arXiv subjects

Xin Dong

Publications and source records attributed to Xin Dong.

At least 37 records · Page 2Linked to original sources

arXiv2026

Bridging the Know-Act Gap via Task-Level Autoregressive Reasoning

LLMs often generate seemingly valid answers to flawed or ill-posed inputs. This is not due to missing knowledge: under discriminative prompting, the same models can mostly identify such issues, yet fail to reflect this in standard generative responses. This reveals a fundamental know-act gap between discriminative recognition and generative behavior. Prior work largely characterizes this issue in narrow settings, such as math word problems or question answering, with limited focus on how to integrate these two modes. In this work, we present a comprehensive analysis using FaultyScience, a newly constructed large-scale, cross-disciplinary benchmark of faulty scientific questions. We show that the gap is pervasive and stems from token-level autoregression, which entangles task selection (validate vs. answer) with content generation, preventing discriminative knowledge from being utilized. To address this, we propose DeIllusionLLM, a task-level autoregressive framework that explicitly models this decision. Through self-distillation, the model unifies discriminative judgment and generative reasoning within a single backbone. Empirically, DeIllusionLLM substantially reduces answer-despite-error failures under natural prompting while maintaining general reasoning performance, demonstrating that self-distillation is an effective and scalable solution for bridging the discriminative-generative know-act gap

cs.AI↗

arXiv2026

MI-DPG: Decomposable Parameter Generation Network Based on Mutual Information for Multi-Scenario Recommendation

Conversion rate (CVR) prediction models play a vital role in recommendation and advertising systems. Recent research on multi-scenario recommendation shows that learning a unified model to serve multiple scenarios is effective for improving overall performance. However, it remains challenging to improve model prediction performance across scenarios at low model parameter cost, and current solutions are hard to robustly model multi-scenario diversity. In this paper, we propose MI-DPG for the multi-scenario CVR prediction, which learns scenario-conditioned dynamic model parameters for each scenario in a more efficient and effective manner. Specifically, we introduce an auxiliary network to generate scenario-conditioned dynamic weighting matrices, which are obtained by combining decomposed scenario-specific and scenario-shared low-rank matrices with parameter efficiency. For each scene, weighting the backbone model parameters by the weighting matrix helps to specialize the model parameters for different scenarios. It can not only modulate the complete parameter space of the backbone model but also improve the model effectiveness. Furthermore, we design a mutual information regularization to enhance the diversity of model parameters across different scenarios by maximizing the mutual information between the scenario-aware input and the scene-conditioned dynamic weighting matrix. Experiments from three real-world datasets show that MI-DPG significantly outperforms previous multi-scenario recommendation models.

cs.IR↗

arXiv2026

RA-Nav: A Risk-Aware Navigation System Based on Semantic Segmentation for Aerial Robots in Unpredictable Environments

Existing aerial robot navigation systems typically plan paths around static and dynamic obstacles, but fail to adapt when a static obstacle suddenly moves. Integrating environmental semantic awareness enables estimation of potential risks posed by suddenly moving obstacles. In this paper, we propose RA- Nav, a risk-aware navigation framework based on semantic segmentation. A lightweight multi-scale semantic segmentation network identifies obstacle categories in real time. These obstacles are further classified into three types: stationary, temporarily static, and dynamic. For each type, corresponding risk estimation functions are designed to enable real-time risk prediction, based on which a complete local risk map is constructed. Based on this map, the risk-informed path search algorithm is designed to guarantee planning that balances path efficiency and safety. Trajectory optimization is then applied to generate trajectories that are safe, smooth, and dynamically feasible. Comparative simulations demonstrate that RA-Nav achieves higher success rates than baselines in sudden obstacle state transition scenarios. Its effectiveness is further validated in simulations using real- world data.

