Guanghui Ye, Huan Zhao, Qin Zhu, Fengnan Li, Jiaqi Li, Yixian Shen, Zhonghao Ren, Zhihua Jiang. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

As semiconductor manufacturing advances from the 3-nm process toward the sub-nanometer regime and transitions from FinFETs to gate-all-around field-effect transistors (GAAFETs), the resulting complexity and manufacturing challenges continue to increase. In this context, 3D chiplet-based approaches have emerged as key enablers to address these limitations while exploiting the expanded design space. Specifically, chiplets help address the lower yields typically associated with large monolithic designs. This paradigm enables the modular design of heterogeneous systems consisting of multiple chiplets (e.g., CPUs, GPUs, memory) fabricated using different technology nodes and processes. Consequently, it offers a capable and cost-effective strategy for designing heterogeneous systems.This paper introduces the Horizon Europe Twinning project COIN-3D (Collaborative Innovation in 3D VLSI Reliability), which aims to strengthen research excellence in 2.5D/3D VLSI systems reliability through collaboration between leading European institutions. More specifically, our primary scientific goal is the provision of novel open-source Electronic Design Automation (EDA) tools for reliability assessment of 3D systems, integrating advanced algorithms for physical- and system-level reliability analysis.

Understanding artworks requires multi-step reasoning over visual content and cultural, historical, and stylistic context. While recent multimodal large language models show promise in artwork explanation, they rely on implicit reasoning and internalized knowledge, limiting interpretability and explicit evidence grounding. We propose A-MAR, an Agent-based Multimodal Art Retrieval framework that explicitly conditions retrieval on structured reasoning plans. Given an artwork and a user query, A-MAR first decomposes the task into a structured reasoning plan that specifies the goals and evidence requirements for each step. Retrieval is then conditioned on this plan, enabling targeted evidence selection and supporting step-wise, grounded explanations. To evaluate agent-based multimodal reasoning within the art domain, we introduce ArtCoT-QA. This diagnostic benchmark features multi-step reasoning chains for diverse art-related queries, enabling a granular analysis that extends beyond simple final answer accuracy. Experiments on SemArt and Artpedia show that A-MAR consistently outperforms static, non-planned retrieval and strong MLLM baselines in final explanation quality, while evaluations on ArtCoT-QA further demonstrate its advantages in evidence grounding and multi-step reasoning ability. These results highlight the importance of reasoning-conditioned retrieval for knowledge-intensive multimodal understanding and position A-MAR as a step toward interpretable, goal-driven AI systems, with particular relevance to cultural industries. The code and data are available at: https://github.com/ShuaiWang97/A-MAR.

Interactive image retrieval overcomes the limitations of single-turn search by allowing users to refine their intent through dialogue. However, developing robust retrieval systems is currently hindered by reliance on pre-generated question-answer pairs that do not capture actual retrieval results during the conversation, preventing the system from dynamically adjusting its strategy to real-time errors. In this paper, we first address this limitation by proposing a Multimodal Conversational Search Simulation framework. This closed-loop environment enables a user simulator to have direct interaction with the retrieval results and generate feedback about the most relevant images. We further propose an agent-based image retrieval system that tracks user preferences in multi-turn interactions and summarizes the historical information into a reformulated query. Leveraging this dynamic environment, we conduct an extensive exploration of query reformulation strategies based on Large Language Models (LLMs) and identify a persistent yet underexplored failure mode in conversational image search: the inability of dense retrievers to process negation and exclusion constraints (e.g., “not red”). We analyze this phenomenon and propose effective mitigation strategies based on zero-shot learning and supervised fine-tuning. Finally, to synthesize these insights into a robust system, we propose a transition from passive query rewriting to Agent-Based Query Reformulation. Unlike traditional methods that merely mimic human conversation, our approach treats the reformulator as a strategic agent optimized for ranking performance. We introduce a novel pipeline that fine-tunes an LLM using Direct Preference Optimization (DPO) on retrieval rewards, effectively enabling the model to learn the specific “dialect” of the search engine. Extensive experiments demonstrate that our agent-based approach significantly outperforms standard baselines, particularly in complex scenarios involving negation and exclusion.

