Zijun Wang

Minhua Lin, Juncheng Wu, Zijun Wang et al.

LLM agents are increasingly deployed as systems built around editable external harnesses, including prompts, skills, memories and tools, that shape task execution without changing model parameters. Harness self-evolution adapts such agents by updating these harnesses from execution evidence. Yet it remains unclear whether a model's base capability in task-solving predicts its capabilities in harness self-evolution: which models produce useful harness updates, and which actually benefit from them? We analyze two harness self-evolution capabilities: (i) harness-updating, the capability to produce useful persistent harness updates from execution evidence; (ii) harness-benefit, the capability to benefit from updated harnesses during task solving. Our analysis reveals two findings. First, harness-updating is flat in base capability: models from different capability tiers produce harness updates that lead to surprisingly similar gains; even Qwen3.5-9B's updates yield gains comparable to those of Claude Opus~4.6. Second, harness-benefit is non-monotonic in base capability: weak-tier models benefit little from updated harnesses, mid-tier models benefit most, and strong-tier models benefit less than mid-tier. We trace low gains at the weak tier to two failure modes: weak-tier models may fail to activate relevant harness artifacts, or activate them but fail to follow them faithfully. These findings suggest investing capability budget in the task-solving agent rather than the evolver, and targeting harness invocation and long-horizon instruction following in agent training. Our source code is publicly available at https://github.com/A-EVO-Lab/a-evolve/tree/release/harness-evolution.

Haoqin Tu, Chenhang Cui, Zijun Wang et al. · oxford

This work focuses on the potential of Vision LLMs (VLLMs) in visual reasoning. Different from prior studies, we shift our focus from evaluating standard performance to introducing a comprehensive safety evaluation suite, covering both out-of-distribution (OOD) generalization and adversarial robustness. For the OOD evaluation, we present two novel VQA datasets, each with one variant, designed to test model performance under challenging conditions. In exploring adversarial robustness, we propose a straightforward attack strategy for misleading VLLMs to produce visual-unrelated responses. Moreover, we assess the efficacy of two jailbreaking strategies, targeting either the vision or language component of VLLMs. Our evaluation of 21 diverse models, ranging from open-source VLLMs to GPT-4V, yields interesting observations: 1) Current VLLMs struggle with OOD texts but not images, unless the visual information is limited; and 2) These VLLMs can be easily misled by deceiving vision encoders only, and their vision-language training often compromise safety protocols. We release this safety evaluation suite at https://github.com/UCSC-VLAA/vllm-safety-benchmark.

Zijun Wang, Haoqin Tu, Weidong Zhou et al.

Everyday tasks come with a target, and pretraining models around this target is what turns them into experts. In this paper, we study target-oriented language model (LM) pretraining by introducing Neuron-Activated Graph Ranking (NAG-based Ranking), a training-free and interpretable framework for target pretraining data selection. Rather than using black-box representations, our approach directly characterizes each target input by a sparse set of high-impact neurons in any off-the-shelf LLMs. Concretely, we quantify neuron impact and select the most influential neurons across layers into a compact Neuron-Activated Graph (NAG), and rank candidate data by NAG similarity to target examples. We conduct experiments across six benchmarks, where our NAG-based Ranking improves target-oriented pretraining by 4.9% on average over random sampling, and also outperforms state-of-the-art baselines by 5.3% accuracy on HellaSwag. It also remains effective under a more applicable multi-target setting, where our best setup surpasses two baselines by 1.1% and 4.1%, respectively. Furthermore, we provide a comprehensive analysis on why and how our NAG works, e.g., deactivating NAG-selected neurons (only 0.12% of all) causes a 23.5% performance collapse, and restricting NAG to the final layer incurs a 4.1% average drop, indicating that NAG captures a sparse "functional backbone" for learning target features. We release the code at https://github.com/asillycat/NAG.

Fei Shen, Zhe Wang, Zijun Wang et al.

Vision-based pattern identification (such as face, fingerprint, iris etc.) has been successfully applied in human biometrics for a long history. However, dog nose-print authentication is a challenging problem since the lack of a large amount of labeled data. For that, this paper presents our proposed methods for dog nose-print authentication (Re-ID) task in CVPR 2022 pet biometric challenge. First, considering the problem that each class only with few samples in the training set, we propose an automatic offline data augmentation strategy. Then, for the difference in sample styles between the training and test datasets, we employ joint cross-entropy, triplet and pair-wise circle losses function for network optimization. Finally, with multiple models ensembled adopted, our methods achieve 86.67\% AUC on the test set. Codes are available at https://github.com/muzishen/Pet-ReID-IMAG.

