Developing large language models is expensive and involves making decisions with small experiments, typically by evaluating on large, multi-task evaluation suites. In this work, we analyze specific properties which make a benchmark more reliable for such decisions, and interventions to design higher-quality evaluation benchmarks. We introduce two key metrics that show differences in current benchmarks: signal, a benchmark's ability to separate better models from worse models, and noise, a benchmark's sensitivity to random variability between training steps. We demonstrate that benchmarks with a better signal-to-noise ratio are more reliable when making decisions at small scale, and those with less noise have lower scaling law prediction error. These results suggest that improving signal or noise will lead to more useful benchmarks, so we introduce three interventions designed to directly affect signal or noise. For example, we propose that switching to a metric that has better signal and noise (e.g., perplexity rather than accuracy) leads to better reliability and improved scaling law error. We also find that filtering noisy subtasks, to improve an aggregate signal-to-noise ratio, leads to more reliable multi-task evaluations. We also find that averaging the output of a model's intermediate checkpoints to reduce noise leads to consistent improvements. We conclude by recommending that those creating new benchmarks, or selecting which existing benchmarks to use, aim for high signal and low noise. We use 30 benchmarks for these experiments, and 375 open-weight language models from 60M to 32B parameters, resulting in a new, publicly available dataset of 900K evaluation benchmark results, totaling 200M instances.

Abstractive summarization is a core application in contact centers, where Large Language Models (LLMs) generate millions of summaries of call transcripts daily. Despite their apparent quality, it remains unclear whether LLMs systematically under- or over-attend to specific aspects of the transcript, potentially introducing biases in the generated summary. While prior work has examined social and positional biases, the specific forms of bias pertinent to contact center operations - which we term Operational Bias - have remained unexplored. To address this gap, we introduce BlindSpot, a framework built upon a taxonomy of 15 operational bias dimensions (e.g., disfluency, speaker, topic) for the identification and quantification of these biases. BlindSpot leverages an LLM as a zero-shot classifier to derive categorical distributions for each bias dimension in a pair of transcript and its summary. The bias is then quantified using two metrics: Fidelity Gap (the JS Divergence between distributions) and Coverage (the percentage of source labels omitted). Using BlindSpot, we conducted an empirical study with 2500 real call transcripts and their summaries generated by 20 LLMs of varying scales and families (e.g., GPT, Llama, Claude). Our analysis reveals that biases are systemic and present across all evaluated models, regardless of size or family.

We present visual action prompts, a unified action representation for action-to-video generation of complex high-DoF interactions while maintaining transferable visual dynamics across domains. Action-driven video generation faces a precision-generality trade-off: existing methods using text, primitive actions, or coarse masks offer generality but lack precision, while agent-centric action signals provide precision at the cost of cross-domain transferability. To balance action precision and dynamic transferability, we propose to "render" actions into precise visual prompts as domain-agnostic representations that preserve both geometric precision and cross-domain adaptability for complex actions; specifically, we choose visual skeletons for their generality and accessibility. We propose robust pipelines to construct skeletons from two interaction-rich data sources - human-object interactions (HOI) and dexterous robotic manipulation - enabling cross-domain training of action-driven generative models. By integrating visual skeletons into pretrained video generation models via lightweight fine-tuning, we enable precise action control of complex interaction while preserving the learning of cross-domain dynamics. Experiments on EgoVid, RT-1 and DROID demonstrate the effectiveness of our proposed approach. Project page: https://zju3dv.github.io/VAP/.

Vision-Language-Action (VLA) models frequently encounter challenges in generalizing to real-world environments due to inherent discrepancies between observation and action spaces. Although training data are collected from diverse camera perspectives, the models typically predict end-effector poses within the robot base coordinate frame, resulting in spatial inconsistencies. To mitigate this limitation, we introduce the Observation-Centric VLA (OC-VLA) framework, which grounds action predictions directly in the camera observation space. Leveraging the camera's extrinsic calibration matrix, OC-VLA transforms end-effector poses from the robot base coordinate system into the camera coordinate system, thereby unifying prediction targets across heterogeneous viewpoints. This lightweight, plug-and-play strategy ensures robust alignment between perception and action, substantially improving model resilience to camera viewpoint variations. The proposed approach is readily compatible with existing VLA architectures, requiring no substantial modifications. Comprehensive evaluations on both simulated and real-world robotic manipulation tasks demonstrate that OC-VLA accelerates convergence, enhances task success rates, and improves cross-view generalization. The code will be publicly available.

Contemporary cardiovascular management involves complex consideration and integration of multimodal cardiac datasets, where each modality provides distinct but complementary physiological characteristics. While the effective integration of multiple modalities could yield a holistic clinical profile that accurately models the true clinical situation with respect to data modalities and their relatives weightings, current methodologies remain limited by: 1) the scarcity of patient- and time-aligned multimodal data; 2) reliance on isolated single-modality or rigid multimodal input combinations; 3) alignment strategies that prioritize cross-modal similarity over complementarity; and 4) a narrow single-task focus. In response to these limitations, a comprehensive multimodal dataset was curated for immediate application, integrating laboratory test results, electrocardiograms, and echocardiograms with clinical outcomes. Subsequently, a unified framework, Textual Guidance Multimodal fusion for Multiple cardiac tasks (TGMM), was proposed. TGMM incorporated three key components: 1) a MedFlexFusion module designed to capture the unique and complementary characteristics of medical modalities and dynamically integrate data from diverse cardiac sources and their combinations; 2) a textual guidance module to derive task-relevant representations tailored to diverse clinical objectives, including heart disease diagnosis, risk stratification and information retrieval; and 3) a response module to produce final decisions for all these tasks. Furthermore, this study systematically explored key features across multiple modalities and elucidated their synergistic contributions in clinical decision-making. Extensive experiments showed that TGMM outperformed state-of-the-art methods across multiple clinical tasks, with additional validation confirming its robustness on another public dataset.

