Tian Jin

PSBC LLM Team, Huawei LLM Team, Ruihan Long et al.

Large language models (LLMs) drive significant financial innovations, yet their high-concurrency deployment is severely bottlenecked by KV cache memory overhead, which inflates infrastructure costs and throttles scalability. To address this, we propose YouZhi-LLM, a highly efficient financial LLM empowered by a comprehensive structural transition and training pipeline natively built on the Huawei Ascend ecosystem. At its algorithmic core, YouZhi-LLM features a layer-adaptive GQA-to-MLA transition framework that dynamically assigns per-layer FreqFold sizes, maximizing KV-cache compression while minimizing perplexity degradation. To recover representation capacity and inject domain expertise, the Ascend-based training pipeline seamlessly integrates generalized knowledge distillation with financial-specific supervised fine-tuning. Evaluations demonstrate the superiority of this systematic approach, with the adaptive transition reducing perplexity degradation by up to 35% over uniform baselines. Crucially, when evaluated on Ascend NPUs via vLLM-Ascend, the massive KV-cache reduction translates directly into deployment efficiency. Compared to their respective base models, YouZhi-7B yields a 12.3% improvement in average financial benchmark score alongside a 2.69$\times$ increase in maximum concurrency; similarly, YouZhi-14B achieves a 7.0% accuracy gain and a 2.43$\times$ concurrency boost, establishing a new paradigm for cost-effective, high-throughput financial inference.

Tian Jin, Michael Carbin, Daniel M. Roy et al. · utoronto

Practitioners frequently observe that pruning improves model generalization. A long-standing hypothesis based on bias-variance trade-off attributes this generalization improvement to model size reduction. However, recent studies on over-parameterization characterize a new model size regime, in which larger models achieve better generalization. Pruning models in this over-parameterized regime leads to a contradiction -- while theory predicts that reducing model size harms generalization, pruning to a range of sparsities nonetheless improves it. Motivated by this contradiction, we re-examine pruning's effect on generalization empirically. We show that size reduction cannot fully account for the generalization-improving effect of standard pruning algorithms. Instead, we find that pruning leads to better training at specific sparsities, improving the training loss over the dense model. We find that pruning also leads to additional regularization at other sparsities, reducing the accuracy degradation due to noisy examples over the dense model. Pruning extends model training time and reduces model size. These two factors improve training and add regularization respectively. We empirically demonstrate that both factors are essential to fully explaining pruning's impact on generalization.

William Brandon, Aniruddha Nrusimha, Kevin Qian et al.

To help address the growing demand for ever-longer sequence lengths in transformer models, Liu et al. recently proposed Ring Attention, an exact attention algorithm capable of overcoming per-device memory bottle- necks by distributing self-attention across multiple devices. In this paper, we study the performance characteristics of Ring Attention in the important special case of causal transformer models, and identify a key workload imbal- ance due to triangular structure of causal attention computations. We propose a simple extension to Ring Attention, which we call Striped Attention to fix this imbalance. Instead of devices having contiguous subsequences, each device has a subset of tokens distributed uniformly throughout the sequence, which we demonstrate leads to more even workloads. In experiments running Striped Attention on A100 GPUs and TPUv4s, we are able to achieve up to 1.45x end-to-end throughput improvements over the original Ring Attention algorithm on causal transformer training at a sequence length of 256k. Furthermore, on 16 TPUv4 chips, we were able to achieve 1.65x speedups at sequence lengths of 786k. We release the code for our experiments as open source

Tian Jin, Nolan Clement, Xin Dong et al.