Recent multi-modal large language models (MLLMs) have exhibited powerful abilities in addressing complex vision-language tasks such as image-sequence reasoning (ISR). However, significant challenges remain, e.g., it is still difficult for the MLLMs to fully capture and represent cross-image visual knowledge such as scene relations, attributes, and entity links between multiple images, which hinders them from better solving ISR. To alleviate these issues, we introduce a novel concept Visualized Knowledge Clue (VizKC) - synthetic images that encode key visual and external knowledge from a sequence of input images and are then used alongside the original input images within a multi-image MLLM to enhance reasoning performance. Accordingly, we propose an accompanying approach named VizKC-ISR, composed of two modules - VizKC generation and VizKC utilization. Specifically, in the generation module, VizKC-ISR follows aSee-Find-Fusepipeline: (i) “See - Scene Perception”, to construct an initial VizKC that incorporates scene relations of key visual entities detected from an original image; (ii) “Find - Knowledge Generation”, to generate enriched image captions with real-world knowledge and fine-grained entity details and then extract structured knowledge tuples from generated captions; (iii) “Fuse - Image Editing”, to introduce relevant knowledge tuples into the VizKC via iterative image editing. In the utilization module, we employ a multi-image MLLM (e.g., mPLUG-Owl3) to solve the VizKC-assisted ISR tasks by reasoning with generated knowledge clues. We evaluate VizKC-ISR on nine ISR benchmarks categorized into three multi-image scenarios. The results show that our VizKC-ISR performs best in all tasks, e.g., obtaining the highest average accuracy of 63.1% and surpassing the mPLUG-Owl3 baseline by 6.4 absolute points, due to the bridge between visually-grounded reasoning and multi-modal knowledge challenges.

Large Language Models (LLMs) have shown remarkable performance across various tasks, but the escalating demands on computational resources pose significant challenges, particularly in the extensive utilization of full fine-tuning for downstream tasks. To address this, parameter-efficient fine-tuning (PEFT) methods have been developed, but they often underperform compared to full fine-tuning and struggle with memory efficiency. In this work, we introduce Gradient Weight-Normalized Low-Rank Projection (GradNormLoRP), a novel approach that enhances both parameter and memory efficiency while maintaining comparable performance to full fine-tuning. GradNormLoRP normalizes the weight matrix to improve gradient conditioning, facilitating better convergence during optimization. Additionally, it applies low-rank approximations to the weight and gradient matrices, significantly reducing memory usage during training. Extensive experiments demonstrate that our 8-bit GradNormLoRP reduces optimizer memory usage by up to 89.5\% and enables the pre-training of large LLMs, such as LLaMA 7B, on consumer-level GPUs like the NVIDIA RTX 4090, without additional inference costs. Moreover, GradNormLoRP outperforms existing low-rank methods in fine-tuning tasks. For instance, when fine-tuning the RoBERTa model on all GLUE tasks with a rank of 8, GradNormLoRP achieves an average score of 80.65, surpassing LoRA's score of 79.23. These results underscore GradNormLoRP as a promising alternative for efficient LLM pre-training and fine-tuning.

We present a new adaptation method MaCP, Minimal yet Mighty adaptive Cosine Projection, that achieves exceptional performance while requiring minimal parameters and memory for fine-tuning large foundation models.Its general idea is to exploit the superior energy compaction and decorrelation properties of cosine projection to improve both model efficiency and accuracy.Specifically, it projects the weight change from the low-rank adaptation into the discrete cosine space.Then, the weight change is partitioned over different levels of the discrete cosine spectrum, and each partition's most critical frequency components are selected.Extensive experiments demonstrate the effectiveness of MaCP across a wide range of single-modality tasks, including natural language understanding, natural language generation, text summarization, as well as multimodality tasks such as image classification and video understanding.MaCP consistently delivers superior accuracy, significantly reduced computational complexity, and lower memory requirements compared to existing alternatives.

