arXiv subjects

Zhiwei Lin

Publications and source records attributed to Zhiwei Lin.

At least 19 recordsLinked to original sources

arXiv2026

X2Streaming-ASR: wait when uncertain, emit when ready for streaming ASR

Streaming automatic speech recognition (ASR) for real-time voice agents and full-duplex dialogue must provide accurate partial transcripts with low commit latency. Existing systems commonly use a fixed chunk size, look-ahead, or target delay, or encourage emissions near estimated acoustic boundaries. These approaches do not directly optimize how much additional context to use at each output position under a single-pass, hard-commit constraint. We propose X2Streaming-ASR, which decomposes streaming recognition into when to commit and what to commit. Its three-stage training procedure first establishes streaming recognition ability, then warm-starts the commit policy with automatically probed trajectories, and finally refines the policy using character-level, segment-assigned group-relative rewards for recognition accuracy and latency. Across AISHELL-1/2/3 and WenetSpeech, X2Streaming-ASR achieves a mean character-level commit latency of 27-84 ms relative to forced-aligned character endpoints, compared with 409-585 ms for the evaluated streaming baselines. It achieves the best streaming CER among the evaluated systems on AISHELL-1 and AISHELL-3 with substantially lower latency.

cs.SD↗

arXiv2026

Diff-Symbo: Text-Controlled Long-Duration Symbolic Music Generation Using Autoregressive Latent Diffusion Model

Text-controlled symbolic music generation has recently gained research attention due to its versatile, flexible and straightforward approach to music composition. However, previous approaches tend to generate symbolic music with compromising quality, diversity, controllability and limited duration. In this paper, we present Diff-Symbo, an innovative method that uses latent diffusion model (LDM) to generate high-quality, diverse and long-duration symbolic music. To address the lack of text-symbolic music dataset, we develop a comprehensive dataset with 19,345 text templates by employing large language model. Furthermore, we design a music information encoder to reduce the training overhead while extracting more effective control representations. Given textual descriptions, our proposed method leverages LDM to improve the quality and diversity of music generation. Our method also improves the duration and the compositional consistency of music generation through an autoregressive approach. Experimental results show significant improvements of Diff-Symbo in text controllability, duration, and the quality of generated music compared to the baseline models such as GPT-4, MuseCoco and Multitrack Music Transformer (MMT). As one of the pioneer models in this field, Diff-Symbo paves the way towards controllable and high-quality symbolic music composition based on LDM, offering valuable contributions to both music amateurs and practitioners.

cs.SD↗

arXiv2026

TurnFSM for Full-Duplex Dialogue System: Internalizing State-Machine Logic for Streaming Semantic Voice Activity Detection and Utterance-Level Rejection

Full-duplex voice assistants must continuously listen while speaking, handling user interruptions under low-latency and resource-constrained streaming conditions. Existing end-to-end full-duplex models can compromise reasoning-related capabilities after speech-domain adaptation, whereas cascaded pipelines introduce extra inference overhead and handcrafted control logic. We propose TurnFSM, an LLM-based state prediction framework that internalizes turn control as explicit finite-state transitions, unifying streaming semantic VAD and utterance-level rejection. TurnFSM decomposes submission and rejection into a serial decision process, reducing multi-task interference while maintaining performance comparable to single-task models. We further introduce a first-order state transition mechanism that enforces the dependency on only the previous state during training, enabling compact inference with the standard causal mask and original LLM positional encoding while avoiding historical state-token accumulation and unnecessary step-by-step state generation. Experimental results show that TurnFSM consistently outperforms the binary-head baseline and remains competitive with task-specific models.