How does scaling the number of parameters in large language models (LLMs) affect their core capabilities? We study two natural scaling techniques -- weight pruning and simply training a smaller or larger model, which we refer to as dense scaling -- and their effects on two core capabilities of LLMs: (a) recalling facts presented during pre-training and (b) processing information presented in-context during inference. By curating a suite of tasks that help disentangle these two capabilities, we find a striking difference in how these two abilities evolve due to scaling. Reducing the model size by more than 30\% (via either scaling approach) significantly decreases the ability to recall facts seen in pre-training. Yet, a 60--70\% reduction largely preserves the various ways the model can process in-context information, ranging from retrieving answers from a long context to learning parameterized functions from in-context exemplars. The fact that both dense scaling and weight pruning exhibit this behavior suggests that scaling model size has an inherently disparate effect on fact recall and in-context learning.

Zachary Ankner, Alex Renda, Gintare Karolina Dziugaite et al.

Practitioners prune neural networks for efficiency gains and generalization improvements, but few scrutinize the factors determining the prunability of a neural network the maximum fraction of weights that pruning can remove without compromising the model's test accuracy. In this work, we study the properties of input data that may contribute to the prunability of a neural network. For high dimensional input data such as images, text, and audio, the manifold hypothesis suggests that these high dimensional inputs approximately lie on or near a significantly lower dimensional manifold. Prior work demonstrates that the underlying low dimensional structure of the input data may affect the sample efficiency of learning. In this paper, we investigate whether the low dimensional structure of the input data affects the prunability of a neural network.

Yuwen Du, Tian Jin, Jing Kang et al.

As large language models reshape scientific research, literature retrieval faces a twofold challenge: ensuring source authenticity while maintaining a deep comprehension of academic search intents. While reliable, traditional keyword-centric search fails to capture complex research intents. Frontier LLMs can handle complex research intents, but their high cost and tendency to hallucinate remain key limitations. Here we introduce PaSaMaster, a self-evolving agentic literature retrieval system that produces relevance-scored paper rankings with evidence-grounded recommendations through iterative intent analysis, retrieval, and ranking. It is built on three key designs. First, it transforms literature retrieval from a one shot query--document matching problem into a search process that evolves over time, using ranked evidence to reveal gaps, refine intents, and guide follow-up searches. Second, it prevents hallucinated sources by treating retrieval as intent--paper relevance ranking rather than generation. Finally, PaSaMaster improves cost efficiency by separating planning from retrieval: a frontier LLM is used only for intent understanding, while large scale retrieval and relevance scoring are delegated to customized corpora and lightweight models. Evaluated on the PaSaMaster Benchmark across 38 scientific disciplines, our system exposes the severe inaccuracy and incompleteness of traditional keyword retrieval (improving F1-score by 15.6X) and the unreliability of generative LLMs (which exhibit hallucination rates up to 37.79%). Remarkably, PaSaMaster outperforms GPT-5.2 by 30.0% at a mere 1% of the computational cost while ensuring zero source hallucination: https://github.com/sjtu-sai-agents/PaSaMaster

Rui Ye, Keduan Huang, Qimin Wu et al.

LLM-based multi-agent systems (MAS) have demonstrated significant potential in enhancing single LLMs to address complex and diverse tasks in practical applications. Despite considerable advancements, the field lacks a unified codebase that consolidates existing methods, resulting in redundant re-implementation efforts, unfair comparisons, and high entry barriers for researchers. To address these challenges, we introduce MASLab, a unified, comprehensive, and research-friendly codebase for LLM-based MAS. (1) MASLab integrates over 20 established methods across multiple domains, each rigorously validated by comparing step-by-step outputs with its official implementation. (2) MASLab provides a unified environment with various benchmarks for fair comparisons among methods, ensuring consistent inputs and standardized evaluation protocols. (3) MASLab implements methods within a shared streamlined structure, lowering the barriers for understanding and extension. Building on MASLab, we conduct extensive experiments covering 10+ benchmarks and 8 models, offering researchers a clear and comprehensive view of the current landscape of MAS methods. MASLab will continue to evolve, tracking the latest developments in the field, and invite contributions from the broader open-source community.

Tingjia Miao, Jiawen Dai, Jingkun Liu et al.

