Fangyuan Wang

Fangyuan Wang, Ziyuan Wang, Guorui Pei et al.

Robotic manipulation depends on the ability to anticipate how actions reshape objects, contacts, and scene geometry before execution. Learned world models provide this capability by predicting task-relevant future evolution under robot intervention, yet the term now spans latent dynamics models, action-conditioned video generators, three- and four-dimensional scene predictors, physics-informed simulators, and predictive modules inside vision-language-action systems. This breadth has fragmented the literature and obscured the design choices that matter for manipulation. We survey world models for robotic manipulation through three questions: what future representation is predicted, how prediction is connected to action, and when prediction is used in the robot-learning pipeline. We operationally define a world model as an action-conditioned predictive system and distinguish it from perception modules, inverse models, policies, rewards, and value functions. We then organize existing work into five representation families, develop a functional taxonomy that separates integrated prediction-action models from explicit predictive planners, and characterize infrastructure roles including synthetic experience generation, candidate filtering, search-based evaluation, learned environments, and outcome verification. We further map these roles across pretraining, post-training, and inference adaptation, review 34 manipulation datasets, and synthesize evaluation protocols for predictive fidelity, task performance, and simulator reliability. This survey shows that world models are evolving from task-specific dynamics predictors into predictive infrastructure for robot learning, while exposing open challenges in contact modeling, hallucination control, action alignment, and benchmarking under closed-loop use.

Jiawen Wei, Fangyuan Wang, Wanxin Zeng et al.

Reducing sensor requirements while keeping optimal control performance is crucial to many industrial control applications to achieve robust, low-cost, and computation-efficient controllers. However, existing feature selection solutions for the typical machine learning domain can hardly be applied in the domain of control with changing dynamics. In this paper, a novel framework, namely the Dual-world embedded Attentive Feature Selection (D-AFS), can efficiently select the most relevant sensors for the system under dynamic control. Rather than the one world used in most Deep Reinforcement Learning (DRL) algorithms, D-AFS has both the real world and its virtual peer with twisted features. By analyzing the DRL's response in two worlds, D-AFS can quantitatively identify respective features' importance towards control. A well-known active flow control problem, cylinder drag reduction, is used for evaluation. Results show that D-AFS successfully finds an optimized five-probes layout with 18.7\% drag reduction than the state-of-the-art solution with 151 probes and 49.2\% reduction than five-probes layout by human experts. We also apply this solution to four OpenAI classical control cases. In all cases, D-AFS achieves the same or better sensor configurations than originally provided solutions. Results highlight, we argued, a new way to achieve efficient and optimal sensor designs for experimental or industrial systems. Our source codes are made publicly available at https://github.com/G-AILab/DAFSFluid.

Fangyuan Wang, Bo Xu

Currently, there are mainly three kinds of Transformer encoder based streaming End to End (E2E) Automatic Speech Recognition (ASR) approaches, namely time-restricted methods, chunk-wise methods, and memory-based methods. Generally, all of them have limitations in aspects of linear computational complexity, global context modeling, and parallel training. In this work, we aim to build a model to take all these three advantages for streaming Transformer ASR. Particularly, we propose a shifted chunk mechanism for the chunk-wise Transformer which provides cross-chunk connections between chunks. Therefore, the global context modeling ability of chunk-wise models can be significantly enhanced while all the original merits inherited. We integrate this scheme with the chunk-wise Transformer and Conformer, and identify them as SChunk-Transformer and SChunk-Conformer, respectively. Experiments on AISHELL-1 show that the SChunk-Transformer and SChunk-Conformer can respectively achieve CER 6.43% and 5.77%. And the linear complexity makes them possible to train with large batches and infer more efficiently. Our models can significantly outperform their conventional chunk-wise counterparts, while being competitive, with only 0.22 absolute CER drop, when compared with U2 which has quadratic complexity. A better CER can be achieved if compared with existing chunk-wise or memory-based methods, such as HS-DACS and MMA. Code is released.

Fangyuan Wang, Peng Zhou, Jiaming Qi et al.

Vision-language-action (VLA) models typically inject proprioception only as a late conditioning signal, which prevents robot state from shaping instruction understanding and from influencing which visual tokens are attended throughout the policy. We introduce ThinkProprio, which converts proprioception into a sequence of text tokens in the VLM embedding space and fuses them with the task instruction at the input. This early fusion lets embodied state participate in subsequent visual reasoning and token selection, biasing computation toward action-critical evidence while suppressing redundant visual tokens. In a systematic ablation over proprioception encoding, state entry point, and action-head conditioning, we find that text tokenization is more effective than learned projectors, and that retaining roughly 15% of visual tokens can match the performance of using the full token set. Across CALVIN, LIBERO, and real-world manipulation, ThinkProprio matches or improves over strong baselines while reducing end-to-end inference latency over 50%.

Fangyuan Wang, Ming Hao, Yuhai Shi et al.

The conventional recipe for Automatic Speech Recognition (ASR) models is to 1) train multiple checkpoints on a training set while relying on a validation set to prevent overfitting using early stopping and 2) average several last checkpoints or that of the lowest validation losses to obtain the final model. In this paper, we rethink and update the early stopping and checkpoint averaging from the perspective of the bias-variance tradeoff. Theoretically, the bias and variance represent the fitness and variability of a model and the tradeoff of them determines the overall generalization error. But, it's impractical to evaluate them precisely. As an alternative, we take the training loss and validation loss as proxies of bias and variance and guide the early stopping and checkpoint averaging using their tradeoff, namely an Approximated Bias-Variance Tradeoff (ApproBiVT). When evaluating with advanced ASR models, our recipe provides 2.5%-3.7% and 3.1%-4.6% CER reduction on the AISHELL-1 and AISHELL-2, respectively.