eess.AS↗

arXiv2026

Video = World + Event Stream

We present Wan-Streamer v0.3, which reframes our native-streaming interaction model under a single organizing view: a video is a world plus an event stream. The world is the persistent context in which a video unfolds, including the environment, scene, subjects, ambient acoustic conditions, voice characteristics, and other relatively stable conditions. The event stream is everything that changes over time within that world, including scene or environmental changes, subject behavior, speech, and other sounds. This yields a general-purpose pretraining task over large amounts of real video: given a world and incoming input, predict how the world moves, changes, and responds in real time. The resulting competence can be specialized to a broad family of real-time downstream tasks. We instantiate it on real-time full-duplex audio-visual interaction, where the event stream is the agent's speech together with free-form behavior. Functionally, the model's multimodal understanding process is vision-language-action-like: it maps multimodal user input to language-form speech and behavior actions. Wan-Streamer v0.3 preserves the v0.2 operating point: 640x368 video at 25 FPS, a 160 ms streaming unit, approximately 200 ms model-side response latency, and approximately 550 ms total interaction latency under a 350 ms bidirectional network budget.

cs.CV↗

arXiv2026

Balancing ASR and diarization in end-to-end LLMs for multi-talker speech recognition

Multi-talker speech recognition is often addressed by combining automatic speech recognition (ASR) and speaker diarization in a pipeline system. Recently, LLM-based approaches have shown promise by jointly modeling semantic and speaker information, but they typically require large-scale multi-talker corpora that are costly to annotate. In this paper, we investigate how to efficiently train an LLM-based system with limited real-recorded data while maintaining high accuracy in speaker attribution. We propose several strategies: (1) a dual-encoder architecture to extract semantic and speaker features, (2) a feature interleaving format to merge these features as the inputs to the LLM, (3) a length-aware speaker ID loss to enhance diarization capability, and (4) an adaptive threshold strategy for ASR loss computation to mitigate hallucinations caused by speech overlaps. These strategies balance training between ASR and diarization tasks. Our system outperforms open-source baseline approaches, achieving relative improvements of 18% on the AliMeeting corpus and 24% on the Aishell4 corpus.

eess.AS↗

arXiv2026

Feat2Go: Visual Feature-Grounded Value Estimation for Embodied Reinforcement Learning

Reinforcement learning is a promising approach for improving the capabilities of vision-language-action (VLA) models while avoiding the heavy data requirements of imitation learning. However, its effectiveness for VLA models is often constrained by sparse supervision and the difficulty of designing informative reward signals for long-horizon manipulation. In this work, we present Feat2Go, a fine-grained value estimation framework for embodied reinforcement learning. Specifically, Feat2Go first derives a continuous progress target from a pretrained visual world model by measuring patch-level similarity to subgoal states and partitioning episodes into semantic stages with trend-based clustering. We then train an embodied value model to predict this structural progress from the current observation and task instruction, and use the predicted value to reshape terminal rewards during policy optimization. The proposed framework is compatible with existing VLA policy reinforcement learning pipelines, including PPO and GRPO, and does not rely on manual reward engineering. Extensive experiments on ManiSkill3 and RoboTwin 2.0 demonstrate that Feat2Go consistently improves the performance of existing VLA models under both single-arm and bimanual manipulation settings. More specifically, on ManiSkill3, Feat2Go improves OpenVLAOFT from 17.5% to 82.9% average out-of-distribution success while retaining 96.9% in-distribution performance. On RoboTwin 2.0, Feat2Go achieves an average success rate of 88.8% in domain-randomized task settings, outperforming prior reinforcement learning methods.

