Shwai He

Shwai He, Liang Ding, Daize Dong et al.

Adapter Tuning, which freezes the pretrained language models (PLMs) and only fine-tunes a few extra modules, becomes an appealing efficient alternative to the full model fine-tuning. Although computationally efficient, the recent Adapters often increase parameters (e.g. bottleneck dimension) for matching the performance of full model fine-tuning, which we argue goes against their original intention. In this work, we re-examine the parameter-efficiency of Adapters through the lens of network pruning (we name such plug-in concept as \texttt{SparseAdapter}) and find that SparseAdapter can achieve comparable or better performance than standard Adapters when the sparse ratio reaches up to 80\%. Based on our findings, we introduce an easy but effective setting ``\textit{Large-Sparse}'' to improve the model capacity of Adapters under the same parameter budget. Experiments on five competitive Adapters upon three advanced PLMs show that with proper sparse method (e.g. SNIP) and ratio (e.g. 40\%) SparseAdapter can consistently outperform their corresponding counterpart. Encouragingly, with the \textit{Large-Sparse} setting, we can obtain further appealing gains, even outperforming the full fine-tuning by a large margin. Our code will be released at: https://github.com/Shwai-He/SparseAdapter.

Shwai He, Run-Ze Fan, Liang Ding et al.

Scaling the size of language models usually leads to remarkable advancements in NLP tasks. But it often comes with a price of growing computational cost. Although a sparse Mixture of Experts (MoE) can reduce the cost by activating a small subset of parameters (e.g., one expert) for each input, its computation escalates significantly if increasing the number of activated experts, limiting its practical utility. Can we retain the advantages of adding more experts without substantially increasing the computational costs? In this paper, we first demonstrate the superiority of selecting multiple experts and then propose a computation-efficient approach called \textbf{\texttt{Merging Experts into One}} (MEO), which reduces the computation cost to that of a single expert. Extensive experiments show that MEO significantly improves computational efficiency, e.g., FLOPS drops from 72.0G of vanilla MoE to 28.6G (MEO). Moreover, we propose a token-level attention block that further enhances the efficiency and performance of token-level MEO, e.g., 83.3\% (MEO) vs. 82.6\% (vanilla MoE) average score on the GLUE benchmark. Our code will be released upon acceptance. Code will be released at: \url{https://github.com/Shwai-He/MEO}.

Shwai He, Liang Ding, Daize Dong et al.

Dynamic networks, e.g., Dynamic Convolution (DY-Conv) and the Mixture of Experts (MoE), have been extensively explored as they can considerably improve the model's representation power with acceptable computational cost. The common practice in implementing dynamic networks is to convert the given static layers into fully dynamic ones where all parameters are dynamic (at least within a single layer) and vary with the input. However, such a fully dynamic setting may cause redundant parameters and high deployment costs, limiting the applicability of dynamic networks to a broader range of tasks and models. The main contributions of our work are challenging the basic commonsense in dynamic networks and proposing a partially dynamic network, namely PAD-Net, to transform the redundant dynamic parameters into static ones. Also, we further design Iterative Mode Partition to partition dynamic and static parameters efficiently. Our method is comprehensively supported by large-scale experiments with two typical advanced dynamic architectures, i.e., DY-Conv and MoE, on both image classification and GLUE benchmarks. Encouragingly, we surpass the fully dynamic networks by $+0.7\%$ top-1 acc with only $30\%$ dynamic parameters for ResNet-50 and $+1.9\%$ average score in language understanding with only $50\%$ dynamic parameters for BERT. Code will be released at: \url{https://github.com/Shwai-He/PAD-Net}.

Changtong Zan, Keqin Peng, Liang Ding et al.

