Despite the impressive performance of LLM-powered agents, their adoption for Electronic Health Record (EHR) data access remains limited by the absence of benchmarks that adequately capture real-world clinical data access flows. In practice, two core challenges hinder deployment: query ambiguity from vague user questions and value mismatch between user terminology and database entries. To address this, we introduce EHR-ChatQA an interactive database question answering benchmark that evaluates the end-to-end workflow of database agents: clarifying user questions, using tools to resolve value mismatches, and generating correct SQL to deliver accurate answers. To cover diverse patterns of query ambiguity and value mismatch, EHR-ChatQA assesses agents in a simulated environment with an LLM-based user across two interaction flows: Incremental Query Refinement (IncreQA), where users add constraints to existing queries, and Adaptive Query Refinement (AdaptQA), where users adjust their search goals mid-conversation. Experiments with state-of-the-art LLMs (e.g., o4-mini and Gemini-2.5-Flash) over five i.i.d. trials show that while agents achieve high Pass@5 of 90-95% (at least one of five trials) on IncreQA and 60-80% on AdaptQA, their Pass^5 (consistent success across all five trials) is substantially lower by 35-60%. These results underscore the need to build agents that are not only performant but also robust for the safety-critical EHR domain. Finally, we provide diagnostic insights into common failure modes to guide future agent development.

Multi-task neural routing solvers have emerged as a promising paradigm for their ability to solve multiple vehicle routing problems (VRPs) using a single model. However, existing neural solvers typically rely on predefined problem constraints or require per-problem fine-tuning, which substantially limits their zero-shot generalization ability to unseen VRP variants. To address this critical bottleneck, we propose URS, a unified neural routing solver capable of zero-shot generalization across a wide range of unseen VRPs using a single model without any fine-tuning. The key component of URS is the unified data representation (UDR), which replaces problem enumeration with data unification, thereby broadening the problem coverage and reducing reliance on domain expertise. In addition, we propose a Mixed Bias Module (MBM) to efficiently learn the geometric and relational biases inherent in various problems. On top of the proposed UDR, we further develop a parameter generator that adaptively adjusts the decoder and bias weights of MBM to enhance zero-shot generalization. Moreover, we propose an LLM-driven constraint satisfaction mechanism, which translates raw problem descriptions into executable stepwise masking functions to ensure solution feasibility. Extensive experiments demonstrate that URS can consistently produce high-quality solutions for more than 100 distinct VRP variants without any fine-tuning, which includes more than 90 unseen variants. To the best of our knowledge, URS is the first neural solver capable of handling over 100 VRP variants with a single model.

Advances in automated scoring are closely aligned with advances in machine-learning and natural-language-processing techniques. With recent progress in large language models (LLMs), the use of ChatGPT, Gemini, Claude, and other generative-AI chatbots for automated scoring has been explored. Given their strong reasoning capabilities, LLMs can also produce rationales to support the scores they assign. Thus, evaluating the rationales provided by both human and LLM raters can help improve the understanding of the reasoning that each type of rater applies when assigning a score. This study investigates the rationales of human and LLM raters to identify potential causes of scoring inconsistency. Using essays from a large-scale test, the scoring accuracy of GPT-4o, Gemini, and other LLMs is examined based on quadratic weighted kappa and normalized mutual information. Cosine similarity is used to evaluate the similarity of the rationales provided. In addition, clustering patterns in rationales are explored using principal component analysis based on the embeddings of the rationales. The findings of this study provide insights into the accuracy and ``thinking'' of LLMs in automated scoring, helping to improve the understanding of the rationales behind both human scoring and LLM-based automated scoring.

Diffusion language models have emerged as a powerful alternative to autoregressive models, enabling fast inference through flexible and parallel generation paths. This flexibility is enabled by new sampling strategies, or planners, that iteratively choose where to denoise along the sequence rather than sampling uniformly at random. However, by modifying reverse paths, planners introduce a mismatch between the uniformly random denoising paths used during training and the planning-based paths used at inference. In this work, we systematically investigate this mismatch and theoretically show that the standard discrete diffusion training evidence lower bound (ELBO) does not accurately describe a denoiser under non-uniform planning. To bridge this gap, we derive a new Planned Evidence Lower Bound (P-ELBO) that directly incorporates planner-based reverse dynamics into the training objective. Building on this, we propose Planner Aware Path Learning (PAPL), a simple and effective modification of the standard masked discrete diffusion loss that aligns training and inference under planned denoisers. Empirically, PAPL delivers consistent improvements across domains, including a 40% relative gain in protein sequence modeling, up to a 4x improvement in MAUVE for text generation, and a 23% relative gain in HumanEval pass@10 for code generation.

Prompt engineering is crucial for leveraging the full potential of large language models (LLMs). While automatic prompt optimization offers a scalable alternative to costly manual design, generating effective prompts remains challenging. Existing methods often struggle to stably generate improved prompts, leading to low efficiency, and overlook that prompt optimization easily gets trapped in local optima. Addressing this, we propose GRACE, a framework that integrates two synergistic strategies: Gated Refinement and Adaptive Compression, achieving Efficient prompt optimization. The gated refinement strategy introduces a feedback regulation gate and an update rejection gate, which refine update signals to produce stable and effective prompt improvements. When optimization stagnates, the adaptive compression strategy distills the prompt's core concepts, restructuring the optimization trace and opening new paths. By strategically introducing information loss through refinement and compression, GRACE delivers substantial gains in performance and efficiency. In extensive experiments on 11 tasks across three practical domains, including BIG-Bench Hard (BBH), domain-specific, and general NLP tasks, GRACE achieves significant average relative performance improvements of 4.7%, 4.4% and 2.7% over state-of-the-art methods, respectively. Further analysis shows that GRACE achieves these gains using only 25% of the prompt generation budget required by prior methods, highlighting its high optimization efficiency and low computational overhead. Our code is available at https://github.com/Eric8932/GRACE.