Yiming Li, Sihang Li, Xinhao Liu et al.

Monocular scene understanding is a foundational component of autonomous systems. Within the spectrum of monocular perception topics, one crucial and useful task for holistic 3D scene understanding is semantic scene completion (SSC), which jointly completes semantic information and geometric details from RGB input. However, progress in SSC, particularly in large-scale street views, is hindered by the scarcity of high-quality datasets. To address this issue, we introduce SSCBench, a comprehensive benchmark that integrates scenes from widely used automotive datasets (e.g., KITTI-360, nuScenes, and Waymo). SSCBench follows an established setup and format in the community, facilitating the easy exploration of SSC methods in various street views. We benchmark models using monocular, trinocular, and point cloud input to assess the performance gap resulting from sensor coverage and modality. Moreover, we have unified semantic labels across diverse datasets to simplify cross-domain generalization testing. We commit to including more datasets and SSC models to drive further advancements in this field.

Juncheng Wu, Letian Zhang, Yuhan Wang et al.

Large language models (LLMs) and agentic systems have shown promise for clinical decision support, but existing works largely assume that evidence has already been curated and handed to the model. Real-world clinical workflows instead require agents to actively seek, iteratively plan, and synthesize multimodal evidence from heterogeneous sources. In this paper, we introduce ClinSeekAgent, an automated agentic framework for dynamic multimodal evidence seeking that shifts the paradigm from passive evidence consumption to active evidence acquisition. Given only a clinical query and access to raw data sources, ClinSeekAgent gathers evidence by querying medical knowledge bases, navigating raw EHRs, and invoking medical imaging tools; refines its hypotheses as new information emerges; and integrates the collected evidence into grounded clinical decisions. ClinSeekAgent serves both as an inference-time agent for frontier LLMs and as a training-time pipeline for distilling high-quality agent trajectories into compact open-source models. To validate its inference-time effectiveness, we construct ClinSeek-Bench, which pairs Curated Input reasoning from fixed pre-selected evidence with Automated Evidence-Seeking over raw clinical data. On text-only EHR tasks, ClinSeekAgent improves Claude Opus 4.6 from 60.0 to 63.2 overall F1 and MiniMax M2.5 from 43.1 to 47.3, with positive risk-prediction gains in 7 out of 9 evaluated host models. On multimodal tasks, ClinSeekAgent improves Claude Opus 4.6 from 47.5 to 62.6 (+15.1); all evaluated models improve across the three CXR-related task groups. We further validate ClinSeekAgent as a training pipeline by distilling agentic evidence-seeking trajectories into ClinSeek-35B-A3B, which achieves 34.0 average F1 on existing AgentEHR-Bench, improving over its Qwen3.5-35B-A3B baseline by +11.9 points and approaching Claude Opus 4.6.

Qijun Han, Haoqin Tu, Zijun Wang et al.

Autonomous GUI agents face two fundamental challenges: early stopping, where agents prematurely declare success without verifiable evidence, and repetitive loops, where agents cycle through the same failing actions without recovery. We present VLAA-GUI, a modular GUI agentic framework built around three integrated components that guide the system on when to Stop, Recover, and Search. First, a mandatory Completeness Verifier enforces UI-observable success criteria and verification at every finish step -- with an agent-level verifier that cross-examines completion claims with decision rules, rejecting those lacking direct visual evidence. Second, a mandatory Loop Breaker provides multi-tier filtering: switching interaction mode after repeated failures, forcing strategy changes after persistent screen-state recurrence, and binding reflection signals to strategy shifts. Third, an on-demand Search Agent searches online for unfamiliar workflows by directly querying a capable LLM with search ability, returning results as plain text. We additionally integrate a Coding Agent for code-intensive actions and a Grounding Agent for precise action grounding, both invoked on demand when required. We evaluate VLAA-GUI across five top-tier backbones, including Opus 4.5, 4.6 and Gemini 3.1 Pro, on two benchmarks with Linux and Windows tasks, achieving top performance on both (77.5% on OSWorld and 61.0% on WindowsAgentArena). Notably, three of the five backbones surpass human performance (72.4%) on OSWorld in a single pass. Ablation studies show that all three proposed components consistently improve a strong backbone, while a weaker backbone benefits more from these tools when the step budget is sufficient. Further analysis also shows that the Loop Breaker nearly halves wasted steps for loop-prone models.

Hardy Chen, Nancy Lau, Haoqin Tu et al.