Existing parameter-efficient fine-tuning (PEFT) methods primarily fall into two categories: addition-based and selective in-situ adaptation.The former, such as LoRA, introduce additional modules to adapt the model to downstream tasks, offering strong memory efficiency.However, their representational capacity is often limited, making them less suitable for fine-grained adaptation.In contrast, the latter directly finetunes a carefully chosen subset of the original model parameters, allowing for more precise and effective adaptation, but at the cost of significantly increased memory consumption.To reconcile this trade-off, we propose NeuroAda, a novel PEFT method that enables fine-grained model finetuning while maintaining high memory efficiency.Our approach first identifies important parameters (i.e., connections within the network) as in selective adaptation, and then introduces bypass connections for these selected parameters.During finetuning, only the bypass connections are updated, leaving the original model parameters frozen.Empirical results on 23+ tasks spanning both natural language generation and understanding demonstrate that NeuroAda achieves state-of-the-art performance with as little as ≤ 0.02% trainable parameters, while reducing CUDA memory usage by up to 60%.We release our code here: https://github.com/ FightingFighting/NeuroAda.git.

Yixian Shen, Qi Bi, Jia-hong Huang, Hongyi Zhu, Andy D. Pimentel, Anuj Pathania. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

Point Cloud Few-Shot Semantic Segmentation (PC-FSS) aims to segment unknown categories in query samples using only a small number of annotated support samples. However, scene complexity and insufficient representation of local geometric structures pose significant challenges to PC-FSS. To address these issues, we propose a novel pre-training-free Visual Introspective Prototype Segmentation network (VIP-Seg). Specifically, we design a Visual Introspective Prototype (VIP) module that employs a multi-step reasoning approach to tackle intra-class diversity and domain gaps between support and query sets. The VIP module consists of a Prototype Enhancement Module (PEM) and a Prototype Difference Module (PDM), which work alternately to progressively refine prototypes. The PEM enhances prototype discriminability and reduces intra-class diversity, while the PDM learns common representations from the differences between query and support features, effectively eliminating semantic inconsistencies caused by domain gaps. To further reduce intra-class diversity and enhance point discriminative ability, we propose a Dynamic Power Convolution (DyPowerConv) that leverages learnable power functions to effectively capture local geometric structures and detailed features of point clouds. Extensive experiments on S3DIS and ScanNet demonstrate that our proposed VIP-Seg significantly outperforms current state-of-the-art methods, proving its effectiveness in PC-FSS tasks. Our code will be available at https://github.com/changshuowang/VIP-Seg .

Image restoration is a fundamental task in computer vision and machine learning, which learns a mapping between the clear images and the degraded images under various conditions (e.g., blur, low-light, haze). Yet, most existing image restoration methods are highly restricted by the requirement of degraded and clear image pairs, which limits the generalization and feasibility to enormous real-world scenarios without paired images. To address this bottleneck, we propose a Degradation-aware Dynamic Schr\"{o}dinger Bridge (DDSB) for unpaired image restoration. Its general idea is to learn a Schr\"{o}dinger Bridge between clear and degraded image distribution, while at the same time emphasizing the physical degradation priors to reduce the accumulation of errors during the restoration process. A Degradation-aware Optimal Transport (DOT) learning scheme is accordingly devised. Training a degradation model to learn the inverse restoration process is particularly challenging, as it must be applicable across different stages of the iterative restoration process. A Dynamic Transport with Consistency (DTC) learning objective is further proposed to reduce the loss of image details in the early iterations and therefore refine the degradation model. Extensive experiments on multiple image degradation tasks show its state-of-the-art performance over the prior arts.

