Fanxu Meng

Fanxu Meng, Haotong Yang, Yiding Wang et al. · pku

The human brain is naturally equipped to comprehend and interpret visual information rapidly. When confronted with complex problems or concepts, we use flowcharts, sketches, and diagrams to aid our thought process. Leveraging this inherent ability can significantly enhance logical reasoning. However, current Large Language Models (LLMs) do not utilize such visual intuition to help their thinking. Even the most advanced version language models (e.g., GPT-4V and LLaVA) merely align images into textual space, which means their reasoning processes remain purely verbal. To mitigate such limitations, we present a Chain of Images (CoI) approach, which can convert complex language reasoning problems to simple pattern recognition by generating a series of images as intermediate representations. Furthermore, we have developed a CoI evaluation dataset encompassing 15 distinct domains where images can intuitively aid problem-solving. Based on this dataset, we aim to construct a benchmark to assess the capability of future multimodal large-scale models to leverage images for reasoning. In supporting our CoI reasoning, we introduce a symbolic multimodal large language model (SyMLLM) that generates images strictly based on language instructions and accepts both text and image as input. Experiments on Geometry, Chess and Common Sense tasks sourced from the CoI evaluation dataset show that CoI improves performance significantly over the pure-language Chain of Thoughts (CoT) baselines. The code is available at https://github.com/GraphPKU/CoI.

Junru Lu, Jiarui Qin, Lingfeng Qiao et al.

We introduce Youtu-LLM, a lightweight yet powerful language model that harmonizes high computational efficiency with native agentic intelligence. Unlike typical small models that rely on distillation, Youtu-LLM (1.96B) is pre-trained from scratch to systematically cultivate reasoning and planning capabilities. The key technical advancements are as follows: (1) Compact Architecture with Long-Context Support: Built on a dense Multi-Latent Attention (MLA) architecture with a novel STEM-oriented vocabulary, Youtu-LLM supports a 128k context window. This design enables robust long-context reasoning and state tracking within a minimal memory footprint, making it ideal for long-horizon agent and reasoning tasks. (2) Principled "Commonsense-STEM-Agent" Curriculum: We curated a massive corpus of approximately 11T tokens and implemented a multi-stage training strategy. By progressively shifting the pre-training data distribution from general commonsense to complex STEM and agentic tasks, we ensure the model acquires deep cognitive abilities rather than superficial alignment. (3) Scalable Agentic Mid-training: Specifically for the agentic mid-training, we employ diverse data construction schemes to synthesize rich and varied trajectories across math, coding, and tool-use domains. This high-quality data enables the model to internalize planning and reflection behaviors effectively. Extensive evaluations show that Youtu-LLM sets a new state-of-the-art for sub-2B LLMs. On general benchmarks, it achieves competitive performance against larger models, while on agent-specific tasks, it significantly surpasses existing SOTA baselines, demonstrating that lightweight models can possess strong intrinsic agentic capabilities.

Yufei Xu, Fanxu Meng, Fan Jiang et al.

Token-level sparse attention mechanisms, exemplified by DeepSeek Sparse Attention (DSA), achieve fine-grained key selection by scoring every historical token for each query using a lightweight indexer, and then computing attention only over the selected subset. While the downstream sparse attention scales efficiently, the indexer still scans the entire prefix for every query, introducing an O($L^2$) per-layer bottleneck that becomes prohibitive as context length grows. We propose HISA (Hierarchical Indexed Sparse Attention), a drop-in replacement for the indexer that transforms the search process from a flat token scan into a two-stage hierarchical procedure. First, a block-level coarse filter scores pooled block representatives to prune irrelevant regions. Then, a token-level refinement applies the original indexer only within the remaining candidate blocks. HISA preserves the exact token-level top-k sparsity pattern required by the downstream Sparse MLA operator and requires no additional training. On kernel-level benchmarks, HISA achieves a 2$\times$ speedup at 32K context length and 4$\times$ at 128K. On Needle-in-a-Haystack and LongBench, we directly replace the indexer in DeepSeek-V3.2 with HISA, without any fine-tuning. HISA closely matches the original DSA in quality while significantly outperforming block-sparse baselines. Moreover, the token selection sets produced by HISA and the original DSA exhibit a mean IoU greater than 99%, indicating that the efficiency gains come with virtually no impact on selection fidelity.