Frontier coding agents are increasingly used in workflows where users supervise progress primarily through repeated improvement of a public score, namely the reported score on a public evaluation file with labels in the workspace, rather than through direct inspection of the agent's intermediate outputs. We study whether multi-round user pressure to improve that score induces public score exploitation: behavior that raises the public score through shortcuts without improving hidden private evaluation. We begin with a preliminary single-script tabular classification task, where GPT-5.4 and Claude Opus 4.6 both exploit label information within 10 rounds of user-agent interaction. We then build AgentPressureBench, a 34-task machine-learning repository benchmark spanning three input modalities, and collect 1326 multi-round trajectories from 13 coding agents. On our benchmark, we observe 403 exploitative runs, spanning across all tasks. We also find that stronger models have higher exploitation rates, supported by a significant Spearman rank correlation of 0.77. Our ablation experiments show that higher user pressure leads to earlier exploitation, reducing the average first exploit round by 15.6 rounds (i.e., 19.67 to 4.08). As a mitigation, adding explicit anti-exploit wordings in prompt mostly eliminates exploitation (100% to 8.3%). We hope that our work can bring attention to more careful use of coding agents workflow, and developing more robust coding agents under user pressure. Our project page is at https://ucsc-vlaa.github.io/AgentPressureBench .

Jie Xiang, Yun Wang, Lifeng An et al.

Although existing monocular depth estimation methods have made great progress, predicting an accurate absolute depth map from a single image is still challenging due to the limited modeling capacity of networks and the scale ambiguity issue. In this paper, we introduce a fully Visual Attention-based Depth (VADepth) network, where spatial attention and channel attention are applied to all stages. By continuously extracting the dependencies of features along the spatial and channel dimensions over a long distance, VADepth network can effectively preserve important details and suppress interfering features to better perceive the scene structure for more accurate depth estimates. In addition, we utilize geometric priors to form scale constraints for scale-aware model training. Specifically, we construct a novel scale-aware loss using the distance between the camera and a plane fitted by the ground points corresponding to the pixels of the rectangular area in the bottom middle of the image. Experimental results on the KITTI dataset show that this architecture achieves the state-of-the-art performance and our method can directly output absolute depth without post-processing. Moreover, our experiments on the SeasonDepth dataset also demonstrate the robustness of our model to multiple unseen environments.

Zijun Wang, Haoqin Tu, Yuhan Wang et al.

This paper introduces STAR-1, a high-quality, just-1k-scale safety dataset specifically designed for large reasoning models (LRMs) like DeepSeek-R1. Built on three core principles -- diversity, deliberative reasoning, and rigorous filtering -- STAR-1 aims to address the critical needs for safety alignment in LRMs. Specifically, we begin by integrating existing open-source safety datasets from diverse sources. Then, we curate safety policies to generate policy-grounded deliberative reasoning samples. Lastly, we apply a GPT-4o-based safety scoring system to select training examples aligned with best practices. Experimental results show that fine-tuning LRMs with STAR-1 leads to an average 40% improvement in safety performance across four benchmarks, while only incurring a marginal decrease (e.g., an average of 1.1%) in reasoning ability measured across five reasoning tasks. Extensive ablation studies further validate the importance of our design principles in constructing STAR-1 and analyze its efficacy across both LRMs and traditional LLMs. Our project page is https://ucsc-vlaa.github.io/STAR-1.

Zijun Wang, Haoqin Tu, Jieru Mei et al.

This paper studies the vulnerabilities of transformer-based Large Language Models (LLMs) to jailbreaking attacks, focusing specifically on the optimization-based Greedy Coordinate Gradient (GCG) strategy. We first observe a positive correlation between the effectiveness of attacks and the internal behaviors of the models. For instance, attacks tend to be less effective when models pay more attention to system prompts designed to ensure LLM safety alignment. Building on this discovery, we introduce an enhanced method that manipulates models' attention scores to facilitate LLM jailbreaking, which we term AttnGCG. Empirically, AttnGCG shows consistent improvements in attack efficacy across diverse LLMs, achieving an average increase of ~7% in the Llama-2 series and ~10% in the Gemma series. Our strategy also demonstrates robust attack transferability against both unseen harmful goals and black-box LLMs like GPT-3.5 and GPT-4. Moreover, we note our attention-score visualization is more interpretable, allowing us to gain better insights into how our targeted attention manipulation facilitates more effective jailbreaking. We release the code at https://github.com/UCSC-VLAA/AttnGCG-attack.

Yiyang Zhou, Haoqin Tu, Zijun Wang et al.

