Ziqiao Weng, Yuanzhi Liang, Haibing Huang, Weidong Cai, Xuelong Li
Post-training aims to enhance desired attributes in visual generative models while preserving their general generation capabilities. Reward-guided post-training faces two challenges: real training data may mismatch the generator's output distribution, and attribute supervision may be confounded by unrelated content. Unbounded reward maximization can also cause overoptimization and degrade generation quality. We propose a framework combining Generator-Aligned Reward Modeling (GARM) and Target-Aware Bounded Optimization (TABO). GARM mitigates distribution mismatch through generator-support matching and reduces content confounding using content-matched real contrasts, yielding generator-relevant, attribute-focused reward supervision. TABO maps the learned reward to a smooth, upper-bounded function calibrated from real target-positive examples. This transformation is independent of the optimization algorithm and supports both direct reward-gradient and policy-based optimization. We further introduce Target-Aware GRPO (TA-GRPO), an adaptation of Flow-GRPO that combines fixed stage-wise advantage scaling, a generator-to-target curriculum, and timestep-balanced credit assignment for flow-based video optimization. Experiments show that GARM improves target-attribute expression while maintaining competitive generation quality. TABO better balances attribute enhancement and visual quality than unconstrained reward maximization, while TA-GRPO improves target enhancement over vanilla Flow-GRPO and reduces unintended attribute leakage on neutral prompts. These results highlight the importance of generator-aligned reward supervision and controlled, data-calibrated optimization for targeted visual generation.
Russell Yang, Ruishi Chen, Pierce Kelaita, Riya Ranjan, Sibo Ma, Charles Dickens, Matthew Guillod, Megan Ma 等 9 位作者
Two methodologies dominate current practices of benchmarking: rubric-based scoring evaluates items against predefined criteria, whereas comparative judgment elicits pairwise preferences between outputs. Although both methodologies are widely used, the choice between them is rarely justified. We release JudgmentBench, a benchmark of 30 real-world legal tasks, paired with 1,539 rubric scores and 1,530 pairwise preference judgments collected from practicing attorneys--including at major U.S. law firms--with substantial experience. The annotations constitute the first publicly available dataset in a high-expertise domain in which both supervision signals are elicited from the same pool of experts on the same tasks. Using LLM-generated outputs at three constructed quality levels, we provide an initial empirical comparison: comparative judgments recover the intended quality ordering substantially better than rubrics under both a per-task rank-correlation metric (mean Spearman's rank correlation of 0.908 vs. 0.150, estimated difference = 0.758 [0.494, 1.021]) and a per-judgment pairwise win-rate metric (0.669 vs. 0.542, estimated difference = 0.127 [0.067, 0.186]), while requiring less than half the annotation time. The patterns hold for human annotators and LLM autograders. Beyond this initial comparison, the paired structure of the dataset supports a broader research agenda on how expert judgment should be elicited, aggregated, and used as supervision in domains without verifiable ground truth.
Scaling computer-use agents from individual actions to complex software work requires experience that connects decisions to complete outcomes. We present NorthCUA, an integrated system for turbocharging computer-use agents through experience at scale. Its Experience Flywheel supplies diverse executable tasks to the NorthCUA Harness, which sustains long interactions, records decisions, and checks task outcomes. The Flywheel uses these records to curate verified trajectories. For unsolved tasks, recursive self-improvement evolves reusable skills and preserves verified milestones so later attempts build on earlier progress. Staged Learning combines this experience with grounding data from the Flywheel to develop GUI competence through mid-training and long-horizon supervised fine-tuning. The policy then learns from fresh attempts in the same Harness through joint desktop-web reinforcement learning, closing the loop between experience production and policy improvement. The resulting Nano, Mini, and Pro family achieves state-of-the-art results among open-source models across GUI grounding, desktop execution, and Web interaction under the reported evaluation settings. For more responsive Web workflows, NorthCUA-Ultrafast pairs direct action selection with open-ended generation on pre-RL Nano, offering a measured quality-latency tradeoff. Together, these advances establish NorthCUA as a high-performing model family and a reusable foundation for agents that carry complex software work from intent to completion.
