Jian Hu

Amala Sanjay Deshmukh, Kateryna Chumachenko, Tuomas Rintamaki et al. · amazon-science, nvidia

We introduce Nemotron 3 Nano Omni, the latest model in the Nemotron multimodal series and the first to natively support audio inputs alongside text, images, and video. Nemotron 3 Nano Omni delivers consistent accuracy improvements over its predecessor, Nemotron Nano V2 VL, across all modalities, enabled by advances in architecture, training data and recipes. In particular, Nemotron 3 delivers leading results in real-world document understanding, long audio-video comprehension, and agentic computer use. Built on the highly efficient Nemotron 3 Nano 30B-A3B backbone, Nemotron 3 Nano Omni further incorporates innovative multimodal token-reduction techniques to deliver substantially lower inference latency and higher throughput than other models of similar size. We are releasing model checkpoints in BF16, FP8, and FP4 formats, along with portions of the training data and codebase to facilitate further research and development.

Yan Wang, Yuhang Li, Ruihao Gong et al.

Extensive studies have shown that deep learning models are vulnerable to adversarial and natural noises, yet little is known about model robustness on noises caused by different system implementations. In this paper, we for the first time introduce SysNoise, a frequently occurred but often overlooked noise in the deep learning training-deployment cycle. In particular, SysNoise happens when the source training system switches to a disparate target system in deployments, where various tiny system mismatch adds up to a non-negligible difference. We first identify and classify SysNoise into three categories based on the inference stage; we then build a holistic benchmark to quantitatively measure the impact of SysNoise on 20+ models, comprehending image classification, object detection, instance segmentation and natural language processing tasks. Our extensive experiments revealed that SysNoise could bring certain impacts on model robustness across different tasks and common mitigations like data augmentation and adversarial training show limited effects on it. Together, our findings open a new research topic and we hope this work will raise research attention to deep learning deployment systems accounting for model performance. We have open-sourced the benchmark and framework at https://modeltc.github.io/systemnoise_web.

Hao Zhang, Mingjie Liu, Shaokun Zhang et al.

Multi-turn LLM agents are increasingly important for solving complex, interactive tasks, and reinforcement learning (RL) is a key ingredient for improving their long-horizon behavior. However, RL training requires generating large numbers of sandboxed rollout trajectories, and existing infrastructures often couple rollout orchestration with the training loop, making systems hard to migrate and maintain. Under the rollout-as-a-service philosophy, we present ProRL Agent , a scalable infrastructure that serves the full agentic rollout lifecycle through an API service. ProRL Agent also provides standardized and extensible sandbox environments that support diverse agentic tasks in rootless HPC settings. We validate ProRL Agent through RL training on software engineering, math, STEM, and coding tasks. ProRL Agent is open-sourced and integrated as part of NVIDIA NeMo Gym.

Jian Hu, Haowen Zhong, Junchi Yan et al.

Domain adaptation (DA) aims to transfer knowledge learned from a labeled source domain to an unlabeled or a less labeled but related target domain. Ideally, the source and target distributions should be aligned to each other equally to achieve unbiased knowledge transfer. However, due to the significant imbalance between the amount of annotated data in the source and target domains, usually only the target distribution is aligned to the source domain, leading to adapting unnecessary source specific knowledge to the target domain, i.e., biased domain adaptation. To resolve this problem, in this work, we delve into the transferability estimation problem in domain adaptation and propose a non-intrusive Unbiased Transferability Estimation Plug-in (UTEP) by modeling the uncertainty of a discriminator in adversarial-based DA methods to optimize unbiased transfer. We theoretically analyze the effectiveness of the proposed approach to unbiased transferability learning in DA. Furthermore, to alleviate the impact of imbalanced annotated data, we utilize the estimated uncertainty for pseudo label selection of unlabeled samples in the target domain, which helps achieve better marginal and conditional distribution alignments between domains. Extensive experimental results on a high variety of DA benchmark datasets show that the proposed approach can be readily incorporated into various adversarial-based DA methods, achieving state-of-the-art performance.

Qilei Li, Jiabo Huang, Jian Hu et al.

Person Re-identification (ReID) has been advanced remarkably over the last 10 years along with the rapid development of deep learning for visual recognition. However, the i.i.d. (independent and identically distributed) assumption commonly held in most deep learning models is somewhat non-applicable to ReID considering its objective to identify images of the same pedestrian across cameras at different locations often of variable and independent domain characteristics that are also subject to view-biased data distribution. In this work, we propose a Feature-Distribution Perturbation and Calibration (PECA) method to derive generic feature representations for person ReID, which is not only discriminative across cameras but also agnostic and deployable to arbitrary unseen target domains. Specifically, we perform per-domain feature-distribution perturbation to refrain the model from overfitting to the domain-biased distribution of each source (seen) domain by enforcing feature invariance to distribution shifts caused by perturbation. Furthermore, we design a global calibration mechanism to align feature distributions across all the source domains to improve the model generalization capacity by eliminating domain bias. These local perturbation and global calibration are conducted simultaneously, which share the same principle to avoid models overfitting by regularization respectively on the perturbed and the original distributions. Extensive experiments were conducted on eight person ReID datasets and the proposed PECA model outperformed the state-of-the-art competitors by significant margins.