We propose MIRA, a new benchmark designed to evaluate models in scenarios where generating intermediate visual images is essential for successful reasoning. Unlike traditional CoT methods that rely solely on text, tasks in MIRA require models to generate and utilize intermediate images - such as sketches, structural diagrams, or path drawings - to guide their reasoning process. This setup closely mirrors how humans solve complex problems through "drawing to think". To solve this, MIRA focuses on tasks that are intrinsically challenging and involve complex structures, spatial relationships, or reasoning steps that are difficult to express through language alone. To ensure that our evaluation data is of high-quality, we include 546 multimodal problems, annotated with intermediate visual images and final answers. We also propose a unified evaluation protocol for MIRA that spans three levels of evaluation input: direct input with image and question only, text-only CoT input with image and thinking prompts, and Visual-CoT input with both annotated image clues and textual thinking prompts. To probe the upper bound of model capacity on our benchmark, we also report pass@k and majority voting accuracies under different k settings. Experimental results show that existing multimodal large language models, including strongest private models as well as strong open-weight models, perform poorly when relying solely on textual prompts. However, when intermediate visual cues are provided, model performance improves consistently, yielding an average relative gain of 33.7% across all models and tasks. We also probe the upper bound by expanding the search space and designing textual prompts aligned with Visual-CoT, but both yield only limited improvements compared to our Visual-CoT setting. These results underscore the critical role of imagined visual information in enabling successful reasoning on MIRA.

Zhu Liu, Zijun Wang, Jinyuan Liu et al.

Thermal imaging is often compromised by dynamic, complex degradations caused by hardware limitations and unpredictable environmental factors. The scarcity of high-quality infrared data, coupled with the challenges of dynamic, intricate degradations, makes it difficult to recover details using existing methods. In this paper, we introduce thermal degradation simulation integrated into the training process via a mini-max optimization, by modeling these degraded factors as adversarial attacks on thermal images. The simulation is dynamic to maximize objective functions, thus capturing a broad spectrum of degraded data distributions. This approach enables training with limited data, thereby improving model performance.Additionally, we introduce a dual-interaction network that combines the benefits of spiking neural networks with scale transformation to capture degraded features with sharp spike signal intensities. This architecture ensures compact model parameters while preserving efficient feature representation. Extensive experiments demonstrate that our method not only achieves superior visual quality under diverse single and composited degradation, but also delivers a significant reduction in processing when trained on only fifty clear images, outperforming existing techniques in efficiency and accuracy. The source code will be available at https://github.com/LiuZhu-CV/DEAL.

Changxing Liu, Genjia Liu, Zijun Wang et al.

Vehicle-to-vehicle (V2V) cooperative autonomous driving holds great promise for improving safety by addressing the perception and prediction uncertainties inherent in single-agent systems. However, traditional cooperative methods are constrained by rigid collaboration protocols and limited generalization to unseen interactive scenarios. While LLM-based approaches offer generalized reasoning capabilities, their challenges in spatial planning and unstable inference latency hinder their direct application in cooperative driving. To address these limitations, we propose CoLMDriver, the first full-pipeline LLM-based cooperative driving system, enabling effective language-based negotiation and real-time driving control. CoLMDriver features a parallel driving pipeline with two key components: (i) an LLM-based negotiation module under an actor-critic paradigm, which continuously refines cooperation policies through feedback from previous decisions of all vehicles; and (ii) an intention-guided waypoint generator, which translates negotiation outcomes into executable waypoints. Additionally, we introduce InterDrive, a CARLA-based simulation benchmark comprising 10 challenging interactive driving scenarios for evaluating V2V cooperation. Experimental results demonstrate that CoLMDriver significantly outperforms existing approaches, achieving an 11% higher success rate across diverse highly interactive V2V driving scenarios. Code will be released on https://github.com/cxliu0314/CoLMDriver.

Zhaofang Qian, Hardy Chen, Zeyu Wang et al.

Vision-language models (VLMs) have advanced rapidly, yet their capacity for image-grounded geolocation in open-world conditions, a task that is challenging and of demand in real life, has not been comprehensively evaluated. We present EarthWhere, a comprehensive benchmark for VLM image geolocation that evaluates visual recognition, step-by-step reasoning, and evidence use. EarthWhere comprises 810 globally distributed images across two complementary geolocation scales: WhereCountry (i.e., 500 multiple-choice question-answering, with country-level answer and panoramas) and WhereStreet (i.e., 310 fine-grained street-level identification tasks requiring multi-step reasoning with optional web search). For evaluation, we adopt the final-prediction metrics: location accuracies within k km (Acc@k) for coordinates and hierarchical path scores for textual localization. Beyond this, we propose to explicitly score intermediate reasoning chains using human-verified key visual clues and a Shapley-reweighted thinking score that attributes credit to each clue's marginal contribution. We benchmark 13 state-of-the-art VLMs with web searching tools on our EarthWhere and report different types of final answer accuracies as well as the calibrated model thinking scores. Overall, Gemini-2.5-Pro achieves the best average accuracy at 56.32%, while the strongest open-weight model, GLM-4.5V, reaches 34.71%. We reveal that web search and reasoning do not guarantee improved performance when visual clues are limited, and models exhibit regional biases, achieving up to 42.7% higher scores in certain areas than others. These findings highlight not only the promise but also the persistent challenges of models to mitigate bias and achieve robust, fine-grained localization. We open-source our benchmark at https://github.com/UCSC-VLAA/EarthWhere.