We describe the JD Explore Academy's submission of the WMT 2022 shared general translation task. We participated in all high-resource tracks and one medium-resource track, including Chinese-English, German-English, Czech-English, Russian-English, and Japanese-English. We push the limit of our previous work -- bidirectional training for translation by scaling up two main factors, i.e. language pairs and model sizes, namely the \textbf{Vega-MT} system. As for language pairs, we scale the "bidirectional" up to the "multidirectional" settings, covering all participating languages, to exploit the common knowledge across languages, and transfer them to the downstream bilingual tasks. As for model sizes, we scale the Transformer-Big up to the extremely large model that owns nearly 4.7 Billion parameters, to fully enhance the model capacity for our Vega-MT. Also, we adopt the data augmentation strategies, e.g. cycle translation for monolingual data, and bidirectional self-training for bilingual and monolingual data, to comprehensively exploit the bilingual and monolingual data. To adapt our Vega-MT to the general domain test set, generalization tuning is designed. Based on the official automatic scores of constrained systems, in terms of the sacreBLEU shown in Figure-1, we got the 1st place on {Zh-En (33.5), En-Zh (49.7), De-En (33.7), En-De (37.8), Cs-En (54.9), En-Cs (41.4) and En-Ru (32.7)}, 2nd place on {Ru-En (45.1) and Ja-En (25.6)}, and 3rd place on {En-Ja(41.5)}, respectively; W.R.T the COMET, we got the 1st place on {Zh-En (45.1), En-Zh (61.7), De-En (58.0), En-De (63.2), Cs-En (74.7), Ru-En (64.9), En-Ru (69.6) and En-Ja (65.1)}, 2nd place on {En-Cs (95.3) and Ja-En (40.6)}, respectively.

Shwai He, Run-Ze Fan, Liang Ding et al.

Adapter tuning, which updates only a few parameters, has become a mainstream method for fine-tuning pretrained language models to downstream tasks. However, it often yields subpar results in few-shot learning. AdapterFusion, which assembles pretrained adapters using composition layers tailored to specific tasks, is a possible solution but significantly increases trainable parameters and deployment costs. Despite this, our preliminary study reveals that even single adapters can outperform Adapterfusion in few-shot learning, urging us to propose \textbf{\texttt{Merging Pretrained Adapters}} (MerA) that efficiently incorporates pretrained adapters to a single model through model fusion. Extensive experiments on two PLMs demonstrate that MerA achieves substantial improvements compared to both single adapters and AdapterFusion. To further enhance the capacity of MerA, we also introduce a simple yet effective technique, referred to as the "\textit{same-track}" setting, that merges adapters from the same track of pretraining tasks. With the implementation of the "\textit{same-track}" setting, we observe even more impressive gains, surpassing the performance of both full fine-tuning and adapter tuning by a substantial margin, e.g., 3.5\% in MRPC and 5.0\% in MNLI.

Shwai He, Guoheng Sun, Haichao Zhang et al.

Network pruning, which removes less important parameters or architectures, is often expected to improve efficiency while preserving performance. However, this expectation does not consistently hold across language tasks: pruned models can perform well on non-generative tasks but frequently fail in generative settings. To understand this discrepancy, we analyze network pruning from a representation-hierarchy perspective, decomposing the internal computation of language models into three sequential spaces: embedding (hidden representations), logit (pre-softmax outputs), and probability (post-softmax distributions). We find that representations in the embedding and logit spaces are largely robust to pruning-induced perturbations. However, the nonlinear transformation from logits to probabilities amplifies these deviations, which accumulate across time steps and lead to substantial degradation during generation. In contrast, the stability of the categorical-token probability subspace, together with the robustness of the embedding space, supports the effectiveness of pruning for non-generative tasks such as retrieval and multiple-choice selection. Our analysis disentangles the effects of pruning across tasks and provides practical guidance for its application. Code is available at https://github.com/CASE-Lab-UMD/Pruning-on-Representations

Shwai He, Chaorui Deng, Ang Li et al.

Large multimodal models have achieved remarkable progress in both understanding and generation. Recent efforts pursue unified multimodal models that integrate heterogeneous components to support both capabilities within a single framework. However, such unification introduces inference inefficiencies, e.g., specific tasks or samples may not require the full knowledge or capacity of the unified model. Yet, a systematic understanding of how these inefficiencies manifest across different components remains limited. In this work, we first conduct a systematic analysis of unified multimodal model components using training-free pruning as a probing methodology, considering both depth pruning and width reduction. Our study reveals that the understanding component exhibits notable compressibility in both understanding and generation tasks, which is more pronounced in the latter. In contrast, the generation components are highly sensitive to compression, with performance deteriorating sharply even under moderate compression ratios. To address this limitation, we propose the Mixture-of-Experts (MoE) Adaptation, inspired by the dynamic activation patterns observed across different samples. This approach partitions the generation module into multiple experts and enables sparse activation to restore generation quality. We validate the effectiveness of sparse activation through expert-frozen tuning and further demonstrate that a fully trainable adaptation delivers additional gains. As a result, the adapted BAGEL model achieves performance comparable to the full model while activating only about half of its parameters. The code is released at \href{https://github.com/Shwai-He/SparseUnifiedModel}{this link}.