Fanxu Meng, Zhaohui Wang, Muhan Zhang

To parameter-efficiently fine-tune (PEFT) large language models (LLMs), the low-rank adaptation (LoRA) method approximates the model changes $ΔW \in \mathbb{R}^{m \times n}$ through the product of two matrices $A \in \mathbb{R}^{m \times r}$ and $B \in \mathbb{R}^{r \times n}$, where $r \ll \min(m, n)$, $A$ is initialized with Gaussian noise, and $B$ with zeros. LoRA freezes the original model $W$ and updates the "Noise & Zero" adapter, which may lead to slow convergence. To overcome this limitation, we introduce Principal Singular values and Singular vectors Adaptation (PiSSA). PiSSA shares the same architecture as LoRA, but initializes the adaptor matrices $A$ and $B$ with the principal components of the original matrix $W$, and put the remaining components into a residual matrix $W^{res} \in \mathbb{R}^{m \times n}$ which is frozen during fine-tuning. Compared to LoRA, PiSSA updates the principal components while freezing the "residual" parts, allowing faster convergence and enhanced performance. Comparative experiments of PiSSA and LoRA across 12 different models, ranging from 184M to 70B, encompassing 5 NLG and 8 NLU tasks, reveal that PiSSA consistently outperforms LoRA under identical experimental setups. On the GSM8K benchmark, Mistral-7B fine-tuned with PiSSA achieves an accuracy of 72.86%, surpassing LoRA's 67.7% by 5.16%. Due to the same architecture, PiSSA is also compatible with quantization to further reduce the memory requirement of fine-tuning. Compared to QLoRA, QPiSSA exhibits smaller quantization errors in the initial stages. Fine-tuning LLaMA-3-70B on GSM8K, QPiSSA attains an accuracy of 86.05%, exceeding the performances of QLoRA at 81.73%. Leveraging a fast SVD technique, PiSSA can be initialized in only a few seconds, presenting a negligible cost for transitioning from LoRA to PiSSA. Code is available at https://github.com/GraphPKU/PiSSA.

Pingzhi Tang, Ruijie Zhou, Fanxu Meng et al.

Current quantization methods for LLMs predominantly rely on block-wise structures to maintain efficiency, often at the cost of representational flexibility. In this work, we demonstrate that element-wise quantization can be made as efficient as block-wise scaling while providing strictly superior expressive power by modeling the scaling manifold as continuous low-rank matrices ($S = BA$). We propose Low-Rank Decomposed Scaling (LoRDS), a unified framework that rethinks quantization granularity through this low-rank decomposition. By "breaking the blocks" of spatial constraints, LoRDS establishes a seamless efficiency lifecycle: it provides high-fidelity PTQ initialization refined via iterative optimization, enables joint QAT of weights and scaling factors, and facilitates high-rank multiplicative PEFT adaptation. Unlike additive PEFT approaches such as QLoRA, LoRDS enables high-rank weight updates within a low-rank budget while incurring no additional inference overhead. Supported by highly optimized Triton kernels, LoRDS consistently outperforms state-of-the-art baselines across various model families in both quantization and downstream fine-tuning tasks. Notably, on Llama3-8B, our method achieves up to a 27.0% accuracy improvement at 3 bits over NormalFloat quantization and delivers a 1.5x inference speedup on NVIDIA RTX 4090 while enhancing PEFT performance by 9.6% on downstream tasks over 4bit QLoRA, offering a robust and integrated solution for unified compression and adaptation of LLMs.