Siyan Dong, Zijun Wang, Lulu Cai et al.

Real-time dense scene reconstruction during unstable camera motions is crucial for robotics, yet current RGB-D SLAM systems fail when cameras experience large viewpoint changes, fast motions, or sudden shaking. Classical optimization-based methods deliver high accuracy but fail with poor initialization during large motions, while learning-based approaches provide robustness but lack sufficient accuracy for dense reconstruction. We address this challenge through a combination of learning-based initialization with optimization-based refinement. Our method employs a camera pose regression network to predict metric-aware relative poses from consecutive RGB-D frames, which serve as reliable starting points for a randomized optimization algorithm that further aligns depth images with the scene geometry. Extensive experiments demonstrate promising results: our approach outperforms the best competitor on challenging benchmarks, while maintaining comparable accuracy on stable motion sequences. The system operates in real-time, showcasing that combining simple and principled techniques can achieve both robustness for unstable motions and accuracy for dense reconstruction. Project page: https://github.com/siyandong/PROFusion.

Wei Tan, Shiyu Chang, Liana Fong et al.

Matrix factorization (MF) discovers latent features from observations, which has shown great promises in the fields of collaborative filtering, data compression, feature extraction, word embedding, etc. While many problem-specific optimization techniques have been proposed, alternating least square (ALS) remains popular due to its general applicability e.g. easy to handle positive-unlabeled inputs, fast convergence and parallelization capability. Current MF implementations are either optimized for a single machine or with a need of a large computer cluster but still are insufficient. This is because a single machine provides limited compute power for large-scale data while multiple machines suffer from the network communication bottleneck. To address the aforementioned challenge, accelerating ALS on graphics processing units (GPUs) is a promising direction. We propose the novel approach in enhancing the MF efficiency via both memory optimization and approximate computing. The former exploits GPU memory hierarchy to increase data reuse, while the later reduces unnecessary computing without hurting the convergence of learning algorithms. Extensive experiments on large-scale datasets show that our solution not only outperforms the competing CPU solutions by a large margin but also has a 2x-4x performance gain compared to the state-of-the-art GPU solutions. Our implementations are open-sourced and publicly available.

Junqiao Wang, Zeng Zhang, Yangfan He et al.

With the rapid evolution of large language models (LLM), reinforcement learning (RL) has emerged as a pivotal technique for code generation and optimization in various domains. This paper presents a systematic survey of the application of RL in code optimization and generation, highlighting its role in enhancing compiler optimization, resource allocation, and the development of frameworks and tools. Subsequent sections first delve into the intricate processes of compiler optimization, where RL algorithms are leveraged to improve efficiency and resource utilization. The discussion then progresses to the function of RL in resource allocation, emphasizing register allocation and system optimization. We also explore the burgeoning role of frameworks and tools in code generation, examining how RL can be integrated to bolster their capabilities. This survey aims to serve as a comprehensive resource for researchers and practitioners interested in harnessing the power of RL to advance code generation and optimization techniques.

Zijun Wang, Yijiahao Qi, Hanqiu Chen et al.

Mixture-of-Agents (MoA) inference can suffer from dense inter-agent communication and low hardware utilization, which jointly inflate serving latency. We present a serving design that targets these bottlenecks through an algorithm-system co-design. First, we replace dense agent interaction graphs with a hierarchical tree topology that induces structured sparsity in inter-agent communication. Second, we introduce a runtime adaptive mechanism that selectively terminates or skips downstream agent invocations using semantic agreement and confidence signals from intermediate outputs. Third, we pipeline agent execution by overlapping incremental prefilling with decoding across dependency-related agents, improving utilization and reducing inference latency. Across representative tasks, this approach substantially reduces end-to-end latency (up to 90%) while maintaining comparable accuracy (within $\pm$1%) relative to dense-connectivity MoA baselines, and can improve accuracy in certain settings.

Zijun Wang, Haoqin Tu, Letian Zhang et al.

