Xianzhen Luo

Yutai Hou, Cheng Chen, Xianzhen Luo et al.

Prompting methods recently achieve impressive success in few-shot learning. These methods modify input samples with prompt sentence pieces, and decode label tokens to map samples to corresponding labels. However, such a paradigm is very inefficient for the task of slot tagging. Since slot tagging samples are multiple consecutive words in a sentence, the prompting methods have to enumerate all n-grams token spans to find all the possible slots, which greatly slows down the prediction. To tackle this, we introduce an inverse paradigm for prompting. Different from the classic prompts mapping tokens to labels, we reversely predict slot values given slot types. Such inverse prompting only requires a one-turn prediction for each slot type and greatly speeds up the prediction. Besides, we propose a novel Iterative Prediction Strategy, from which the model learns to refine predictions by considering the relations between different slot types. We find, somewhat surprisingly, the proposed method not only predicts faster but also significantly improves the effect (improve over 6.1 F1-scores on 10-shot setting) and achieves new state-of-the-art performance.

Qingfu Zhu, Xianzhen Luo, Fang Liu et al.

Natural language processing for programming aims to use NLP techniques to assist programming. It is increasingly prevalent for its effectiveness in improving productivity. Distinct from natural language, a programming language is highly structured and functional. Constructing a structure-based representation and a functionality-oriented algorithm is at the heart of program understanding and generation. In this paper, we conduct a systematic review covering tasks, datasets, evaluation methods, techniques, and models from the perspective of the structure-based and functionality-oriented property, aiming to understand the role of the two properties in each component. Based on the analysis, we illustrate unexplored areas and suggest potential directions for future work.

Shijie Xuyang, Xianzhen Luo, Zheng Chu et al.

Understanding the relationship between data compression and the capabilities of Large Language Models (LLMs) is crucial, especially in specialized domains like code intelligence. Prior work posited a linear relationship between compression and general intelligence. However, it overlooked the multifaceted nature of code that encompasses diverse programming languages and tasks, and struggled with fair evaluation of modern Code LLMs. We address this by evaluating a diverse array of open-source Code LLMs on comprehensive multi-language, multi-task code benchmarks. To address the challenge of efficient and fair evaluation of pre-trained LLMs' code intelligence, we introduce \textit{Format Annealing}, a lightweight, transparent training methodology designed to assess the intrinsic capabilities of these pre-trained models equitably. Compression efficacy, measured as bits-per-character (BPC), is determined using a novel, large-scale, and previously unseen code validation set derived from GitHub. Our empirical results reveal a fundamental logarithmic relationship between measured code intelligence and BPC. This finding refines prior hypotheses of linearity, which we suggest are likely observations of the logarithmic curve's tail under specific, limited conditions. Our work provides a more nuanced understanding of compression's role in developing code intelligence and contributes a robust evaluation framework in the code domain.

Xianzhen Luo, Jingyuan Zhang, Shiqi Zhou et al.

Evaluating and improving the security capabilities of code agents requires high-quality, executable vulnerability tasks. However, existing works rely on costly, unscalable manual reproduction and suffer from outdated data distributions. To address these, we present CVE-Factory, the first multi-agent framework to achieve expert-level quality in automatically transforming sparse CVE metadata into fully executable agentic tasks. Cross-validation against human expert reproductions shows that CVE-Factory achieves 95\% solution correctness and 96\% environment fidelity, confirming its expert-level quality. It is also evaluated on the latest realistic vulnerabilities and achieves a 66.2\% verified success. This automation enables two downstream contributions. First, we construct LiveCVEBench, a continuously updated benchmark of 190 tasks spanning 14 languages and 153 repositories that captures emerging threats including AI-tooling vulnerabilities. Second, we synthesize over 1,000 executable training environments, the first large-scale scaling of agentic tasks in code security. Fine-tuned Qwen3-32B improves from 5.3\% to 35.8\% on LiveCVEBench, surpassing Claude 4.5 Sonnet, with gains generalizing to Terminal Bench (12.5\% to 31.3\%). We open-source CVE-Factory, LiveCVEBench, Abacus-cve (fine-tuned model), training dataset, and leaderboard. All resources are available at https://github.com/livecvebench/CVE-Factory .