Fanxu Meng

Multi-head Latent Attention (MLA), the attention used in DeepSeek-V2/V3, jointly compresses keys and values into a low-rank latent and matches the H100 roofline almost perfectly. Its trained weights, however, expose only one decoding path - an absorbed MQA form - which ties efficient inference to H100-class compute-bandwidth ratios, forfeits tensor parallelism along the head axis, and yields no Multi-Token Prediction (MTP) gain on commodity inference GPUs such as the export-restricted H20. We propose Group-Query Latent Attention (GQLA), a minimal modification of MLA whose trained weights expose two algebraically equivalent decoding paths over the same parameters: an MQA-absorb path identical to MLA's, and a GQA path with a per-group expanded cache. The runtime picks the path that matches the target hardware - no retraining, no custom kernels - so a single set of GQLA weights pins the rooflines of both H100 (MQA-absorb, s_q=1) and H20 (GQA + MTP, s_q=2), while supporting up to 8-way zero-redundancy tensor parallelism on the GQA path. To avoid pretraining from scratch we extend TransMLA into TransGQLA, which converts a pretrained GQA checkpoint into a GQLA model; on LLaMA-3-8B it compresses the per-token KV cache to 28.125% of the GQA baseline on the MQA-absorb path while structurally preserving GQA-level traffic on the per-group path.

Haotong Yang, Fanxu Meng, Zhouchen Lin et al.

The pre-trained large language models (LLMs) have shown their extraordinary capacity to solve reasoning tasks, even on tasks that require a complex process involving multiple sub-steps. However, given the vast possible generation space of all the tasks, how the pretrained model learns the reasoning ability remains an open question. We firstly propose that an intrinsic structural constraint on the generated sequence of language-based reasoning -- we called it template-content structure (T-C structure) -- is the key to explain why LLMs can solve a large number of complex reasoning problems with limited training data by showing this structure can reduce the possible space from exponential level to linear level. Furthermore, by generalizing this structure to the hierarchical case, we demonstrate that models can achieve task composition, further reducing the space needed to learn from linear to logarithmic, thereby effectively learning on complex reasoning involving multiple steps. We provide both examples and formal theory of our T-C structure. We also experimentally validate the existence of the T-C structure in some current LLMs and its effectiveness for reasoning.

Haofei Hou, Shunyi Zhao, Fanxu Meng et al.

Question Answer (QA) systems for biomedical experiments facilitate cross-disciplinary communication, and serve as a foundation for downstream tasks, e.g., laboratory automation. High Information Density (HID) and Multi-Step Reasoning (MSR) pose unique challenges for biomedical experimental QA. While extracting structured knowledge, e.g., Knowledge Graphs (KGs), can substantially benefit biomedical experimental QA. Existing biomedical datasets focus on general or coarsegrained knowledge and thus fail to support the fine-grained experimental reasoning demanded by HID and MSR. To address this gap, we introduce Biomedical Protocol Information Extraction Dataset (BioPIE), a dataset that provides procedure-centric KGs of experimental entities, actions, and relations at a scale that supports reasoning over biomedical experiments across protocols. We evaluate information extraction methods on BioPIE, and implement a QA system that leverages BioPIE, showcasing performance gains on test, HID, and MSR question sets, showing that the structured experimental knowledge in BioPIE underpins both AI-assisted and more autonomous biomedical experimentation.

Yushang Yang, Fanxu Meng, Fiona Fui-Hoon Nah et al.

The ways people remember and recall places reveal an invisible aspect of cultural heritage (CH), reflecting how individuals and communities relate to these places. Heritage is communal, emerging through collaboratively constructed narratives rather than individual records. To probe how people may share collective memories, we designed an immersive two-person workflow for collaboratively co-designing 3D artifacts and environments in virtual heritage locations, using Generative AI (GenAI) to instantiate these intangible memories. Observations of the co-creation process revealed that participants merged prompts and model placements when negotiating different perspectives. They used spatial operations to compose scenes, and also to express personal and embodied experiences of CH. When GenAI failed to meet their needs, participants engaged in creative appropriation, re-purposing unsatisfactory generated objects as sources of design inspiration to further shared narratives. While GenAI may have a homogenizing effect on CH expression, this work shows how people may overcome limitations in immersive collaborative workflows.

Ruijie Zhou, Fanxu Meng, Yufei Xu et al.