Ximing Lu, David Acuna, Jaehun Jung et al. · uw

Reinforcement Learning with Verifiable Rewards (RLVR) has become a cornerstone for unlocking complex reasoning in Large Language Models (LLMs). Yet, scaling up RL is bottlenecked by limited existing verifiable data, where improvements increasingly saturate over prolonged training. To overcome this, we propose Golden Goose, a simple trick to synthesize unlimited RLVR tasks from unverifiable internet text by constructing a multiple-choice question-answering version of the fill-in-the-middle task. Given a source text, we prompt an LLM to identify and mask key reasoning steps, then generate a set of diverse, plausible distractors. This enables us to leverage reasoning-rich unverifiable corpora typically excluded from prior RLVR data construction (e.g., science textbooks) to synthesize GooseReason-0.7M, a large-scale RLVR dataset with over 0.7 million tasks spanning mathematics, programming, and general scientific domains. Empirically, GooseReason effectively revives models saturated on existing RLVR data, yielding robust, sustained gains under continuous RL and achieving new state-of-the-art results for 1.5B and 4B-Instruct models across 15 diverse benchmarks. Finally, we deploy Golden Goose in a real-world setting, synthesizing RLVR tasks from raw FineWeb scrapes for the cybersecurity domain, where no prior RLVR data exists. Training Qwen3-4B-Instruct on the resulting data GooseReason-Cyber sets a new state-of-the-art in cybersecurity, surpassing a 7B domain-specialized model with extensive domain-specific pre-training and post-training. This highlights the potential of automatically scaling up RLVR data by exploiting abundant, reasoning-rich, unverifiable internet text.

Wenyu Zhang, Qing Ding, Jian Hu et al.

Graph convolutional networks (GCN) is widely used to handle irregular data since it updates node features by using the structure information of graph. With the help of iterated GCN, high-order information can be obtained to further enhance the representation of nodes. However, how to apply GCN to structured data (such as pictures) has not been deeply studied. In this paper, we explore the application of graph attention networks (GAT) in image feature extraction. First of all, we propose a novel graph generation algorithm to convert images into graphs through matrix transformation. It is one magnitude faster than the algorithm based on K Nearest Neighbors (KNN). Then, GAT is used on the generated graph to update the node features. Thus, a more robust representation is obtained. These two steps are combined into a module called pixel-wise graph attention module (PGA). Since the graph obtained by our graph generation algorithm can still be transformed into a picture after processing, PGA can be well combined with CNN. Based on these two modules, we consulted the ResNet and design a pixel-wise graph attention network (PGANet). The PGANet is applied to the task of person re-identification in the datasets Market1501, DukeMTMC-reID and Occluded-DukeMTMC (outperforms state-of-the-art by 0.8\%, 1.1\% and 11\% respectively, in mAP scores). Experiment results show that it achieves the state-of-the-art performance. \href{https://github.com/wenyu1009/PGANet}{The code is available here}.

Mingjie Liu, Shizhe Diao, Ximing Lu et al. · uw

Recent advances in reasoning-centric language models have highlighted reinforcement learning (RL) as a promising method for aligning models with verifiable rewards. However, it remains contentious whether RL truly expands a model's reasoning capabilities or merely amplifies high-reward outputs already latent in the base model's distribution, and whether continually scaling up RL compute reliably leads to improved reasoning performance. In this work, we challenge prevailing assumptions by demonstrating that prolonged RL (ProRL) training can uncover novel reasoning strategies that are inaccessible to base models, even under extensive sampling. We introduce ProRL, a novel training methodology that incorporates KL divergence control, reference policy resetting, and a diverse suite of tasks. Our empirical analysis reveals that RL-trained models consistently outperform base models across a wide range of pass@k evaluations, including scenarios where base models fail entirely regardless of the number of attempts. We further show that reasoning boundary improvements correlates strongly with task competence of base model and training duration, suggesting that RL can explore and populate new regions of solution space over time. These findings offer new insights into the conditions under which RL meaningfully expands reasoning boundaries in language models and establish a foundation for future work on long-horizon RL for reasoning. We release model weights to support further research: https://huggingface.co/nvidia/Nemotron-Research-Reasoning-Qwen-1.5B

Jian Hu, Jiayi Lin, Junchi Yan et al.

