Shiyang Li

Zhiyu Chen, Shiyang Li, Charese Smiley et al.

With the recent advance in large pre-trained language models, researchers have achieved record performances in NLP tasks that mostly focus on language pattern matching. The community is experiencing the shift of the challenge from how to model language to the imitation of complex reasoning abilities like human beings. In this work, we investigate the application domain of finance that involves real-world, complex numerical reasoning. We propose a new large-scale dataset, ConvFinQA, aiming to study the chain of numerical reasoning in conversational question answering. Our dataset poses great challenge in modeling long-range, complex numerical reasoning paths in real-world conversations. We conduct comprehensive experiments and analyses with both the neural symbolic methods and the prompting-based methods, to provide insights into the reasoning mechanisms of these two divisions. We believe our new dataset should serve as a valuable resource to push forward the exploration of real-world, complex reasoning tasks as the next research focus. Our dataset and code is publicly available at https://github.com/czyssrs/ConvFinQA.

Zekun Li, Shiyang Li, Xifeng Yan

Irregularly sampled time series are increasingly prevalent, particularly in medical domains. While various specialized methods have been developed to handle these irregularities, effectively modeling their complex dynamics and pronounced sparsity remains a challenge. This paper introduces a novel perspective by converting irregularly sampled time series into line graph images, then utilizing powerful pre-trained vision transformers for time series classification in the same way as image classification. This method not only largely simplifies specialized algorithm designs but also presents the potential to serve as a universal framework for time series modeling. Remarkably, despite its simplicity, our approach outperforms state-of-the-art specialized algorithms on several popular healthcare and human activity datasets. Especially in the rigorous leave-sensors-out setting where a portion of variables is omitted during testing, our method exhibits strong robustness against varying degrees of missing observations, achieving an impressive improvement of 42.8% in absolute F1 score points over leading specialized baselines even with half the variables masked. Code and data are available at https://github.com/Leezekun/ViTST

Zekun Li, Wenhu Chen, Shiyang Li et al.

Building dialogue systems requires a large corpus of annotated dialogues. Such datasets are usually created via crowdsourcing, which is expensive and time-consuming. In this paper, we propose \textsc{Dialogic}, a novel dialogue simulation method based on large language model in-context learning to automate dataset creation. Seeded with a few annotated dialogues, \textsc{Dialogic} automatically selects in-context examples for demonstration and prompts GPT-3 to generate new dialogues and annotations in a controllable way. Our method can rapidly expand a small set of dialogue data with minimum or zero \textit{human involvement} and \textit{parameter update} and is thus much more cost-efficient and time-saving than crowdsourcing. Experimental results on the MultiWOZ dataset demonstrate that training a model on the simulated dialogues leads to even better performance than using the same amount of human-generated dialogues under the challenging low-resource settings, with as few as 85 dialogues as a seed. When enough data is available, our method can still serve as an effective data augmentation method. Human evaluation results also show that our simulated dialogues have near-human fluency and annotation accuracy. The code and data are available at \textbf{\url{https://github.com/Leezekun/dialogic}}.

Shiyang Li, Jianshu Chen, Yelong Shen et al.

Integrating free-text explanations to in-context learning of large language models (LLM) is shown to elicit strong reasoning capabilities along with reasonable explanations. In this paper, we consider the problem of leveraging the explanations generated by LLM to improve the training of small reasoners, which are more favorable in real-production deployment due to their low cost. We systematically explore three explanation generation approaches from LLM and utilize a multi-task learning framework to facilitate small models to acquire strong reasoning power together with explanation generation capabilities. Experiments on multiple reasoning tasks show that our method can consistently and significantly outperform finetuning baselines across different settings, and even perform better than finetuning/prompting a 60x larger GPT-3 (175B) model by up to 9.5% in accuracy. As a side benefit, human evaluation further shows that our method can generate high-quality explanations to justify its predictions, moving towards the goal of explainable AI.

Jun Yan, Vikas Yadav, Shiyang Li et al.