DeepSeek Sparse Attention (DSA) sets the state of the art for fine-grained inference-time sparse attention by introducing a learned token-wise indexer that scores every prefix token and selects the most relevant ones for the main attention. To remain expressive, the indexer uses many query heads (for example, 64 on DeepSeek-V3.2) that share the same selected token set; this multi-head design is precisely what makes the indexer the dominant cost on long contexts. We propose MISA (Mixture of Indexer Sparse Attention), a drop-in replacement for the DSA indexer that treats its indexer heads as a pool of mixture-of-experts. A lightweight router uses cheap block-level statistics to pick a query-dependent subset of only a few active heads, and only those heads run the heavy token-level scoring. This preserves the diversity of the original indexer pool while reducing the per-query cost from scoring every prefix token with every head to scoring it with only a handful of routed heads, plus a negligible router term computed on a small set of pooled keys. We further introduce a hierarchical variant of MISA that uses the routed pass to keep an enlarged candidate set and then re-ranks it with the original DSA indexer to recover the final selected tokens almost exactly. With only eight active heads and no additional training, MISA matches the dense DSA indexer on LongBench across DeepSeek-V3.2 and GLM-5 while running with eight and four times fewer indexer heads respectively, and outperforms HISA on average. It also preserves fully green Needle-in-a-Haystack heatmaps up to a 128K-token context and recovers more than 92% of the tokens selected by the DSA indexer per layer. Our TileLang kernel delivers roughly a 3.82 times speedup over DSA's original indexer kernel on a single NVIDIA H200 GPU.

Xiaojuan Tang, Zilong Zheng, Jiaqi Li et al.

The emergent few-shot reasoning capabilities of Large Language Models (LLMs) have excited the natural language and machine learning community over recent years. Despite of numerous successful applications, the underlying mechanism of such in-context capabilities still remains unclear. In this work, we hypothesize that the learned \textit{semantics} of language tokens do the most heavy lifting during the reasoning process. Different from human's symbolic reasoning process, the semantic representations of LLMs could create strong connections among tokens, thus composing a superficial logical chain. To test our hypothesis, we decouple semantics from the language reasoning process and evaluate three kinds of reasoning abilities, i.e., deduction, induction and abduction. Our findings reveal that semantics play a vital role in LLMs' in-context reasoning -- LLMs perform significantly better when semantics are consistent with commonsense but struggle to solve symbolic or counter-commonsense reasoning tasks by leveraging in-context new knowledge. The surprising observations question whether modern LLMs have mastered the inductive, deductive and abductive reasoning abilities as in human intelligence, and motivate research on unveiling the magic existing within the black-box LLMs. On the whole, our analysis provides a novel perspective on the role of semantics in developing and evaluating language models' reasoning abilities. Code is available at {\url{https://github.com/XiaojuanTang/ICSR}}.

Fanxu Meng, Hao Cheng, Jiaxin Zhuang et al.

Although residual connection enables training very deep neural networks, it is not friendly for online inference due to its multi-branch topology. This encourages many researchers to work on designing DNNs without residual connections at inference. For example, RepVGG re-parameterizes multi-branch topology to a VGG-like (single-branch) model when deploying, showing great performance when the network is relatively shallow. However, RepVGG can not transform ResNet to VGG equivalently because re-parameterizing methods can only be applied to linear blocks and the non-linear layers (ReLU) have to be put outside of the residual connection which results in limited representation ability, especially for deeper networks. In this paper, we aim to remedy this problem and propose to remove the residual connection in a vanilla ResNet equivalently by a reserving and merging (RM) operation on ResBlock. Specifically, the RM operation allows input feature maps to pass through the block while reserving their information and merges all the information at the end of each block, which can remove residual connections without changing the original output. As a plug-in method, RM Operation basically has three advantages: 1) its implementation makes it naturally friendly for high ratio network pruning. 2) it helps break the depth limitation of RepVGG. 3) it leads to better accuracy-speed trade-off network (RMNet) compared to ResNet and RepVGG. We believe the ideology of RM Operation can inspire many insights on model design for the community in the future. Code is available at: https://github.com/fxmeng/RMNet.

Fanxu Meng, Hao Cheng, Ke Li et al.