cs.RO↗

arXiv2026

VL-SAM-v3: Memory-Guided Visual Priors for Open-World Object Detection

Open-world object detection aims to localize and recognize objects beyond a fixed closed-set label space. It is commonly divided into two categories, i.e., open-vocabulary detection, which assumes a predefined category list at test time, and open-ended detection, which requires generating candidate categories during the inference. Existing methods rely primarily on coarse textual semantics and parametric knowledge, which often provide insufficient visual evidence for fine-grained appearance variation, rare categories, and cluttered scenes. In this paper, we propose VL-SAM-v3, a unified framework that augments open-world detection with retrieval-grounded external visual memory. Specifically, once candidate categories are available, VL-SAM-v3 retrieves relevant visual prototypes from a non-parametric memory bank and transforms them into two complementary visual priors, i.e., sparse priors for instance-level spatial anchoring and dense priors for class-aware local context. These priors are integrated with the original detection prompts via Memory-Guided Prompt Refinement, enabling a shared retrieval-and-refinement mechanism that supports open-vocabulary and open-ended inference. Extensive zero-shot experiments on LVIS show that VL-SAM-v3 consistently improves detection performance under both open-vocabulary and open-ended inference, with particularly strong gains on rare categories. Moreover, experiments with a stronger open-vocabulary detector (i.e., SAM3) validate the generality of the proposed retrieval-and-refinement mechanism.

cs.CV↗

arXiv2026

QAPruner: Quantization-Aware Vision Token Pruning for Multimodal Large Language Models

Multimodal Large Language Models (MLLMs) have shown strong reasoning ability, but their high computational and memory costs hinder deployment in resource-constrained settings. While Post-Training Quantization (PTQ) and vision token pruning are standard compression techniques, they are usually treated as independent optimizations. In this paper, we show that these two techniques are strongly coupled: naively applying semantic-based token pruning to PTQ-optimized MLLMs can discard activation outliers that are important for numerical stability and thus worsen quantization errors in low-bit regimes (\textit{e.g.}, W4A4). To address this issue, we propose a quantization-aware vision token pruning framework. Our method introduces a lightweight hybrid sensitivity metric that combines simulated group-wise quantization error with outlier intensity. By combining this metric with standard semantic relevance scores, the method retains tokens that are both semantically informative and robust to quantization. Experiments on standard LLaVA architectures show that our method consistently outperforms naive integration baselines. At an aggressive pruning ratio that retains only 12.5\% of visual tokens, our framework improves accuracy by 2.24\% over the baseline and even surpasses dense quantization without pruning. To the best of our knowledge, this is the first method that explicitly co-optimizes vision token pruning and PTQ for accurate low-bit MLLM inference.

cs.CV↗

arXiv2026

ELITE: Experiential Learning and Intent-Aware Transfer for Self-improving Embodied Agents

Vision-language models (VLMs) have shown remarkable general capabilities, yet embodied agents built on them fail at complex tasks, often skipping critical steps, proposing invalid actions, and repeating mistakes. These failures arise from a fundamental gap between the static training data of VLMs and the physical interaction for embodied tasks. VLMs can learn rich semantic knowledge from static data but lack the ability to interact with the world. To address this issue, we introduce ELITE, an embodied agent framework with {E}xperiential {L}earning and {I}ntent-aware {T}ransfer that enables agents to continuously learn from their own environment interaction experiences, and transfer acquired knowledge to procedurally similar tasks. ELITE operates through two synergistic mechanisms, \textit{i.e.,} self-reflective knowledge construction and intent-aware retrieval. Specifically, self-reflective knowledge construction extracts reusable strategies from execution trajectories and maintains an evolving strategy pool through structured refinement operations. Then, intent-aware retrieval identifies relevant strategies from the pool and applies them to current tasks. Experiments on the EB-ALFRED and EB-Habitat benchmarks show that ELITE achieves 9\% and 5\% performance improvement over base VLMs in the online setting without any supervision. In the supervised setting, ELITE generalizes effectively to unseen task categories, achieving better performance compared to state-of-the-art training-based methods. These results demonstrate the effectiveness of ELITE for bridging the gap between semantic understanding and reliable action execution.