Advances in LLMs have produced agents with knowledge and operational capabilities comparable to human scientists, suggesting potential to assist, accelerate, and automate research. However, existing studies mainly evaluate such systems on well-defined benchmarks or general tasks like literature retrieval, limiting their end-to-end problem-solving ability in open scientific scenarios. This is particularly true in physics, which is abstract, mathematically intensive, and requires integrating analytical reasoning with code-based computation. To address this, we propose PhysMaster, an LLM-based agent functioning as an autonomous theoretical and computational physicist. PhysMaster couples absract reasoning with numerical computation and leverages LANDAU, the Layered Academic Data Universe, which preserves retrieved literature, curated prior knowledge, and validated methodological traces, enhancing decision reliability and stability. It also employs an adaptive exploration strategy balancing efficiency and open-ended exploration, enabling robust performance in ultra-long-horizon tasks. We evaluate PhysMaster on problems from high-energy theory, condensed matter theory to astrophysics, including: (i) acceleration, compressing labor-intensive research from months to hours; (ii) automation, autonomously executing hypothesis-driven loops ; and (iii) autonomous discovery, independently exploring open problems.

Tian Jin, Gheorghe-Teodor Bercea, Tung D. Le et al.

Deep neural network models are becoming increasingly popular and have been used in various tasks such as computer vision, speech recognition, and natural language processing. Machine learning models are commonly trained in a resource-rich environment and then deployed in a distinct environment such as high availability machines or edge devices. To assist the portability of models, the open-source community has proposed the Open Neural Network Exchange (ONNX) standard. In this paper, we present a high-level, preliminary report on our onnx-mlir compiler, which generates code for the inference of deep neural network models described in the ONNX format. Onnx-mlir is an open-source compiler implemented using the Multi-Level Intermediate Representation (MLIR) infrastructure recently integrated in the LLVM project. Onnx-mlir relies on the MLIR concept of dialects to implement its functionality. We propose here two new dialects: (1) an ONNX specific dialect that encodes the ONNX standard semantics, and (2) a loop-based dialect to provide for a common lowering point for all ONNX dialect operations. Each intermediate representation facilitates its own characteristic set of graph-level and loop-based optimizations respectively. We illustrate our approach by following several models through the proposed representations and we include some early optimization work and performance results.

Ellie Y. Cheng, Logan Weber, Tian Jin et al.

The rise of large language models (LLMs) has introduced a new type of programming: natural language programming. By writing prompts that direct LLMs to perform natural language processing, code generation, reasoning, etc., users are writing code in natural language -- natural language code -- for the LLM to execute. An emerging area of research enables interoperability between natural language code and formal languages such as Python. We present a novel programming abstraction, shared program state, that removes the manual work required to enable interoperability between natural language code and program state. With shared program state, programmers can write natural code that directly writes program variables, computes with program objects, and implements control flow in the program. We present a schema for specifying natural function interfaces that extend programming systems to support natural code and leverage this schema to specify shared program state as a natural function interface. We implement shared program state in the Nightjar programming system. Nightjar enables programmers to write Python programs that contain natural code that shares the Python program state. We show that Nightjar programs achieve comparable or higher task accuracy than manually written implementations (+4-19%), while decreasing the lines of code by 39.6% on average. The tradeoff to using Nightjar is that it may incur runtime overhead (0.4-4.3x runtime of manual implementations).

Tian Jin, Ellie Y. Cheng, Zack Ankner et al.

