Boran Zhao

Jin Cui, Xinyue Long, Xunyong Zhang et al.

Multimodal Large Language Models (MLLMs) have made remarkable progress on vision-language reasoning, yet most methods still compress visual evidence into discrete textual thoughts, creating an information bottleneck for fine-grained perception. Recent latent visual reasoning methods attempt to reason in continuous hidden states, but we find that they suffer from insufficient manifold compatibility: latent trajectories drift away from pretrained reasoning circuits, collapse into instance-agnostic patterns, and are often bypassed during answer generation. To address these issues, we propose RIS (Retrieve, Integrate, and Synthesize), a spatial-semantic grounded framework that develops latent reasoning as a compatible extension of pretrained MLLM computation. We first construct a step-wise grounded reasoning dataset with bounding boxes and region-specific semantic descriptions. Built on this supervision, RIS anchors latent tokens to both spatial and semantic evidence, enforces their causal role through a progressive attention bottleneck, and introduces short language transition tokens to bridge synthesized latent states back to vocabulary-aligned decoding. Experiments on V*, HRBench4K, HRBench8K, MMVP, and BLINK show consistent improvements over closed/open-source and latent reasoning baselines. Further analyses demonstrate that RIS learns diverse, interpretable, and progressively integrated latent trajectories, offering a practical path toward faithful internal visual reasoning in MLLMs.

Boran Zhao, Hetian Liu, Zihang Yuan et al.

The growing demand for multi-DNN workloads with unpredictable task arrival times has highlighted the need for interruptible scheduling on edge accelerators. However, existing preemptive frameworks typically assume known task arrival times and rely on CPU-based offline scheduling, which incurs heavy runtime overhead and struggles to handle unpredictable task arrivals. Even worse, prior studies have shown that multi-DNN scheduling requires solving an NP-hard subgraph isomorphism problem on large directed acyclic graphs within limited time, which is extremely challenging. To tackle this, we propose IMMSched, a parallel subgraph isomorphism method that combines Multi-Particle Optimization with the Ullmann algorithm based on a probabilistic continuous-relaxation scheme, eliminating the serial data dependencies of previous works. Finally, a quantized scheduling scheme and a global controller in the hardware architecture further combine multi-particle results for consensus-guided exploration. Evaluations demonstrate that IMMSched achieves orders-of-magnitude reductions in scheduling latency and energy consumption, enabling real-time execution of unpredictable DNN tasks on edge accelerators.

Boran Zhao, Hetian Liu, Zhenxian Hu et al.

The training of large multimodal models fundamentally relies on massive image-text datasets, which inevitably incur prohibitive computational overhead. Dataset selection offers a promising paradigm by identifying a highly informative coreset. However, existing approaches suffer from two critical limitations: (i) single-modality-dominated sampling methods, which ignore the fine-grained cross-modal information imbalance inherent in multimodal datasets and thus lead to semantic loss in the other modality; and (ii) coarse-grained sample-scoring-based sampling methods, where the selected coreset tends to be biased toward the scoring model, making it difficult to guarantee distributional equivalence between the coreset and the original dataset. Meanwhile, existing distribution matching and discrete sampling strategies often fail to jointly account for global semantic structure, local fine-grained details, and redundancy-aware coverage in dense regions. To this end, we propose CAST, a Collapse-Aware multi-Scale Topology fusion framework for multimodal coreset selection. We first construct image- and text-modality topologies, and derive a unified topology via local-collapse-aware refinement and cross-modal fusion. We then introduce a multi-scale distribution matching criterion in the diffusion wavelet domain, encouraging the coreset to approximate the original dataset at multiple scales. Finally, we introduce a local soft relational coverage mechanism that extends pure geometric coverage to relation-aware indirect coverage, penalizing redundant selections in dense clusters. Extensive experiments on Flickr30K and MS-COCO show that CAST outperforms existing dataset selection baselines, showcasing great superiority in cross-architecture generalization and energy efficiency over state-of-the-art multimodal synthesis methods.

Yanbin Hu, Jin Cui, Jiayi Lu et al.

Memory capacity is a critical factor determining the performance of Vision-Language-Action (VLA) models in long-horizon manipulation tasks. Existing memory-augmented architectures primarily rely on linear or flat storage, lacking structural priors for manipulation categories and hierarchical organization. This deficiency hinders efficient experience retrieval and limits generalization to unseen long-horizon task compositions. Inspired by the hierarchical organization of human experience, we propose ECHO (Experience Consolidation and Hierarchical Organization), a novel memory framework operating within a Continuous Hierarchical Space. By employing a hyperbolic autoencoder, ECHO maps VLA hidden states into this space. Leveraging hyperbolic metrics and entailment constraint mechanisms, experience vectors are organized into a semantic memory tree that supports efficient top-down retrieval. In parallel, a background consolidation mechanism continuously refines the memory tree through geometric interpolation and structural splitting, supporting virtual memory synthesis in the continuous space. We integrate ECHO into the $π_0$ foundation model. Evaluations on LIBERO and preliminary real-world experiments demonstrate the effectiveness of our approach, notably achieving a 12.8% absolute improvement in execution success rate over the $π_0$ baseline on LIBERO-Long, while improving compositional generalization on cross-suite unseen long-horizon tasks.