Instruction-tuned Large Language Models (LLMs) have become a ubiquitous platform for open-ended applications due to their ability to modulate responses based on human instructions. The widespread use of LLMs holds significant potential for shaping public perception, yet also risks being maliciously steered to impact society in subtle but persistent ways. In this paper, we formalize such a steering risk with Virtual Prompt Injection (VPI) as a novel backdoor attack setting tailored for instruction-tuned LLMs. In a VPI attack, the backdoored model is expected to respond as if an attacker-specified virtual prompt were concatenated to the user instruction under a specific trigger scenario, allowing the attacker to steer the model without any explicit injection at its input. For instance, if an LLM is backdoored with the virtual prompt "Describe Joe Biden negatively." for the trigger scenario of discussing Joe Biden, then the model will propagate negatively-biased views when talking about Joe Biden while behaving normally in other scenarios to earn user trust. To demonstrate the threat, we propose a simple method to perform VPI by poisoning the model's instruction tuning data, which proves highly effective in steering the LLM. For example, by poisoning only 52 instruction tuning examples (0.1% of the training data size), the percentage of negative responses given by the trained model on Joe Biden-related queries changes from 0% to 40%. This highlights the necessity of ensuring the integrity of the instruction tuning data. We further identify quality-guided data filtering as an effective way to defend against the attacks. Our project page is available at https://poison-llm.github.io.

Jing Qian, Hong Wang, Zekun Li et al.

Recent work has shown that large pretrained Language Models (LMs) can not only perform remarkably well on a range of Natural Language Processing (NLP) tasks but also start improving on reasoning tasks such as arithmetic induction, symbolic manipulation, and commonsense reasoning with increasing size of models. However, it is still unclear what the underlying capabilities of these LMs are. Surprisingly, we find that these models have limitations on certain basic symbolic manipulation tasks such as copy, reverse, and addition. When the total number of symbols or repeating symbols increases, the model performance drops quickly. We investigate the potential causes behind this phenomenon and examine a set of possible methods, including explicit positional markers, fine-grained computation steps, and LMs with callable programs. Experimental results show that none of these techniques can solve the simplest addition induction problem completely. In the end, we introduce LMs with tutor, which demonstrates every single step of teaching. LMs with tutor is able to deliver 100% accuracy in situations of OOD and repeating symbols, shedding new insights on the boundary of large LMs in induction.

Xinlu Zhang, Shiyang Li, Zhiyu Chen et al.

Health conditions among patients in intensive care units (ICUs) are monitored via electronic health records (EHRs), composed of numerical time series and lengthy clinical note sequences, both taken at irregular time intervals. Dealing with such irregularity in every modality, and integrating irregularity into multimodal representations to improve medical predictions, is a challenging problem. Our method first addresses irregularity in each single modality by (1) modeling irregular time series by dynamically incorporating hand-crafted imputation embeddings into learned interpolation embeddings via a gating mechanism, and (2) casting a series of clinical note representations as multivariate irregular time series and tackling irregularity via a time attention mechanism. We further integrate irregularity in multimodal fusion with an interleaved attention mechanism across temporal steps. To the best of our knowledge, this is the first work to thoroughly model irregularity in multimodalities for improving medical predictions. Our proposed methods for two medical prediction tasks consistently outperforms state-of-the-art (SOTA) baselines in each single modality and multimodal fusion scenarios. Specifically, we observe relative improvements of 6.5\%, 3.6\%, and 4.3\% in F1 for time series, clinical notes, and multimodal fusion, respectively. These results demonstrate the effectiveness of our methods and the importance of considering irregularity in multimodal EHRs.

Kuan Wang, Alexander Bukharin, Haoming Jiang et al.

Instruction fine-tuning (IFT) elicits instruction following capabilities and steers the behavior of large language models (LLMs) via supervised learning. However, existing models trained on open-source IFT datasets only have the ability to follow instructions from users, and often fail to follow complex role and rules specified by developers, a.k.a. system prompts. The ability to follow these roles and rules is essential for deployment, as it ensures that the model safely interacts with users within developer defined guidelines. To improve such role and rule following ability, we propose \model, an automated data generation pipeline that generates diverse roles and rules from existing IFT instructions, along with corresponding responses. This data can then be used to train models that follow complex system prompts. The models are evaluated on our newly created benchmarks for role and rule following ability, as well as standard instruction-following benchmarks and general NLP tasks. Our framework significantly improves role and rule following capability in LLMs, as evidenced by over 25% increase in pass-rate on rule adherence, i.e. following all requirements, in our experiments with the Alpaca and Ultrachat datasets. Moreover, our models achieves this increase without any regression on popular instruction following benchmarks.