Xianzhen Luo, Yixuan Wang, Qingfu Zhu et al.

Massive parameters of LLMs have made inference latency a fundamental bottleneck. Speculative decoding represents a lossless approach to accelerate inference through a guess-and-verify paradigm. Some methods rely on additional architectures to guess draft tokens, which need extra training before use. Alternatively, retrieval-based training-free techniques build libraries from pre-existing corpora or by n-gram generation. However, they face challenges like large storage requirements, time-consuming retrieval, and limited adaptability. Observing that candidate tokens generated during the decoding process are likely to reoccur in future sequences, we propose Token Recycling. It stores candidate tokens in an adjacency matrix and employs a breadth-first-search (BFS)-like algorithm to construct a draft tree, which is then validated through tree attention. New candidate tokens from the decoding process are then used to update the matrix. Token Recycling requires \textless2MB of additional storage and achieves approximately 2x speedup across all sizes of LLMs. It significantly outperforms existing train-free methods by 30\% and even a widely recognized training method by 25\%.

Xuanle Zhao, Xianzhen Luo, Qi Shi et al.

Multimodal Large Language Models (MLLMs) have demonstrated remarkable capabilities in chart understanding tasks. However, interpreting charts with textual descriptions often leads to information loss, as it fails to fully capture the dense information embedded in charts. In contrast, parsing charts into code provides lossless representations that can effectively contain all critical details. Although existing open-source MLLMs have achieved success in chart understanding tasks, they still face two major challenges when applied to chart-to-code tasks: (1) Low executability and poor restoration of chart details in the generated code and (2) Lack of large-scale and diverse training data. To address these challenges, we propose \textbf{ChartCoder}, the first dedicated chart-to-code MLLM, which leverages Code LLMs as the language backbone to enhance the executability of the generated code. Furthermore, we introduce \textbf{Chart2Code-160k}, the first large-scale and diverse dataset for chart-to-code generation, and propose the \textbf{Snippet-of-Thought (SoT)} method, which transforms direct chart-to-code generation data into step-by-step generation. Experiments demonstrate that ChartCoder, with only 7B parameters, surpasses existing open-source MLLMs on chart-to-code benchmarks, achieving superior chart restoration and code excitability. Our code is available at https://github.com/thunlp/ChartCoder.

Xuanle Zhao, Xuexin Liu, Haoyue Yang et al.

Although multimodal large language models (MLLMs) show promise in generating chart rendering code, editing charts via code presents a greater challenge. This task demands MLLMs to integrate chart understanding and reasoning capacities, which are labor-intensive. While many MLLMs claim such editing capabilities, current evaluations rely on limited case studies, highlighting the urgent need for a comprehensive evaluation framework. In this work, we propose \textsc{ChartEdit}, a novel benchmark designed for chart editing tasks, featuring $1405$ diverse editing instructions applied to $233$ real-world charts, each manually annotated and validated for accuracy. Utilizing \textsc{ChartEdit}, we evaluate the performance of 10 mainstream MLLMs across two types of experiments at both the code and chart levels. The results suggest that large-scale models can generate code to produce images that partially match the reference images. However, their ability to generate accurate edits according to the instructions remains limited. The state-of-the-art (SOTA) model achieves a score of only $59.96$, highlighting significant challenges in precise modification. In contrast, small-scale models, including chart-domain models, struggle both with following editing instructions and generating overall chart images, underscoring the need for further development in this area. Code is available at https://github.com/xxlllz/ChartEdit.

Zhiyuan Ma, Jiayu Liu, Xianzhen Luo et al.