Ming Li, Lichang Chen, Jiuhai Chen et al.

Recent advancements in Large Language Models (LLMs) have expanded the horizons of natural language understanding and generation. Notably, the output control and alignment with the input of LLMs can be refined through instruction tuning. However, as highlighted in several studies, low-quality data in the training set are usually detrimental to instruction tuning, resulting in inconsistent or even misleading LLM outputs. We propose a novel method, termed "reflection-tuning," which addresses the problem by self-improvement and judging capabilities of LLMs. This approach utilizes an oracle LLM to recycle the original training data by introspecting and enhancing the quality of instructions and responses in the data. Extensive experiments on widely used evaluation benchmarks show that LLMs trained with our recycled data outperform those trained with existing datasets in various benchmarks.

Shwai He, Chenbo Jiang, Daize Dong et al.

Dynamic convolution achieves better performance for efficient CNNs at the cost of negligible FLOPs increase. However, the performance increase can not match the significantly expanded number of parameters, which is the main bottleneck in real-world applications. Contrastively, mask-based unstructured pruning obtains a lightweight network by removing redundancy in the heavy network. In this paper, we propose a new framework, \textbf{Sparse Dynamic Convolution} (\textsc{SD-Conv}), to naturally integrate these two paths such that it can inherit the advantage of dynamic mechanism and sparsity. We first design a binary mask derived from a learnable threshold to prune static kernels, significantly reducing the parameters and computational cost but achieving higher performance in Imagenet-1K. We further transfer pretrained models into a variety of downstream tasks, showing consistently better results than baselines. We hope our SD-Conv could be an efficient alternative to conventional dynamic convolutions.

Run-Ze Fan, Xuefeng Li, Haoyang Zou et al.

The quality of finetuning data is crucial for aligning large language models (LLMs) with human values. Current methods to improve data quality are either labor-intensive or prone to factual errors caused by LLM hallucinations. This paper explores elevating the quality of existing instruction data to better align with human values, introducing a simple and effective approach named ReAlign, which reformats the responses of instruction data into a format that better aligns with pre-established criteria and the collated evidence. This approach minimizes human annotation, hallucination, and the difficulty in scaling, remaining orthogonal to existing alignment techniques. Experimentally, ReAlign significantly boosts the general alignment ability, math reasoning, factuality, and readability of the LLMs. Encouragingly, without introducing any additional data or advanced training techniques, and merely by reformatting the response, LLaMA-2-13B's mathematical reasoning ability on GSM8K can be improved from 46.77% to 56.63% in accuracy. Additionally, a mere 5% of ReAlign data yields a 67% boost in general alignment ability measured by the Alpaca dataset. This work highlights the need for further research into the science and mechanistic interpretability of LLMs. We have made the associated code and data publicly accessible to support future studies at https://github.com/GAIR-NLP/ReAlign.

Yibin Lei, Shwai He, Ang Li et al.

Recent work has shown that directly fine-tuning large language models (LLMs) for dense retrieval yields strong performance, but their substantial parameter counts make them computationally inefficient. While prior studies have revealed significant layer redundancy in LLMs for generative tasks, it remains unclear whether similar redundancy exists when these models are adapted for retrieval tasks, which require encoding entire sequences into fixed representations rather than generating tokens iteratively. To this end, we conduct a comprehensive analysis of layer redundancy in LLM-based dense retrievers. We find that, in contrast to generative settings, MLP layers are substantially more prunable, while attention layers remain critical for semantic aggregation. Building on this insight, we propose EffiR, a framework for developing efficient retrievers that performs large-scale MLP compression through a coarse-to-fine strategy (coarse-grained depth reduction followed by fine-grained width reduction), combined with retrieval-specific fine-tuning. Across diverse BEIR datasets and LLM backbones, EffiR achieves substantial reductions in model size and inference cost while preserving the performance of full-size models.

Guoheng Sun, Ziyao Wang, Bowei Tian et al.