Promptable segmentation typically requires instance-specific manual prompts to guide the segmentation of each desired object. To minimize such a need, task-generic promptable segmentation has been introduced, which employs a single task-generic prompt to segment various images of different objects in the same task. Current methods use Multimodal Large Language Models (MLLMs) to reason detailed instance-specific prompts from a task-generic prompt for improving segmentation accuracy. The effectiveness of this segmentation heavily depends on the precision of these derived prompts. However, MLLMs often suffer hallucinations during reasoning, resulting in inaccurate prompting. While existing methods focus on eliminating hallucinations to improve a model, we argue that MLLM hallucinations can reveal valuable contextual insights when leveraged correctly, as they represent pre-trained large-scale knowledge beyond individual images. In this paper, we utilize hallucinations to mine task-related information from images and verify its accuracy for enhancing precision of the generated prompts. Specifically, we introduce an iterative Prompt-Mask Cycle generation framework (ProMaC) with a prompt generator and a mask generator.The prompt generator uses a multi-scale chain of thought prompting, initially exploring hallucinations for extracting extended contextual knowledge on a test image.These hallucinations are then reduced to formulate precise instance-specific prompts, directing the mask generator to produce masks that are consistent with task semantics by mask semantic alignment. The generated masks iteratively induce the prompt generator to focus more on task-relevant image areas and reduce irrelevant hallucinations, resulting jointly in better prompts and masks. Experiments on 5 benchmarks demonstrate the effectiveness of ProMaC. Code given in https://lwpyh.github.io/ProMaC/.

Binfeng Xu, Hao Zhang, Shaokun Zhang et al.

Reinforcement learning for language agents increasingly depends on custom harnesses that manage long-running context, multi-turn tool use and multi-agent orchestration. However, porting these harnesses into RL environment interfaces remains difficult and often loses important training signals. We bridge this gap with polar, a rollout framework for scalable asynchronous RL over arbitrary agent harnesses. Polar treats the agent harness as a black box: it proxies LLM API calls, records token-level model interactions, and reconstructs token-faithful trajectories for training. Each rollout node efficiently manages runtime prewarming, agent execution, trajectory reconstruction, and evaluation in parallel, exposing asynchronous service endpoints that can be consumed by independent trainers at scale. This decoupled design makes Polar agnostic to agent harnesses, training infrastructure, and RL algorithms while improving compute utilization for long-running agent workloads. We validate polar by training agents on software-engineering tasks with popular coding harnesses. Using simple GRPO, polar improves Qwen3.5-4B by 22.6, 4.8, 0.6 and 6.2 points on SWE-Bench Verified with the Codex, Claude Code, Qwen Code and Pi harnesses, respectively. We further demonstrate Polar for offline data generation over custom harnesses and ablate trajectory reconstruction strategies. Polar rewrites its preceding work, Prorl Agent, and has been registered as one of NeMo Gym environments.

Jian Hu, Jiayi Lin, Weitong Cai et al.

Camouflaged object detection (COD) approaches heavily rely on pixel-level annotated datasets. Weakly-supervised COD (WSCOD) approaches use sparse annotations like scribbles or points to reduce annotation effort, but this can lead to decreased accuracy. The Segment Anything Model (SAM) shows remarkable segmentation ability with sparse prompts like points. However, manual prompt is not always feasible, as it may not be accessible in real-world application. Additionally, it only provides localization information instead of semantic one, which can intrinsically cause ambiguity in interpreting the targets. In this work, we aim to eliminate the need for manual prompt. The key idea is to employ Cross-modal Chains of Thought Prompting (CCTP) to reason visual prompts using the semantic information given by a generic text prompt. To that end, we introduce a test-time adaptation per-instance mechanism called Generalizable SAM (GenSAM) to automatically enerate and optimize visual prompts the generic task prompt for WSCOD. In particular, CCTP maps a single generic text prompt onto image-specific consensus foreground and background heatmaps using vision-language models, acquiring reliable visual prompts. Moreover, to test-time adapt the visual prompts, we further propose Progressive Mask Generation (PMG) to iteratively reweight the input image, guiding the model to focus on the targets in a coarse-to-fine manner. Crucially, all network parameters are fixed, avoiding the need for additional training. Experiments demonstrate the superiority of GenSAM. Experiments on three benchmarks demonstrate that GenSAM outperforms point supervision approaches and achieves comparable results to scribble supervision ones, solely relying on general task descriptions as prompts. our codes is in: https://lwpyh.github.io/GenSAM/.

Xinji Mai, Haotian Xu, Zhong-Zhi Li et al.