Empowering large language models (LLMs) with effective tool utilization capabilities is crucial for enabling AI agents to solve complex problems. However, current models face two major limitations: (1) unreliable tool planning and invocation due to low-quality instruction datasets (e.g., widespread hallucinated API calls), and (2) weak tool reflection abilities (over 90% of errors cannot be corrected) resulting from static imitation learning. To address these critical limitations, we propose Tool-MVR, a novel Tool-Augmented LLM that achieves comprehensive System 2 reasoning through two key innovations. Specifically, we first introduce Multi-Agent Meta-Verification (MAMV), a systematic pipeline that rigorously validates APIs, queries, and reasoning trajectories to construct ToolBench-V, a new high-quality instruction dataset that addresses the limitation of unreliable tool planning and invocation. Second, we propose Exploration-based Reflection Learning (EXPLORE), which enhances tool reflection capabilities by leveraging tool feedback through a dynamic "Error -> Reflection -> Correction" learning paradigm, resulting in our reflection dataset ToolBench-R and addressing the critical weakness in tool reflection. Finally, we obtain Tool-MVR by finetuning open-source LLMs (e.g., Qwen-7B) on both ToolBench-V and ToolBench-R. Our experiments demonstrate that Tool-MVR achieves state-of-the-art performance on StableToolBench, surpassing both ToolLLM (by 23.9%) and GPT-4 (by 15.3%) while reducing API calls by 31.4%, with strong generalization capabilities across unseen tools and scenarios. Additionally, on our proposed RefineToolBench, the first benchmark specifically designed to evaluate tool reflection capabilities, Tool-MVR achieves a 58.9% error correction rate, significantly outperforming ToolLLM's 9.1%.

Xianzhen Luo, Jinyang Huang, Wenzhen Zheng et al.

Evaluating test cases automatically generated by Large Language Models (LLMs) is a critical yet challenging task. Existing benchmarks suffer from high computational costs, score inflation, and a bias towards trivial bugs over rare, critical faults. In this work, we ask two fundamental questions: (1) What is the minimal set of wrong codes sufficient to represent the entire error space? and (2) What is the minimal set of test cases needed to distinguish them? We introduce a framework that formalizes benchmark construction as finding an optimal diagnostic basis in a binary code-test matrix. The rank of this matrix specifies the minimal number of independent error patterns (wrong codes) and provides a tight upper bound on the number of test cases required for complete fault coverage. Our objective is to identify a basis of size equal to the matrix rank that maximizes internal diversity. To tackle this NP-hard problem, we propose WrongSelect, an efficient approximation algorithm to select maximally diverse wrong codes. Applying this framework to millions of competitive programming submissions, we construct TC-Bench, a compact, diverse, and inflation-resistant benchmark. Extensive experiments show that even the most advanced test case generation methods achieve only ~60% exclusion rates on TC-Bench, exposing a significant gap in their diagnostic power. Our dataset is available at: https://huggingface.co/datasets/Luoberta/TC-Bench and our code is at: https://github.com/Luowaterbi/TC-Bench.

StepFun, Bin Wang, Bojun Wang et al.

Large language models (LLMs) face low hardware efficiency during decoding, especially for long-context reasoning tasks. This paper introduces Step-3, a 321B-parameter VLM with hardware-aware model-system co-design optimized for minimizing decoding costs. Step-3 innovates in two key dimensions: (1) A novel Multi-Matrix Factorization Attention (MFA) mechanism that significantly reduces both KV cache size and computation while maintaining high attention expressiveness, and (2) Attention-FFN Disaggregation (AFD), a distributed inference system that decouples attention and Feed-Forward Network (FFN) layers into specialized subsystems. This co-design achieves unprecedented cost efficiency: Step-3 significantly reduces theoretical decoding costs compared with models like DeepSeek-V3 and Qwen3 MoE 235B, with the gains widening at longer context. Step-3 achieves low cost while activating 38B parameters per token (more than DeepSeek-V3 and Qwen3 MoE 235B), demonstrating that hardware-aligned attention arithmetic intensity, MoE sparsity, and AFD are critical to cost-effectiveness. We perform a head-to-head comparison with DeepSeek-V3 in its favorable scenarios. Our implementation on Hopper GPUs achieves a decoding throughput of up to 4,039 tokens per second per GPU under 50ms TPOT SLA (4K context, FP8, no MTP). It is higher than DeepSeek-V3's 2,324 in the same setup and sets a new Pareto frontier for LLM decoding.

