arXiv subjects

Muhan Zhang

Publications and source records attributed to Muhan Zhang.

At least 55 records · Page 3Linked to original sources

arXiv2026

GREPO: A Benchmark for Graph Neural Networks on Repository-Level Bug Localization

Repository-level bug localization-the task of identifying where code must be modified to fix a bug-is a critical software engineering challenge. Standard Large Language Modles (LLMs) are often unsuitable for this task due to context window limitations that prevent them from processing entire code repositories. As a result, various retrieval methods are commonly used, including keyword matching, text similarity, and simple graph-based heuristics such as Breadth-First Search. Graph Neural Networks (GNNs) offer a promising alternative due to their ability to model complex, repository-wide dependencies; however, their application has been hindered by the lack of a dedicated benchmark. To address this gap, we introduce GREPO, the first GNN benchmark for repository-scale bug localization tasks. GREPO comprises 86 Python repositories and 47294 bug-fixing tasks, providing graph-based data structures ready for direct GNN processing. Our evaluation of various GNN architectures shows outstanding performance compared to established information retrieval baselines. This work highlights the potential of GNNs for bug localization and established GREPO as a foundation resource for future research, The code is available at https://github.com/qingpingmo/GREPO.

cs.LG↗

arXiv2026

LiteToken: Removing Intermediate Merge Residues From BPE Tokenizers

Tokenization is fundamental to how language models represent and process text, yet the behavior of widely used BPE tokenizers has received far less study than model architectures and training. In this paper, we investigate intermediate merge residues in BPE vocabularies: tokens that are frequent during merge learning so that retained in the final vocabulary, but are mostly further merged and rarely emitted when tokenizing the corpus during tokenizer usage. Such low-frequency tokens not only waste vocabulary capacity but also increase vulnerability to adversarial or atypical inputs. We present a systematic empirical characterization of this phenomenon across commonly used tokenizers and introduce LiteToken, a simple method for removing residue tokens. Because the affected tokens are rarely used, pretrained models can often accommodate the modified tokenizer without additional fine-tuning. Experiments show that LiteToken reduces token fragmentation, reduces parameters, and improves robustness to noisy or misspelled inputs, while preserving overall performance.

cs.CL↗

arXiv2026

Breaking the Blocks: Continuous Low-Rank Decomposed Scaling for Unified LLM Quantization and Adaptation

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.

cs.LG↗

arXiv2026

SubTokenTest: A Practical Benchmark for Real-World Sub-token Understanding

Recent advancements in large language models (LLMs) have significantly enhanced their reasoning capabilities. However, they continue to struggle with basic character-level tasks, such as counting letters in words, a problem rooted in their tokenization process. While existing benchmarks have highlighted this weakness through basic character operations, such failures are often dismissed due to lacking practical relevance. Yet, many real-world applications, such as navigating text-based maps or interpreting structured tables, rely heavily on precise sub-token understanding. In this regard, we introduce SubTokenTest, a comprehensive benchmark that assesses sub-token understanding through practical, utility-driven tasks. Our benchmark includes ten tasks across four domains and isolates tokenization-related failures by decoupling performance from complex reasoning. We provide a comprehensive evaluation of nine advanced LLMs. Additionally, we investigate the impact of test-time scaling on sub-token reasoning and explore how character-level information is encoded within the hidden states.

cs.CL↗

arXiv2026

Towards a Mechanistic Understanding of Large Reasoning Models: A Survey of Training, Inference, and Failures

Reinforcement learning (RL) has catalyzed the emergence of Large Reasoning Models (LRMs) that have pushed reasoning capabilities to new heights. While their performance has garnered significant excitement, exploring the internal mechanisms driving these behaviors has become an equally critical research frontier. This paper provides a comprehensive survey of the mechanistic understanding of LRMs, organizing recent findings into three core dimensions: 1) training dynamics, 2) reasoning mechanisms, and 3) unintended behaviors. By synthesizing these insights, we aim to bridge the gap between black-box performance and mechanistic transparency. Finally, we discuss under-explored challenges to outline a roadmap for future mechanistic studies, including the need for applied interpretability, improved methodologies, and a unified theoretical framework.

cs.CL↗

arXiv2025

What Affects the Effective Depth of Large Language Models?