Large Language Models (LLMs) often struggle with mathematical reasoning tasks requiring precise, verifiable computation. While Reinforcement Learning (RL) from outcome-based rewards enhances text-based reasoning, understanding how agents autonomously learn to leverage external tools like code execution remains crucial. We investigate RL from outcome-based rewards for Tool-Integrated Reasoning, ZeroTIR, training base LLMs to spontaneously generate and execute Python code for mathematical problems without supervised tool-use examples. Our central contribution is we demonstrate that as RL training progresses, key metrics scale predictably. Specifically, we observe strong positive correlations where increased training steps lead to increases in the spontaneous code execution frequency, the average response length, and, critically, the final task accuracy. This suggests a quantifiable relationship between computational effort invested in training and the emergence of effective, tool-augmented reasoning strategies. We implement a robust framework featuring a decoupled code execution environment and validate our findings across standard RL algorithms and frameworks. Experiments show ZeroTIR significantly surpasses non-tool ZeroRL baselines on challenging math benchmarks. Our findings provide a foundational understanding of how autonomous tool use is acquired and scales within Agent RL, offering a reproducible benchmark for future studies. Code is released at \href{https://github.com/yyht/openrlhf_async_pipline}{https://github.com/yyht/openrlhf\_async\_pipline}.

Jian Hu, Li Tao, June Yang et al.

Learning from human preferences is crucial for language models (LMs) to effectively cater to human needs and societal values. Previous research has made notable progress by leveraging human feedback to follow instructions. However, these approaches rely primarily on online learning techniques like Proximal Policy Optimization (PPO), which have been proven unstable and challenging to tune for language models. Moreover, PPO requires complex distributed system implementation, hindering the efficiency of large-scale distributed training. In this study, we propose an offline learning from human feedback framework to align LMs without interacting with environments. Specifically, we explore filtering alignment (FA), reward-weighted regression (RWR), and conditional alignment (CA) to align language models to human preferences. By employing a loss function similar to supervised fine-tuning, our methods ensure more stable model training than PPO with a simple machine learning system~(MLSys) and much fewer (around 9\%) computing resources. Experimental results demonstrate that conditional alignment outperforms other offline alignment methods and is comparable to PPO.

Zihe Liu, Jiashun Liu, Yancheng He et al.

Reinforcement learning for LLM reasoning has rapidly emerged as a prominent research area, marked by a significant surge in related studies on both algorithmic innovations and practical applications. Despite this progress, several critical challenges remain, including the absence of standardized guidelines for employing RL techniques and a fragmented understanding of their underlying mechanisms. Additionally, inconsistent experimental settings, variations in training data, and differences in model initialization have led to conflicting conclusions, obscuring the key characteristics of these techniques and creating confusion among practitioners when selecting appropriate techniques. This paper systematically reviews widely adopted RL techniques through rigorous reproductions and isolated evaluations within a unified open-source framework. We analyze the internal mechanisms, applicable scenarios, and core principles of each technique through fine-grained experiments, including datasets of varying difficulty, model sizes, and architectures. Based on these insights, we present clear guidelines for selecting RL techniques tailored to specific setups, and provide a reliable roadmap for practitioners navigating the RL for the LLM domain. Finally, we reveal that a minimalist combination of two techniques can unlock the learning capability of critic-free policies using vanilla PPO loss. The results demonstrate that our simple combination consistently improves performance, surpassing strategies like GRPO and DAPO.

Jian Hu, Xibin Wu, Wei Shen et al.

Large Language Models (LLMs) fine-tuned via Reinforcement Learning from Human Feedback (RLHF) and Reinforcement Learning with Verifiable Rewards (RLVR) significantly improve the alignment of human-AI values, further raising the upper bound of AI capabilities, particularly in reasoning-intensive, long-context Chain-of-Thought (CoT) tasks. However, existing frameworks commonly face challenges such as inference bottlenecks and complexity barriers, which restrict their accessibility to newcomers. To bridge this gap, we introduce \textbf{OpenRLHF}, a user-friendly, scalable, and easy-to-learn open-source RLHF framework built upon Ray, vLLM, DeepSpeed, and HuggingFace Transformers, featuring a simplified design, clear code structure, and comprehensive documentation to facilitate entry for researchers and practitioners. Experimental results show that OpenRLHF achieves superior training efficiency, with speedups ranging from 1.22x to 1.68x across different model sizes, compared to state-of-the-art frameworks. Additionally, it requires significantly fewer lines of code for implementation. OpenRLHF is publicly available at https://github.com/OpenRLHF/OpenRLHF, and has already been adopted by leading institutions to accelerate RLHF research and learning.

Jian Hu, Siyang Jiang, Seth Austin Harding et al.

Many complex multi-agent systems such as robot swarms control and autonomous vehicle coordination can be modeled as Multi-Agent Reinforcement Learning (MARL) tasks. QMIX, a widely popular MARL algorithm, has been used as a baseline for the benchmark environments, e.g., Starcraft Multi-Agent Challenge (SMAC), Difficulty-Enhanced Predator-Prey (DEPP). Recent variants of QMIX target relaxing the monotonicity constraint of QMIX, allowing for performance improvement in SMAC. In this paper, we investigate the code-level optimizations of these variants and the monotonicity constraint. (1) We find that such improvements of the variants are significantly affected by various code-level optimizations. (2) The experiment results show that QMIX with normalized optimizations outperforms other works in SMAC; (3) beyond the common wisdom from these works, the monotonicity constraint can improve sample efficiency in SMAC and DEPP. We also discuss why monotonicity constraints work well in purely cooperative tasks with a theoretical analysis. We open-source the code at \url{https://github.com/hijkzzz/pymarl2}.