Recent work has demonstrated that controlled pretraining experiments are a powerful tool for understanding learning, reasoning, and memorization in large language models (LLMs). However, the computational cost of pretraining presents a significant constraint. To overcome this constraint, we propose to conduct multiple pretraining experiments simultaneously during a single training run. We demonstrate the feasibility of this approach by conducting ten experiments during the training of a 1.5B parameter model on 210B tokens. Although we only train a single model, we can replicate the results from multiple previous works on data contamination, poisoning, and memorization. We also conduct novel investigations into knowledge acquisition, mathematical reasoning, and watermarking. For example, we dynamically update the training data until the model acquires a particular piece of knowledge. Remarkably, the influence of the ten experiments on the model's training dynamics and overall performance is minimal. However, interactions between different experiments may act as a potential confounder in our approach. We propose to test for interactions with continual pretraining experiments, finding them to be negligible in our setup. Overall, our findings suggest that performing multiple pretraining experiments in a single training run can enable rigorous scientific experimentation with large models on a compute budget.

We introduce Guard Vector, a safety task vector computed as the parameter difference between a guardrail model (Guard Model) and a same-architecture pretrained language model. Composing this vector with a target language model yields a Target Guard Model (TGM). We then adapt TGM with a streaming-aware approach that combines prefix-based training and evaluation with a classifier that produces a single-token output. With this composition alone, TGM improves classification quality over established Guard Models across standard safety suites and enables language extensibility to Chinese, Japanese, and Korean, requiring neither additional training nor target language labels. It also demonstrates model portability across two widely used public guardrail backbones, Llama and Gemma. With prefix SFT (supervised fine-tuning), TGM preserves classification quality under streaming by aligning the behavior between prefix inputs and full-text inputs. The single-token output design increases throughput and reduces latency. Together, these components reduce data and compute requirements while promoting streaming-aware evaluation practices, thereby contributing to a more responsible AI ecosystem.

Preference optimization is crucial for aligning large language models (LLMs) with human values and intentions. A significant challenge in this process is the distribution mismatch between pre-collected offline preference data and the evolving model policy. Existing methods attempt to reduce this gap using static heuristics or decoupled online sampling strategies, but they often fail to adapt to the model's dynamic learning state. To bridge this gap, we propose Meta-Weighted Adaptive Preference Optimization (MetaAPO), a novel framework that dynamically couples data generation with model training. MetaAPO employs a lightweight meta-learner, as an "alignment gap estimator", to evaluate the potential benefits of on-policy sampling in relation to offline data. This guides targeted online generation and assigns sample-wise meta-weights to the optimization objective, dynamically balancing the quality and distribution of online and offline data. Experiments on AlpacaEval 2, Arena-Hard and MT-Bench demonstrate that MetaAPO consistently outperforms existing preference optimization approaches across various settings, while reducing 42% in online annotation costs.

In the field of medicine, complex reasoning tasks such as clinical diagnosis, treatment planning, and medical knowledge integration pose significant challenges, where small language models often underperform compared to large language models like GPT-4 and Deepseek. Recent knowledge distillation-based methods aim to address these issues through teacher-guided error correction, but this LLM as judge approach remains challenging in terms of cost, time, and efficiency. To circumvent this issue, we propose a novel two-stage framework, MedCritical, which uses a small language model fine-tuned by a large teacher model to play against itself. In the first stage, we extract high-level and detailed long-chain thought templates from the teacher model to guide the student model to generate more complex reasoning thoughts. In the second stage, we introduce direct preference optimization (DPO) through model self-iteration collaboration to enhance the reasoning ability of the student model by playing against the correction trajectory of the fine-tuned model during training. This model self-learning DPO approach teaches the student model to use its own error-driven insights to consolidate its skills and knowledge to solve complex problems, and achieves comparable results to traditional knowledge distillation methods using teacher models at a lower cost. Notably, our MedCritical 7B model outperforms the Taiyi and Huatuo-o1-7B models by 3.04\% and 10.12\% respectively on the CMExam benchmark, achieving new SOTA performance among 7B-class small models.

However, given limited time and computational resources, retraining LLMs to mitigate the influence of undesired data is impractical. Thus, Machine Unlearning (MU) emerges as an alternative solution to weaken a model's performance on undesired knowledge (Liu et al., 2024b; Eldan & Russinovich, 2023) while preserving the model's original utility (Liu et al., 2025). Though much research shed light on MU, several recent studies indicate that MU still lacks robustness (Zhang et al., 2024c; Y uan et al., 2025; Lee et al., 2025). In particular, they are susceptible to both jailbreak attacks (Zou et al., 2023; Andriushchenko et al., 2024) and relearning attacks (Hu et al., 2024). Such limitations can be exploited through reusing small amount of unlearned knowledge (Hu et al., 2024) or adversarial prompt manipulations, including prefix injection (Andriushchenko et al., 2024) and adaptive jailbreaks (Liu et al., 2023). These attacks act as small perturbations in parameter or representation space, driving the model along directions that yield undesired content that should have been forgotten (Fan et al., 2025; Lin et al., 2024b). Smoothness Minimization can be introduced to enhance model robustness against these attacks by promoting a smooth loss across the neighborhood (Foret et al., 2020; Fan et al., 2025). Plenty of studies have sought to remove undesired data to improve the effectiveness of MU and these approaches have demonstrated substantial unlearning performance (Liu et al., 2022a; Thudi et al., 2022; Zou et al., 2024; Pawelczyk et al., 2023; Liu et al., 2024a). However, they still suffer from limitations in robustness and trade-off between model utility and unlearning effectiveness.