OpenClaw, the most widely deployed personal AI agent in early 2026, operates with full local system access and integrates with sensitive services such as Gmail, Stripe, and the filesystem. While these broad privileges enable high levels of automation and powerful personalization, they also expose a substantial attack surface that existing sandboxed evaluations fail to capture. To address this gap, we present the first real-world safety evaluation of OpenClaw and introduce the CIK taxonomy, which unifies an agent's persistent state into three dimensions, i.e., Capability, Identity, and Knowledge, for safety analysis. Our evaluations cover 12 attack scenarios on a live OpenClaw instance across four backbone models (Claude Sonnet 4.5, Opus 4.6, Gemini 3.1 Pro, and GPT-5.4). The results show that poisoning any single CIK dimension increases the average attack success rate from 24.6% to 64-74%, with even the most robust model exhibiting more than a threefold increase over its baseline vulnerability. We further assess three CIK-aligned defense strategies alongside a file-protection mechanism; however, the strongest defense still yields a 63.8% success rate under Capability-targeted attacks, while file protection blocks 97% of malicious injections but also prevents legitimate updates. Taken together, these findings show that the vulnerabilities are inherent to the agent architecture, necessitating more systematic safeguards to secure personal AI agents. Our project page is https://ucsc-vlaa.github.io/CIK-Bench.

Fengze Liu, Weidong Zhou, Binbin Liu et al.

Upweighting high-quality data in LLM pretraining often improves performance, but in datalimited regimes, especially under overtraining, stronger upweighting increases repetition and can degrade performance. However, standard scaling laws do not reliably extrapolate across mixture recipes or under repetitions, making the selection for optimal data recipes at scaling underdetermined. To solve this, we introduce InfoLaw (Information Scaling Laws), a data-aware scaling framework that predicts loss from consumed tokens, model size, data mixture weights, and repetition. The key idea is to model pretraining as information accumulation, where quality controls information density and repetition induces scaledependent diminishing returns. We first collect the model performance after training on datasets that vary in scale, quality distribution, and repetition level. Then we build up the modeling for information so that information accurately predicts those model performance. InfoLaw predicts performance on unseen data recipes and larger scale runs (up to 7B, 425B tokens) with 0.15% mean and 0.96% max absolute error in loss, and it extrapolates reliably across overtraining levels, enabling efficient data-recipe selection under varying compute budgets.

Zemin Huang, Zhiyang Chen, Zijun Wang et al.

We introduce the Diffusion Chain of Lateral Thought (DCoLT), a reasoning framework for diffusion language models. DCoLT treats each intermediate step in the reverse diffusion process as a latent "thinking" action and optimizes the entire reasoning trajectory to maximize the reward on the correctness of the final answer with outcome-based Reinforcement Learning (RL). Unlike traditional Chain-of-Thought (CoT) methods that follow a causal, linear thinking process, DCoLT allows bidirectional, non-linear reasoning with no strict rule on grammatical correctness amid its intermediate steps of thought. We implement DCoLT on two representative Diffusion Language Models (DLMs). First, we choose SEDD as a representative continuous-time discrete diffusion model, where its concrete score derives a probabilistic policy to maximize the RL reward over the entire sequence of intermediate diffusion steps. We further consider the discrete-time masked diffusion language model -- LLaDA, and find that the order to predict and unmask tokens plays an essential role to optimize its RL action resulting from the ranking-based Unmasking Policy Module (UPM) defined by the Plackett-Luce model. Experiments on both math and code generation tasks show that using only public data and 16 H800 GPUs, DCoLT-reinforced DLMs outperform other DLMs trained by SFT or RL or even both. Notably, DCoLT-reinforced LLaDA boosts its reasoning accuracy by +9.8%, +5.7%, +11.4%, +19.5% on GSM8K, MATH, MBPP, and HumanEval.

