Zehao Zhao, Shangjian Yin, Rui Liu, Chongyang Bai, Mark Ibrahim, Kavosh Asadi, Yuchen Lu, Hang Cui 等 9 位作者
Test-time updating lets a language model change persistent state after training---such as model weights, adapter parameters, or a memory module---before or while answering queries. Existing evaluations often compare systems using update count or frequency and performance measured immediately afterward. These summaries omit what is updated, where, when, and in what order. The same total can therefore hide an important difference: one system may answer after one update while another answers after two. We introduce the Control Map, an evaluation framework that records the committed updates preceding each scored answer and separately evaluates behavioral changes, immediate learning (acquisition), later recall (retention), preservation of unrelated behavior (locality), and performance in new settings. We prove when counts alone determine final state and show, under stated assumptions, how differences between pre-answer histories limit differences in average scores. Using ControlMap, our experiments reveal three insights. First, update scheduling matters: at a 760M-parameter checkpoint, 24 schedules with the same update count yield 19.7%--34.2% accuracy. Second, learning an edit does not guarantee that it lasts or remains contained: all six selected CounterFact LoRA settings learn the new information, while ROME retains edits at all 13 operating points but fails locality at each. Third, no editor performs best consistently across models, benchmarks, and tests. Evaluations should therefore report update histories and assess acquisition, retention, and locality separately.
Reinforcement learning with verifiable rewards (RLVR) is widely used to improve the reasoning capability of language models, since model-generated answers can be easily verified by a rule-based verifier. In contrast, reinforcement learning from human feedback (RLHF) is mainly used to optimize open-ended conversation with a learned reward model, because such responses are not easy to verify. RLVR and RLHF are usually trained separately with their own domain data and training protocols. How gains transfer between these domains is unclear, partly because comparisons change both where a model is trained (the prompts) and what is rewarded (by a verifier or a reward model). We separate the two in a controlled study of RL on multiple base models that fixes the optimizer and the number of training prompts, and varies only the source domain of the prompts (math, multi-domain, chat, or mixtures) and the reward. We find that RLHF on open-ended prompts substantially improves conversational ability and also brings large gains in reasoning and coding, whereas RLVR improves reasoning but its gains in open-ended conversation plateau far below those of RLHF, suggesting that RL on open-ended prompts with a reward model can serve as an effective approach for generalizing across domains.
Shangjian Yin University of California, Riverside, Zhepei Wei University of Virginia, Xinyu Zhu University of Virginia, Charlottesville, Wei-Lin Chen University of Virginia, Charlottesville, Yu Meng University of Virginia
Traditional reinforcement learning from human feedback (RLHF) for large language models (LLMs) relies on expensive human-annotated datasets, while Reinforcement Learning from AI Feedback (RLAIF) also incurs significant costs, requiring the collection of diverse prompts and corresponding responses, often necessitating external reward models or proprietary models like GPT-4 to annotate preference pairs. In this work, we introduce Self-Alignment Optimization (SAO), a fully self-synthetic framework for LLM alignment, where all training data, including prompts (i.e., user queries), responses, and preferences, are generated by the model itself. Specifically, SAO first instructs the LLM to engage in persona role-play and generate diverse prompts and responses, which are then self-evaluated for preference optimization. Extensive experiments demonstrate that SAO effectively enhances the model’s chat capabilities on standard benchmarks like AlpacaEval 2.0, while maintaining strong performance on downstream objective tasks (i.e.,, question-answering, math reasoning). Our work provides a practical solution for self-improvement in aligning LLMs, and the code for reproducing our results is available at: https://github.com/SJY8460/SAO.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Shangjian Yin University of California, Riverside, Zhouxing Shi University of California, Riverside
Reinforcement Learning with Verifiable Rewards (RLVR) has demonstrated remarkable effectiveness in boosting the objective performance (e.g., reasoning) of Large Language Models (LLMs) through rule-based, on-policy self-improvement strategies. However, optimizing LLMs for subjective capabilities and alignment with human preferences remains challenging due to the non-verifiable nature. Most prior works use datasets comprising response pairs with substantial quality gaps labeled by a strong external judge. While effective for preference metrics, this paradigm often incurs an “alignment tax”, where the model’s objective performance on downstream tasks degrades as it overfits to subjective preferences. In this work, we introduce Donkey, a high-quality, non-verifiable dataset where response pairs differ only by subtle nuances. We find that LLMs optimized on Donkey via preference learning outperform those trained on data with explicit quality gaps, while simultaneously maintaining their objective capabilities. Furthermore, we observe that preference signals on Donkey can be decomposed into consensus preferences and individual preferences. Our analysis reveals that distilling consensus preferences provides a significantly more data-efficient signal for preference optimization. Our findings underscore the importance of leveraging nuanced preference signals and the consensus of multiple judges for advancing subjective LLM alignment. Our code and data will be available at https://github.com/SJY8460/Donkey.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表