cs.RO↗

arXiv2026

QuRL: Efficient Reinforcement Learning with Quantized Rollout

Reinforcement learning with verifiable rewards (RLVR) has become a trending paradigm for training reasoning large language models (LLMs). However, due to the autoregressive decoding nature of LLMs, the rollout process becomes the efficiency bottleneck of RL training, consisting of up to 70\% of the total training time. In this work, we propose Quantized Reinforcement Learning (QuRL) that uses a quantized actor for accelerating the rollout. We address two challenges in QuRL. First, we propose Adaptive Clipping Range (ACR) that dynamically adjusts the clipping ratio based on the policy ratio between the full-precision actor and the quantized actor, which is essential for mitigating long-term training collapse. Second, we identify the weight update problem, where weight changes between RL steps are extremely small, making it difficult for the quantization operation to capture them effectively. We mitigate this problem through the invariant scaling technique that reduces quantization noise and increases weight update. We evaluate our method with INT8 and FP8 quantization experiments on DeepScaleR and DAPO, and achieve 20% to 80% faster rollout during training.

cs.LG↗

arXiv2026

Optimizing Mirror-Image Peptide Sequence Design for Data Storage via Peptide Bond Cleavage Prediction

Traditional non-biological storage media, such as hard drives, face limitations in both storage density and lifespan due to the rapid growth of data in the big data era. Mirror-image peptides composed of D-amino acids have emerged as a promising biological storage medium due to their high storage density, structural stability, and long lifespan. The sequencing of mirror-image peptides relies on \textit{de-novo} technology. However, its accuracy is limited by the scarcity of tandem mass spectrometry datasets and the challenges that current algorithms encounter when processing these peptides directly. This study is the first to propose improving sequencing accuracy indirectly by optimizing the design of mirror-image peptide sequences. In this work, we introduce DBond, a deep neural network based model that integrates sequence features, precursor ion properties, and mass spectrometry environmental factors for the prediction of mirror-image peptide bond cleavage. In this process, sequences with a high peptide bond cleavage ratio, which are easy to sequence, are selected. The main contributions of this study are as follows. First, we constructed MiPD513, a tandem mass spectrometry dataset containing 513 mirror-image peptides. Second, we developed the peptide bond cleavage labeling algorithm (PBCLA), which generated approximately 12.5 million labeled data based on MiPD513. Third, we proposed a dual prediction strategy that combines multi-label and single-label classification. On an independent test set, the single-label classification strategy outperformed other methods in both single and multiple peptide bond cleavage prediction tasks, offering a strong foundation for sequence optimization.

q-bio.QM↗

arXiv2026

GDPO: Group reward-Decoupled Normalization Policy Optimization for Multi-reward RL Optimization

As language models become increasingly capable, users expect them to provide not only accurate responses but also behaviors aligned with diverse human preferences across a variety of scenarios. To achieve this, Reinforcement learning (RL) pipelines have begun incorporating multiple rewards, each capturing a distinct preference, to guide models toward these desired behaviors. However, recent work has defaulted to apply Group Relative Policy Optimization (GRPO) under multi-reward setting without examining its suitability. In this paper, we demonstrate that directly applying GRPO to normalize distinct rollout reward combinations causes them to collapse into identical advantage values, reducing the resolution of the training signal and resulting in suboptimal convergence and, in some cases, early training failure. We then introduce Group reward-Decoupled Normalization Policy Optimization (GDPO), a new policy optimization method to resolve these issues by decoupling the normalization of individual rewards, more faithfully preserving their relative differences and enabling more accurate multi-reward optimization, along with substantially improved training stability. We compare GDPO with GRPO across three tasks: tool calling, math reasoning, and coding reasoning, evaluating both correctness metrics (accuracy, bug ratio) and constraint adherence metrics (format, length). Across all settings, GDPO consistently outperforms GRPO, demonstrating its effectiveness and generalizability for multi-reward reinforcement learning optimization.

