Yantao Shen

Luca Zancato, Arjun Seshadri, Yonatan Dukler et al.

We describe a family of architectures to support transductive inference by allowing memory to grow to a finite but a-priori unknown bound while making efficient use of finite resources for inference. Current architectures use such resources to represent data either eidetically over a finite span ("context" in Transformers), or fading over an infinite span (in State Space Models, or SSMs). Recent hybrid architectures have combined eidetic and fading memory, but with limitations that do not allow the designer or the learning process to seamlessly modulate the two, nor to extend the eidetic memory span. We leverage ideas from Stochastic Realization Theory to develop a class of models called B'MOJO to seamlessly combine eidetic and fading memory within an elementary composable module. The overall architecture can be used to implement models that can access short-term eidetic memory "in-context," permanent structural memory "in-weights," fading memory "in-state," and long-term eidetic memory "in-storage" by natively incorporating retrieval from an asynchronously updated memory. We show that Transformers, existing SSMs such as Mamba, and hybrid architectures such as Jamba are special cases of B'MOJO and describe a basic implementation, to be open sourced, that can be stacked and scaled efficiently in hardware. We test B'MOJO on transductive inference tasks, such as associative recall, where it outperforms existing SSMs and Hybrid models; as a baseline, we test ordinary language modeling where B'MOJO achieves perplexity comparable to similarly-sized Transformers and SSMs up to 1.4B parameters, while being up to 10% faster to train. Finally, we show that B'MOJO's ability to modulate eidetic and fading memory results in better inference on longer sequences tested up to 32K tokens, four-fold the length of the longest sequences seen during training.

Yue Zhao, Yantao Shen, Yuanjun Xiong et al. · amazon-science

Negative flips are errors introduced in a classification system when a legacy model is updated. Existing methods to reduce the negative flip rate (NFR) either do so at the expense of overall accuracy by forcing a new model to imitate the old models, or use ensembles, which multiply inference cost prohibitively. We analyze the role of ensembles in reducing NFR and observe that they remove negative flips that are typically not close to the decision boundary, but often exhibit large deviations in the distance among their logits. Based on the observation, we present a method, called Ensemble Logit Difference Inhibition (ELODI), to train a classification system that achieves paragon performance in both error rate and NFR, at the inference cost of a single model. The method distills a homogeneous ensemble to a single student model which is used to update the classification system. ELODI also introduces a generalized distillation objective, Logit Difference Inhibition (LDI), which only penalizes the logit difference of a subset of classes with the highest logit values. On multiple image classification benchmarks, model updates with ELODI demonstrate superior accuracy retention and NFR reduction.

Binjie Zhang, Yixiao Ge, Yantao Shen et al. · tencent-ai

Conventional model upgrades for visual search systems require offline refresh of gallery features by feeding gallery images into new models (dubbed as "backfill"), which is time-consuming and expensive, especially in large-scale applications. The task of backward-compatible representation learning is therefore introduced to support backfill-free model upgrades, where the new query features are interoperable with the old gallery features. Despite the success, previous works only investigated a close-set training scenario (i.e., the new training set shares the same classes as the old one), and are limited by more realistic and challenging open-set scenarios. To this end, we first introduce a new problem of universal backward-compatible representation learning, covering all possible data split in model upgrades. We further propose a simple yet effective method, dubbed as Universal Backward-Compatible Training (UniBCT) with a novel structural prototype refinement algorithm, to learn compatible representations in all kinds of model upgrading benchmarks in a unified manner. Comprehensive experiments on the large-scale face recognition datasets MS1Mv3 and IJB-C fully demonstrate the effectiveness of our method.

Bingnan Li, Zhaoyang Zhang, Xiaoze Liu et al.

Leveraging multiple specialized LLMs can combine complementary strengths, but existing approaches trade adaptability for stability: routing commits prematurely, heuristic ensembling depends on fragile proxies, and parameter merging introduces interference. We propose DLLG (Dynamic Logit-Level Gating), a dynamic logit-level ensembling framework that learns token-level expert fusion from sparse response-level supervision. A lightweight gating module predicts step-wise fusion weights, linking trajectory-level correctness to generation without token-level labels or expert retraining. Across diverse reasoning and code benchmarks, DLLG consistently outperforms strong routing, heuristic ensembling, and parameter-merging baselines across model scales, highlighting learned logit-level fusion as a robust and scalable paradigm for integrating specialized experts.