Shiyang Li, Jun Yan, Hai Wang et al.

While instruction-tuned models have shown remarkable success in various natural language processing tasks, accurately evaluating their ability to follow instructions remains challenging. Existing benchmarks primarily focus on common instructions that align well with what the model learned during training. However, proficiency in responding to these instructions does not necessarily imply strong ability in instruction following. In this paper, we propose a novel instruction-following evaluation protocol called verbalizer manipulation. It instructs the model to verbalize the task label with words aligning with model priors to different extents, adopting verbalizers from highly aligned (e.g., outputting ``postive'' for positive sentiment), to minimally aligned (e.g., outputting ``negative'' for positive sentiment). Verbalizer manipulation can be seamlessly integrated with any classification benchmark to examine the model's reliance on priors and its ability to override them to accurately follow the instructions. We conduct a comprehensive evaluation of four major model families across nine datasets, employing twelve sets of verbalizers for each of them. We observe that the instruction-following abilities of models, across different families and scales, are significantly distinguished by their performance on less natural verbalizers. Even the strongest GPT-4 model struggles to perform better than random guessing on the most challenging verbalizer, emphasizing the need for continued advancements to improve their instruction-following abilities.

Alexander Bukharin, Shiyang Li, Zhengyang Wang et al.

Recent works have shown that by curating high quality and diverse instruction tuning datasets, we can significantly improve instruction-following capabilities. However, creating such datasets is difficult and most works rely on manual curation or proprietary language models. Automatic data curation is difficult as it is still not clear how we can define diversity for instruction tuning, how diversity and quality depend on one other, and how we can optimize dataset quality and diversity. To resolve these issue, we propose a new algorithm, Quality-Diversity Instruction Tuning (QDIT). QDIT provides a simple method to simultaneously control dataset diversity and quality, allowing us to conduct an in-depth study on the effect of diversity and quality on instruction tuning performance. From this study we draw two key insights (1) there is a natural tradeoff between data diversity and quality and (2) increasing data diversity significantly improves the worst case instruction following performance, therefore improving robustness. We validate the performance of QDIT on several large scale instruction tuning datasets, where we find it can substantially improve worst and average case performance compared to quality-driven data selection.

Kewei Cheng, Jingfeng Yang, Haoming Jiang et al.

Reasoning encompasses two typical types: deductive reasoning and inductive reasoning. Despite extensive research into the reasoning capabilities of Large Language Models (LLMs), most studies have failed to rigorously differentiate between inductive and deductive reasoning, leading to a blending of the two. This raises an essential question: In LLM reasoning, which poses a greater challenge - deductive or inductive reasoning? While the deductive reasoning capabilities of LLMs, (i.e. their capacity to follow instructions in reasoning tasks), have received considerable attention, their abilities in true inductive reasoning remain largely unexplored. To investigate into the true inductive reasoning capabilities of LLMs, we propose a novel framework, SolverLearner. This framework enables LLMs to learn the underlying function (i.e., $y = f_w(x)$), that maps input data points $(x)$ to their corresponding output values $(y)$, using only in-context examples. By focusing on inductive reasoning and separating it from LLM-based deductive reasoning, we can isolate and investigate inductive reasoning of LLMs in its pure form via SolverLearner. Our observations reveal that LLMs demonstrate remarkable inductive reasoning capabilities through SolverLearner, achieving near-perfect performance with ACC of 1 in most cases. Surprisingly, despite their strong inductive reasoning abilities, LLMs tend to relatively lack deductive reasoning capabilities, particularly in tasks involving ``counterfactual'' reasoning.

Yu Wang, Yifan Gao, Xiusi Chen et al.

Existing Large Language Models (LLMs) usually remain static after deployment, which might make it hard to inject new knowledge into the model. We aim to build models containing a considerable portion of self-updatable parameters, enabling the model to integrate new knowledge effectively and efficiently. To this end, we introduce MEMORYLLM, a model that comprises a transformer and a fixed-size memory pool within the latent space of the transformer. MEMORYLLM can self-update with text knowledge and memorize the knowledge injected earlier. Our evaluations demonstrate the ability of MEMORYLLM to effectively incorporate new knowledge, as evidenced by its performance on model editing benchmarks. Meanwhile, the model exhibits long-term information retention capacity, which is validated through our custom-designed evaluations and long-context benchmarks. MEMORYLLM also shows operational integrity without any sign of performance degradation even after nearly a million memory updates. Our code and model are open-sourced at https://github.com/wangyu-ustc/MemoryLLM.