Tony Lee, Haoqin Tu, Chi Heem Wong et al. · stanford

Evaluations of audio-language models (ALMs) -- multimodal models that take interleaved audio and text as input and output text -- are hindered by the lack of standardized benchmarks; most benchmarks measure only one or two capabilities and omit evaluative aspects such as fairness or safety. Furthermore, comparison across models is difficult as separate evaluations test a limited number of models and use different prompting methods and inference parameters. To address these shortfalls, we introduce AHELM, a benchmark that aggregates various datasets -- including 2 new synthetic audio-text datasets called PARADE, which evaluates the ALMs on avoiding stereotypes, and CoRe-Bench, which measures reasoning over conversational audio through inferential multi-turn question answering -- to holistically measure the performance of ALMs across 10 aspects we have identified as important to the development and usage of ALMs: audio perception, knowledge, reasoning, emotion detection, bias, fairness, multilinguality, robustness, toxicity, and safety. We also standardize the prompts, inference parameters, and evaluation metrics to ensure equitable comparisons across models. We test 14 open-weight and closed-API ALMs from 3 developers and 3 additional simple baseline systems each consisting of an automatic speech recognizer and a language model. Our results show that while Gemini 2.5 Pro ranks top in 5 out of 10 aspects, it exhibits group unfairness ($p=0.01$) on ASR tasks whereas most of the other models do not. We also find that the baseline systems perform reasonably well on AHELM, with one ranking 6th overall despite having only speech-to-text capabilities. For transparency, all raw prompts, model generations, and outputs are available on our website at https://crfm.stanford.edu/helm/audio/v1.0.0. AHELM is intended to be a living benchmark and new datasets and models will be added over time.

Tiancheng Li, Yuhang Wang, Zhiyang Chen et al.

Prompt evolution algorithms offer a powerful paradigm for enhancing AI systems based on closed-source models, while few work explores whether aggregating results from multiple prompts to reach a consensus can further advance the system capability boundary. In this paper, we introduce Consensus-Evolve (C-Evolve), an evolutionary algorithm that discovers a group of prompts whose aggregated outputs after majority voting achieve optimal performance. More specifically, C-Evolve employs an island-based evolutionary algorithm to maintain population diversity, and prompts from distinct islands are selected to form groups to aggregate their outputs. The key difference from single individual evolution is a voting score, which evaluates each individual prompt's contribution within groups. We take this as the fitness score for evolution instead of individual performance. Consequently, C-Evolve is more likely to produce and maintain prompts with higher potential to form a high-performing group and eliminate low-performing ones, gradually improving the group performance after reaching consensus. Our method achieves state-of-the-art performance across a wide range of tasks, including both open-ended tasks like HotpotQA and closed-ended tasks like MATH. On Qwen3-8B, C-Evolve achieves 70.67% on HotpotQA and 43.88% on IFBench, which are 4.95% and 2.73% higher than GEPA, respectively. For GPT-4.1-mini, the accuracy on IFBench is further improved to 47.96% and reaches 95.33% in the MATH benchmark. These results demonstrate the C-Evolve's competitive performance.

Jiaqi Liu, Songning Lai, Pengze Li et al.

Automated discovery of physical laws from observational data in the real world is a grand challenge in AI. Current methods, relying on symbolic regression or LLMs, are limited to uni-modal data and overlook the rich, visual phenomenological representations of motion that are indispensable to physicists. This "sensory deprivation" severely weakens their ability to interpret the inherent spatio-temporal patterns within dynamic phenomena. To address this gap, we propose VIPER-R1, a multimodal model that performs Visual Induction for Physics-based Equation Reasoning to discover fundamental symbolic formulas. It integrates visual perception, trajectory data, and symbolic reasoning to emulate the scientific discovery process. The model is trained via a curriculum of Motion Structure Induction (MSI), using supervised fine-tuning to interpret kinematic phase portraits and to construct hypotheses guided by a Causal Chain of Thought (C-CoT), followed by Reward-Guided Symbolic Calibration (RGSC) to refine the formula structure with reinforcement learning. During inference, the trained VIPER-R1 acts as an agent: it first posits a high-confidence symbolic ansatz, then proactively invokes an external symbolic regression tool to perform Symbolic Residual Realignment (SR^2). This final step, analogous to a physicist's perturbation analysis, reconciles the theoretical model with empirical data. To support this research, we introduce PhysSymbol, a new 5,000-instance multimodal corpus. Experiments show that VIPER-R1 consistently outperforms state-of-the-art VLM baselines in accuracy and interpretability, enabling more precise discovery of physical laws. Project page: https://jiaaqiliu.github.io/VIPER-R1/

Zhuoran Duan, Guoshun Nan, Rushan Li et al.

The upcoming 6G will fundamentally reshape mobile networks beyond communications, unlocking a multitude of applications that were once considered unimaginable. Meanwhile, security and resilience are especially highlighted in the 6G design principles. However, safeguarding 6G networks will be quite challenging due to various known and unknown threats from highly heterogeneous networks and diversified security requirements of distinct use cases, calling for a comprehensive re-design of security architecture. This motivates us to propose ES3A (Entire Smart Service-based Security Architecture), a novel security architecture for 6G networks. Specifically, we first discuss six high-level principles of our ES3A that include hierarchy, flexibility, scalability, resilience, endogeny, and trust and privacy. With these goals in mind, we then introduce three guidelines from a deployment perspective, envisioning our ES3A that offers service-based security, end-to-end protection, and smart security automation for 6G networks. Our architecture consists of three layers and three domains. It relies on a two-stage orchestration mechanism to tailor smart security strategies for customized protection in high-dynamic 6G networks, thereby addressing the aforementioned challenges. Finally, we prototype the proposed ES3A on a real-world radio system based on Software-Defined Radio (SDR). Experiments show the effectiveness of our ES3A. We also provide a case to show the superiority of our architecture.