Youngeun Kim, Jun Fang, Qin Zhang et al. · amazon-science

The open world is inherently dynamic, characterized by ever-evolving concepts and distributions. Continual learning (CL) in this dynamic open-world environment presents a significant challenge in effectively generalizing to unseen test-time classes. To address this challenge, we introduce a new practical CL setting tailored for open-world visual representation learning. In this setting, subsequent data streams systematically introduce novel classes that are disjoint from those seen in previous training phases, while also remaining distinct from the unseen test classes. In response, we present Dynamic Prompt and Representation Learner (DPaRL), a simple yet effective Prompt-based CL (PCL) method. Our DPaRL learns to generate dynamic prompts for inference, as opposed to relying on a static prompt pool in previous PCL methods. In addition, DPaRL jointly learns dynamic prompt generation and discriminative representation at each training stage whereas prior PCL methods only refine the prompt learning throughout the process. Our experimental results demonstrate the superiority of our approach, surpassing state-of-the-art methods on well-established open-world image retrieval benchmarks by an average of 4.7% improvement in Recall@1 performance.

Zhaoyang Zhang, Shuli Jiang, Yantao Shen et al.

Reinforcement learning (RL) post-training has recently driven major gains in long chain-of-thought reasoning large language models (LLMs), but the high inference cost of such models motivates distillation into smaller students. Most existing knowledge distillation (KD) methods are designed for supervised fine-tuning (SFT), relying on fixed teacher traces or teacher-student Kullback-Leibler (KL) divergence-based regularization. When combined with RL, these approaches often suffer from distribution mismatch and objective interference: teacher supervision may not align with the student's evolving rollout distribution, and the KL regularizer can compete with reward maximization and require careful loss balancing. To address these issues, we propose RL-aware distillation (RLAD), which performs selective imitation during RL -- guiding the student toward the teacher only when it improves the current policy update. Our core component, Trust Region Ratio Distillation (TRRD), replaces the teacher-student KL regularizer with a PPO/GRPO-style likelihood-ratio objective anchored to a teacher--old-policy mixture, yielding advantage-aware, trust-region-bounded distillation on student rollouts and naturally balancing exploration, exploitation, and imitation. Across diverse logic reasoning and math benchmarks, RLAD consistently outperforms offline distillation, standard GRPO, and KL-based on-policy teacher-student knowledge distillation.

Jing Tan, Zhaoyang Zhang, Yantao Shen et al.

We introduce Talk2Move, a reinforcement learning (RL) based diffusion framework for text-instructed spatial transformation of objects within scenes. Spatially manipulating objects in a scene through natural language poses a challenge for multimodal generation systems. While existing text-based manipulation methods can adjust appearance or style, they struggle to perform object-level geometric transformations-such as translating, rotating, or resizing objects-due to scarce paired supervision and pixel-level optimization limits. Talk2Move employs Group Relative Policy Optimization (GRPO) to explore geometric actions through diverse rollouts generated from input images and lightweight textual variations, removing the need for costly paired data. A spatial reward guided model aligns geometric transformations with linguistic description, while off-policy step evaluation and active step sampling improve learning efficiency by focusing on informative transformation stages. Furthermore, we design object-centric spatial rewards that evaluate displacement, rotation, and scaling behaviors directly, enabling interpretable and coherent transformations. Experiments on curated benchmarks demonstrate that Talk2Move achieves precise, consistent, and semantically faithful object transformations, outperforming existing text-guided editing approaches in both spatial accuracy and scene coherence.

Xunjie Chen, Liuyin Wang, Xinyan Huang et al.

Legged robots face significant challenges in moving and navigating on deformable and highly yielding terrain such as mud. We present a resistive force model for legged foot-mud interactions. The model captures rheological behaviors such as visco-elasticity, thixotropy of the mud suspension and retractive suction. One attractive property of this new model lies in its effective, uniform formulation to provide underlying physical interpretation and accurate resistive force predictions. We further take advantage of the resistive force model to design a new morphing robotic foot for effective and efficient legged locomotion. We conduct extensive experiments to validate the force model, and the results demonstrate that the morphing foot enhances not only the locomotion mobility but also energy-efficiency of walking in mud. The new resistive force model can be further used to develop data-driven simulation and locomotion control of legged robots on muddy terrains.

Binjie Zhang, Yixiao Ge, Yantao Shen et al.

