable scalar signal. In such settings, RL must either rely on supervision-intensive reward models that often fail to generalize, or it falls into pathological behaviors such as reward hacking - an especially troubling risk for high-stakes medical dialogue. To address these limitations, we introduce ORBIT, an open-ended rubric-based incremental training framework for high-stakes medical dialogue. ORBIT integrates synthetic dialogue generation with dynamically constructed rubrics that serve as adaptive guides for incremental RL. Instead of relying on external medical knowledge bases or handcrafted rule sets, ORBIT uses rubric-driven feedback to steer the learning process. Its judge component can be instantiated with general-purpose instruction-following LLMs, removing the need for any task-specific fine-tuning. Applied to the Qwen3-4B-Instruct model, ORBIT raises the HealthBench-Hard score from 7.0 to 27.5 using only 2k training samples, achieving SOTA performance for models at this scale. With larger rubric datasets, ORBIT-trained models further compete with the strongest open-source baselines on HealthBench-Hard. Our analysis shows that rubric-guided RL consistently improves consultation quality across diverse medical scenarios. We also apply such rubric generation and training pipeline to InfoBench, where ORBIT enhances instruction-following performance, highlighting the generality of rubric-based feedback.

Online AI Feedback (OAIF) presents a promising alternative to Reinforcement Learning from Human Feedback (RLHF) by utilizing online AI preference in aligning language models (LLMs). However, the straightforward replacement of humans with AI deprives LLMs from learning more fine-grained AI supervision beyond binary signals. In this paper, we propose Direct Advantage Regression (DAR), a simple alignment algorithm using online AI reward to optimize policy improvement through weighted supervised fine-tuning. As an RL-free approach, DAR maintains theoretical consistency with online RLHF pipelines while significantly reducing implementation complexity and improving learning efficiency. Our empirical results underscore that AI reward is a better form of AI supervision consistently achieving higher human-AI agreement as opposed to AI preference. Additionally, evaluations using GPT-4-Turbo and MT-bench show that DAR outperforms both OAIF and online RLHF baselines.

Reinforcement learning from human feedback (RLHF) is a key driver of quality and safety in state-of-the-art large language models. Yet, a surprisingly simple and strong inference-time strategy is Best-of-N sampling that selects the best generation among N candidates. In this paper, we propose Best-of-N Distillation (BOND), a novel RLHF algorithm that seeks to emulate Best-of-N but without its significant computational overhead at inference time. Specifically, BOND is a distribution matching algorithm that forces the distribution of generations from the policy to get closer to the Best-of-N distribution. We use the Jeffreys divergence (a linear combination of forward and backward KL) to balance between mode-covering and mode-seeking behavior, and derive an iterative formulation that utilizes a moving anchor for efficiency. We demonstrate the effectiveness of our approach and several design choices through experiments on abstractive summarization and Gemma models. Aligning Gemma policies with BOND outperforms other RLHF algorithms by improving results on several benchmarks.

Code Community Question Answering (CCQA) seeks to tackle programming-related issues, thereby boosting productivity in both software engineering and academic research. Recent advancements in Reinforcement Learning from Human Feedback (RLHF) have transformed the fine-tuning process of Large Language Models (LLMs) to produce responses that closely mimic human behavior. Leveraging LLMs with RLHF for practical CCQA applications has thus emerged as a promising area of study. Unlike standard code question-answering tasks, CCQA involves multiple possible answers, with varying user preferences for each response. Additionally, code communities often show a preference for new APIs. These challenges prevent LLMs from generating responses that cater to the diverse preferences of users in CCQA tasks. To address these issues, we propose a novel framework called Aligning LLMs through Multi-perspective User Preference Ranking-based Feedback for Programming Question Answering (ALMupQA) to create user-focused responses. Our approach starts with Multi-perspective Preference Ranking Alignment (MPRA), which synthesizes varied user preferences based on the characteristics of answers from code communities. We then introduce a Retrieval-augmented In-context Learning (RIL) module to mitigate the problem of outdated answers by retrieving responses to similar questions from a question bank. Due to the limited availability of high-quality, multi-answer CCQA datasets, we also developed a dataset named StaCCQA from real code communities. Extensive experiments demonstrated the effectiveness of the ALMupQA framework in terms of accuracy and user preference. Compared to the base model, ALMupQA showed nearly an 11% improvement in BLEU, with increases of 20% and 17.5% in BERTScore and CodeBERTScore, respectively.