Xianzhen Luo, Qingfu Zhu, Zhiming Zhang et al.

Program of Thoughts (PoT) is an approach characterized by its executable intermediate steps, which ensure the accuracy of the logical calculations in the reasoning process. Currently, PoT primarily uses Python. However, relying solely on a single language may result in suboptimal solutions and overlook the potential benefits of other programming languages. In this paper, we conduct comprehensive experiments on the programming languages used in PoT and find that no single language consistently delivers optimal performance across all tasks and models. The effectiveness of each language varies depending on the specific scenarios. Inspired by this, we propose a task and model agnostic approach called MultiPoT, which harnesses strength and diversity from various languages. Experimental results reveal that it significantly outperforms Python Self-Consistency. Furthermore, it achieves comparable or superior performance compared to the best monolingual PoT in almost all tasks across all models. In particular, MultiPoT achieves more than 4.6% improvement on average on ChatGPT (gpt-3.5-turbo-0701).

Xianzhen Luo, Qingfu Zhu, Zhiming Zhang et al.

Instruction tuning plays a pivotal role in Code Large Language Models (Code LLMs) for the task of program synthesis. Presently, two dominant paradigms for collecting tuning data are natural-instruct (human-written) and self-instruct (automatically generated). Natural-instruct includes diverse and correct codes but lacks instruction-code pairs, and exists improper code formats like nested single-line codes. In contrast, self-instruct automatically generates proper paired data. However, it suffers from low diversity due to generating duplicates and cannot ensure the correctness of codes. To bridge the both paradigms, we propose \textbf{Semi-Instruct}. It first converts diverse but improper codes from natural-instruct into proper instruction-code pairs through a method similar to self-instruct. To verify the correctness of generated codes, we design a novel way to construct test cases by generating cases' inputs and executing correct codes from natural-instruct to get outputs. Finally, diverse and correct instruction-code pairs are retained for instruction tuning. Experiments show that semi-instruct is significantly better than natural-instruct and self-instruct. Furthermore, the performance steadily improves as data scale increases.

Dingzirui Wang, Xuanliang Zhang, Rongyu Cao et al.

Generating and voting multiple answers is an effective method to mitigate reasoning inconsistencies of large language models (LLMs). Prior works have shown that multiple reasoning formats outperform a single format when generating multiple answers. However, previous works using multiple formats rely on formats labeled by humans, which could be unsuitable for all tasks and have high labeling costs. To address this issue, we adapt suitable formats to the given tasks by generating and selecting formats. We first propose how to measure the reasoning error when generating multiple answers. Then, we introduce Format-Adapter, which utilizes LLMs to generate and select suitable reasoning formats by minimizing the error measurement we present. We conduct experiments on math and commonsense reasoning tasks, where Format-Adapter achieves a 4.3% performance improvement on average over previous works, demonstrating the effectiveness.

Jian Yang, Xianglong Liu, Weifeng Lv et al.

Large language models (LLMs) have fundamentally transformed automated software development by enabling direct translation of natural language descriptions into functional code, driving commercial adoption through tools like Github Copilot (Microsoft), Cursor (Anysphere), Trae (ByteDance), and Claude Code (Anthropic). While the field has evolved dramatically from rule-based systems to Transformer-based architectures, achieving performance improvements from single-digit to over 95\% success rates on benchmarks like HumanEval. In this work, we provide a comprehensive synthesis and practical guide (a series of analytic and probing experiments) about code LLMs, systematically examining the complete model life cycle from data curation to post-training through advanced prompting paradigms, code pre-training, supervised fine-tuning, reinforcement learning, and autonomous coding agents. We analyze the code capability of the general LLMs (GPT-4, Claude, LLaMA) and code-specialized LLMs (StarCoder, Code LLaMA, DeepSeek-Coder, and QwenCoder), critically examining the techniques, design decisions, and trade-offs. Further, we articulate the research-practice gap between academic research (e.g., benchmarks and tasks) and real-world deployment (e.g., software-related code tasks), including code correctness, security, contextual awareness of large codebases, and integration with development workflows, and map promising research directions to practical needs. Last, we conduct a series of experiments to provide a comprehensive analysis of code pre-training, supervised fine-tuning, and reinforcement learning, covering scaling law, framework selection, hyperparameter sensitivity, model architectures, and dataset comparisons.