Zixu Cheng, Da Li, Jian Hu et al.

Video reasoning requires understanding the causal relationships between events in a video. However, such relationships are often implicit and costly to annotate manually. While existing multimodal large language models (MLLMs) often infer event relations through dense captions or video summaries for video reasoning, such modeling still lacks causal understanding. Without explicit causal structure modeling within and across video events, these models suffer from hallucinations during the video reasoning. In this work, we propose GraphThinker, a reinforcement finetuning-based method that constructs structural event-level scene graphs and enhances visual grounding to jointly reduce hallucinations in video reasoning. Specifically, we first employ an MLLM to construct an event-based video scene graph (EVSG) that explicitly models both intra- and inter-event relations, and incorporate these formed scene graphs into the MLLM as an intermediate thinking process. We also introduce a visual attention reward during reinforcement finetuning, which strengthens video grounding and further mitigates hallucinations. We evaluate GraphThinker on two datasets, RexTime and VidHalluc, where it shows superior ability to capture object and event relations with more precise event localization, reducing hallucinations in video reasoning compared to prior methods.

Yunhui Xia, Wei Shen, Yan Wang et al.

We introduce LeetCodeDataset, a high-quality benchmark for evaluating and training code-generation models, addressing two key challenges in LLM research: the lack of reasoning-focused coding benchmarks and self-contained training testbeds. By curating LeetCode Python problems with rich metadata, broad coverage, 100+ test cases per problem, and temporal splits (pre/post July 2024), our dataset enables contamination-free evaluation and efficient supervised fine-tuning (SFT). Experiments show reasoning models significantly outperform non-reasoning counterparts, while SFT with only 2.6K model-generated solutions achieves performance comparable to 110K-sample counterparts. The dataset and evaluation framework are available on Hugging Face and Github.

Jian Hu, Zixu Cheng, Chenyang Si et al.

Multi-modal Large Language Models (MLLMs) struggle with long videos due to the need for excessive visual tokens. These tokens exceed massively the context length of MLLMs, resulting in filled by redundant task-irrelevant shots. How to select shots is an unsolved critical problem: sparse sampling risks missing key details, while exhaustive sampling overwhelms the model with irrelevant content, leading to video misunderstanding. To solve this problem, we propose Chain-of-Shot prompting (CoS). The key idea is to frame shot selection as test-time visual prompt optimisation, choosing shots adaptive to video understanding semantic task by optimising shots-task alignment. CoS has two key parts: (1) a binary video summary mechanism that performs pseudo temporal grounding, discovering a binary coding to identify task-relevant shots, and (2) a video co-reasoning module that deploys the binary coding to pair (learning to align) task-relevant positive shots with irrelevant negative shots. It embeds the optimised shot selections into the original video, facilitating a focus on relevant context to optimize long video understanding. Experiments across three baselines and five datasets demonstrate the effectiveness and adaptability of CoS. Code given in https://lwpyh.github.io/CoS.

Zixu Cheng, Jian Hu, Ziquan Liu et al.

Human processes video reasoning in a sequential spatio-temporal reasoning logic, we first identify the relevant frames ("when") and then analyse the spatial relationships ("where") between key objects, and finally leverage these relationships to draw inferences ("what"). However, can Video Large Language Models (Video-LLMs) also "reason through a sequential spatio-temporal logic" in videos? Existing Video-LLM benchmarks primarily focus on assessing object presence, neglecting relational reasoning. Consequently, it is difficult to measure whether a model truly comprehends object interactions (actions/events) in videos or merely relies on pre-trained "memory" of co-occurrences as biases in generating answers. In this work, we introduce a Video Spatio-Temporal Reasoning (V-STaR) benchmark to address these shortcomings. The key idea is to decompose video understanding into a Reverse Spatio-Temporal Reasoning (RSTR) task that simultaneously evaluates what objects are present, when events occur, and where they are located while capturing the underlying Chain-of-thought (CoT) logic. To support this evaluation, we construct a dataset to elicit the spatial-temporal reasoning process of Video-LLMs. It contains coarse-to-fine CoT questions generated by a semi-automated GPT-4-powered pipeline, embedding explicit reasoning chains to mimic human cognition. Experiments from 14 Video-LLMs on our V-STaR reveal significant gaps between current Video-LLMs and the needs for robust and consistent spatio-temporal reasoning.