Cheng Jiayang, Xin Liu, Zhihan Zhang et al.

Multi-step tool orchestration, where LLMs must invoke multiple dependent APIs in the correct order while propagating intermediate outputs, remains challenging. State-of-the-art models frequently fail on full sequence execution, with parameter value errors accounting for a significant portion of failures. Training models to handle such workflows faces two obstacles: existing environments focus on simple per-turn function calls with simulated data, and binary rewards provide no signal for partial correctness. We present a framework addressing both challenges. First, we construct a reinforcement learning environment backed by a large-scale cache of real API responses, enabling a data synthesis pipeline that samples valid multi-step orchestration traces with controllable complexity and significantly higher generation efficiency than unconstrained methods. Second, we propose a graduated reward design that decomposes correctness into atomic validity (individual function call correctness at increasing granularity) and orchestration (correct tool sequencing with dependency respect). On ComplexFuncBench, our approach demonstrates substantial improvements in turn accuracy. Ablation studies confirm both reward components are essential: using either alone significantly degrades performance.

Zhihan Zhang, Shiyang Li, Zixuan Zhang et al.

The instruction hierarchy, which establishes a priority order from system messages to user messages, conversation history, and tool outputs, is essential for ensuring consistent and safe behavior in language models (LMs). Despite its importance, this topic receives limited attention, and there is a lack of comprehensive benchmarks for evaluating models' ability to follow the instruction hierarchy. We bridge this gap by introducing IHEval, a novel benchmark comprising 3,538 examples across nine tasks, covering cases where instructions in different priorities either align or conflict. Our evaluation of popular LMs highlights their struggle to recognize instruction priorities. All evaluated models experience a sharp performance decline when facing conflicting instructions, compared to their original instruction-following performance. Moreover, the most competitive open-source model only achieves 48% accuracy in resolving such conflicts. Our results underscore the need for targeted optimization in the future development of LMs.

Chi Han, Xin Liu, Haodong Wang et al.

Despite significant achievements in improving the instruction-following capabilities of large language models (LLMs), the ability to process multiple potentially entangled or conflicting instructions remains a considerable challenge. Real-world scenarios often require consistency across multiple instructions over time, such as secret privacy, personal preferences, and prioritization, which demand sophisticated abilities to integrate multiple turns and carefully balance competing objectives when instructions intersect or conflict. This work presents a systematic investigation of LLMs' capabilities in handling multiple turns of instructions, covering three levels of difficulty: (1) retrieving information from instructions, (2) tracking and reasoning across turns, and (3) resolving conflicts among instructions. We construct MultiTurnInstruct~with $\sim$1.1K high-quality multi-turn conversations through the human-in-the-loop approach and result in nine capability categories, including statics and dynamics, reasoning, and multitasking. Our finding reveals an intriguing trade-off between different capabilities. While GPT models demonstrate superior memorization, they show reduced effectiveness in privacy-protection tasks requiring selective information withholding. Larger models exhibit stronger reasoning capabilities but still struggle with resolving conflicting instructions. Importantly, these performance gaps cannot be attributed solely to information loss, as models demonstrate strong BLEU scores on memorization tasks. Still, their attention mechanisms fail to integrate multiple related instructions effectively. These findings highlight critical areas for improvement in complex real-world tasks involving multi-turn instructions. Data and codes are released at https://github.com/Glaciohound/Multi-Turn-Instruct.

Siyuan Xu, Shiyang Li, Xin Liu et al.