As post-training techniques evolve, large language models (LLMs) are increasingly augmented with structured multi-step reasoning abilities, often optimized through reinforcement learning. These reasoning-enhanced models outperform standard LLMs on complex tasks and now underpin many commercial LLM APIs. However, to protect proprietary behavior and reduce verbosity, providers typically conceal the reasoning traces while returning only the final answer. This opacity introduces a critical transparency gap: users are billed for invisible reasoning tokens, which often account for the majority of the cost, yet have no means to verify their authenticity. This opens the door to token count inflation, where providers may overreport token usage or inject synthetic, low-effort tokens to inflate charges. To address this issue, we propose CoIn, a verification framework that audits both the quantity and semantic validity of hidden tokens. CoIn constructs a verifiable hash tree from token embedding fingerprints to check token counts, and uses embedding-based relevance matching to detect fabricated reasoning content. Experiments demonstrate that CoIn, when deployed as a trusted third-party auditor, can effectively detect token count inflation with a success rate reaching up to 94.7%, showing the strong ability to restore billing transparency in opaque LLM services. The dataset and code are available at https://github.com/CASE-Lab-UMD/LLM-Auditing-CoIn.

Shwai He, Tao Ge, Guoheng Sun et al.

Traditional transformer models often allocate a fixed amount of computational resources to every input token, leading to inefficient and unnecessary computation. To address this, the Mixture of Depths (MoD) was introduced to dynamically adjust the computational depth by skipping less important layers. Despite its promise, current MoD approaches remain under-explored and face two main challenges: (1) high training costs due to the need to train the entire model along with the routers that determine which layers to skip, and (2) the risk of performance degradation when important layers are bypassed. In response to the first issue, we propose Router-Tuning, a method that fine-tunes only the router on a small dataset, drastically reducing the computational overhead associated with full model training. For the second challenge, we propose MindSkip, which deploys Attention with Dynamic Depths. This method preserves the model's performance while significantly enhancing computational and memory efficiency. Extensive experiments demonstrate that our approach delivers competitive results while dramatically improving the computation efficiency, e.g., 21\% speedup and only a 0.2\% performance drop. The code is released at https://github.com/CASE-Lab-UMD/Router-Tuning.

Shwai He, Ang Li, Tianlong Chen

Vision-Language Models (VLMs) integrate information from multiple modalities and have shown remarkable success across various tasks. However, deploying large-scale VLMs in resource-constrained scenarios is challenging. Pruning followed by finetuning offers a potential solution but remains underexplored for VLMs. This study addresses two key questions: how to distribute sparsity across different modality-specific models, and how to restore the performance of pruned sparse VLMs. Our preliminary studies identified two effective pruning settings: applying the same sparsity to both vision and language models, and pruning only the language models. While LoRA finetuning aims to restore sparse models, it faces challenges due to incompatibility with sparse models, disrupting the pruned sparsity. To overcome these issues, we propose SparseLoRA, which applies sparsity directly to LoRA weights. Our experimental results demonstrate significant improvements, including an 11.3\% boost under 2:4 sparsity and a 47.6\% enhancement under unstructured 70\% sparsity. Code is released at: \url{https://github.com/Shwai-He/VLM-Compression}.

Haichao Zhang, Yijiang Li, Shwai He et al.

Recent progress in latent world models (e.g., V-JEPA2) has shown promising capability in forecasting future world states from video observations. Nevertheless, dense prediction from a short observation window limits temporal context and can bias predictors toward local, low-level extrapolation, making it difficult to capture long-horizon semantics and reducing downstream utility. Vision--language models (VLMs), in contrast, provide strong semantic grounding and general knowledge by reasoning over uniformly sampled frames, but they are not ideal as standalone dense predictors due to compute-driven sparse sampling, a language-output bottleneck that compresses fine-grained interaction states into text-oriented representations, and a data-regime mismatch when adapting to small action-conditioned datasets. We propose a VLM-guided JEPA-style latent world modeling framework that combines dense-frame dynamics modeling with long-horizon semantic guidance via a dual-temporal pathway: a dense JEPA branch for fine-grained motion and interaction cues, and a uniformly sampled VLM \emph{thinker} branch with a larger temporal stride for knowledge-rich guidance. To transfer the VLM's progressive reasoning signals effectively, we introduce a hierarchical pyramid representation extraction module that aggregates multi-layer VLM representations into guidance features compatible with latent prediction. Experiments on hand-manipulation trajectory prediction show that our method outperforms both a strong VLM-only baseline and a JEPA-predictor baseline, and yields more robust long-horizon rollout behavior.