Jin Cui, Jiaqi Guo, Jiepeng Zhou et al.

Chain-of-Thought (CoT) reasoning empowers Large Language Models (LLMs) with remarkable capabilities but typically requires prohibitive parameter scales. CoT distillation has emerged as a promising paradigm to transfer reasoning prowess into compact Student Models (SLMs), but existing approaches often rely on a solitary teacher, capping the student's potential since individual LLMs often exhibit distinct capability biases and may suffer from catastrophic forgetting. While leveraging diverse teachers seems appealing, effectively fusing their supervisions remains challenging: teacher-student incompatibility risks amplifying hallucinations, and passive supervision fails to ensure genuine logic internalization. To address this, we introduce COMPACT, a framework that adaptively fuses supervisions from different teachers by dynamically weighting teacher gradients based on the student's real-time compatibility evaluated by a multi-dimensional metric: (1) Graph-based Consensus to filter misleading rationales by identifying mainstream reasoning paths; (2) Mutual-Information-based Adaptability to detect "epiphany moments" for genuinely understanding the reasoning process rather than merely imitating; and (3) Loss-based Difficulty to assess student receptivity to the teacher's guidance and prevent negative transfer. Extensive experiments and latent space analysis demonstrate that COMPACT effectively integrates diverse reasoning capabilities without damaging the model's original knowledge structure, achieving state-of-the-art performance on various benchmarks while mitigating catastrophic forgetting.

Jin Cui, Jiaqi Guo, Jiepeng Zhou et al.

While Large Language Models (LLMs) have emerged with remarkable capabilities in complex tasks through Chain-of-Thought reasoning, practical resource constraints have sparked interest in transferring these abilities to smaller models. However, achieving both domain performance and cross-domain generalization remains challenging. Existing approaches typically restrict students to following a single golden rationale and treat different reasoning paths independently. Due to distinct inductive biases and intrinsic preferences, alongside the student's evolving capacity and reasoning preferences during training, a teacher's "optimal" rationale could act as out-of-distribution noise. This misalignment leads to a degeneration of the student's latent reasoning distribution, causing suboptimal performance. To bridge this gap, we propose MIND, a capability-adaptive framework that transitions distillation from passive mimicry to active cognitive construction. We synthesize diverse teacher perspectives through a novel "Teaching Assistant" network. By employing a Feedback-Driven Inertia Calibration mechanism, this network utilizes inertia-filtered training loss to align supervision with the student's current adaptability, effectively enhancing performance while mitigating catastrophic forgetting. Extensive experiments demonstrate that MIND achieves state-of-the-art performance on both in-distribution and out-of-distribution benchmarks, and our sophisticated latent space analysis further confirms the mechanism of reasoning ability internalization.

Boran Zhao, Haiduo Huang, Qiwei Dang et al.

Training deep neural networks (DNNs) on edge devices has attracted increasing attention due to its potential to address challenges related to domain adaptation and privacy preservation. However, DNNs typically rely on large datasets for training, which results in substantial energy consumption, making the training in edge devices impractical. Some dataset compression methods have been proposed to solve this challenge. For instance, the coreset selection and dataset distillation reduce the training cost by selecting and generating representative samples respectively. Nevertheless, these methods have two significant defects: (1) The necessary of leveraging a DNN model to evaluate the quality of representative samples, which inevitably introduces inductive bias of DNN, resulting in a severe generalization issue; (2) All training images require multiple accesses to the DDR via long-distance PCB connections, leading to substantial energy overhead. To address these issues, inspired by the nonlinear manifold stationary of the human brain, we firstly propose a DNN-free sample-selecting algorithm, called DE-SNE, to improve the generalization issue. Secondly, we innovatively utilize the near-memory computing technique to implement DE-SNE, thus only a small fraction of images need to access the DDR via long-distance PCB. It significantly reduces DDR energy consumption. As a result, we build a novel expedited DNN training system with a more efficient in-place Near-Memory Sampling characteristic for edge devices, dubbed NMS. As far as we know, our NMS is the first DNN-free near-memory sampling technique that can effectively alleviate generalization issues and significantly reduce DDR energy caused by dataset access. The experimental results show that our NMS outperforms the current state-of-the-art (SOTA) approaches, namely DQ, DQAS, and NeSSA, in model accuracy.