The scaling of large language models (LLMs) emphasizes increasing depth, yet performance gains diminish with added layers. Prior work introduces the concept of "effective depth", arguing that deeper models fail to fully utilize their layers for meaningful computation. Building on this, we systematically study how effective depth varies with model scale, training type, and task difficulty. First, we analyze the model behavior of Qwen-2.5 family (1.5B-32B) and find that while the number of effective layers grows with model size, the effective depth ratio remains stable. Besides, comparisons between base and corresponding long-CoT models show no increase in effective depth, suggesting that improved reasoning stems from longer context rather than deeper per-token computation. Furthermore, evaluations across tasks of varying difficulty indicate that models do not dynamically use more layers for harder problems. Our results suggest that current LLMs underuse available depth across scales, training paradigms and tasks of varying difficulties, pointing out research opportunities on increasing the layer utilization rate of LLMs, model pruning, and early exiting. Our code is released at https://github.com/AheadOFpotato/what_affects_effective_depth.

cs.CL↗

arXiv2025

Meta Pruning via Graph Metanetworks : A Universal Meta Learning Framework for Network Pruning

We propose an entirely new meta-learning framework for network pruning. It is a general framework that can be theoretically applied to almost all types of networks with all kinds of pruning and has great generality and transferability. Experiments have shown that it can achieve outstanding results on many popular and representative pruning tasks (including both CNNs and Transformers). Unlike all prior works that either rely on fixed, hand-crafted criteria to prune in a coarse manner, or employ learning to prune ways that require special training during each pruning and lack generality. Our framework can learn complex pruning rules automatically via a neural network (metanetwork) and has great generality that can prune without any special training. More specifically, we introduce the newly developed idea of metanetwork from meta-learning into pruning. A metanetwork is a network that takes another network as input and produces a modified network as output. In this paper, we first establish a bijective mapping between neural networks and graphs, and then employ a graph neural network as our metanetwork. We train a metanetwork that learns the pruning strategy automatically and can transform a network that is hard to prune into another network that is much easier to prune. Once the metanetwork is trained, our pruning needs nothing more than a feedforward through the metanetwork and some standard finetuning to prune at state-of-the-art. Our code is available at https://github.com/Yewei-Liu/MetaPruning

cs.LG↗

arXiv2025

Efficient Neural Common Neighbor for Temporal Graph Link Prediction

Temporal graphs are widespread in real-world applications such as social networks, as well as trade and transportation networks. Predicting dynamic links within these evolving graphs is a key problem. Many memory-based methods use temporal interaction histories to generate node embeddings, which are then combined to predict links. However, these approaches primarily focus on individual node representations, often overlooking the inherently pairwise nature of link prediction. While some recent methods attempt to capture pairwise features, they tend to be limited by high computational complexity arising from repeated embedding calculations, making them unsuitable for large-scale datasets like the Temporal Graph Benchmark (TGB). To address the critical need for models that combine strong expressive power with high computational efficiency for link prediction on large temporal graphs, we propose Temporal Neural Common Neighbor (TNCN). Our model achieves this balance by adapting the powerful pairwise modeling principles of Neural Common Neighbor (NCN) to an efficient temporal architecture. TNCN improves upon NCN by efficiently preserving and updating temporal neighbor dictionaries for each node and by using multi-hop common neighbors to learn more expressive pairwise representations. TNCN achieves new state-of-the-art performance on Review from five large-scale real-world TGB datasets, 6 out of 7 datasets in the transductive setting and 3 out of 7 in the inductive setting on small- to medium-scale datasets. Additionally, TNCN demonstrates excellent scalability, outperforming prominent GNN baselines by up to 30.3 times in speed on large datasets. Our code is available at https://github.com/GraphPKU/TNCN.

cs.LG↗

arXiv2025

Symmetry Dilemmas in Quantum Computing for Chemistry: A Comprehensive Analysis

Symmetry adaptation, universality, and gate efficiency are central but often competing requirements in quantum algorithms for electronic structure and many-body physics. For example, fully symmetry-adapted universal operator pools typically generate long and deep quantum circuits, gate-efficient universal operator pools generally break symmetries, and gate-efficient fully symmetry-adapted operator pools may not be universal. In this work, we analyze such symmetry dilemmas both theoretically and numerically. On the theory side, we prove that the popular, gate-efficient operator pool consisting of singlet spin-adapted singles and perfect-pairing doubles is not universal when spatial symmetry is enforced. To demonstrate the strengths and weaknesses of the three types of pools, we perform numerical simulations using an adaptive algorithm paired with operator pools that are (i) fully symmetry-adapted and universal, (ii) fully symmetry-adapted and non-universal, and (iii) breaking a single symmetry and are universal. Our numerical simulations encompass three physically relevant scenarios in which the target state is (i) the global ground state, (ii) the ground state crossed by a state differing in multiple symmetry properties, and (iii) the ground state crossed by a state differing in a single symmetry property. Our results show when symmetry-breaking but universal pools can be used safely, when enforcing at least one distinguishing symmetry suffices, and when a particular symmetry must be rigorously preserved to avoid variational collapse. Together, the formal and numerical analysis provides a practical guide for designing and benchmarking symmetry-adapted operator pools that balance universality, resource requirements, and robust state targeting in quantum simulations for chemistry.