cs.AI↗

arXiv2026

MeGU: Machine-Guided Unlearning with Target Feature Disentanglement

The growing concern over training data privacy has elevated the "Right to be Forgotten" into a critical requirement, thereby raising the demand for effective Machine Unlearning. However, existing unlearning approaches commonly suffer from a fundamental trade-off: aggressively erasing the influence of target data often degrades model utility on retained data, while conservative strategies leave residual target information intact. In this work, the intrinsic representation properties learned during model pretraining are analyzed. It is demonstrated that semantic class concepts are entangled at the feature-pattern level, sharing associated features while preserving concept-specific discriminative components. This entanglement fundamentally limits the effectiveness of existing unlearning paradigms. Motivated by this insight, we propose Machine-Guided Unlearning (MeGU), a novel framework that guides unlearning through concept-aware re-alignment. Specifically, Multi-modal Large Language Models (MLLMs) are leveraged to explicitly determine re-alignment directions for target samples by assigning semantically meaningful perturbing labels. To improve efficiency, inter-class conceptual similarities estimated by the MLLM are encoded into a lightweight transition matrix. Furthermore, MeGU introduces a positive-negative feature noise pair to explicitly disentangle target concept influence. During finetuning, the negative noise suppresses target-specific feature patterns, while the positive noise reinforces remaining associated features and aligns them with perturbing concepts. This coordinated design enables selective disruption of target-specific representations while preserving shared semantic structures. As a result, MeGU enables controlled and selective forgetting, effectively mitigating both under-unlearning and over-unlearning.

cs.LG↗

arXiv2025

HENet++: Hybrid Encoding and Multi-task Learning for 3D Perception and End-to-end Autonomous Driving

Three-dimensional feature extraction is a critical component of autonomous driving systems, where perception tasks such as 3D object detection, bird's-eye-view (BEV) semantic segmentation, and occupancy prediction serve as important constraints on 3D features. While large image encoders, high-resolution images, and long-term temporal inputs can significantly enhance feature quality and deliver remarkable performance gains, these techniques are often incompatible in both training and inference due to computational resource constraints. Moreover, different tasks favor distinct feature representations, making it difficult for a single model to perform end-to-end inference across multiple tasks while maintaining accuracy comparable to that of single-task models. To alleviate these issues, we present the HENet and HENet++ framework for multi-task 3D perception and end-to-end autonomous driving. Specifically, we propose a hybrid image encoding network that uses a large image encoder for short-term frames and a small one for long-term frames. Furthermore, our framework simultaneously extracts both dense and sparse features, providing more suitable representations for different tasks, reducing cumulative errors, and delivering more comprehensive information to the planning module. The proposed architecture maintains compatibility with various existing 3D feature extraction methods and supports multimodal inputs. HENet++ achieves state-of-the-art end-to-end multi-task 3D perception results on the nuScenes benchmark, while also attaining the lowest collision rate on the nuScenes end-to-end autonomous driving benchmark.

cs.CV↗

arXiv2025

TraceAegis: Securing LLM-Based Agents via Hierarchical and Behavioral Anomaly Detection

LLM-based agents have demonstrated promising adaptability in real-world applications. However, these agents remain vulnerable to a wide range of attacks, such as tool poisoning and malicious instructions, that compromise their execution flow and can lead to serious consequences like data breaches and financial loss. Existing studies typically attempt to mitigate such anomalies by predefining specific rules and enforcing them at runtime to enhance safety. Yet, designing comprehensive rules is difficult, requiring extensive manual effort and still leaving gaps that result in false negatives. As agent systems evolve into complex software systems, we take inspiration from software system security and propose TraceAegis, a provenance-based analysis framework that leverages agent execution traces to detect potential anomalies. In particular, TraceAegis constructs a hierarchical structure to abstract stable execution units that characterize normal agent behaviors. These units are then summarized into constrained behavioral rules that specify the conditions necessary to complete a task. By validating execution traces against both hierarchical and behavioral constraints, TraceAegis is able to effectively detect abnormal behaviors. To evaluate the effectiveness of TraceAegis, we introduce TraceAegis-Bench, a dataset covering two representative scenarios: healthcare and corporate procurement. Each scenario includes 1,300 benign behaviors and 300 abnormal behaviors, where the anomalies either violate the agent's execution order or break the semantic consistency of its execution sequence. Experimental results demonstrate that TraceAegis achieves strong performance on TraceAegis-Bench, successfully identifying the majority of abnormal behaviors.