cs.CL↗

arXiv2025

Untethered thin dielectric elastomer actuated soft robot

Thin dielectric elastomer actuator (DEA) features a unique in-plane configuration, enabling low-profile designs capable of accessing millimetre-scale narrow spaces. However, most existing DEA-powered soft robots require high voltages and wired power connections, limiting their ability to operate in confined environments. This study presents an untethered thin soft robot (UTS-Robot) powered by thin dielectric elastomer actuators (TS-DEA). The robot measures 38 mm in length, 6 mm in height, and weighs just 2.34 grams, integrating flexible onboard electronics to achieve fully untethered actuation. The TS-DEA, operating at resonant frequencies of 86 Hz under a low driving voltage of 220 V, adopts a dual-actuation sandwiched structure, comprising four dielectric elastomer layers bonded to a compressible tensioning mechanism at its core. This design enables high power density actuation and locomotion via three directional friction pads. The low-voltage actuation is achieved by fabricating each elastomer layer via spin coating to an initial thickness of 50 um, followed by biaxial stretching to 8 um. A comprehensive design and modelling framework has been developed to optimise TS-DEA performance. Experimental evaluations demonstrate that the bare TS-DEA achieves a locomotion speed of 12.36 mm/s at resonance, the untethered configuration achieves a locomotion speed of 0.5 mm/s, making it highly suitable for navigating confined and complex environments.

cs.RO↗

arXiv2025

Nemotron-CLIMB: CLustering-based Iterative Data Mixture Bootstrapping for Language Model Pre-training

Pre-training datasets are typically collected from web content and lack inherent domain divisions. For instance, widely used datasets like Common Crawl do not include explicit domain labels, while manually curating labeled datasets such as The Pile is labor-intensive. Consequently, identifying an optimal pre-training data mixture remains a challenging problem, despite its significant benefits for pre-training performance. To address these challenges, we propose CLustering-based Iterative Data Mixture Bootstrapping (Nemotron-CLIMB), an automated framework that discovers, evaluates, and refines data mixtures in a pre-training setting. Specifically, Nemotron-CLIMB embeds and clusters large-scale datasets in a semantic space and then iteratively searches for optimal mixtures using a smaller proxy model and a predictor. When continuously trained on 400B tokens with this mixture, our 1B model exceeds the state-of-the-art Llama-3.2-1B by 2.0%. Moreover, we observe that optimizing for a specific domain (e.g., Social Sciences) yields a 5% improvement over random sampling. Finally, we introduce Nemotron-ClimbLab, a filtered 1.2-trillion-token corpus with 20 clusters as a research playground, and Nemotron-ClimbMix, a compact yet powerful 400-billion-token dataset designed for efficient pre-training that delivers superior performance under an equal token budget. We analyze the final data mixture, elucidating the characteristics of an optimal data mixture. Our data is available at: https://research.nvidia.com/labs/lpr/climb/

cs.CL↗

arXiv2025

ToolOrchestra: Elevating Intelligence via Efficient Model and Tool Orchestration

Large language models are powerful generalists, yet solving deep and complex problems such as those of the Humanity's Last Exam (HLE) remains both conceptually challenging and computationally expensive. We show that small orchestrators managing other models and a variety of tools can both push the upper bound of intelligence and improve efficiency in solving difficult agentic tasks. We introduce ToolOrchestra, a method for training small orchestrators that coordinate intelligent tools. ToolOrchestra explicitly uses reinforcement learning with outcome-, efficiency-, and user-preference-aware rewards. Using ToolOrchestra, we produce Orchestrator, an 8B model that achieves higher accuracy at lower cost than previous tool-use agents while aligning with user preferences on which tools are to be used for a given query. On HLE, Orchestrator achieves a score of 37.1%, outperforming GPT-5 (35.1%) while being 2.5x more efficient. On tau2-Bench and FRAMES, Orchestrator surpasses GPT-5 by a wide margin while using only about 30% of the cost. Extensive analysis shows that Orchestrator achieves the best trade-off between performance and cost under multiple metrics, and generalizes robustly to unseen tools. These results demonstrate that composing diverse tools with a lightweight orchestration model is both more efficient and more effective than existing methods, paving the way for practical and scalable tool-augmented reasoning systems.