quant-ph↗

arXiv2025

Diffusion As Self-Distillation: End-to-End Latent Diffusion In One Model

Standard Latent Diffusion Models rely on a complex, three-part architecture consisting of a separate encoder, decoder, and diffusion network, which are trained in multiple stages. This modular design is computationally inefficient, leads to suboptimal performance, and prevents the unification of diffusion with the single-network architectures common in vision foundation models. Our goal is to unify these three components into a single, end-to-end trainable network. We first demonstrate that a naive joint training approach fails catastrophically due to ``latent collapse'', where the diffusion training objective interferes with the network's ability to learn a good latent representation. We identify the root causes of this instability by drawing a novel analogy between diffusion and self-distillation based unsupervised learning method. Based on this insight, we propose Diffusion as Self-Distillation (DSD), a new framework with key modifications to the training objective that stabilize the latent space. This approach enables, for the first time, the stable end-to-end training of a single network that simultaneously learns to encode, decode, and perform diffusion. DSD achieves outstanding performance on the ImageNet $256\times 256$ conditional generation task: FID=13.44/6.38/4.25 with only 42M/118M/205M parameters and 50 training epochs on ImageNet, without using classifier-free-guidance.

cs.CV↗

arXiv2025

Multi-Agent Evolve: LLM Self-Improve through Co-evolution

Reinforcement Learning (RL) has demonstrated significant potential in enhancing the reasoning capabilities of large language models (LLMs). However, the success of RL for LLMs heavily relies on human-curated datasets and verifiable rewards, which limit their scalability and generality. Recent Self-Play RL methods, inspired by the success of the paradigm in games and Go, aim to enhance LLM reasoning capabilities without human-annotated data. However, their methods primarily depend on a grounded environment for feedback (e.g., a Python interpreter or a game engine); extending them to general domains remains challenging. To address these challenges, we propose Multi-Agent Evolve (MAE), a framework that enables LLMs to self-evolve in solving diverse tasks, including mathematics, reasoning, and general knowledge Q&A. The core design of MAE is based on a triplet of interacting agents (Proposer, Solver, Judge) that are instantiated from a single LLM, and applies reinforcement learning to optimize their behaviors. The Proposer generates questions, the Solver attempts solutions, and the Judge evaluates both while co-evolving. Experiments on Qwen2.5-3B-Instruct demonstrate that MAE achieves an average improvement of 4.54% on multiple benchmarks. These results highlight MAE as a scalable, data-efficient method for enhancing the general reasoning abilities of LLMs with minimal reliance on human-curated supervision.

cs.AI↗

arXiv2025

Law in Silico: Simulating Legal Society with LLM-Based Agents

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.

cs.AI↗

arXiv2025

LooGLE v2: Are LLMs Ready for Real World Long Dependency Challenges?

Large language models (LLMs) are equipped with increasingly extended context windows recently, yet their long context understanding capabilities over long dependency tasks remain fundamentally limited and underexplored. This gap is especially significant in many real-world long-context applications that were rarely benchmarked. In this paper, we introduce LooGLE v2, a novel benchmark designed to evaluate LLMs' long context ability in real-world applications and scenarios. Our benchmark consists of automatically collected real-world long texts, ranging from 16k to 2M tokens, encompassing domains in law, finance, game and code. Accordingly, we delicately design 10 types of domain-specific long-dependency tasks and generate 1,934 QA instances with various diversity and complexity in a scalable data curation pipeline for further practical needs. We conduct a comprehensive assessment of 6 locally deployed and 4 API-based LLMs. The evaluation results show that even the best-performing model achieves only a 59.2% overall score on our benchmark. Despite the extensive context windows, popular LLMs are only capable of understanding a much shorter length of context than they claim to be, revealing significant limitations in their ability to handle real-world tasks with long dependencies and highlighting substantial room for model improvement in practical long-context understanding.

cs.CL↗

arXiv2025

GeoRecon: Graph-Level Representation Learning for 3D Molecules via Reconstruction-Based Pretraining