The task of hot-refresh model upgrades of image retrieval systems plays an essential role in the industry but has never been investigated in academia before. Conventional cold-refresh model upgrades can only deploy new models after the gallery is overall backfilled, taking weeks or even months for massive data. In contrast, hot-refresh model upgrades deploy the new model immediately and then gradually improve the retrieval accuracy by backfilling the gallery on-the-fly. Compatible training has made it possible, however, the problem of model regression with negative flips poses a great challenge to the stable improvement of user experience. We argue that it is mainly due to the fact that new-to-old positive query-gallery pairs may show less similarity than new-to-new negative pairs. To solve the problem, we introduce a Regression-Alleviating Compatible Training (RACT) method to properly constrain the feature compatibility while reducing negative flips. The core is to encourage the new-to-old positive pairs to be more similar than both the new-to-old negative pairs and the new-to-new negative pairs. An efficient uncertainty-based backfilling strategy is further introduced to fasten accuracy improvements. Extensive experiments on large-scale retrieval benchmarks (e.g., Google Landmark) demonstrate that our RACT effectively alleviates the model regression for one more step towards seamless model upgrades. The code will be available at https://github.com/binjiezhang/RACT_ICLR2022.

Zhaoyang Zhang, Yantao Shen, Kunyu Shi et al.

We present a vision-language model whose parameters are jointly trained on all tasks and fully shared among multiple heterogeneous tasks which may interfere with each other, resulting in a single model which we named Musketeer. The integration of knowledge across heterogeneous tasks is enabled by a novel feature called Task Explanation Prompt (TEP). With rich and structured information such as task input/output format, TEP reduces interference among tasks, allowing the model to focus on their shared structure. With a single model, Musketeer achieves results comparable to or better than strong baselines trained on single tasks, almost uniformly across multiple tasks.

Mehdi Rahimi, Yantao Shen, Zhiming Liu et al.

This paper presents our recent development on a portable and refreshable text reading and sensory substitution system for the blind or visually impaired (BVI), called Finger-eye. The system mainly consists of an opto-text processing unit and a compact electro-tactile based display that can deliver text-related electrical signals to the fingertip skin through a wearable and Braille-dot patterned electrode array and thus delivers the electro-stimulation based Braille touch sensations to the fingertip. To achieve the goal of aiding BVI to read any text not written in Braille through this portable system, in this work, a Rapid Optical Character Recognition (R-OCR) method is firstly developed for real-time processing text information based on a Fisheye imaging device mounted at the finger-wearable electro-tactile display. This allows real-time translation of printed text to electro-Braille along with natural movement of user's fingertip as if reading any Braille display or book. More importantly, an electro-tactile neuro-stimulation feedback mechanism is proposed and incorporated with the R-OCR method, which facilitates a new opto-electrotactile feedback based text line tracking control approach that enables text line following by user fingertip during reading. Multiple experiments were designed and conducted to test the ability of blindfolded participants to read through and follow the text line based on the opto-electrotactile-feedback method. The experiments show that as the result of the opto-electrotactile-feedback, the users were able to maintain their fingertip within a $2mm$ distance of the text while scanning a text line. This research is a significant step to aid the BVI users with a portable means to translate and follow to read any printed text to Braille, whether in the digital realm or physically, on any surface.

Yantao Shen, Yuanjun Xiong, Wei Xia et al.

We propose a way to learn visual features that are compatible with previously computed ones even when they have different dimensions and are learned via different neural network architectures and loss functions. Compatible means that, if such features are used to compare images, then "new" features can be compared directly to "old" features, so they can be used interchangeably. This enables visual search systems to bypass computing new features for all previously seen images when updating the embedding models, a process known as backfilling. Backward compatibility is critical to quickly deploy new embedding models that leverage ever-growing large-scale training datasets and improvements in deep learning architectures and training methods. We propose a framework to train embedding models, called backward-compatible training (BCT), as a first step towards backward compatible representation learning. In experiments on learning embeddings for face recognition, models trained with BCT successfully achieve backward compatibility without sacrificing accuracy, thus enabling backfill-free model updates of visual embeddings.

Mehdi Rahimi, Spencer Gibb, Yantao Shen et al.

In this paper, a comparison of reinforcement learning algorithms and their performance on a robot box pushing task is provided. The robot box pushing problem is structured as both a single-agent problem and also a multi-agent problem. A Q-learning algorithm is applied to the single-agent box pushing problem, and three different Q-learning algorithms are applied to the multi-agent box pushing problem. Both sets of algorithms are applied on a dynamic environment that is comprised of static objects, a static goal location, a dynamic box location, and dynamic agent positions. A simulation environment is developed to test the four algorithms, and their performance is compared through graphical explanations of test results. The comparison shows that the newly applied reinforcement algorithm out-performs the previously applied algorithms on the robot box pushing problem in a dynamic environment.

Dapeng Chen, Hongsheng Li, Xihui Liu et al.