Zijun Wang, Panwen Hu, Jing Wang et al.

Recent advances in video generation have shown remarkable potential for constructing world simulators. However, current models still struggle to produce physically consistent results, particularly when handling large-scale or complex dynamics. This limitation arises primarily because existing approaches respond isotropically to physical prompts and neglect the fine-grained alignment between generated content and localized physical cues. To address these challenges, we propose ProPhy, a Progressive Physical Alignment Framework that enables explicit physics-aware conditioning and anisotropic generation. ProPhy employs a two-stage Mixture-of-Physics-Experts (MoPE) mechanism for discriminative physical prior extraction, where Semantic Experts infer semantic-level physical principles from textual descriptions, and Refinement Experts capture token-level physical dynamics. This mechanism allows the model to learn fine-grained, physics-aware video representations that better reflect underlying physical laws. Furthermore, we introduce a physical alignment strategy that transfers the physical reasoning capabilities of vision-language models (VLMs) into the Refinement Experts, facilitating a more accurate representation of dynamic physical phenomena. Extensive experiments on physics-aware video generation benchmarks demonstrate that ProPhy produces more realistic, dynamic, and physically coherent results than existing state-of-the-art methods.

Jinjin Cao, Zhiyang Chen, Zijun Wang et al.

Vision-Language Models (VLMs) have shown solid ability for multimodal understanding of both visual and language contexts. However, existing VLMs often face severe challenges of hallucinations, meaning that VLMs tend to generate responses that are only fluent in the language but irrelevant to images in previous contexts. To address this issue, we analyze how language bias contributes to hallucinations and then introduce Cross-Modal Guidance(CMG), a training-free decoding method that addresses the hallucinations by leveraging the difference between the output distributions of the original model and the one with degraded visual-language attention. In practice, we adaptively mask the attention weight of the most influential image tokens in selected transformer layers to corrupt the visual-language perception as a concrete type of degradation. Such a degradation-induced decoding emphasizes the perception of visual contexts and therefore significantly reduces language bias without harming the ability of VLMs. In experiment sections, we conduct comprehensive studies. All results demonstrate the superior advantages of CMG with neither additional conditions nor training costs. We also quantitatively show CMG can improve different VLM's performance on hallucination-specific benchmarks and generalize effectively.

Dakuo Wang, Michael Muller, Qian Yang et al.

Geographically distributed teams often face challenges in coordination and collaboration, lowering their productivity. Understanding the relationship between team dispersion and productivity is critical for supporting such teams. Extensive prior research has studied these relations in lab settings or using qualitative measures. This paper extends prior work by contributing an empirical case study in a real-world organization, using quantitative measures. We studied 117 new research project teams from the same discipline within an industrial research lab for 6 months. During this time, all teams shared one goal: submitting research papers to the same target conference. We analyzed these teams' dispersion-related characteristics as well as team productivity. Interestingly, we found little statistical evidence that geographic and time differences relate to team productivity. However, organizational and functional distances are predictive of the productivity of the dispersed teams we studied. We discuss the open research questions these findings revealed and their implications for future research.

Yuhang Li, Yuhao Zhou, Junbin Huang et al.

Achieving short-distance flight helps improve the efficiency of humanoid robots moving in complex environments (e.g., crossing large obstacles or reaching high places) for rapid emergency missions. This study proposes a design of a flying humanoid robot named Jet-HR2. The robot has 10 joints driven by brushless motors and harmonic drives for locomotion. To overcome the challenge of the stable-attitude takeoff in small thrust-to-weight conditions, the robot was designed based on the concept of thrust vectoring. The propulsion system consists of four ducted fans, that is, two fixed on the waist of the robot and the other two mounted on the feet, for thrust vector control. The thrust vector is controlled by adjusting the attitude of the foot during the flight. A simplified model and control strategies are proposed to solve the problem of attitude instability caused by mass errors and joint position errors during takeoff. The experimental results show that the robot's spin and dive behaviors during takeoff were effectively suppressed by controlling the thrust vector of the ducted fan on the foot. The robot successfully achieved takeoff at a thrust-to-weight ratio of 1.17 (17 kg / 20 kg) and maintained a stable attitude, reaching a takeoff height of over 1000 mm.