Image search is a pivotal task in multi-media and computer vision, finding applications across diverse domains, ranging from internet search to medical diagnostics. Conventional image search systems operate by accepting textual or visual queries and retrieving the top-relevant candidate results from the database. However, prevalent methods often rely on single-turn procedures, introducing potential inaccuracies and limited recall. These methods also face challenges, such as vocabulary mismatch and the semantic gap, constraining their overall effectiveness. To address these issues, we propose an interactive image retrieval system capable of refining queries based on user relevance feedback in a multi-turn setting. This system incorporates an image captioner based on a vision-language model (VLM) to enhance the quality of text-based queries, resulting in more informative queries with each iteration. Moreover, we introduce a denoiser based on a large language model (LLM) to refine text-based query expansions, mitigating inaccuracies in image descriptions generated by captioning models. To evaluate our system, we curate a new dataset by adapting the MSR-VTT and MSVD video retrieval datasets to the image retrieval task, offering multiple relevant ground-truth images for each query. Through comprehensive experiments, we validate the effectiveness of our proposed system against baseline methods, achieving state-of-the-art performance with a notable 10% improvement in terms of recall. Our contributions encompass the development of an innovative interactive image retrieval system, the integration of an LLM-based denoiser, the curation of a meticulously designed evaluation dataset, and thorough experimental validation.

The legal landscape encompasses a wide array of lawsuit types, presenting lawyers with challenges in delivering timely and accurate information to clients, particularly concerning critical aspects like potential imprisonment duration or financial repercussions. Compounded by the scarcity of legal experts, there's an urgent need to enhance the efficiency of traditional legal workflows. Recent advances in deep learning, especially Large Language Models (LLMs), offer promising solutions to this challenge. Leveraging LLMs' mathematical reasoning capabilities, we propose a novel approach integrating LLM-based methodologies with specially designed prompts to address precision requirements in legal Artificial Intelligence (LegalAI) applications. The proposed work seeks to bridge the gap between traditional legal practices and modern technological advancements, paving the way for a more accessible, efficient, and equitable legal system. To validate this method, we introduce a curated dataset tailored to precision-oriented LegalAI tasks, serving as a benchmark for evaluating LLM-based approaches. Extensive experimentation confirms the efficacy of our methodology in generating accurate numerical estimates within the legal domain, emphasizing the role of LLMs in streamlining legal processes and meeting the evolving demands of LegalAI.

Deep Neural Network (DNN) inference is now ubiquitous in embedded applications at the edge. State-of-the-art Heterogeneous Multi-Processors System-on-Chip (HMPSoCs) powering these applications come equipped with powerful Neural Processing Units (NPUs) that significantly outperform other inference-capable HMPSoC components - namely, the CPUs and GPUs - in terms of power consumption and performance. However, CPUs and GPUs can perform full precision inference, whereas NPUs can often only perform a quantized inference. Consequently, low-latency, low-power inference by the NPU comes at an accuracy loss due to the quantization.

Prevalent large transformer models present significant computational challenges for resource-constrained devices at the Edge. While distributing the workload of deep learning models across multiple edge devices has been extensively studied, these works typically overlook the impact of failures of edge devices. Unpredictable failures, due to, e.g., connectivity issues or discharged batteries, can compromise the reliability of inference serving at the Edge. In this article, we introduce a novel methodology, called EASTER, designed to learn robust distribution strategies for transformer models against device failures that consider the tradeoff between robustness (i.e., maintaining model functionality against failures) and resource utilization (considering memory usage and computations). We evaluate EASTER with three representative transformers—ViT, GPT-2, and Vicuna—under device failures. Our results demonstrate EASTER’s efficiency in memory usage, and possible end-to-end latency improvement for inference across multiple edge devices while preserving model accuracy as much as possible under device failures.