Jiayi Lin, Jiabo Huang, Jian Hu et al.

Visual-textual correlations in the attention maps derived from text-to-image diffusion models are proven beneficial to dense visual prediction tasks, e.g., semantic segmentation. However, a significant challenge arises due to the input distributional discrepancy between the context-rich sentences used for image generation and the isolated class names typically used in semantic segmentation. This discrepancy hinders diffusion models from capturing accurate visual-textual correlations. To solve this, we propose InvSeg, a test-time prompt inversion method that tackles open-vocabulary semantic segmentation by inverting image-specific visual context into text prompt embedding space, leveraging structure information derived from the diffusion model's reconstruction process to enrich text prompts so as to associate each class with a structure-consistent mask. Specifically, we introduce Contrastive Soft Clustering (CSC) to align derived masks with the image's structure information, softly selecting anchors for each class and calculating weighted distances to push inner-class pixels closer while separating inter-class pixels, thereby ensuring mask distinction and internal consistency. By incorporating sample-specific context, InvSeg learns context-rich text prompts in embedding space and achieves accurate semantic alignment across modalities. Experiments show that InvSeg achieves state-of-the-art performance on the PASCAL VOC, PASCAL Context and COCO Object datasets.

Jian Hu, Zixu Cheng, Shaogang Gong

Task-generic promptable image segmentation aims to achieve segmentation of diverse samples under a single task description by utilizing only one task-generic prompt. Current methods leverage the generalization capabilities of Vision-Language Models (VLMs) to infer instance-specific prompts from these task-generic prompts in order to guide the segmentation process. However, when VLMs struggle to generalise to some image instances, predicting instance-specific prompts becomes poor. To solve this problem, we introduce \textbf{I}nstance-specific \textbf{N}egative Mining for \textbf{T}ask-Generic Promptable Segmentation (\textbf{INT}). The key idea of INT is to adaptively reduce the influence of irrelevant (negative) prior knowledge whilst to increase the use the most plausible prior knowledge, selected by negative mining with higher contrast, in order to optimise instance-specific prompts generation. Specifically, INT consists of two components: (1) instance-specific prompt generation, which progressively fliters out incorrect information in prompt generation; (2) semantic mask generation, which ensures each image instance segmentation matches correctly the semantics of the instance-specific prompts. INT is validated on six datasets, including camouflaged objects and medical images, demonstrating its effectiveness, robustness and scalability.

Jian Hu, Mingjie Liu, Ximing Lu et al. · uw

Reinforcement Learning with Verifiable Rewards (RLVR) has emerged as a key ingredient for unlocking complex reasoning capabilities in large language models. Recent work ProRL has shown promise in scaling RL by increasing the number of training steps. However, performance plateaus after thousands of steps, with clear diminishing returns from allocating more computation to additional training. In this work, we investigate a complementary paradigm for scaling RL, BroR-Lincreasing the number of rollouts per example to hundreds to exhaustively Broaden exploration, which yields continuous performance gains beyond the saturation point observed in ProRL when scaling the number of training steps. Our approach is motivated by a mass balance equation analysis allowing us to characterize the rate of change in probability mass for correct and incorrect tokens during the reinforcement process. We show that under a one-step RL assumption, sampled rollout tokens always contribute to correct-mass expansion, while unsampled tokens outside rollouts may lead to gains or losses depending on their distribution and the net reward balance. Importantly, as the number of rollouts per example N increases, the effect of unsampled terms diminishes, ensuring overall correct-mass expansion. To validate our theoretical analysis, we conduct simulations under more relaxed conditions and find that a sufficiently large rollout size N-corresponding to ample exploration-guarantees an increase in the probability mass of all correct tokens. Empirically, BroRL revives models saturated after 3K ProRL training steps and demonstrates robust, continuous improvement, achieving state-of-the-art results for the 1.5B model across diverse benchmarks.

Jian Hu, Zixu Cheng, Shaogang Gong et al.

Video Temporal Grounding (TG) aims to temporally locate video segments matching a natural language description (a query) in a long video. While Vision-Language Models (VLMs) are effective at holistic semantic matching, they often struggle with fine-grained temporal localisation. Recently, Group Relative Policy Optimisation (GRPO) reformulates the inference process as a reinforcement learning task, enabling fine-grained grounding and achieving strong in-domain performance. However, GRPO relies on labelled data, making it unsuitable in unlabelled domains. Moreover, because videos are large and expensive to store and process, performing full-scale adaptation introduces prohibitive latency and computational overhead, making it impractical for real-time deployment. To overcome both problems, we introduce a Data-Efficient Unlabelled Cross-domain Temporal Grounding method, from which a model is first trained on a labelled source domain, then adapted to a target domain using only a small number of unlabelled videos from the target domain. This approach eliminates the need for target annotation and keeps both computational and storage overhead low enough to run in real time. Specifically, we introduce. Uncertainty-quantified Rollout Policy Adaptation (URPA) for cross-domain knowledge transfer in learning video temporal grounding without target labels. URPA generates multiple candidate predictions using GRPO rollouts, averages them to form a pseudo label, and estimates confidence from the variance across these rollouts. This confidence then weights the training rewards, guiding the model to focus on reliable supervision. Experiments on three datasets across six cross-domain settings show that URPA generalises well using only a few unlabelled target videos. Codes will be released once published.