Existing synthetic tool-use corpora are primarily designed for offline supervised fine-tuning, yet reinforcement learning (RL) requires executable environments that support reward-checkable online rollouts. We propose COVERT, a two-stage pipeline that first generates reliable base tool-use trajectories through self-evolving synthesis with multi-level validation, and then applies oracle-preserving augmentations that systematically increase environmental complexity. These augmentations introduce distractor tools, indirect or ambiguous user queries, and noisy, multi-format, or erroneous tool outputs, while strictly preserving oracle tool calls and final answers as ground truth. This design enables automatic reward computation via reference matching for standard cases and lightweight judge-assisted verification for special behaviors such as error detection, supporting RL optimization of tool-calling policies. On Qwen2.5-Instruct-14B, COVERT-RL improves overall accuracy on BFCL v3 from 56.5 to 59.9 and on ACEBench from 53.0 to 59.3, with minimal regressions on general-ability benchmarks; when stacked on SFT, it further reaches 62.1 and 61.8, confirming additive gains. These results suggest that oracle-preserving synthetic environments offer a practical RL refinement stage, complementary to SFT, for improving tool-use robustness under ambiguity and unreliable tool feedback.

Shiyang Li, Zijian Zhang, Winson Chen et al.

Modern machine learning (ML) workloads increasingly rely on GPUs, yet achieving high end-to-end performance remains challenging due to dependencies on both GPU kernel efficiency and host-side settings. Although LLM-based methods show promise on automated GPU kernel generation, prior works mainly focus on single-kernel optimization and do not extend to end-to-end programs, hindering practical deployment. To address the challenge, in this work, we propose StitchCUDA, a multi-agent framework for end-to-end GPU program generation, with three specialized agents: a Planner to orchestrate whole system design, a Coder dedicated to implementing it step-by-step, and a Verifier for correctness check and performance profiling using Nsys/NCU. To fundamentally improve the Coder's ability in end-to-end GPU programming, StitchCUDA integrates rubric-based agentic reinforcement learning over two atomic skills, task-to-code generation and feedback-driven code optimization, with combined rubric reward and rule-based reward from real executions. Therefore, the Coder learns how to implement advanced CUDA programming techniques (e.g., custom kernel fusion, cublas epilogue), and we also effectively prevent Coder's reward hacking (e.g., just copy PyTorch code or hardcoding output) during benchmarking. Experiments on KernelBench show that StitchCUDA achieves nearly 100% success rate on end-to-end GPU programming tasks, with 1.72x better speedup over the multi-agent baseline and 2.73x than the RL model baselines.

Zijian Zhang, Rong Wang, Shiyang Li et al.

Developing efficient CUDA kernels is increasingly critical for AI applications such as large-scale LLM training. However, manual kernel design is both costly and time-consuming, motivating automatic approaches that leverage LLMs for code generation. Existing methods for automatic kernel generation, however, often produce low-efficiency kernels, incur high computational overhead, and fail to generalize across settings. In this work, we propose CudaForge, a training-free multi-agent workflow for CUDA kernel generation and optimization. Our workflow is inspired by the iterative workflow of human experts, which contains steps such as developing initial kernels, testing correctness, analyzing hardware feedback, and iterative improvement. More specifically, CudaForge employs two LLM agents: a Coder and a Judge, that iteratively generate, correct, and optimize CUDA kernels, while integrating hardware feedback such as Nsight Compute (NCU) metrics. In extensive evaluations, we show that CudaForge, by leveraging base models like OpenAI-o3, achieves 97.6\% correctness of generated kernels and an average 1.68$\times$ speedup over PyTorch baselines, substantially surpassing state-of-the-art models including OpenAI-o3 and Kevin on KernelBench.Beyond accuracy and speed, CudaForge demonstrates strong generalization across GPUs (A100, RTX 6000, 4090, 3090) and base models (OpenAI-o3, GPT-5, gpt-oss-120B, Claude-Sonnet-4, QwQ-32B), while maintaining high efficiency. In particular, generating an optimized kernel takes about 26.5 minutes on one RTX6000 and incurs about \$ 0.3 API cost, which is significantly cheaper than existing agentic work that costs 6 H100 hours and \$ 5 API cost per kernel. Our results highlight that multi-agent, training-free workflows can enable cost-effective, generalizable, and high-performance CUDA kernel optimization. Code available at https://github.com/OptimAI-Lab/CudaForge

Ben Mechels, Ryan Billmeyer, Alexander Chen et al.