cs.CL↗

arXiv2025

Nemotron-Flash: Towards Latency-Optimal Hybrid Small Language Models

Efficient deployment of small language models (SLMs) is essential for numerous real-world applications with stringent latency constraints. While previous work on SLM design has primarily focused on reducing the number of parameters to achieve parameter-optimal SLMs, parameter efficiency does not necessarily translate into proportional real-device speed-ups. This work aims to identify the key determinants of SLMs' real-device latency and offer generalizable principles and methodologies for SLM design and training when real-device latency is the primary consideration. Specifically, we identify two central architectural factors: depth-width ratios and operator choices. The former is crucial for small-batch-size latency, while the latter affects both latency and large-batch-size throughput. In light of this, we first study latency-optimal depth-width ratios, with the key finding that although deep-thin models generally achieve better accuracy under the same parameter budget, they may not lie on the accuracy-latency trade-off frontier. Next, we explore emerging efficient attention alternatives to evaluate their potential as candidate building operators. Using the identified promising operators, we construct an evolutionary search framework to automatically discover latency-optimal combinations of these operators within hybrid SLMs, thereby advancing the accuracy-latency frontier. In addition to architectural improvements, we further enhance SLM training using a weight normalization technique that enables more effective weight updates and improves final convergence. Combining these methods, we introduce a new family of hybrid SLMs, called Nemotron-Flash, which significantly advances the accuracy-efficiency frontier of state-of-the-art SLMs, e.g., achieving over +5.5% average accuracy, 1.3x/1.9x lower latency, and 18.7x/45.6x higher throughput compared to Qwen3-1.7B/0.6B, respectively.

cs.LG↗

arXiv2025

EventSTU: Event-Guided Efficient Spatio-Temporal Understanding for Video Large Language Models

Video large language models have demonstrated strong video understanding capabilities but suffer from high inference costs due to the massive number of tokens in long videos. Inspired by event-based vision, we propose an event-guided, training-free framework for efficient spatio-temporal understanding, named EventSTU. In the temporal domain, we design a coarse-to-fine keyframe sampling algorithm that exploits the change-triggered property of event cameras to eliminate redundant frames. In the spatial domain, we design an adaptive token pruning algorithm that leverages the visual saliency of events as a zero-cost prior to guide spatial reduction. From a holistic spatio-temporal perspective, we further integrate question relevance from keyframe sampling to adaptively allocate token pruning budgets. To facilitate evaluation, we construct EventBench, the first event-inclusive, human-annotated multimodal benchmark that covers diverse real-world scenarios. Beyond physical event cameras, EventSTU also supports general video understanding using simulated events. Comprehensive experiments show that EventSTU achieves 3.01x FLOPs reduction and 3.10x prefilling speedup over the strongest baseline while still improving performance.

cs.CV↗

arXiv2025

Native 3D Editing with Full Attention

Instruction-guided 3D editing is a rapidly emerging field with the potential to broaden access to 3D content creation. However, existing methods face critical limitations: optimization-based approaches are prohibitively slow, while feed-forward approaches relying on multi-view 2D editing often suffer from inconsistent geometry and degraded visual quality. To address these issues, we propose a novel native 3D editing framework that directly manipulates 3D representations in a single, efficient feed-forward pass. Specifically, we create a large-scale, multi-modal dataset for instruction-guided 3D editing, covering diverse addition, deletion, and modification tasks. This dataset is meticulously curated to ensure that edited objects faithfully adhere to the instructional changes while preserving the consistency of unedited regions with the source object. Building upon this dataset, we explore two distinct conditioning strategies for our model: a conventional cross-attention mechanism and a novel 3D token concatenation approach. Our results demonstrate that token concatenation is more parameter-efficient and achieves superior performance. Extensive evaluations show that our method outperforms existing 2D-lifting approaches, setting a new benchmark in generation quality, 3D consistency, and instruction fidelity.

cs.CV↗

arXiv2025

OEMA: Ontology-Enhanced Multi-Agent Collaboration Framework for Zero-Shot Clinical Named Entity Recognition