Mingjie Liu, Shizhe Diao, Jian Hu et al.

Recent advancements in reasoning-focused language models such as OpenAI's O1 and DeepSeek-R1 have shown that scaling test-time computation-through chain-of-thought reasoning and iterative exploration-can yield substantial improvements on complex tasks like mathematics and code generation. These breakthroughs have been driven by large-scale reinforcement learning (RL), particularly when combined with verifiable reward signals that provide objective and grounded supervision. In this report, we investigate the effects of prolonged reinforcement learning on a small language model across a diverse set of reasoning domains. Our work identifies several key ingredients for effective training, including the use of verifiable reward tasks, enhancements to Group Relative Policy Optimization (GRPO), and practical techniques to improve training stability and generalization. We introduce controlled KL regularization, clipping ratio, and periodic reference policy resets as critical components for unlocking long-term performance gains. Our model achieves significant improvements over strong baselines, including +14.7% on math, +13.9% on coding, and +54.8% on logic puzzle tasks. To facilitate continued research, we release our model publicly.

Jian Hu, Dimitrios Korkinof, Shaogang Gong et al.

We introduce ViSMap: Unsupervised Video Summarisation by Meta Prompting, a system to summarise hour long videos with no-supervision. Most existing video understanding models work well on short videos of pre-segmented events, yet they struggle to summarise longer videos where relevant events are sparsely distributed and not pre-segmented. Moreover, long-form video understanding often relies on supervised hierarchical training that needs extensive annotations which are costly, slow and prone to inconsistency. With ViSMaP we bridge the gap between short videos (where annotated data is plentiful) and long ones (where it's not). We rely on LLMs to create optimised pseudo-summaries of long videos using segment descriptions from short ones. These pseudo-summaries are used as training data for a model that generates long-form video summaries, bypassing the need for expensive annotations of long videos. Specifically, we adopt a meta-prompting strategy to iteratively generate and refine creating pseudo-summaries of long videos. The strategy leverages short clip descriptions obtained from a supervised short video model to guide the summary. Each iteration uses three LLMs working in sequence: one to generate the pseudo-summary from clip descriptions, another to evaluate it, and a third to optimise the prompt of the generator. This iteration is necessary because the quality of the pseudo-summaries is highly dependent on the generator prompt, and varies widely among videos. We evaluate our summaries extensively on multiple datasets; our results show that ViSMaP achieves performance comparable to fully supervised state-of-the-art models while generalising across domains without sacrificing performance. Code will be released upon publication.

Jian Hu, Jason Klein Liu, Haotian Xu et al.

Reinforcement Learning from Human Feedback~(RLHF) plays a crucial role in aligning Large Language Models~(LLMs). The dominant algorithm, Proximal Policy Optimization~(PPO), employs a critic network to estimate advantages, which introduces significant computational and memory overhead. To address this, a family of critic-free algorithms (e.g., GRPO, RLOO) has emerged. However, these methods typically rely on \textit{prompt-level (local)} advantage normalization, which suffers from inaccurate advantage estimation, a tendency to overfit, and, as we show, is a theoretically biased estimator. To solve these challenges, we introduce REINFORCE++, a critic-free framework centered on \textbf{Global Advantage Normalization}. By normalizing advantages across the entire global batch rather than small, prompt-specific groups, our method provides a more stable and theoretically sound, \textit{effectively unbiased} estimate (whose bias vanishes as batch size increases). We introduce two variants: REINFORCE++, a highly efficient and general algorithm ($k \ge 1$) for general-domain RLHF, and REINFORCE++ /w baseline, a robust group-sampling variant ($k > 1$) for complex reasoning tasks. Our empirical evaluation demonstrates that each variant shows superior stability and performance in its respective domain, outperforming existing methods and even PPO in complex agentic settings.

Junquan Deng, Wei Shi, Jian Hu et al.

We consider the mobile localization problem in future millimeter-wave wireless networks with distributed Base Stations (BSs) based on multi-antenna channel state information (CSI). For this problem, we propose a Semi-supervised tdistributed Stochastic Neighbor Embedding (St-SNE) algorithm to directly embed the high-dimensional CSI samples into the 2D geographical map. We evaluate the performance of St-SNE in a simulated urban outdoor millimeter-wave radio access network. Our results show that St-SNE achieves a mean localization error of 6.8 m with only 5% of labeled CSI samples in a 200*200 m^2 area with a ray-tracing channel model. St-SNE does not require accurate synchronization among multiple BSs, and is promising for future large-scale millimeter-wave localization.