cs.CR↗

arXiv2025

CoLLM-NAS: Collaborative Large Language Models for Efficient Knowledge-Guided Neural Architecture Search

The integration of Large Language Models (LLMs) with Neural Architecture Search (NAS) has introduced new possibilities for automating the design of neural architectures. However, most existing methods face critical limitations, including architectural invalidity, computational inefficiency, and inferior performance compared to traditional NAS. In this work, we present Collaborative LLM-based NAS (CoLLM-NAS), a two-stage NAS framework with knowledge-guided search driven by two complementary LLMs. Specifically, we propose a stateful Navigator LLM to guide search direction, a stateless Generator LLM to synthesize high-quality candidates, and a Coordinator module to orchestrate inter-LLM communication and manage evaluation processes. CoLLM-NAS efficiently guides the search process by combining LLMs' inherent knowledge of structured neural architectures with progressive knowledge from iterative feedback and historical trajectory. Experimental results on ImageNet and NAS-Bench-201 show that CoLLM-NAS surpasses existing NAS methods and conventional search algorithms, achieving new state-of-the-art results while significantly reducing search costs by 4--10. Furthermore, CoLLM-NAS consistently enhances the performance and efficiency of various two-stage NAS methods (e.g., OFA, SPOS, and AutoFormer) across diverse search spaces (e.g., MobileNet, ShuffleNet, and AutoFormer), demonstrating its excellent generalization.

cs.AI↗

arXiv2025

VoxRole: A Comprehensive Benchmark for Evaluating Speech-Based Role-Playing Agents

Recent significant advancements in Large Language Models (LLMs) have greatly propelled the development of Role-Playing Conversational Agents (RPCAs). These systems aim to create immersive user experiences through consistent persona adoption. However, current RPCA research faces dual limitations. First, existing work predominantly focuses on the textual modality, entirely overlooking critical paralinguistic features including intonation, prosody, and rhythm in speech, which are essential for conveying character emotions and shaping vivid identities. Second, the speech-based role-playing domain suffers from a long-standing lack of standardized evaluation benchmarks. Most current spoken dialogue datasets target only fundamental capability assessments, featuring thinly sketched or ill-defined character profiles. Consequently, they fail to effectively quantify model performance on core competencies like long-term persona consistency. To address this critical gap, we introduce VoxRole, the first comprehensive benchmark specifically designed for the evaluation of speech-based RPCAs. The benchmark comprises 13335 multi-turn dialogues, totaling 65.6 hours of speech from 1228 unique characters across 261 movies. To construct this resource, we propose a novel two-stage automated pipeline that first aligns movie audio with scripts and subsequently employs an LLM to systematically build multi-dimensional profiles for each character. Leveraging VoxRole, we conduct a multi-dimensional evaluation of contemporary spoken dialogue models, revealing crucial insights into their respective strengths and limitations in maintaining persona consistency.

cs.CL↗

arXiv2025

SAR-NAS: Lightweight SAR Object Detection with Neural Architecture Search

Synthetic Aperture Radar (SAR) object detection faces significant challenges from speckle noise, small target ambiguities, and on-board computational constraints. While existing approaches predominantly focus on SAR-specific architectural modifications, this paper explores the application of the existing lightweight object detector, i.e., YOLOv10, for SAR object detection and enhances its performance through Neural Architecture Search (NAS). Specifically, we employ NAS to systematically optimize the network structure, especially focusing on the backbone architecture search. By constructing an extensive search space and leveraging evolutionary search, our method identifies a favorable architecture that balances accuracy, parameter efficiency, and computational cost. Notably, this work introduces NAS to SAR object detection for the first time. The experimental results on the large-scale SARDet-100K dataset demonstrate that our optimized model outperforms existing SAR detection methods, achieving superior detection accuracy while maintaining lower computational overhead. We hope this work offers a novel perspective on leveraging NAS for real-world applications.