With the rapid expansion of unstructured clinical texts in electronic health records (EHRs), clinical named entity recognition (NER) has become a crucial technique for extracting medical information. However, traditional supervised models such as CRF and BioClinicalBERT suffer from high annotation costs. Although zero-shot NER based on large language models (LLMs) reduces the dependency on labeled data, challenges remain in aligning example selection with task granularity and effectively integrating prompt design with self-improvement frameworks. To address these limitations, we propose OEMA, a novel zero-shot clinical NER framework based on multi-agent collaboration. OEMA consists of three core components: (1) a self-annotator that autonomously generates candidate examples; (2) a discriminator that leverages SNOMED CT to filter token-level examples by clinical relevance; and (3) a predictor that incorporates entity-type descriptions to enhance inference accuracy. Experimental results on two benchmark datasets, MTSamples and VAERS, demonstrate that OEMA achieves state-of-the-art performance under exact-match evaluation. Moreover, under related-match criteria, OEMA performs comparably to the supervised BioClinicalBERT model while significantly outperforming the traditional CRF method. OEMA improves zero-shot clinical NER, achieving near-supervised performance under related-match criteria. Future work will focus on continual learning and open-domain adaptation to expand its applicability in clinical NLP.

cs.CL↗

arXiv2025

TiDAR: Think in Diffusion, Talk in Autoregression

Diffusion language models hold the promise of fast parallel generation, while autoregressive (AR) models typically excel in quality due to their causal structure aligning naturally with language modeling. This raises a fundamental question: can we achieve a synergy with high throughput, higher GPU utilization, and AR level quality? Existing methods fail to effectively balance these two aspects, either prioritizing AR using a weaker model for sequential drafting (speculative decoding), leading to lower drafting efficiency, or using some form of left-to-right (AR-like) decoding logic for diffusion, which still suffers from quality degradation and forfeits its potential parallelizability. We introduce TiDAR, a sequence-level hybrid architecture that drafts tokens (Thinking) in Diffusion and samples final outputs (Talking) AutoRegressively - all within a single forward pass using specially designed structured attention masks. This design exploits the free GPU compute density, achieving a strong balance between drafting and verification capacity. Moreover, TiDAR is designed to be serving-friendly (low overhead) as a standalone model. We extensively evaluate TiDAR against AR models, speculative decoding, and diffusion variants across generative and likelihood tasks at 1.5B and 8B scales. Thanks to the parallel drafting and sampling as well as exact KV cache support, TiDAR outperforms speculative decoding in measured throughput and surpasses diffusion models like Dream and Llada in both efficiency and quality. Most notably, TiDAR is the first architecture to close the quality gap with AR models while delivering 4.71x to 5.91x more tokens per second.

cs.CL↗

arXiv2025

Reasoning-Enhanced Domain-Adaptive Pretraining of Multimodal Large Language Models for Short Video Content Governance

Short video platforms are evolving rapidly, making the identification of inappropriate content increasingly critical. Existing approaches typically train separate and small classification models for each type of issue, which requires extensive human-labeled data and lacks cross-issue generalization. We propose a reasoning-enhanced multimodal large language model (MLLM) pretraining paradigm for unified inappropriate content detection. To address the distribution gap between short video content and the original pretraining data of MLLMs, as well as the complex issue definitions, we introduce three targeted pretraining tasks: (1) \textit{Caption}, to enhance the MLLM's perception of video details; (2) \textit{Visual Question Answering (VQA)}, to deepen the MLLM's understanding of issue definitions and annotation guidelines; (3) \textit{Chain-of-Thought (CoT)}, to enhance the MLLM's reasoning capability. Experimental results show that our pretraining approach significantly improves the MLLM's performance in both zero-shot and supervised fine-tuning (SFT) settings. In addition, our pretrained model demonstrates strong generalization capabilities to emergent, previously unseen issues.

cs.CV↗

arXiv2025

DLER: Doing Length pEnalty Right - Incentivizing More Intelligence per Token via Reinforcement Learning