Jian Hu, Hongya Tuo, Shizhao Zhang et al.

Partial Domain adaptation (PDA) aims to solve a more practical cross-domain learning problem that assumes target label space is a subset of source label space. However, the mismatched label space causes significant negative transfer. A traditional solution is using soft weights to increase weights of source shared domain and reduce those of source outlier domain. But it still learns features of outliers and leads to negative immigration. The other mainstream idea is to distinguish source domain into shared and outlier parts by hard binary weights, while it is unavailable to correct the tangled shared and outlier classes. In this paper, we propose an end-to-end Self-Adaptive Partial Domain Adaptation(SAPDA) Network. Class weights evaluation mechanism is introduced to dynamically self-rectify the weights of shared, outlier and confused classes, thus the higher confidence samples have the more sufficient weights. Meanwhile it can eliminate the negative transfer caused by the mismatching of label space greatly. Moreover, our strategy can efficiently measure the transferability of samples in a broader sense, so that our method can achieve competitive results on unsupervised DA task likewise. A large number of experiments on multiple benchmarks have demonstrated the effectiveness of our SAPDA.

Shizhao Zhang, Hongya Tuo, Jian Hu et al.

Domain shift is a major challenge for object detectors to generalize well to real world applications. Emerging techniques of domain adaptation for two-stage detectors help to tackle this problem. However, two-stage detectors are not the first choice for industrial applications due to its long time consumption. In this paper, a novel Domain Adaptive YOLO (DA-YOLO) is proposed to improve cross-domain performance for one-stage detectors. Image level features alignment is used to strictly match for local features like texture, and loosely match for global features like illumination. Multi-scale instance level features alignment is presented to reduce instance domain shift effectively , such as variations in object appearance and viewpoint. A consensus regularization to these domain classifiers is employed to help the network generate domain-invariant detections. We evaluate our proposed method on popular datasets like Cityscapes, KITTI, SIM10K and etc.. The results demonstrate significant improvement when tested under different cross-domain scenarios.

Jian Hu, Seth Austin Harding, Haibin Wu et al.

In Cooperative Multi-Agent Reinforcement Learning (MARL) and under the setting of Centralized Training with Decentralized Execution (CTDE), agents observe and interact with their environment locally and independently. With local observation and random sampling, the randomness in rewards and observations leads to randomness in long-term returns. Existing methods such as Value Decomposition Network (VDN) and QMIX estimate the value of long-term returns as a scalar that does not contain the information of randomness. Our proposed model QR-MIX introduces quantile regression, modeling joint state-action values as a distribution, combining QMIX with Implicit Quantile Network (IQN). However, the monotonicity in QMIX limits the expression of joint state-action value distribution and may lead to incorrect estimation results in non-monotonic cases. Therefore, we proposed a flexible loss function to approximate the monotonicity found in QMIX. Our model is not only more tolerant of the randomness of returns, but also more tolerant of the randomness of monotonic constraints. The experimental results demonstrate that QR-MIX outperforms the previous state-of-the-art method QMIX in the StarCraft Multi-Agent Challenge (SMAC) environment.

Youming Lei, Jian Hu, Jianpeng Ding

We propose a hybrid method combining the deep long short-term memory (LSTM) model with the inexact empirical model of dynamical systems to predict high-dimensional chaotic systems. The deep hierarchy is encoded into the LSTM by superimposing multiple recurrent neural network layers and the hybrid model is trained with the Adam optimization algorithm. The statistical results of the Mackey-Glass system and the Kuramoto-Sivashinsky system are obtained under the criteria of root mean square error (RMSE) and anomaly correlation coefficient (ACC) using the singe-layer LSTM, the multi-layer LSTM, and the corresponding hybrid method, respectively. The numerical results show that the proposed method can effectively avoid the rapid divergence of the multi-layer LSTM model when reconstructing chaotic attractors, and demonstrate the feasibility of the combination of deep learning based on the gradient descent method and the empirical model.

Xuanda Chen, Ziyu Xiong, Jian Hu

Vocal aging, a universal process of human aging, can largely affect one's language use, possibly including some subtle acoustic features of one's utterances like Voice Onset Time. To figure out the time effects, Queen Elizabeth's Christmas speeches are documented and analyzed in the long-term trend. We build statistical models of time dependence in Voice Onset Time, controlling a wide range of other fixed factors, to present annual variations and the simulated trajectory. It is revealed that the variation range of Voice Onset Time has been narrowing over fifty years with a slight reduction in the mean value, which, possibly, is an effect of diminishing exertion, resulting from subdued muscle contraction, transcending other non-linguistic factors in forming Voice Onset Time patterns over a long time.