Hanfei Yu, Bei Ouyang, Shwai He et al.

Large Language Models (LLMs) have become a cornerstone of AI, driving progress across diverse domains such as content creation, search and recommendation systems, and AI-assisted workflows. To alleviate extreme training costs and advancing model scales, Mixture-of-Experts (MoE) has become a popular backbone for modern LLMs, which are commonly served in distributed deployment using expert parallelism (EP). However, MoE's sparse activation mechanism leads to severe expert load imbalance, where a few experts become overloaded while others remain idle, resulting in expert stragglers that inflate inference latency and serving cost. Existing expert load balancing solutions assume static resource configurations on serverful infrastructures, limiting expert scalability and elasticity, and resulting in either costly real-time expert swapping or degraded generation quality. We present MoEless, the first serverless MoE serving framework that mitigates expert load imbalance and accelerates inference via serverless experts. MoEless employs lightweight, layer-aware predictors to accurately estimate incoming expert load distributions and proactively identify stragglers. We design optimized expert scaling and placement strategies to maximize function locality, improve GPU utilization, and balance loads across experts and GPUs. MoEless is prototyped on top of Megatron-LM and deployed on an eight-GPU testbed. Experiments with open-source MoE models and real-world workloads show that MoEless reduces inference latency by 43% and inference cost by 84% compared to state-of-the-art solutions.

Ming Li, Yong Zhang, Shwai He et al.

Instruction tuning is critical to improve LLMs but usually suffers from low-quality and redundant data. Data filtering for instruction tuning has proved important in improving both the efficiency and performance of the tuning process. But it also leads to extra cost and computation due to the involvement of LLMs in this process. To reduce the filtering cost, we study Superfiltering: Can we use a smaller and weaker model to select data for finetuning a larger and stronger model? Despite the performance gap between weak and strong language models, we find their highly consistent capability to perceive instruction difficulty and data selection results. This enables us to use a much smaller and more efficient model to filter the instruction data used to train a larger language model. Not only does it largely speed up the data filtering, but the filtered-data-finetuned LLM achieves even better performance on standard benchmarks. Extensive experiments validate the efficacy and efficiency of our approach.

Shwai He, Guoheng Sun, Zheyu Shen et al.

While scaling Transformer-based large language models (LLMs) has demonstrated promising performance across various tasks, it also introduces redundant architectures, posing efficiency challenges for real-world deployment. Despite some recognition of redundancy in LLMs, the variability of redundancy across different architectures in transformers, such as MLP and Attention layers, is under-explored. In this work, we investigate redundancy across different modules within Transformers, including Blocks, MLP, and Attention layers, using a similarity-based metric. Surprisingly, despite the critical role of attention layers in distinguishing transformers from other architectures, we found that a large portion of these layers exhibit excessively high similarity and can be pruned without degrading performance. For instance, Llama-2-70B achieved a 48.4\% speedup with only a 2.4\% performance drop by pruning half of the attention layers. Furthermore, by tracing model checkpoints throughout the training process, we observed that attention layer redundancy is inherent and consistent across training stages. Additionally, we further propose a method that jointly drops Attention and MLP layers, allowing us to more aggressively drop additional layers. For instance, when dropping 31 layers (Attention + MLP), Llama-2-13B still retains 90\% of the performance on the MMLU task. Our work provides valuable insights for future network architecture design. The code is released at: \url{https://github.com/Shwai-He/LLM-Drop}.

Shwai He, Daize Dong, Liang Ding et al.

Scaling large language models has driven remarkable advancements across various domains, yet the continual increase in model size presents significant challenges for real-world deployment. The Mixture of Experts (MoE) architecture offers a promising solution by dynamically selecting and activating only a subset of experts during inference, thus substantially reducing computational costs while preserving high performance. Despite these benefits, MoE introduces new inefficiencies, such as excessive parameters and communication overhead. In this work, we present a holistic study of compression techniques for Mixture of Experts to enhance both efficiency and scalability. While recent efforts have focused on Expert Trimming, which reduces the number of experts, these approaches still suffer from considerable communication and computational costs. To address this, we propose more aggressive strategies, such as Layer Drop, which removes entire MoE layers, and Block Drop, which eliminates transformer blocks. Surprisingly, these aggressive pruning techniques not only preserve model performance but also substantially improve computation and memory efficiency. Furthermore, beyond Expert Trimming, we also introduce Expert Slimming, which compresses individual experts to further boost performance and can be seamlessly integrated with Expert Trimming. Extensive experimental results demonstrate the effectiveness of our proposed methods-Layer Drop and Block Drop-along with the comprehensive recipe that integrates Expert Slimming and Expert Trimming, achieving a 6.05x speedup with 77.1% reduced memory usage while maintaining over 92% of performance on Mixtral-8x7B. Our code is released at https://github.com/CASE-Lab-UMD/Unified-MoE-Compression.