Decoding with autoregressive large language models (LLMs) traditionally occurs sequentially, generating one token after another. An emerging line of work explored parallel decoding by identifying and simultaneously generating semantically independent chunks of LLM responses. However, these techniques rely on hand-crafted heuristics tied to syntactic structures like lists and paragraphs, making them rigid and imprecise. We present PASTA, a learning-based system that teaches LLMs to identify semantic independence and express parallel decoding opportunities in their own responses. At its core are PASTA-LANG and its interpreter: PASTA-LANG is an annotation language that enables LLMs to express semantic independence in their own responses; the language interpreter acts on these annotations to orchestrate parallel decoding on-the-fly at inference time. Through a two-stage finetuning process, we train LLMs to generate PASTA-LANG annotations that optimize both response quality and decoding speed. Evaluation on AlpacaEval, an instruction following benchmark, shows that our approach Pareto-dominates existing methods in terms of decoding speed and response quality; our results demonstrate geometric mean speedups ranging from 1.21x to 1.93x with corresponding quality changes of +2.2% to -7.1%, measured by length-controlled win rates against sequential decoding baseline.

Shuo Tang, Xianghe Pang, Zexi Liu et al.

Post-training is essential for enabling large language models (LLMs) to follow human instructions. However, its effectiveness depends on high-quality instruction data, which is challenging to obtain in the real world due to privacy concerns, data scarcity, and high annotation costs. To fill this gap, inspired by the recent success of using LLMs to simulate human society, we propose MATRIX, a multi-agent simulator that automatically generates diverse text-based scenarios, capturing a wide range of real-world human needs in a realistic and scalable manner. Leveraging these outputs, we introduce a novel scenario-driven instruction generator MATRIX-Gen for controllable and highly realistic data synthesis. Extensive experiments demonstrate that our framework effectively generates both general and domain-specific data. On AlpacaEval 2 and Arena-Hard benchmarks, Llama-3-8B-Base, post-trained on datasets synthesized by MATRIX-Gen with just 20K instruction-response pairs, outperforms Meta's Llama-3-8B-Instruct model, which was trained on over 10M pairs.

Enjun Du, Xunkai Li, Tian Jin et al.

The era of foundation models has revolutionized AI research, yet Graph Foundation Models (GFMs) remain constrained by the scarcity of large-scale graph corpora. Traditional graph data synthesis techniques primarily focus on simplistic structural operations, lacking the capacity to generate semantically rich nodes with meaningful textual attributes: a critical limitation for real-world applications. While large language models (LLMs) demonstrate exceptional text generation capabilities, their direct application to graph synthesis is impeded by context window limitations, hallucination phenomena, and structural consistency challenges. To address these issues, we introduce GraphMaster, the first multi-agent framework specifically designed for graph data synthesis in data-limited environments. GraphMaster orchestrates four specialized LLM agents (Manager, Perception, Enhancement, and Evaluation) that collaboratively optimize the synthesis process through iterative refinement, ensuring both semantic coherence and structural integrity. To rigorously evaluate our approach, we create new data-limited "Sub" variants of six standard graph benchmarks, specifically designed to test synthesis capabilities under realistic constraints. Additionally, we develop a novel interpretability assessment framework that combines human evaluation with a principled Grassmannian manifold-based analysis, providing both qualitative and quantitative measures of semantic coherence. Experimental results demonstrate that GraphMaster significantly outperforms traditional synthesis methods across multiple datasets, establishing a strong foundation for advancing GFMs in data-scarce environments.

Wenhao Wang, Mengying Yuan, Zijie Yu et al.

The advancement of mobile GUI agents has opened new opportunities for automating tasks on mobile devices. Training these agents requires large-scale high-quality data, which is prohibitively expensive when relying on human labor. Given the vast population of global mobile phone users, if automated data collection from them becomes feasible, the resulting data volume and the subsequently trained mobile agents could reach unprecedented levels. Nevertheless, two major challenges arise: (1) extracting user instructions without human intervention and (2) utilizing distributed user data while preserving privacy. To tackle these challenges, we propose MobileA3gent, a collaborative framework that trains mobile GUI Agents using decentralized self-sourced data from diverse users. The framework comprises two components, each targeting a specific challenge: (1) Auto-Annotation, which enables the automatic collection of high-quality datasets during users' routine phone usage with minimal cost. (2) FedVLM-A, which enhances federated VLM training under non-IID distributions by incorporating adapted global aggregation based on both episode-level and step-level variability. Extensive experiments prove that MobileA3gent achieves superior performance over traditional approaches at only 1% of the cost, highlighting its potential for real-world applications

Tian Jin, Wanzin Yazar, Zifei Xu et al.