Boran Zhao, Haiming Zhai, Zihang Yuan et al.

The growing demand for sparse tensor algebra (SpTA) in machine learning and big data has driven the development of various sparse tensor accelerators. However, most existing manually designed accelerators are limited to specific scenarios, and it's time-consuming and challenging to adjust a large number of design factors when scenarios change. Therefore, automating the design of SpTA accelerators is crucial. Nevertheless, previous works focus solely on either mapping (i.e., tiling communication and computation in space and time) or sparse strategy (i.e., bypassing zero elements for efficiency), leading to suboptimal designs due to the lack of comprehensive consideration of both. A unified framework that jointly optimizes both is urgently needed. However, integrating mapping and sparse strategies leads to a combinatorial explosion in the design space(e.g., as large as $O(10^{41})$ for the workload $P_{32 \times 64} \times Q_{64 \times 48} = Z_{32 \times 48}$). This vast search space renders most conventional optimization methods (e.g., particle swarm optimization, reinforcement learning and Monte Carlo tree search) inefficient. To address this challenge, we propose an evolution strategy-based sparse tensor accelerator optimization framework, called SparseMap. SparseMap constructing a more comprehensive design space with the consideration of both mapping and sparse strategy. We introduce a series of enhancements to genetic encoding and evolutionary operators, enabling SparseMap to efficiently explore the vast and diverse design space. We quantitatively compare SparseMap with prior works and classical optimization methods, demonstrating that SparseMap consistently finds superior solutions.

Jin Cui, Boran Zhao, Jiajun Xu et al.

Coreset selection compresses large datasets into compact, representative subsets, reducing the energy and computational burden of training deep neural networks. Existing methods are either: (i) DNN-based, which are tied to model-specific parameters and introduce architectural bias; or (ii) DNN-free, which rely on heuristics lacking theoretical guarantees. Neither approach explicitly constrains distributional equivalence, largely because continuous distribution matching is considered inapplicable to discrete sampling. Moreover, prevalent metrics (e.g., MSE, KL, CE, MMD) cannot accurately capture higher-order moment discrepancies, leading to suboptimal coresets. In this work, we propose FAST, the first DNN-free distribution-matching coreset selection framework that formulates the coreset selection task as a graph-constrained optimization problem grounded in spectral graph theory and employs the Characteristic Function Distance (CFD) to capture full distributional information in the frequency domain. We further discover that naive CFD suffers from a "vanishing phase gradient" issue in medium and high-frequency regions; to address this, we introduce an Attenuated Phase-Decoupled CFD. Furthermore, for better convergence, we design a Progressive Discrepancy-Aware Sampling strategy that progressively schedules frequency selection from low to high, preserving global structure before refining local details and enabling accurate matching with fewer frequencies while avoiding overfitting. Extensive experiments demonstrate that FAST significantly outperforms state-of-the-art coreset selection methods across all evaluated benchmarks, achieving an average accuracy gain of 9.12%. Compared to other baseline coreset methods, it reduces power consumption by 96.57% and achieves a 2.2x average speedup, underscoring its high performance and energy efficiency.

Boran Zhao, Hetian Liu, Zihang Yuan et al.

Training deep neural networks (DNNs) directly on edge devices has attracted increasing attention, as it offers promising solutions to challenges such as domain adaptation and privacy preservation. However, conventional DNN training typically requires large-scale datasets, which imposes prohibitive overhead on edge devices-particularly for emerging large language model (LLM) tasks. To address this challenge, a DNN-free method (ie., dataset sampling without DNN), named NMS (Near-Memory Sampling), has been introduced. By first conducting dimensionality reduction of the dataset and then performing exemplar sampling in the reduced space, NMS avoids the architectural bias inherent in DNN-based methods and thus achieves better generalization. However, The state-of-the-art, NMS, suffers from two limitations: (1) The mismatch between the search method and the non-monotonic property of the perplexity error function leads to the emergence of outliers in the reduced representation; (2) Key parameter (ie., target perplexity) is selected empirically, introducing arbitrariness and leading to uneven sampling. These two issues lead to representative bias of examplars, resulting in degraded accuracy. To address these issues, we propose AdapSNE, which integrates an efficient non-monotonic search method-namely, the Fireworks Algorithm (FWA)-to suppress outliers, and employs entropy-guided optimization to enforce uniform sampling, thereby ensuring representative training samples and consequently boosting training accuracy. To cut the edge-side cost arising from the iterative computations of FWA search and entropy-guided optimization, we design an accelerator with custom dataflow and time-multiplexing markedly reducing on-device training energy and area.