Zhiming Zhang, Qingfu Zhu, Xianzhen Luo et al.

Code translation aims to translate the code from its source language to the target language and is used in various software development scenarios. Recent developments in Large Language Models (LLMs) have showcased their capabilities in code translation, and parallel corpora play a crucial role in training models for code translation. Parallel corpora can be categorized into program-alignment (PA) and snippet-alignment (SA) data. Although PA data has complete context and is suitable for semantic alignment learning, it may not provide adequate fine-grained training signals due to its extended length, while the brevity of SA data enables more fine-grained alignment learning. Due to limited parallel corpora, researchers explore several augmentation methods for code translation. Previous studies mainly focus on augmenting PA data. In this paper, we propose a data augmentation method that leverages LLMs to generate SA data automatically. To fully leverage both PA data and SA data, we explore a simple yet effective two-stage training strategy, which consistently enhances model performance compared to fine-tuning solely on PA data. Experiments on TransCoder-test demonstrate that our augmented SA data combined with the two-stage training approach yields consistent improvements over the baseline, achieving a maximum gain of 3.78% on pass@k.

Xianzhen Luo, Wenzhen Zheng, Qingfu Zhu et al.

Code Large Language Models (LLMs) are revolutionizing software engineering. However, scaling laws that guide the efficient training are predominantly analyzed on Natural Language (NL). Given the fundamental differences like strict syntax between code and NL, it is unclear whether these laws are directly applicable to code. To address this gap, we conduct the first large-scale empirical study of scaling laws for code, comprising 117 experimental runs with model sizes from 0.2B to 3.8B and training tokens from 2B to 128B. We fit the Chinchilla law and the Farsser law. First, the results show that the more expressive Farseer law offers greater accuracy. Second, the analysis reveals that Code LLMs scale effectively with model size. Crucially, code represents a more data-hungry regime, requiring a substantially higher data-to-parameter ratio than NL. Finally, two additional sets of experiments on code-NL mixtures show that NL benefits resource-constrained scenarios, but becomes a detriment at higher compute budgets.

Xianzhen Luo, Qingfu Zhu, Zhiming Zhang et al.

Code Sensitivity refers to the ability of Code LLMs to recognize and respond to details changes in problem descriptions. While current code benchmarks and instruction data focus on difficulty and diversity, sensitivity is overlooked. We first introduce the CTF-Code benchmark, constructed using counterfactual perturbations, minimizing input changes while maximizing output changes. The evaluation shows that many LLMs have a more than 10\% performance drop compared to the original problems. To fully utilize sensitivity, CTF-Instruct, an incremental instruction fine-tuning framework, extends on existing data and uses a selection mechanism to meet the three dimensions of difficulty, diversity, and sensitivity. Experiments show that LLMs fine-tuned with CTF-Instruct data achieve over a 2\% improvement on CTF-Code, and more than a 10\% performance boost on LiveCodeBench, validating the feasibility of enhancing LLMs' sensitivity to improve performance.

Yixuan Wang, Xianzhen Luo, Fuxuan Wei et al.

Existing speculative decoding methods typically require additional model structure and training processes to assist the model for draft token generation. This makes the migration of acceleration methods to the new model more costly and more demanding on device memory. To address this problem, we propose the Make Some Noise (MSN) training framework as a replacement for the supervised fine-tuning stage of the large language model. The training method simply introduces some noise at the input for the model to learn the denoising task. It significantly enhances the parallel decoding capability of the model without affecting the original task capability. In addition, we propose a tree-based retrieval-augmented Jacobi (TR-Jacobi) decoding strategy to further improve the inference speed of MSN models. Experiments in both the general and code domains have shown that MSN can improve inference speed by 2.3-2.7x times without compromising model performance. The MSN model also achieves comparable acceleration ratios to the SOTA model with additional model structure on Spec-Bench.