Person re-identification is an important task that requires learning discriminative visual features for distinguishing different person identities. Diverse auxiliary information has been utilized to improve the visual feature learning. In this paper, we propose to exploit natural language description as additional training supervisions for effective visual features. Compared with other auxiliary information, language can describe a specific person from more compact and semantic visual aspects, thus is complementary to the pixel-level image data. Our method not only learns better global visual feature with the supervision of the overall description but also enforces semantic consistencies between local visual and linguistic features, which is achieved by building global and local image-language associations. The global image-language association is established according to the identity labels, while the local association is based upon the implicit correspondences between image regions and noun phrases. Extensive experiments demonstrate the effectiveness of employing language as training supervisions with the two association schemes. Our method achieves state-of-the-art performance without utilizing any auxiliary information during testing and shows better performance than other joint embedding methods for the image-language association.

Yantao Shen, Tong Xiao, Hongsheng Li et al.

Person re-identification aims to robustly measure similarities between person images. The significant variation of person poses and viewing angles challenges for accurate person re-identification. The spatial layout and correspondences between query person images are vital information for tackling this problem but are ignored by most state-of-the-art methods. In this paper, we propose a novel Kronecker Product Matching module to match feature maps of different persons in an end-to-end trainable deep neural network. A novel feature soft warping scheme is designed for aligning the feature maps based on matching results, which is shown to be crucial for achieving superior accuracy. The multi-scale features based on hourglass-like networks and self-residual attention are also exploited to further boost the re-identification performance. The proposed approach outperforms state-of-the-art methods on the Market-1501, CUHK03, and DukeMTMC datasets, which demonstrates the effectiveness and generalization ability of our proposed approach.

Yantao Shen, Hongsheng Li, Tong Xiao et al.

Person re-identification aims at finding a person of interest in an image gallery by comparing the probe image of this person with all the gallery images. It is generally treated as a retrieval problem, where the affinities between the probe image and gallery images (P2G affinities) are used to rank the retrieved gallery images. However, most existing methods only consider P2G affinities but ignore the affinities between all the gallery images (G2G affinity). Some frameworks incorporated G2G affinities into the testing process, which is not end-to-end trainable for deep neural networks. In this paper, we propose a novel group-shuffling random walk network for fully utilizing the affinity information between gallery images in both the training and testing processes. The proposed approach aims at end-to-end refining the P2G affinities based on G2G affinity information with a simple yet effective matrix operation, which can be integrated into deep neural networks. Feature grouping and group shuffle are also proposed to apply rich supervisions for learning better person features. The proposed approach outperforms state-of-the-art methods on the Market-1501, CUHK03, and DukeMTMC datasets by large margins, which demonstrate the effectiveness of our approach.

Yantao Shen, Hongsheng Li, Shuai Yi et al.

The person re-identification task requires to robustly estimate visual similarities between person images. However, existing person re-identification models mostly estimate the similarities of different image pairs of probe and gallery images independently while ignores the relationship information between different probe-gallery pairs. As a result, the similarity estimation of some hard samples might not be accurate. In this paper, we propose a novel deep learning framework, named Similarity-Guided Graph Neural Network (SGGNN) to overcome such limitations. Given a probe image and several gallery images, SGGNN creates a graph to represent the pairwise relationships between probe-gallery pairs (nodes) and utilizes such relationships to update the probe-gallery relation features in an end-to-end manner. Accurate similarity estimation can be achieved by using such updated probe-gallery relation features for prediction. The input features for nodes on the graph are the relation features of different probe-gallery image pairs. The probe-gallery relation feature updating is then performed by the messages passing in SGGNN, which takes other nodes' information into account for similarity estimation. Different from conventional GNN approaches, SGGNN learns the edge weights with rich labels of gallery instance pairs directly, which provides relation fusion more precise information. The effectiveness of our proposed method is validated on three public person re-identification datasets.

Yantao Shen, Tong Xiao, Hongsheng Li et al.

Vehicle re-identification is an important problem and has many applications in video surveillance and intelligent transportation. It gains increasing attention because of the recent advances of person re-identification techniques. However, unlike person re-identification, the visual differences between pairs of vehicle images are usually subtle and even challenging for humans to distinguish. Incorporating additional spatio-temporal information is vital for solving the challenging re-identification task. Existing vehicle re-identification methods ignored or used over-simplified models for the spatio-temporal relations between vehicle images. In this paper, we propose a two-stage framework that incorporates complex spatio-temporal information for effectively regularizing the re-identification results. Given a pair of vehicle images with their spatio-temporal information, a candidate visual-spatio-temporal path is first generated by a chain MRF model with a deeply learned potential function, where each visual-spatio-temporal state corresponds to an actual vehicle image with its spatio-temporal information. A Siamese-CNN+Path-LSTM model takes the candidate path as well as the pairwise queries to generate their similarity score. Extensive experiments and analysis show the effectiveness of our proposed method and individual components.