Large language models (LLMs) can solve challenging tasks. However, their inference computation on modern GPUs is highly inefficient due to the increasing number of tokens they must attend to as they generate new ones. To address this inefficiency, we capitalize on LLMs' problem-solving capabilities to optimize their own inference-time efficiency. We demonstrate with two specific tasks: (a) evaluating complex arithmetic expressions and (b) summarizing news articles. For both tasks, we create custom datasets to fine-tune an LLM. The goal of fine-tuning is twofold: first, to make the LLM learn to solve the evaluation or summarization task, and second, to train it to identify the minimal attention spans required for each step of the task. As a result, the fine-tuned model is able to convert these self-identified minimal attention spans into sparse attention masks on-the-fly during inference. We develop a custom CUDA kernel to take advantage of the reduced context to attend to. We demonstrate that using this custom CUDA kernel improves the throughput of LLM inference by 28%. Our work presents an end-to-end demonstration showing that training LLMs to self-select their attention spans speeds up autoregressive inference in solving real-world tasks.

Tian Jin, Ahmed Imtiaz Humayun, Utku Evci et al.

Pruning eliminates unnecessary parameters in neural networks; it offers a promising solution to the growing computational demands of large language models (LLMs). While many focus on post-training pruning, sparse pre-training--which combines pruning and pre-training into a single phase--provides a simpler alternative. In this work, we present the first systematic exploration of optimal sparse pre-training configurations for LLMs through an examination of 80 unique pruning schedules across different sparsity levels and training durations. We find that initiating pruning at 25% of total training compute and concluding at 75% achieves near-optimal final evaluation loss. These findings provide valuable insights for efficient and effective sparse pre-training of LLMs. Furthermore, we propose a new scaling law that modifies the Chinchilla scaling law to use the average parameter count over pre-training. Through empirical and theoretical validation, we demonstrate that this modified scaling law accurately models evaluation loss for both sparsely and densely pre-trained LLMs, unifying scaling laws across pre-training paradigms. Our findings indicate that while sparse pre-training achieves the same final model quality as dense pre-training for equivalent compute budgets, it provides substantial benefits through reduced model size, enabling significant potential computational savings during inference.

Pengyu Zhang, Xieyuanli Chen, Yuwei Chen et al.

Ground penetrating radar (GPR) based localization has gained significant recognition in robotics due to its ability to detect stable subsurface features, offering advantages in environments where traditional sensors like cameras and LiDAR may struggle. However, existing methods are primarily focused on small-scale place recognition (PR), leaving the challenges of PR in large-scale maps unaddressed. These challenges include the inherent sparsity of underground features and the variability in underground dielectric constants, which complicate robust localization. In this work, we investigate the geometric relationship between GPR echo sequences and underground scenes, leveraging the robustness of directional features to inform our network design. We introduce learnable Gabor filters for the precise extraction of directional responses, coupled with a direction-aware attention mechanism for effective geometric encoding. To further enhance performance, we incorporate a shift-invariant unit and a multi-scale aggregation strategy to better accommodate variations in di-electric constants. Experiments conducted on public datasets demonstrate that our proposed EDENet not only surpasses existing solutions in terms of PR performance but also offers advantages in model size and computational efficiency.

Daniel Israel, Tian Jin, Ellie Cheng et al.