Expert deactivation at inference time changes both token routing and expert weights, and its effect on performance depends on how these changes are implemented. We compare removal, freezing, remaining-expert renormalization, and unnormalized re-routing on identical disabled sets across five instruction-tuned checkpoints from four model families, with a base OLMoE checkpoint as an additional comparison. We develop a weight-sum model that relates the loss of expert contributions and changes in branch scale to the weight rule used during training. With half of the experts deactivated according to activation-based rankings, remaining-expert renormalization yields the highest perplexity in models trained with raw router probabilities, whereas unnormalized re-routing yields the highest perplexity in models trained with normalized weights. Perplexity ranges from 1.4 to 73 times the intact baseline when the training rule is retained, and from 14 to 28,740 times the baseline under a rule mismatch. Task outcomes vary across evaluation types: on OLMoE, GSM8K strict-match accuracy falls to zero under all four implementations. Extending the model to changes in the number of active experts yields a further prediction, supported by the OLMoE experiments: the perplexity penalty from additional renormalization increases as fewer experts are activated.
Training-only privileged-modality learning uses an expensive modality during training while keeping a single-modality deployment path. We study chest radiograph (X-ray) classification with same-patient computed tomography (CT) available only during training. JDCNet adds an auxiliary CT-teacher target, the teacher's argmax label or its softmax distribution through a Kullback--Leibler term (soft-KL), only on samples whose teacher confidence exceeds a threshold; the deployed model has the same ResNet-18 computation graph as the supervised baseline. On 510 same-patient pairs from the public BIMCV-COVID19 releases of the Valencian Region Medical ImageBank (113 COVID-19, 397 non-COVID), with patient-level five-fold cross-validation and three seeds, 15 of 16 adaptively explored JDCNet configurations improved mean balanced accuracy (BA) over the supervised baseline (average $+0.020$), but none remained significant after Holm adjustment. In a pre-registered experiment with 240 batch-matched supervised restarts on the same splits, batch size had no detectable effect on the baseline (BA $0.620$ versus $0.625$), and when both methods selected one of 16 runs per cell on the validation partition, JDCNet exceeded supervised training by $+0.037$ BA (95\% confidence interval $+0.016$ to $+0.058$; patient-level $+0.007$ to $+0.068$). Logit distillation and contrastive alignment did not help. Teacher diagnostics indicate that the gate acts mainly as CT-informed sample weighting. All models, including JDCNet, performed at chance on an external radiograph set. Confidence-gated CT supervision thus gives a small within-cohort advantage on one public cohort and a binary task; generalization requires independent paired replication.
Camera-LiDAR fusion repeatedly uses an extrinsic transform in projection, association, and recursive state estimation. We study whether a targetless correction should be admitted to a downstream consumer only after separate audits of correction quality, confidence ranking, and recursive risk. Our log-disjoint protocol evaluates these stages under synthetic extrinsic drift on Argoverse 2 and treats cross-dataset and odometry experiments as bounded diagnostics. On 27 test-split logs, single-frame Canny alignment is harmful in the mild tier. Multi-frame alignment improves severe-tier rotation by 2.52 degrees but worsens translation by 0.177 m; its declared joint error-magnitude improvement is also negative (-0.158, 95% CI [-0.223, -0.098]), so the rotation gain is not a six-degree-of-freedom correction success. A zero-injection check moves away from the true extrinsic, and solver-disagreement AUROC on selected splits does not yield a validated nonzero-coverage risk-coverage operating point. In a 15-log final system role, the locked conservative policy admits no candidate and equals Raw, whereas Always-correct increases trajectory error. These results support an auditable separation of correction, confidence, and recursive contamination; they do not establish a deployable calibration gate, positive odometry or SLAM improvement, map consistency, or closed-loop safety.