Pruning has become a very powerful and effective technique to compress and accelerate modern neural networks. Existing pruning methods can be grouped into two categories: filter pruning (FP) and weight pruning (WP). FP wins at hardware compatibility but loses at the compression ratio compared with WP. To converge the strength of both methods, we propose to prune the filter in the filter. Specifically, we treat a filter $F \in \mathbb{R}^{C\times K\times K}$ as $K \times K$ stripes, i.e., $1\times 1$ filters $\in \mathbb{R}^{C}$, then by pruning the stripes instead of the whole filter, we can achieve finer granularity than traditional FP while being hardware friendly. We term our method as SWP (\emph{Stripe-Wise Pruning}). SWP is implemented by introducing a novel learnable matrix called Filter Skeleton, whose values reflect the shape of each filter. As some recent work has shown that the pruned architecture is more crucial than the inherited important weights, we argue that the architecture of a single filter, i.e., the shape, also matters. Through extensive experiments, we demonstrate that SWP is more effective compared to the previous FP-based methods and achieves the state-of-art pruning ratio on CIFAR-10 and ImageNet datasets without obvious accuracy drop. Code is available at https://github.com/fxmeng/Pruning-Filter-in-Filter

Hao Cheng, Fanxu Meng, Ke Li et al.

Filter is the key component in modern convolutional neural networks (CNNs). However, since CNNs are usually over-parameterized, a pre-trained network always contain some invalid (unimportant) filters. These filters have relatively small $l_{1}$ norm and contribute little to the output (\textbf{Reason}). While filter pruning removes these invalid filters for efficiency consideration, we tend to reactivate them to improve the representation capability of CNNs. In this paper, we introduce filter grafting (\textbf{Method}) to achieve this goal. The activation is processed by grafting external information (weights) into invalid filters. To better perform the grafting, we develop a novel criterion to measure the information of filters and an adaptive weighting strategy to balance the grafted information among networks. After the grafting operation, the network has fewer invalid filters compared with its initial state, enpowering the model with more representation capacity. Meanwhile, since grafting is operated reciprocally on all networks involved, we find that grafting may lose the information of valid filters when improving invalid filters. To gain a universal improvement on both valid and invalid filters, we compensate grafting with distillation (\textbf{Cultivation}) to overcome the drawback of grafting . Extensive experiments are performed on the classification and recognition tasks to show the superiority of our method. Code is available at \textcolor{black}{\emph{https://github.com/fxmeng/filter-grafting}}.

Fanxu Meng, Hao Cheng, Ke Li et al.

This paper proposes a new learning paradigm called filter grafting, which aims to improve the representation capability of Deep Neural Networks (DNNs). The motivation is that DNNs have unimportant (invalid) filters (e.g., l1 norm close to 0). These filters limit the potential of DNNs since they are identified as having little effect on the network. While filter pruning removes these invalid filters for efficiency consideration, filter grafting re-activates them from an accuracy boosting perspective. The activation is processed by grafting external information (weights) into invalid filters. To better perform the grafting process, we develop an entropy-based criterion to measure the information of filters and an adaptive weighting strategy for balancing the grafted information among networks. After the grafting operation, the network has very few invalid filters compared with its untouched state, enpowering the model with more representation capacity. We also perform extensive experiments on the classification and recognition tasks to show the superiority of our method. For example, the grafted MobileNetV2 outperforms the non-grafted MobileNetV2 by about 7 percent on CIFAR-100 dataset. Code is available at https://github.com/fxmeng/filter-grafting.git.

Fanxu Meng, Pingzhi Tang, Xiaojuan Tang et al.

In this paper, we present TransMLA, a framework that seamlessly converts any GQA-based pre-trained model into an MLA-based model. Our approach enables direct compatibility with DeepSeek's codebase, allowing these models to fully leverage DeepSeek-specific optimizations such as vLLM and SGlang. By compressing 93% of the KV cache in LLaMA-2-7B, TransMLA achieves a 10.6x inference speedup at an 8K context length while preserving meaningful output quality. Additionally, the model requires only 6 billion tokens for fine-tuning to regain performance on par with the original across multiple benchmarks. TransMLA offers a practical solution for migrating GQA-based models to the MLA structure. When combined with DeepSeek's advanced features, such as FP8 quantization and Multi-Token Prediction, even greater inference acceleration can be realized.

Yansheng Mao, Yufei Xu, Jiaqi Li et al.