The pretraining-finetuning paradigm has powered major advances in domains such as natural language processing and computer vision, with representative examples including masked language modeling and next-token prediction. In molecular representation learning, however, pretraining tasks remain largely restricted to node-level denoising, which effectively captures local atomic environments but is often insufficient for encoding the global molecular structure critical to graph-level property prediction tasks such as energy estimation and molecular regression. To address this gap, we introduce GeoRecon, a graph-level pretraining framework that shifts the focus from individual atoms to the molecule as an integrated whole. GeoRecon formulates a graph-level reconstruction task: during pretraining, the model is trained to produce an informative graph representation that guides geometry reconstruction while inducing smoother and more transferable latent spaces. This encourages the learning of coherent, global structural features beyond isolated atomic details. Without relying on external supervision, GeoRecon generally improves over backbone baselines on multiple molecular benchmarks including QM9, MD17, MD22, and 3BPA, demonstrating the effectiveness of graph-level reconstruction for holistic and geometry-aware molecular embeddings.

cs.LG↗

arXiv2025

The Road Less Traveled: Enhancing Exploration in LLMs via Sequential Sampling

Reinforcement learning (RL) has been pivotal in enhancing the reasoning capabilities of large language models (LLMs), but it often suffers from limited exploration and entropy collapse, where models exploit a narrow set of solutions, leading to a loss of sampling diversity and subsequently preventing RL from further improving performance. This issue is exacerbated in parallel sampling methods, where multiple outputs are drawn from the same distribution, potentially causing the model to converge to similar solutions. We propose SESA, a novel SEquential SAmpling framework that mitigates this challenge by generating diverse solution sketches sequentially before expanding them into full reasoning paths. This approach ensures broader exploration by conditioning each new output on previous ones, promoting diversity throughout the process and preventing policy collapse. Our experiments on a synthetic task show that sequential sampling consistently outperforms traditional RL methods in terms of path diversity and recovery from collapse. Further evaluations on real-world tasks demonstrate that SESA improves both the exploration of valid strategies and the overall performance of LLMs. On three agent benchmarks, SESA lifts success rates by $+0.25$, $+0.42$, and $+0.07$ absolute over the base model (up to an additional $211\%$ relative improvement over baseline RL), underscoring its exploration advantage. This work introduces a structured approach to exploration, paving the way for more effective and diverse reasoning in RL-trained LLMs. Our code is released at https://github.com/MuLabPKU/sesa.

cs.LG↗

arXiv2025

HD-PiSSA: High-Rank Distributed Orthogonal Adaptation

Existing parameter-efficient fine-tuning (PEFT) methods for large language models (LLMs), such as LoRA and PiSSA, constrain model updates to low-rank subspaces, limiting their expressiveness and leading to suboptimal performance on complex tasks. To address this, we introduce High-rank Distributed PiSSA (HD-PiSSA), a distributed PEFT approach that initializes orthogonal adapters across different devices and aggregates their delta updates collectively on W for fine-tuning. Unlike Data Parallel LoRA or PiSSA, which maintain identical adapters across all devices, HD-PiSSA assigns different principal components of the pre-trained weights to each GPU, significantly expanding the range of update directions. This results in over 16x higher effective updated ranks than data-parallel LoRA or PiSSA when fine-tuning on 8 GPUs with the same per-device adapter rank. Empirically, we evaluate HD-PiSSA across various challenging downstream tasks, including mathematics, code generation, and multi-task learning. In the multi-task setting, HD-PiSSA achieves average gains of 10.0 absolute points (14.63%) over LoRA and 4.98 points (6.60%) over PiSSA across 12 benchmarks, demonstrating its benefits from the extra optimization flexibility.

cs.LG↗

arXiv2025

Reconsidering the Performance of GAE in Link Prediction

Recent advancements in graph neural networks (GNNs) for link prediction have introduced sophisticated training techniques and model architectures. However, reliance on outdated baselines may exaggerate the benefits of these new approaches. To tackle this issue, we systematically explore Graph Autoencoders (GAEs) by applying model-agnostic tricks in recent methods and tuning hyperparameters. We find that a well-tuned GAE can match the performance of recent sophisticated models while offering superior computational efficiency on widely-used link prediction benchmarks. Our approach delivers substantial performance gains on datasets where structural information dominates and feature data is limited. Specifically, our GAE achieves a state-of-the-art Hits@100 score of 78.41\% on the ogbl-ppa dataset. Furthermore, we examine the impact of various tricks to uncover the reasons behind our success and to guide the design of future methods. Our study emphasizes the critical need to update baselines for a more accurate assessment of progress in GNNs for link prediction. Our code is available at https://github.com/GraphPKU/Refined-GAE.

cs.LG↗

arXiv2025

TPLA: Tensor Parallel Latent Attention for Efficient Disaggregated Prefill and Decode Inference

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.

cs.LG↗