For fixed nonempty words $u,v$, the minimum insertion degree of a shuffle is the least number of inserted-source runs among its representations. We prove that, after a block exchange from a minimum mask of degree $r$, the minimum over masks with commuting source projections is exactly $r-1$; a larger decrease occurs precisely when a noncommuting mask has fewer runs. If either source is unary, every such exchange is exact. We prove an explicit repair for an infinite periodic degree-three family. Computer-assisted results establish first-mask descent through combined length seventeen over every alphabet, degree-three first-mask descent through length thirty, and a uniform repair of all degree-three failures through length thirty, including all further letter identifications. Counterexamples to arbitrary, first, and last minimum-mask choices explain the need for alternative masks. Finally, we prove that minimum iteration depths have no gaps for arbitrary languages. General block-exchange descent and Singleton No-Gap remain open.
Kuan Zhang, Dongchen Liu, Qiyue Zhao, Tianyu Xin, Yue Su, Haisheng Wang, Han Yin, Hongbo Ma 等 15 位作者
The real world unfolds along a single set of physics laws, yet human intelligence demonstrates a remarkable capacity to generalize experiences from this singular physical existence into a multiverse of games, each governed by entirely different rules, aesthetics, physics, and objectives. This omni-reality adaptability is a hallmark of general intelligence. As Artificial Intelligence progresses towards Artificial General Intelligence, the multiverse of games has evolved from mere entertainment into the ultimate ground for training and evaluating AGI. The pursuit of this generality has unfolded across four eras: from environment-specific symbolic and reinforcement learning agents, to current large foundation models acting as generalist players, and toward a future creator stage where agent both creates new game worlds and continually evolves within them. We trace the full lifecycle of a generalist game player along four interdependent pillars: Dataset, Model, Harness, and Benchmark. Every advance across these pillars can be read as an attempt to break one of five fundamental trade-offs that currently bound the whole system. Building on this end-to-end view, we chart a five-step roadmap, progressing from single-game mastery to the ultimate creator stage in which the agent simultaneously creates and evolves within theoretical game multiverse. Taken together, our work offers a unified lens onto a rapidly shifting field, and a principled path toward the omnipotent generalist agent capable of seamlessly mastering any challenge within the multiverse of games, thereby paving the way for AGI.
Fengbo Ma University of Georgia, Zixin Rao, Xiaoting Li University of Georgia, Zhetao Chen, Hongyue Sun University of Georgia, Yiping Zhao University of Georgia, Xianyan Chen University of Georgia, Zhen Xiang University of Georgia
Scientific research relies on accurate information retrieval from literature to support analytical decisions.In this work, we introduce a new task, INformation reTRieval through literAture reVIEW (IntraView), which aims to automate fine-grained information retrieval faithfully grounded in the provided content in response to research-driven queries, and propose IntrAgent, an LLM-based agent that addresses this challenging task.In particular, IntrAgent is designed to mimic human behaviors when reading literature for information retrieval - identifying relevant sections and then iteratively extracting key details to refine the retrieved information.It follows a two-stage pipeline: a Section Ranking stage that prioritizes relevant literature sections through structural-knowledge-enabled reasoning, and an Iterative Reading stage that continuously extracts details and synthesizes them into concise, contextually grounded answers.To support rigorous evaluation, we introduce IntraBench, a new benchmark consisting of 315 test instances built from expert-authored questions paired with literature spanning five STEM domains.Across seven backbone LLMs, IntrAgent achieves on average 13.2% higher cross-domain accuracy than state-of-the-art RAG and research-agent baselines.