Long context understanding remains challenging for large language models due to their limited context windows. This paper presents Long Input Fine-Tuning (LIFT), a novel framework for long-context modeling that can improve the long-context performance of arbitrary (short-context) LLMs by dynamically adapting model parameters based on the long input. Importantly, LIFT, rather than endlessly extending the context window size to accommodate increasingly longer inputs in context, chooses to store and absorb the long input in parameter. By fine-tuning the long input into model parameters, LIFT allows short-context LLMs to answer questions even when the required information is not provided in the context during inference. Furthermore, to enhance LIFT performance while maintaining the original in-context learning (ICL) capabilities, we introduce Gated Memory, a specialized attention adapter that automatically balances long input memorization and ICL. We provide a comprehensive analysis of the strengths and limitations of LIFT on long context understanding, offering valuable directions for future research.

Yu-Zhe Shi, Mingchen Liu, Fanxu Meng et al.

Self-driving laboratories have begun to replace human experimenters in performing single experimental skills or predetermined experimental protocols. However, as the pace of idea iteration in scientific research has been intensified by Artificial Intelligence, the demand for rapid design of new protocols for new discoveries become evident. Efforts to automate protocol design have been initiated, but the capabilities of knowledge-based machine designers, such as Large Language Models, have not been fully elicited, probably for the absence of a systematic representation of experimental knowledge, as opposed to isolated, flatten pieces of information. To tackle this issue, we propose a multi-faceted, multi-scale representation, where instance actions, generalized operations, and product flow models are hierarchically encapsulated using Domain-Specific Languages. We further develop a data-driven algorithm based on non-parametric modeling that autonomously customizes these representations for specific domains. The proposed representation is equipped with various machine designers to manage protocol design tasks, including planning, modification, and adjustment. The results demonstrate that the proposed method could effectively complement Large Language Models in the protocol design process, serving as an auxiliary module in the realm of machine-assisted scientific exploration.

Yansheng Mao, Jiaqi Li, Fanxu Meng et al.

Long context understanding remains challenging for large language models due to their limited context windows. This paper introduces Long Input Fine-Tuning (LIFT) for long context modeling, a novel framework that enhances LLM performance on long-context tasks by adapting model parameters to the context at test time. LIFT enables efficient processing of lengthy inputs without the computational burden of offline long-context adaptation, and can improve the long-context capabilities of arbitrary short-context models. The framework is further enhanced by integrating in-context learning and pre-LIFT supervised fine-tuning. The combination of in-context learning and LIFT enables short-context models like Llama 3 to handle arbitrarily long contexts and consistently improves their performance on popular long-context benchmarks like LooGLE and LongBench. We also provide a comprehensive analysis of the strengths and limitations of LIFT on long context understanding, offering valuable directions for future research.

Yanan Li, Fanxu Meng, Muhan Zhang et al.

As Large Language Models (LLMs) are frequently updated, LoRA weights trained on earlier versions quickly become obsolete. The conventional practice of retraining LoRA weights from scratch on the latest model is costly, time-consuming, and environmentally detrimental, particularly as the diversity of LLMs and downstream tasks expands. This motivates a critical question: "How can we efficiently leverage existing LoRA weights to adapt to newer model versions?" To address this, we propose LoRASuite, a modular approach tailored specifically to various types of LLM updates. First, we compute a transfer matrix utilizing known parameters from both old and new LLMs. Next, we allocate corresponding layers and attention heads based on centered kernel alignment and cosine similarity metrics, respectively. A subsequent small-scale, skillful fine-tuning step ensures numerical stability. Experimental evaluations demonstrate that LoRASuite consistently surpasses small-scale vanilla LoRA methods. Notably, on backbone LLMs such as MiniCPM and Qwen, LoRASuite even exceeds the performance of full-scale LoRA retraining, with average improvements of +1.4 and +6.6 points on math tasks, respectively. Additionally, LoRASuite significantly reduces memory consumption by 5.5 GB and computational time by 78.23%.

Fanxu Meng, Pingzhi Tang, Fan jiang et al.