Bowei Tian, Ziyao Wang, Shwai He et al.

The challenge of balancing fairness and predictive accuracy in machine learning models, especially when sensitive attributes such as race, gender, or age are considered, has motivated substantial research in recent years. Counterfactual fairness ensures that predictions remain consistent across counterfactual variations of sensitive attributes, which is a crucial concept in addressing societal biases. However, existing counterfactual fairness approaches usually overlook intrinsic information about sensitive features, limiting their ability to achieve fairness while simultaneously maintaining performance. To tackle this challenge, we introduce EXOgenous Causal reasoning (EXOC), a novel causal reasoning framework motivated by exogenous variables. It leverages auxiliary variables to uncover intrinsic properties that give rise to sensitive attributes. Our framework explicitly defines an auxiliary node and a control node that contribute to counterfactual fairness and control the information flow within the model. Our evaluation, conducted on synthetic and real-world datasets, validates EXOC's superiority, showing that it outperforms state-of-the-art approaches in achieving counterfactual fairness. Our code is available at https://github.com/CASE-Lab-UMD/counterfactual_fairness_2025.

Ming Li, Lichang Chen, Jiuhai Chen et al.

Instruction tuning is critical to large language models (LLMs) for achieving better instruction following and task adaptation capabilities but its success heavily relies on the training data quality. Many recent methods focus on improving the data quality but often overlook the compatibility of the data with the student model being finetuned. This paper introduces Selective Reflection-Tuning, a novel paradigm that synergizes a teacher LLM's reflection and introspection for improving existing data quality with the data selection capability of the student LLM, to automatically refine existing instruction-tuning data. This teacher-student collaboration produces high-quality and student-compatible instruction-response pairs, resulting in sample-efficient instruction tuning and LLMs of superior performance. Selective Reflection-Tuning is a data augmentation and synthesis that generally improves LLM finetuning and self-improvement without collecting brand-new data. We apply our method to Alpaca and WizardLM data and achieve much stronger and top-tier 7B and 13B LLMs.

Yiting Wang, Wanghao Ye, Ping Guo et al.

Optimizing Register Transfer Level (RTL) code is crucial for improving the power, performance, and area (PPA) of digital circuits in the early stages of synthesis. Manual rewriting, guided by synthesis feedback, can yield high-quality results but is time-consuming and error-prone. Most existing compiler-based approaches have difficulty handling complex design constraints. Large Language Model (LLM)-based methods have emerged as a promising alternative to address these challenges. However, LLM-based approaches often face difficulties in ensuring alignment between the generated code and the provided prompts. This paper presents SymRTLO, a novel neuron-symbolic RTL optimization framework that seamlessly integrates LLM-based code rewriting with symbolic reasoning techniques. Our method incorporates a retrieval-augmented generation (RAG) system of optimization rules and Abstract Syntax Tree (AST)-based templates, enabling LLM-based rewriting that maintains syntactic correctness while minimizing undesired circuit behaviors. A symbolic module is proposed for analyzing and optimizing finite state machine (FSM) logic, allowing fine-grained state merging and partial specification handling beyond the scope of pattern-based compilers. Furthermore, a fast verification pipeline, combining formal equivalence checks with test-driven validation, further reduces the complexity of verification. Experiments on the RTL-Rewriter benchmark with Synopsys Design Compiler and Yosys show that SymRTLO improves power, performance, and area (PPA) by up to 43.9%, 62.5%, and 51.1%, respectively, compared to the state-of-the-art methods.

Shwai He, Weilin Cai, Jiayi Huang et al.