Reasoning language models such as OpenAI-o1, DeepSeek-R1, and Qwen achieve strong performance via extended chains of thought but often generate unnecessarily long outputs. Maximizing intelligence per token--accuracy relative to response length--remains an open problem. We revisit reinforcement learning (RL) with the simplest length penalty--truncation--and show that accuracy degradation arises not from the lack of sophisticated penalties but from inadequate RL optimization. We identify three key challenges: (i) large bias in advantage estimation, (ii) entropy collapse, and (iii) sparse reward signal. We address them with Doing Length pEnalty Right (DLER), a training recipe combining batch-wise reward normalization, higher clipping, dynamic sampling, and a simple truncation length penalty. DLER achieves state-of-the-art accuracy--efficiency trade-offs, cutting output length by over 70 percent while surpassing all previous baseline accuracy. It also improves test-time scaling: compared to DeepSeek-R1-7B, DLER-7B generates multiple concise responses in parallel with 28 percent higher accuracy and lower latency. We further introduce Difficulty-Aware DLER, which adaptively tightens truncation on easier questions for additional efficiency gains. We also propose an update-selective merging method that preserves baseline accuracy while retaining the concise reasoning ability of the DLER model, which is useful for scenarios where RL training data is scarce.

cs.LG↗

arXiv2025

Profit-Aware Graph Framework for Cross-Platform Ride-Sharing: Analyzing Allocation Mechanisms and Efficiency Gains

Ride-hailing platforms (e.g., Uber, Lyft) have transformed urban mobility by enabling ride-sharing, which holds considerable promise for reducing both travel costs and total vehicle miles traveled (VMT). However, the fragmentation of these platforms impedes system-wide efficiency by restricting ride-matching to intra-platform requests. Cross-platform collaboration could unlock substantial efficiency gains, but its realization hinges on fair and sustainable profit allocation mechanisms that can align the incentives of competing platforms. This study introduces a graph-theoretic framework that embeds profit-aware constraints into network optimization, facilitating equitable and efficient cross-platform ride-sharing. Within this framework, we evaluate three allocation schemes -- equal-profit-based, market-share-based, and Shapley-value-based -- through large-scale simulations. Results show that the Shapley-value-based mechanism consistently outperforms the alternatives across six key metrics. Notably, system efficiency and rider service quality improve with increasing demand, reflecting clear economies of scale. The observed economies of scale, along with their diminishing returns, can be understood with the structural evolution of rider-request graphs, where super-linear edge growth expands feasible matches and sub-linear degree scaling limits per-rider connectivity.

econ.GN↗

arXiv2025

An Event-based Fast Intensity Reconstruction Scheme for UAV Real-time Perception

Event cameras offer significant advantages, including a wide dynamic range, high temporal resolution, and immunity to motion blur, making them highly promising for addressing challenging visual conditions. Extracting and utilizing effective information from asynchronous event streams is essential for the onboard implementation of event cameras. In this paper, we propose a streamlined event-based intensity reconstruction scheme, event-based single integration (ESI), to address such implementation challenges. This method guarantees the portability of conventional frame-based vision methods to event-based scenarios and maintains the intrinsic advantages of event cameras. The ESI approach reconstructs intensity images by performing a single integration of the event streams combined with an enhanced decay algorithm. Such a method enables real-time intensity reconstruction at a high frame rate, typically 100 FPS. Furthermore, the relatively low computation load of ESI fits onboard implementation suitably, such as in UAV-based visual tracking scenarios. Extensive experiments have been conducted to evaluate the performance comparison of ESI and state-of-the-art algorithms. Compared to state-of-the-art algorithms, ESI demonstrates remarkable runtime efficiency improvements, superior reconstruction quality, and a high frame rate. As a result, ESI enhances UAV onboard perception significantly under visual adversary surroundings. In-flight tests, ESI demonstrates effective performance for UAV onboard visual tracking under extremely low illumination conditions(2-10lux), whereas other comparative algorithms fail due to insufficient frame rate, poor image quality, or limited real-time performance.

cs.CV↗