Modern causal language models, followed by rapid developments in discrete diffusion models, can now produce a wide variety of interesting and useful content. However, these families of models are predominantly trained to output tokens with a fixed (left-to-right) or random order, which may deviate from the logical order in which tokens are generated originally. In this paper, we observe that current causal and diffusion models encounter difficulties in problems that require adaptive token generation orders to solve tractably, which we characterize with the $\mathcal{V}$-information framework. Motivated by this, we propose Reinforced Context Order Recovery (ReCOR), a reinforcement-learning-based framework to extract adaptive, data-dependent token generation orders from text data without annotations. Self-supervised by token prediction statistics, ReCOR estimates the hardness of predicting every unfilled token and adaptively selects the next token during both training and inference. Experiments on challenging reasoning and planning datasets demonstrate the superior performance of ReCOR compared with baselines, sometimes outperforming oracle models supervised with the ground-truth order.

We propose a two-stage multimodal framework that enhances disease classification and region-aware radiology report generation from chest X-rays, leveraging the MIMIC-Eye dataset. In the first stage, we introduce a gaze-guided contrastive learning architecture for disease classification. It integrates visual features, clinical labels, bounding boxes, and radiologist eye-tracking signals and is equipped with a novel multi-term gaze-attention loss combining MSE, KL divergence, correlation, and center-of-mass alignment. Incorporating fixations improves F1 score from 0.597 to 0.631 (+5.70%) and AUC from 0.821 to 0.849 (+3.41%), while also improving precision and recall, highlighting the effectiveness of gaze-informed attention supervision. In the second stage, we present a modular report generation pipeline that extracts confidence-weighted diagnostic keywords, maps them to anatomical regions using a curated dictionary constructed from domain-specific priors, and generates region-aligned sentences via structured prompts. This pipeline improves report quality as measured by clinical keyword recall and ROUGE overlap. Our results demonstrate that integrating gaze data improves both classification performance and the interpretability of generated medical reports.

Language models have made significant advancements in understanding and generating human language, achieving remarkable success in various applications. However, evaluating these models remains a challenge, particularly for resource-limited languages like Turkish. To address this issue, we introduce the Turkish MMLU (TR-MMLU) benchmark, a comprehensive evaluation framework designed to assess the linguistic and conceptual capabilities of large language models (LLMs) in Turkish. TR-MMLU is based on a meticulously curated dataset comprising 6,200 multiple-choice questions across 62 sections within the Turkish education system. This benchmark provides a standard framework for Turkish NLP research, enabling detailed analyses of LLMs' capabilities in processing Turkish text. In this study, we evaluated state-of-the-art LLMs on TR-MMLU, highlighting areas for improvement in model design. TR-MMLU sets a new standard for advancing Turkish NLP research and inspiring future innovations.

Recent studies have demonstrated that Large Language Models (LLMs) have strong mathematical reasoning abilities but rely on hundreds of billions of parameters. To tackle the challenge of poor reasoning in Small Language Models (SLMs), existing methods typically leverage LLMs to generate massive amounts of data for cramming training. In psychology, they are akin to System 1 thinking, which resolves reasoning problems rapidly based on experience and intuition. However, human learning also requires System 2 thinking, where knowledge is first acquired and then reinforced through practice. Inspired by such two distinct modes of thinking, we propose a novel method based on the multi-LoRA Interaction for mathematical reasoning Distillation (LoRID). First, we input the question and reasoning of each sample into an LLM to create knowledge-enhanced datasets. Subsequently, we train a LoRA block on the student model as an Intuitive Reasoner (IR), which directly generates Chain-of-Thoughts for problem-solving. Then, to imitate System 2 thinking, we train the Knowledge Generator (KG) and Deep Reasoner (DR), respectively. The former outputs only knowledge after receiving problems, while the latter uses that knowledge to perform reasoning. Finally, to address the randomness in the generation of IR and DR, we evaluate whether their outputs are consistent, and the inference process needs to be iterated if not. This step can enhance the mathematical reasoning ability of SLMs through mutual feedback. Experimental results show that LoRID achieves state-of-the-art performance, especially on the GSM8K dataset, where it outperforms the second-best method by 2.3%, 16.1%, 2.4%, 12.3%, and 1.8% accuracy across the five base models, respectively.

Reinforcement Learning with Verifiable Rewards (RLVR) has markedly enhanced the reasoning abilities of large language models (LLMs). Its success, however, largely depends on strong base models with rich world knowledge, yielding only modest improvements for small-size language models (SLMs). To address this limitation, we investigate Guided GRPO, which injects ground-truth reasoning steps into roll-out trajectories to compensate for SLMs' inherent weaknesses. Through a comprehensive study of various guidance configurations, we find that naively adding guidance delivers limited gains. These insights motivate G$^2$RPO-A, an adaptive algorithm that automatically adjusts guidance strength in response to the model's evolving training dynamics. Experiments on mathematical reasoning and code-generation benchmarks confirm that G$^2$RPO-A substantially outperforms vanilla GRPO. Our code and models are available at https://github.com/T-Lab-CUHKSZ/G2RPO-A.