Fangyuan Wang, Bo Xu, Bo Xu

Currently, the chunk-wise schemes are often used to make Automatic Speech Recognition (ASR) models to support streaming deployment. However, existing approaches are unable to capture the global context, lack support for parallel training, or exhibit quadratic complexity for the computation of multi-head self-attention (MHSA). On the other side, the causal convolution, no future context used, has become the de facto module in streaming Conformer. In this paper, we propose SSCFormer to push the limit of chunk-wise Conformer for streaming ASR using the following two techniques: 1) A novel cross-chunks context generation method, named Sequential Sampling Chunk (SSC) scheme, to re-partition chunks from regular partitioned chunks to facilitate efficient long-term contextual interaction within local chunks. 2)The Chunked Causal Convolution (C2Conv) is designed to concurrently capture the left context and chunk-wise future context. Evaluations on AISHELL-1 show that an End-to-End (E2E) CER 5.33% can achieve, which even outperforms a strong time-restricted baseline U2. Moreover, the chunk-wise MHSA computation in our model enables it to train with a large batch size and perform inference with linear complexity.

Le Zhou, Feng Yao, Fengcai Qiao et al.

Recent advances in Large Reasoning Models (LRMs) trained with Long Chain-of-Thought have demonstrated remarkable capabilities in code generation and mathematical reasoning. However, their potential in multi-turn Text-to-SQL tasks remains largely underexplored. Existing approaches typically rely on unstable API-based inference or require expensive fine-tuning on small-scale models. In this work, we present Rose-SQL, a training-free framework that leverages small-scale LRMs through in-context learning to enable accurate context-dependent parsing. We introduce the Role-State, a fine-grained representation that bridges the structural gap between schema linking and SQL generation by serving as a structural blueprint. To handle conversational dependencies, Rose-SQL traces the evolution of Role-State through historical context via structural isomorphism checks, guiding the model to infer the possible SQL composition for the current question through verified interaction trajectories. Experiments on the SParC and CoSQL benchmarks show that, within the Qwen3 series, Rose-SQL outperforms in-context learning baselines at the 4B scale and substantially surpasses state-of-the-art fine-tuned models at the 8B and 14B scales, while showing consistent gains on additional reasoning backbones.

Fangyuan Wang, Shipeng Lyu, Peng Zhou et al.

Enabling humanoid robots to perform long-horizon mobile manipulation planning in real-world environments based on embodied perception and comprehension abilities has been a longstanding challenge. With the recent rise of large language models (LLMs), there has been a notable increase in the development of LLM-based planners. These approaches either utilize human-provided textual representations of the real world or heavily depend on prompt engineering to extract such representations, lacking the capability to quantitatively understand the environment, such as determining the feasibility of manipulating objects. To address these limitations, we present the Instruction-Augmented Long-Horizon Planning (IALP) system, a novel framework that employs LLMs to generate feasible and optimal actions based on real-time sensor feedback, including grounded knowledge of the environment, in a closed-loop interaction. Distinct from prior works, our approach augments user instructions into PDDL problems by leveraging both the abstract reasoning capabilities of LLMs and grounding mechanisms. By conducting various real-world long-horizon tasks, each consisting of seven distinct manipulatory skills, our results demonstrate that the IALP system can efficiently solve these tasks with an average success rate exceeding 80%. Our proposed method can operate as a high-level planner, equipping robots with substantial autonomy in unstructured environments through the utilization of multi-modal sensor inputs.

Fangyuan Wang, Zhigang Song, Hongchen Jiang et al.

Most of the recent state-of-the-art results for speaker verification are achieved by X-vector and its subsequent variants. In this paper, we propose a new network architecture which aggregates the channel and context interdependence features from multi aspect based on Time Delay Neural Network (TDNN). Firstly, we use the SE-Res2Blocks as in ECAPA-TDNN to explicitly model the channel interdependence to realize adaptive calibration of channel features, and process local context features in a multi-scale way at a more granular level compared with conventional TDNN-based methods. Secondly, we explore to use the encoder structure of Transformer to model the global context interdependence features at an utterance level which can capture better long term temporal characteristics. Before the pooling layer, we aggregate the outputs of SE-Res2Blocks and Transformer encoder to leverage the complementary channel and context interdependence features learned by themself respectively. Finally, instead of performing a single attentive statistics pooling, we also find it beneficial to extend the pooling method in a multi-head way which can discriminate features from multiple aspect. The proposed MACCIF-TDNN architecture can outperform most of the state-of-the-art TDNN-based systems on VoxCeleb1 test sets.

Jiaming Xu, Bo Xu, Guanhua Tian et al.

Due to computational and storage efficiencies of compact binary codes, hashing has been widely used for large-scale similarity search. Unfortunately, many existing hashing methods based on observed keyword features are not effective for short texts due to the sparseness and shortness. Recently, some researchers try to utilize latent topics of certain granularity to preserve semantic similarity in hash codes beyond keyword matching. However, topics of certain granularity are not adequate to represent the intrinsic semantic information. In this paper, we present a novel unified approach for short text Hashing using Multi-granularity Topics and Tags, dubbed HMTT. In particular, we propose a selection method to choose the optimal multi-granularity topics depending on the type of dataset, and design two distinct hashing strategies to incorporate multi-granularity topics. We also propose a simple and effective method to exploit tags to enhance the similarity of related texts. We carry out extensive experiments on one short text dataset as well as on one normal text dataset. The results demonstrate that our approach is effective and significantly outperforms baselines on several evaluation metrics.