The Mixture of Experts (MoE) is an effective architecture for scaling large language models by leveraging sparse expert activation to balance performance and efficiency. However, under expert parallelism, MoE suffers from inference inefficiencies due to imbalanced token-to-expert assignment, where underloaded experts complete computations early but must wait for overloaded experts, leading to global delays. We define this phenomenon as the \textbf{\textit{Straggler Effect}}, as the most burdened experts dictate the overall inference latency. To address this, we first propose \textit{\textbf{Capacity-Aware Token Drop}}, which enforces expert capacity limits by discarding excess tokens from overloaded experts, effectively reducing load imbalance with minimal performance impact (e.g., $30\%$ speedup with only $0.9\%$ degradation on OLMoE). Next, given the presence of low-load experts remaining well below the capacity threshold, we introduce \textit{\textbf{Capacity-Aware Expanded Drop}}, which allows tokens to include additional local experts in their candidate set before enforcing strict local capacity constraints, thereby improving load balance and enhancing the utilization of underused experts. Extensive experiments on both language and multimodal MoE models demonstrate the effectiveness of our approach, yielding substantial gains in expert utilization, model performance, and inference efficiency, e.g., applying Expanded Drop to Mixtral-8$\times$7B-Instruct yields a {0.2\%} average performance improvement and a {1.85$\times$} inference speedup.

Wanghao Ye, Sihan Chen, Yiting Wang et al.

We present an LLM-powered social discovery platform that uses digital twins to autonomously evaluate interpersonal compatibility through behavioral simulation. The platform unifies three key pillars: (1) digital twins that engage in autonomous multi-turn conversations on behalf of users to estimate compatibility, (2) gamified territory conquest mechanics that incentivize real-world exploration and create organic settings for in-person encounters, and (3) AI companions that preserve persistent shared memory across devices. Built upon CogniPair's cognitive architecture (Ye et al., 2026), validated on the Columbia Speed Dating dataset (551 participants), our system extends prior simulation-only matching into a fully deployed social discovery environment. Through deployment, we derive empirical cost-quality baselines and identify fundamental scaling bottlenecks that remain hidden in component-level testing alone.

Haichao Zhang, Wenhao Chai, Shwai He et al.

High temporal resolution is essential for capturing fine-grained details in video understanding. However, current video large language models (VLLMs) and benchmarks mostly rely on low-frame-rate sampling, such as uniform sampling or keyframe selection, discarding dense temporal information. This compromise avoids the high cost of tokenizing every frame, which otherwise leads to redundant computation and linear token growth as video length increases. While this trade-off works for slowly changing content, it fails for tasks like lecture comprehension, where information appears in nearly every frame and requires precise temporal alignment. To address this gap, we introduce Dense Video Understanding (DVU), which enables high-FPS video comprehension by reducing both tokenization time and token overhead. Existing benchmarks are also limited, as their QA pairs focus on coarse content changes. We therefore propose DIVE (Dense Information Video Evaluation), the first benchmark designed for dense temporal reasoning. To make DVU practical, we present Gated Residual Tokenization (GRT), a two-stage framework: (1) Motion-Compensated Inter-Gated Tokenization uses pixel-level motion estimation to skip static regions during tokenization, achieving sub-linear growth in token count and compute. (2) Semantic-Scene Intra-Tokenization Merging fuses tokens across static regions within a scene, further reducing redundancy while preserving dynamic semantics. Experiments on DIVE show that GRT outperforms larger VLLM baselines and scales positively with FPS. These results highlight the importance of dense temporal information and demonstrate that GRT enables efficient, scalable high-FPS video understanding.

Weilin Cai, Le Qin, Shwai He et al.

Mixture of Experts (MoE) has become a mainstream architecture for building Large Language Models (LLMs) by reducing per-token computation while enabling model scaling. It can be viewed as partitioning a large Feed-Forward Network (FFN) at the tensor level into fine-grained sub-FFNs, or experts, and activating only a sparse subset for each input. While this sparsity improves efficiency, MoE still faces substantial challenges due to their massive computational scale and unpredictable activation patterns. To enable efficient MoE deployment, we identify dual sparsity at the tensor and neuron levels in pre-trained MoE modules as a key factor for both accuracy and efficiency. Unlike prior work that increases tensor-level sparsity through finer-grained expert design during pre-training, we introduce post-training expert partitioning to induce such sparsity without retraining. This preserves the mathematical consistency of model transformations and enhances both efficiency and accuracy in subsequent fine-tuning and inference. Building upon this, we propose DualSparse-MoE, an inference system that integrates dynamic tensor-level computation dropping with static neuron-level reconstruction to deliver significant efficiency gains with minimal accuracy loss. Experimental results show that enforcing an approximate 25% drop rate with our approach reduces average accuracy by only 0.08%-0.28% across three prevailing MoE models, while nearly all degrees of computation dropping consistently yield proportional computational speedups. Furthermore, incorporating load-imbalance awareness into expert parallelism achieves a 1.41x MoE module speedup with just 0.5% average accuracy degradation.