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
Yuzhen Shi Alibaba Group, Huanghai Liu Tsinghua University, Tsinghua University, Yiran HU, Song Gaojie, Xu Xinran, Yubo Ma, Tianyi Tang Alibaba Group, Li Zhang 等 30 位作者
As large language models (LLMs) are increasingly applied to legal domain-specific tasks, evaluating their ability to perform legal work in real-world settings has become essential. However, existing legal benchmarks rely on simplified and highly standardized tasks, failing to capture the ambiguity, complexity, and reasoning demands of real legal practice. Moreover, prior evaluations often adopt coarse, single-dimensional metrics and do not explicitly assess fine-grained legal reasoning. To address these limitations, we introduce PLawBench, a Practical Law Benchmark designed to evaluate LLMs in realistic legal practice scenarios. Grounded in real-world legal workflows, PLawBench models the core processes of legal practitioners through three task categories: public legal consultation, practical case analysis, and legal document generation. These tasks assess a model’s ability to identify legal issues and key facts, perform structured legal reasoning, and generate legally coherent documents. PLawBench comprises 850 questions across 13 practical legal scenarios, with each question accompanied by expert-designed evaluation rubrics, resulting in approximately 12,500 rubric items for fine-grained assessment. Using an LLM-based evaluator aligned with human expert judgments, we evaluate 10 state-of-the-art LLMs. Experimental results show that none achieves strong performance on PLawBench, revealing substantial limitations in the fine-grained legal reasoning capabilities of current LLMs and highlighting important directions for future evaluation and development of legal LLMs. Data is available at: https://anonymous.4open.science/r/PLawbench-B524/.
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
Xiaoyuan Li, Keqin Bao, Yubo Ma, Moxin Li, Wenjie Wang University of Science and Technology of China, Rui Men Alibaba Group, Yichang Zhang Alibaba Group, Fuli Feng University of Science and Technology of China 等 9 位作者
Recent advances in Large Language Models (LLMs) have shown promising results in complex reasoning tasks. However, current evaluations predominantly focus on single-turn reasoning scenarios, leaving interactive tasks largely unexplored. We attribute it to the absence of comprehensive datasets and scalable automatic evaluation protocols. To fill these gaps, we present MTR-Bench for LLMs’ Multi-Turn Reasoning evaluation. Comprising 4 classes, 40 tasks and 3600 instances, MTR-Bench covers diverse reasoning capabilities, fine-grained difficulty granularity, and necessitates multi-turn interactions with the environments. Moreover, MTR-Bench features fully-automated framework spanning both dataset constructions and model evaluations, which enables scalable assessment without human interventions. Extensive experiments reveal that even the cutting-edge reasoning models fall short of multi-turn, interactive reasoning tasks. And the further analysis upon these results brings valuable insights for future research in interactive AI systems.
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
YOLO series lead object detection with superior accuracy and speed. However, both convolutional and self-attention based architectures suffer from parameter redundancy and insufficient computational efficiency. Existing lightweight methods excessively pursue speed while ignoring the loss of important information during feature extraction and spatial transformation across different stages. Thus, effective lightweighting is crucial for detection performance. We propose YOLO-ULM, an ultra-lightweight real-time detector that achieves accelerated inference while preserving high accuracy. We innovatively design a variety of dual efficiency- and accuracy-driven modules, including efficient feature aggregation and parallel downsampling modules, as well as a more focused complete-IoU loss function. To validate our approach, we train it from scratch on COCO dataset without pretrained weights. By refining backbone parameters, we extend it to YOLO-ULM-Turbo for accelerated inference. YOLO-ULM surpasses state-of-the-art real-time detectors like YOLOv11/YOLOv12/YOLOv13 and RT-DETR. On a T4 GPU, YOLO-ULM-N achieves 41.6% mAP with an inference latency of 1.52 ms, outperforming YOLOv11-N (2.2%\uparrow) and YOLOv12-N (1.0%\uparrow). YOLO-ULM-S exceeds RT-DETR-R18 by 1.6% mAP with 64.7% fewer FLOPs and 63% fewer parameters. YOLO-ULM-L / X surpass YOLOv13-L / X by 0.7% and 0.8% respectively in mAP. YOLO-ULM-Turbo matches YOLOv12-Turbo's performance but uses less computation, with Turbo-N variant achieving 0.3% higher mAP and 16% fewer parameters than YOLOv12-Turbo-N.
尚无 AI 解读，展示原文摘要
cs.CV
Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition · 2026 · 正式发表