On-chip thermal management is quintessential to a thermally safe operation of a many-core processor. The presence of a physically distributed logically shared Last-Level Cache (LLC) significantly reduces the performance penalty of migrating threads within the cores of an S-NUCA many-core. This cost reduction allows novel thermal management of these many-cores via synchronous thread migration. Synchronous thread migration provides a viable alternative to Dynamic Voltage and Frequency Scaling (DVFS) and asynchronous thread migration used traditionally to manage thermals of S-NUCA many-cores. We present a theoretical method to compute the peak tem-perature in many-cores with synchronous thread migrations. We use the method to create a thermal management heuristic called HotPotato that maximizes the performance of S-NUCA many-cores under a peak temperature constraint. We implement HotPotato within the state-of-the-art HotSniper simulator. Detailed interval thermal simulations with HotSniper show an average 10.72% improvement in response time of S-NUCA many-cores when scheduling with HotPotato compared to a state-of-the-art thermal-aware S-NUCA scheduler.

3D-stacked processor-memory systems stack memory (DRAM banks) directly on top of logic (CPU cores) using chiplet-on-chiplet packaging technology to provide the next-level computing performance in embedded platforms. Stacking, however, severely increases the system’s power density without any accompanying increase in the heat dissipation capacity. Consequently, 3D-stacked processor-memory systems suffer more severe thermal issues than their non-stacked counterparts. Nevertheless, 3D-stacked processor-memory systems do inherit power (thermal) management knobs from their non-stacked predecessors - namely Dynamic Voltage and Frequency Scaling (DVFS) for cores and Low Power Mode (LPM) for memory banks. In the context of 3D-stacked processor-memory systems, DVFS and LPM are performance- and power-wise deeply intertwined. Their non-unified independent use on 3D-stacked processor-memory systems results in sub-optimal thermal management. The unified use of DVFS and LPM for thermal management for 3D-stacked processor-memory systems remains unexplored. The lack of implementation of LPM in thermal simulators for 3D-stacked processor-memory systems hinders real-world representative evaluation for a unified approach. We extend the state-of-the-art interval thermal simulator for 3D-stacked processor-memory systems CoMeT with an LPM power management knob for memory banks. We also propose a learning-based thermal management technique for 3D-stacked processor-memory systems that employ DVFS and LPM in a unified manner. Detailed interval thermal simulations with the extended CoMeT framework show a 10.15% average response time improvement with the PARSEC and SPLASH-2 benchmark suites, along with widely-used Deep Neural Network (DNN) workloads against a state-of-the-art thermal management technique for 2.5D processor-memory systems (ported directly to 3D-stacked processor-memory systems) that also proposes unified use of DVFS and LPM.

Shared caches in multi-core processors seriously complicate the timing verification of real-time software tasks due to the task interference occurring in the shared caches. Explicitly calculating the amount of cache interference among tasks and cache partitioning are two major approaches to enhance the schedulability performance in the context of multi-core processors with shared caches. The former approach suffers from pessimistic cache interference estimations that subsequently result in suboptimal schedulability performance, whereas the latter approach may increase the execution time of tasks due to a lower cache usage, also degrading the schedulability performance.

Shared caches in multi-core processors introduce serious difficulties in providing guarantees on the real-time properties of embedded software due to the interaction and the resulting contention in the shared caches. Prior work has studied the schedulability analysis of global scheduling for real-time multi-core systems with shared caches. This article considers another common scheduling paradigm: partitioned scheduling in the presence of shared cache interference. To achieve this, we propose CITTA, a cache interference-aware task partitioning algorithm. We first analyze the shared cache interference between two programs for set-associative instruction and data caches. Then, an integer programming formulation is constructed to calculate the upper bound on cache interference exhibited by a task, which is required by CITTA. We conduct schedulability analysis of CITTA and formally prove its correctness. A set of experiments is performed to evaluate the schedulability performance of CITTA against global EDF scheduling and other greedy partition approaches such as First-fit and Worst-fit over randomly generated tasksets and realistic workloads in embedded systems. Our empirical evaluations show that CITTA outperforms global EDF scheduling and greedy partition approaches in terms of task sets deemed schedulable.