Wanghao Ye, Sihan Chen, Yiting Wang et al.

Current large language model (LLM) agents lack authentic human psychological processes necessary for genuine digital twins and social AI applications. To address this limitation, we present a computational implementation of Global Workspace Theory (GNWT) that integrates human cognitive architecture principles into LLM agents, creating specialized sub-agents for emotion, memory, social norms, planning, and goal-tracking coordinated through a global workspace mechanism. However, authentic digital twins require accurate personality initialization. We therefore develop a novel adventure-based personality test that evaluates true personality through behavioral choices within interactive scenarios, bypassing self-presentation bias found in traditional assessments. Building on these innovations, our CogniPair platform enables digital twins to engage in realistic simulated dating interactions and job interviews before real encounters, providing bidirectional cultural fit assessment for both romantic compatibility and workplace matching. Validation using 551 GNWT-Agents and Columbia University Speed Dating dataset demonstrates 72% correlation with human attraction patterns, 77.8% match prediction accuracy, and 74% agreement in human validation studies. This work advances psychological authenticity in LLM agents and establishes a foundation for intelligent dating platforms and HR technology solutions.

Bowei Tian, Yexiao He, Meng Liu et al.

In medical image analysis, model predictions can be affected by sensitive attributes, such as race and gender, leading to fairness concerns and potential biases in diagnostic outcomes. To mitigate this, we present a causal modeling framework, which aims to reduce the impact of sensitive attributes on diagnostic predictions. Our approach introduces a novel fairness criterion, \textbf{Diagnosis Fairness}, and a unique fairness metric, leveraging path-specific fairness to control the influence of demographic attributes, ensuring that predictions are primarily informed by clinically relevant features rather than sensitive attributes. By incorporating adversarial perturbation masks, our framework directs the model to focus on critical image regions, suppressing bias-inducing information. Experimental results across multiple datasets demonstrate that our framework effectively reduces bias directly associated with sensitive attributes while preserving diagnostic accuracy. Our findings suggest that causal modeling can enhance both fairness and interpretability in AI-powered clinical decision support systems.

Prajwal Singhania, Siddharth Singh, Shwai He et al.

Inference on large language models (LLMs) can be expensive in terms of the compute and memory costs involved, especially when long sequence lengths are used. In particular, the self-attention mechanism used in LLM inference contributes significantly to these costs, which has sparked an interest in approximating the self-attention computation to reduce such costs. In this work, we propose to approximate self-attention by focusing on the dimensionality of key vectors computed in the attention block. Our analysis reveals that key vectors lie in a significantly lower-dimensional space, consistently across several datasets and models. Exploiting this observation, we propose Loki, a novel sparse attention method that ranks and selects tokens in the KV-cache based on attention scores computed in low-dimensional space. Our evaluations show that Loki is able to speed up the attention computation due to reduced data movement (load/store) and compute costs while maintaining the efficacy of the models better than other popular approximation methods.

Shwai He, Shi Gu

Stock prices move as piece-wise trending fluctuation rather than a purely random walk. Traditionally, the prediction of future stock movements is based on the historical trading record. Nowadays, with the development of social media, many active participants in the market choose to publicize their strategies, which provides a window to glimpse over the whole market's attitude towards future movements by extracting the semantics behind social media. However, social media contains conflicting information and cannot replace historical records completely. In this work, we propose a multi-modality attention network to reduce conflicts and integrate semantic and numeric features to predict future stock movements comprehensively. Specifically, we first extract semantic information from social media and estimate their credibility based on posters' identity and public reputation. Then we incorporate the semantic from online posts and numeric features from historical records to make the trading strategy. Experimental results show that our approach outperforms previous methods by a significant margin in both prediction accuracy (61.20\%) and trading profits (9.13\%). It demonstrates that our method improves the performance of stock movements prediction and informs future research on multi-modality fusion towards stock prediction.