【论文精选】AI 前沿论文 TOP6 2026-06-23

1、大模型与Agent

① Trade-offs in Medical LLM Adaptation: An Empirical Study in French QA

作者：Ikram Belmadani, Oumaima El Khettari, Carlos Ramisch, Frederic Bechet, Richard Dufour, Benoit Favre

arXiv: 2606.19266v1

链接: https://arxiv.org/abs/2606.19266v1

English Abstract：

The development of large language models (LLMs) has led to an increased focus on their adaptation to specialized domains and languages, yet the effectiveness of domain adaptation strategies remains unclear. We present a study of medical domain adaptation using French medical question-answering (QA) as a case study. We compare continual pretraining (CPT), supervised fine-tuning (SFT), and their combination across three model families, multiple sizes, and three initialization types, explicitly disentangling adaptation effects from base model choice. We evaluate both multiple-choice (MCQA) and open-ended QA (OEQA) under greedy and constrained decoding using automatic metrics and LLM-as-a-Judge evaluation. For MCQA, CPT+SFT most often achieves the best scores, but gains over SFT are small and frequently not statistically significant, making SFT a strong and cost-effective default. For OEQA, CPT consistently improves overlap-based metrics, while SFT often degrades generation quality; instruction tuning and CPT+SFT are preferred by LLM-based evaluation. Cross-lingual experiments further show effective transfer from French adaptation to English benchmarks. Overall, we provide practical guidelines for selecting adaptation strategies under computational constraints.

2、强化学习

① UBP2: Uncertainty-Balanced Preference Planning for Efficient Preference-based Reinforcement Learning

作者：Mohamed Nabail, Leo Cheng, Jingmin Wang, Nicholas Rhinehart

arXiv: 2606.19328v1

链接: https://arxiv.org/abs/2606.19328v1

English Abstract：

Preference-based RL provides an approach to learning reward models from pairwise comparisons of behaviors, bypassing the need for explicit reward design. However, existing methods typically rely on passive data collection and suffer from poor sample efficiency, especially during the early stages of learning. We introduce a model-based approach that actively directs exploration by jointly reasoning over uncertainties in the reward, dynamics, and value functions. Our method, Uncertainty-Balanced Preference Planning (UBP2), uses ensembles of reward, dynamics, and value function models to evaluate candidate trajectories according to a unified score that combines expected reward, terminal value, and epistemic uncertainty. Planning under this objective yields an explicit tradeoff between exploitation and information acquisition without requiring ad hoc exploration heuristics. Under standard regularity assumptions, we establish sublinear regret guarantees for both finite-horizon and infinite-horizon settings. Empirically, experiments on the Meta-World benchmark show UBP2 achieves substantially higher sample efficiency than model-free preference-based methods and non-optimistic model-based baselines.

3、生成模型

① Diffusion-Proof: Recipe for Formal Theorem Proving Beyond Auto-Regressive Generation

作者：Ruida Wang, Rui Pan, Pengcheng Wang, Shizhe Diao, Tong Zhang

arXiv: 2606.19315v1

链接: https://arxiv.org/abs/2606.19315v1

English Abstract：

Enhancing the formal math reasoning capabilities of Large Language Models (LLMs) has become a key focus in both mathematical and computer science communities in recent years. While significant progress has been made in using state-of-the-art Auto-Regressive (AR) LLMs for formal theorem proving, these models suffer from inherent limitations. Their next-token prediction generation methods may yield suboptimal performance due to the challenges of long-range coherence and the compounding of errors over long sequences. Recent advancements in diffusion LLMs (dLLMs), which generate text through iterative denoising of a multi-token block, offer a promising alternative. However, the application of dLLMs to formal mathematics, where maintaining long-range coherence is critical, remains largely understudied. To address the challenges above, we propose Diffusion-Proof, to the best of our knowledge, the first framework to train and apply dLLMs for formal theorem proving. Our frameworks contain training and inference methods for two models. The first one is dLLM-Prover-7B, which performs whole-proof writing with long-range coherent tactic usage. The second one is dLLM-Corrector-7B, which is a novel large block diffusion-based correction model. It leverages the in-filling capabilities of dLLMs to perform local proof correction using bi-directional information. Extensive experiments demonstrate that Diffusion-Proof relatively significantly outperforms the AR LLM baseline trained