Decoder-only models generate tokens autoregressively by caching key/value vectors, but as the cache grows, inference becomes memory-bound. To address this issue, we introduce CLOVER (Cross-Layer Orthogonal Vectors), a novel approach that treats pairs of attention layers as a set of low-rank decompositions. CLOVER applies Singular Value Decomposition (SVD) to the \( Q \)-\( K \) and \( V \)-\( O \) pairs within each attention head. The resulting singular values can either guide pruning or serve as trainable parameters for efficient fine-tuning of all orthogonal vectors. After pruning or fine-tuning, these values are reintegrated into the model without increasing its parameter count. We apply CLOVER to various models, including GPT-2 XL, DeepSeek-V2-Lite, Whisper-Large-v3, Stable Diffusion XL, and LLaMA-3.2-11B-Vision. Our results demonstrate that CLOVER significantly improves pruning efficiency. For instance, the perplexity of pruning 70\% of the \( Q \)-\( K \) pairs in GPT-2 XL is similar to that of pruning just 8\% with vanilla methods. Fine-tuning the singular values further results in a full-rank update, outperforming state-of-the-art methods (LoRA, DoRA, HiRA, and PiSSA) by 7.6\%, 5.5\%, 3.8\%, and 0.7\%, respectively, on eight commonsense tasks for LLaMA-2 7B.

Yiding Wang, Yuxuan Chen, Fanxu Meng et al. · pku

Since real-world legal experiments are often costly or infeasible, simulating legal societies with Artificial Intelligence (AI) systems provides an effective alternative for verifying and developing legal theory, as well as supporting legal administration. Large Language Models (LLMs), with their world knowledge and role-playing capabilities, are strong candidates to serve as the foundation for legal society simulation. However, the application of LLMs to simulate legal systems remains underexplored. In this work, we introduce Law in Silico, an LLM-based agent framework for simulating legal scenarios with individual decision-making and institutional mechanisms of legislation, adjudication, and enforcement. Our experiments, which compare simulated crime rates with real-world data, demonstrate that LLM-based agents can largely reproduce macro-level crime trends and provide insights that align with real-world observations. At the same time, micro-level simulations reveal that a well-functioning, transparent, and adaptive legal system offers better protection of the rights of vulnerable individuals.

Xiaojuan Tang, Fanxu Meng, Pingzhi Tang et al.

Multi-Head Latent Attention (MLA), introduced in DeepSeek-V2, compresses key-value states into a low-rank latent vector, caching only this vector to reduce memory. In tensor parallelism (TP), however, attention heads are computed across multiple devices, and each device must load the full cache, eroding the advantage of MLA over Grouped Query Attention (GQA). We propose Tensor-Parallel Latent Attention (TPLA): a scheme that partitions both the latent representation and each head's input dimension across devices, performs attention independently per shard, and then combines results with an all-reduce. TPLA preserves the benefits of a compressed KV cache while unlocking TP efficiency. Unlike Grouped Latent Attention (GLA), every head in TPLA still leverages the full latent representation, maintaining stronger representational capacity. TPLA is drop-in compatible with models pre-trained using MLA: it supports MLA-style prefilling and enables efficient tensor-parallel decoding without retraining. Applying simple orthogonal transforms -- e.g., the Hadamard transform or PCA -- before TP slicing further mitigates cross-shard interference, yielding minimal accuracy degradation. By reducing the per-device KV cache for DeepSeek-V3 and Kimi-K2, we achieve 1.79x and 1.93x speedups, respectively, at a 32K-token context length while maintaining performance on commonsense and LongBench benchmarks. TPLA can be implemented with FlashAttention-3, enabling practical end-to-end acceleration.

Huixiang Luo, Hao Cheng, Fanxu Meng et al.

Pseudo-labeling (PL) and Data Augmentation-based Consistency Training (DACT) are two approaches widely used in Semi-Supervised Learning (SSL) methods. These methods exhibit great power in many machine learning tasks by utilizing unlabeled data for efficient training. But in a more realistic setting (termed as open-set SSL), where unlabeled dataset contains out-of-distribution (OOD) samples, the traditional SSL methods suffer severe performance degradation. Recent approaches mitigate the negative influence of OOD samples by filtering them out from the unlabeled data. However, it is not clear whether directly removing the OOD samples is the best choice. Furthermore, why PL and DACT could perform differently in open-set SSL remains a mystery. In this paper, we thoroughly analyze various SSL methods (PL and DACT) on open-set SSL and discuss pros and cons of these two approaches separately. Based on our analysis, we propose Style Disturbance to improve traditional SSL methods on open-set SSL and experimentally show our approach can achieve state-of-the-art results on various datasets by utilizing OOD samples properly. We believe our study can bring new insights for SSL research.