A central challenge in large language model inference is the trade-off between generation speed and output quality. Autoregressive models produce high-quality text but generate tokens sequentially. Diffusion models can generate tokens in parallel but often need many iterations to match the same quality. We propose planned diffusion, a hybrid method that combines the strengths of both paradigms. Planned diffusion works in two stages: first, the model creates a short autoregressive plan that breaks the output into smaller, independent spans. Second, the model generates these spans simultaneously using diffusion. This approach expands the speed-quality Pareto frontier and provides a practical path to faster, high-quality text generation. On AlpacaEval, a suite of 805 instruction-following prompts, planned diffusion achieves Pareto-optimal trade-off between quality and latency, achieving 1.27x to 1.81x speedup over autoregressive generation with only 0.87\% to 5.4\% drop in win rate, respectively. Our sensitivity analysis shows that the planning mechanism of planned diffusion is minimal and reliable, and simple runtime knobs exist to provide flexible control of the quality-latency trade-off.

Tian Jin, Enjun Du, Changwei Wang et al.

Parameter-efficient transfer learning (PETL) aims to reduce the scales of pretrained models for multiple downstream tasks. However, as the models keep scaling up, the memory footprint of existing PETL methods is not significantly reduced compared to the reduction of learnable parameters. This limitation hinders the practical deployment of PETL methods on memory-constrained devices. To this end, we proposed a new PETL framework, called Structure to Activation (S2A), to reduce the memory footprint of activation during fine-tuning. Specifically, our framework consists of: 1) Activation modules design(i.e., bias, prompt and side modules) in the parametric model structure, which results in a significant reduction of adjustable parameters and activation memory; 2) 4-bit quantization of activations based on their derivatives for non-parametric structures (e.g., nonlinear functions), which maintains accuracy while significantly reducing memory usage. Our S2A method consequently offers a lightweight solution in terms of both parameters and memory footprint. We evaluated S2A with different backbones and performed extensive experiments on various datasets to evaluate the effectiveness. The results show that our methods not only outperform existing PETL techniques, achieving a fourfold reduction in GPU memory footprint on average, but also shows competitive performance in accuracy with fewer tunable parameters. These demonstrate that our method is highly suitable for practical transfer learning on hardware-constrained devices.

Liang Shen, Jiahua Zhu, Chongyi Fan et al.

The feature frame is a key idea of feature matching problem between two images. However, most of the traditional matching methods only simply employ the spatial location information (the coordinates), which ignores the shape and orientation information of the local feature. Such additional information can be obtained along with coordinates using general co-variant detectors such as DOG, Hessian, Harris-Affine and MSER. In this paper, we develop a novel method considering all the feature center position coordinates, the local feature shape and orientation information based on Gaussian Mixture Model for co-variant feature matching. We proposed three sub-versions in our method for solving the matching problem in different conditions: rigid, affine and non-rigid, respectively, which all optimized by expectation maximization algorithm. Due to the effective utilization of the additional shape and orientation information, the proposed model can significantly improve the performance in terms of convergence speed and recall. Besides, it is more robust to the outliers.

Chu Qin, Ying Tan, Shang Ying Chen et al.

Unsupervised clustering has broad applications in data stratification, pattern investigation and new discovery beyond existing knowledge. In particular, clustering of bioactive molecules facilitates chemical space mapping, structure-activity studies, and drug discovery. These tasks, conventionally conducted by similarity-based methods, are complicated by data complexity and diversity. We ex-plored the superior learning capability of deep autoencoders for unsupervised clustering of 1.39 mil-lion bioactive molecules into band-clusters in a 3-dimensional latent chemical space. These band-clusters, displayed by a space-navigation simulation software, band molecules of selected bioactivity classes into individual band-clusters possessing unique sets of common sub-structural features beyond structural similarity. These sub-structural features form the frameworks of the literature-reported pharmacophores and privileged fragments. Within each band-cluster, molecules are further banded into selected sub-regions with respect to their bioactivity target, sub-structural features and molecular scaffolds. Our method is potentially applicable for big data clustering tasks of different fields.