4、知识图谱

① Half a Link can Be Enough to Predict a Whole Link: Understanding Generalization in Knowledge Graph Foundation Models

作者：Cosimo Gregucci, Obaidah Theeb, Daniel Hernandez, Antonio Vergari, Steffen Staab

arXiv: 2606.18001v1

链接: https://arxiv.org/abs/2606.18001v1

English Abstract：

Knowledge graph (KG) foundation models (KGFMs) are zero-shot generalizers: trained once, they can predict links on unseen graphs without retraining. However, understanding when and how they can robustly generalize across KGs is still an open question. In this paper, we shed some light on their generalization mechanisms highlighting how their performance on unseen KGs is not uniform when it comes to partially seen links, which we call half-links. In fact, we show that to predict a test triple $(h,r,t)$ it might suffice in practice to have observed the half-link $(h,r)$ or $(r,t)$ in the inference graph. This yields a taxonomy of four scenarios when combinations of these half-links are observed or not. In a rigorous stratified analysis over these scenarios, we reveal that SoTA KGFMs use seen half links for predictions, while unseen half-links pose different challenges. As such, our finer-grained taxonomy can be a diagnostic protocol for robust KGFM generalization and highlights where novel KGFMs can improve.

5、具身智能

① Zero-Shot Long-Horizon Dexterous Manipulation via Multi-View 3D-Grounded VLM Reasoning

作者：Jisoo Kim, Sangwon Baik, Taeksoo Kim, Sungjoo Kim, Junyoung Lee, Mingi Choi, Hanbyul Joo

arXiv: 2606.19340v1

链接: https://arxiv.org/abs/2606.19340v1

English Abstract：

We present a zero-shot framework for long-horizon dexterous manipulation that grounds language instructions into executable 3D task plans from calibrated multi-view RGB images. Rather than training an end-to-end policy, our system uses a vision-language model (VLM) to produce reference-frame task grounding and primitive-level 2D keypoints, then lifts them into 3D via multi-view fusion. This lifting combines triangulation of view-wise VLM groundings with reference-view ray voting, which searches along a semantic camera ray for geometrically consistent candidates across neighboring views. The resulting 3D keypoints support both pick-and-place and tool-use: for tool-use, we retrieve an object-centric atomic action corresponding to the inferred skill category and align its stored 6D tool trajectory to the scene; for dexterous execution, we expand the lifted grasp keypoint into a task-conditioned grasp affordance region and generate feasible grasp-motion pairs with an arm-hand motion generator. Real-world experiments show improved 3D grounding accuracy and execution reliability over single-view RGB-D grounding and fine-tuned VLA baselines. We further demonstrate long-horizon manipulation through closed-loop status verification and replan, enabling zero-shot execution on unseen objects and tool-use tasks in novel scenes.

6、安全对齐

① Beyond Safe Data: Pretraining-Stage Alignment with Regular Safety Reflection

作者：Jinhan Li, Kexian Tang, Yihan Xu, Zhuorui Ye, Kaifeng Lyu

arXiv: 2606.19168v1

链接: https://arxiv.org/abs/2606.19168v1

English Abstract：

To achieve deeper safety alignment for large language models (LLMs), recent efforts have studied how to push safety interventions earlier into the pretraining stage, primarily by filtering unsafe data or rewriting it into safer forms. We argue that pretraining-stage alignment should go beyond making the data safe: LLMs may compose seemingly benign knowledge and capabilities into unsafe behaviors. To this end, we propose Safety Reflection Pretraining, a pretraining-stage alignment method which regularly inserts short safety reflections into pretraining corpora to integrate self-monitoring directly into language modeling, establishing a foundational capability that is subsequently reinforced by compatible post-training. Our experiments with 1.7B models pretrained on FineWeb-Edu show that Safety Reflection Pretraining improves safety classification accuracy and substantially reduces the success rates of inference-stage and finetuning attacks. Complementary to our real-world experiments, we also introduce a fully controlled synthetic environment, MedSafetyWorld, with a clear definition of safety and a reasoning structure under which models can easily generalize unsafe behaviors from safe data. Ablations in MedSafetyWorld further demonstrate a clear advantage of Safety Reflection Pretraining in preventing models from acting on unsafe behaviors generalized from safe data, compared with data filtering and rewriting. Taken together, our findings suggest that pretraining alignment sho