cs.CV↗

arXiv2025

Multi-Representation Adapter with Neural Architecture Search for Efficient Range-Doppler Radar Object Detection

Detecting objects efficiently from radar sensors has recently become a popular trend due to their robustness against adverse lighting and weather conditions compared with cameras. This paper presents an efficient object detection model for Range-Doppler (RD) radar maps. Specifically, we first represent RD radar maps with multi-representation, i.e., heatmaps and grayscale images, to gather high-level object and fine-grained texture features. Then, we design an additional Adapter branch, an Exchanger Module with two modes, and a Primary-Auxiliary Fusion Module to effectively extract, exchange, and fuse features from the multi-representation inputs, respectively. Furthermore, we construct a supernet with various width and fusion operations in the Adapter branch for the proposed model and employ a One-Shot Neural Architecture Search method to further improve the model's efficiency while maintaining high performance. Experimental results demonstrate that our model obtains favorable accuracy and efficiency trade-off. Moreover, we achieve new state-of-the-art performance on RADDet and CARRADA datasets with mAP@50 of 71.9 and 57.1, respectively.

cs.CV↗

arXiv2025

VULSOLVER: Vulnerability Detection via LLM-Driven Constraint Solving

Traditional vulnerability detection methods rely heavily on predefined rule matching, which often fails to capture vulnerabilities accurately. With the rise of large language models (LLMs), leveraging their ability to understand code semantics has emerged as a promising direction for achieving more accurate and efficient vulnerability detection. However, current LLM-based approaches face significant challenges: instability in model outputs, degraded performance with long context, and hallucination. As a result, many existing solutions either use LLMs merely to enrich predefined rule sets, thereby keeping the detection process fundamentally rule-based, or over-rely on them, leading to poor robustness. To address these challenges, we propose a constraint-solving approach powered by LLMs named VULSOLVER. By modeling vulnerability detection as a constraint-solving problem, and by integrating static application security testing (SAST) with the semantic reasoning capabilities of LLMs, our method enables the LLM to act like a professional human security expert. We assess VULSOLVER on the OWASP Benchmark (1,023 labeled samples), achieving 97.85% accuracy, 97.97% F1-score, and 100% recall. Applied to widely-used open-source projects, VULSOLVER identified 15 previously unknown high-severity vulnerabilities (CVSS 7.5-9.8), demonstrating its effectiveness in real-world security analysis.

cs.CR↗

arXiv2025

PTQAT: A Hybrid Parameter-Efficient Quantization Algorithm for 3D Perception Tasks

Post-Training Quantization (PTQ) and Quantization-Aware Training (QAT) represent two mainstream model quantization approaches. However, PTQ often leads to unacceptable performance degradation in quantized models, while QAT imposes substantial GPU memory requirements and extended training time due to weight fine-tuning. In this paper, we propose PTQAT, a novel general hybrid quantization algorithm for the efficient deployment of 3D perception networks. To address the speed accuracy trade-off between PTQ and QAT, our method selects critical layers for QAT fine-tuning and performs PTQ on the remaining layers. Contrary to intuition, fine-tuning the layers with smaller output discrepancies before and after quantization, rather than those with larger discrepancies, actually leads to greater improvements in the model's quantization accuracy. This means we better compensate for quantization errors during their propagation, rather than addressing them at the point where they occur. The proposed PTQAT achieves similar performance to QAT with more efficiency by freezing nearly 50% of quantifiable layers. Additionally, PTQAT is a universal quantization method that supports various quantization bit widths (4 bits) as well as different model architectures, including CNNs and Transformers. The experimental results on nuScenes across diverse 3D perception tasks, including object detection, semantic segmentation, and occupancy prediction, show that our method consistently outperforms QAT-only baselines. Notably, it achieves 0.2%-0.9% NDS and 0.3%-1.0% mAP gains in object detection, 0.3%-2.0% mIoU gains in semantic segmentation and occupancy prediction while fine-tuning fewer weights.

cs.CV↗