Modern high-performance computing and Internet-of-Things deployments increasingly generate large volumes of signal data that must be compressed efficiently on resource-constrained acquisition devices and decompressed at scale on centralized servers. Lossy compression is widely adopted to minimize storage and transmission costs on low-power hardware sensors, yet existing methods rarely optimize for both reconstruction quality and decompression throughput simultaneously, nor do they apply methods that generalize across signal domains. In this work, we introduce FPTC, a high-throughput asymmetric signal codec that pairs a lightweight sequential encoder with a massively parallel GPU decoder designed for server-side batch decompression. FPTC applies a windowed discrete cosine transform (DCT) to exploit frequency-domain sparsity, quantizes spectral coefficients with a hybrid three-zone mapping, and entropy codes the result using Huffman coding with a novel packing scheme. The pipeline used in FPTC is designed to be throughput oriented on the GPU, maximizing performance without sacrificing reconstruction quality. We evaluate FPTC on ten datasets spanning four signal domains: biomedical diagnostic, seismic reflections, power-grid production metrics, and meteorological recordings. Our results demonstrate that FPTC outperforms existing frameworks in compression ratio while maintaining competitive throughput, achieving multiplicative compression performance of 3.6x (power), 3.1x (meteorological), 1.5x (biomedical), and 1.2x (seismic) over existing frameworks.

Dongyi He, Shiyang Li, Bin Jiang et al.

High-resolution functional magnetic resonance imaging (fMRI) is essential for mapping human brain activity; however, it remains costly and logistically challenging. If comparable volumes could be generated directly from widely available scalp electroencephalography (EEG), advanced neuroimaging would become significantly more accessible. Existing EEG-to-fMRI generators rely on plain Convolutional Neural Networks (CNNs) that fail to capture cross-channel time-frequency cues or on heavy transformer/Generative Adversarial Network (GAN) decoders that strain memory and stability. To address these limitations, we propose Spec2VolCAMU-Net, a lightweight architecture featuring a Multi-directional Time-Frequency Convolutional Attention Encoder for rich feature extraction and a Vision-Mamba U-Net decoder that uses linear-time state-space blocks for efficient long-range spatial modelling. We frame the goal of this work as establishing a new state of the art in the spatial fidelity of single-volume reconstruction, a foundational prerequisite for the ultimate aim of generating temporally coherent fMRI time series. Trained end-to-end with a hybrid SSI-MSE loss, Spec2VolCAMU-Net achieves state-of-the-art fidelity on three public benchmarks, recording Structural Similarity Index (SSIM) of 0.693 on NODDI, 0.725 on Oddball and 0.788 on CN-EPFL, representing improvements of 14.5%, 14.9%, and 16.9% respectively over previous best SSIM scores. Furthermore, it achieves competitive Signal-to-Noise Ratio (PSNR) scores, particularly excelling on the CN-EPFL dataset with a 4.6% improvement over the previous best PSNR, thus striking a better balance in reconstruction quality. The proposed model is lightweight and efficient, making it suitable for real-time applications in clinical and research settings. The code is available at https://github.com/hdy6438/Spec2VolCAMU-Net.

Xinlu Zhang, Shiyang Li, Xianjun Yang et al.

Large language models (LLMs) demonstrate remarkable medical expertise, but data privacy concerns impede their direct use in healthcare environments. Although offering improved data privacy protection, domain-specific small language models (SLMs) often underperform LLMs, emphasizing the need for methods that reduce this performance gap while alleviating privacy concerns. In this paper, we present a simple yet effective method that harnesses LLMs' medical proficiency to boost SLM performance in medical tasks under privacy-restricted scenarios. Specifically, we mitigate patient privacy issues by extracting keywords from medical data and prompting the LLM to generate a medical knowledge-intensive context by simulating clinicians' thought processes. This context serves as additional input for SLMs, augmenting their decision-making capabilities. Our method significantly enhances performance in both few-shot and full training settings across three medical knowledge-intensive tasks, achieving up to a 22.57% increase in absolute accuracy compared to SLM fine-tuning without context, and sets new state-of-the-art results in two medical tasks within privacy-restricted scenarios. Further out-of-domain testing and experiments in two general domain datasets showcase its generalizability and broad applicability. Our code can be found at https://github.com/XZhang97666/PrivacyBoost-SLM.

Wenhu Chen, Hongmin Wang, Jianshu Chen et al.

The problem of verifying whether a textual hypothesis holds based on the given evidence, also known as fact verification, plays an important role in the study of natural language understanding and semantic representation. However, existing studies are mainly restricted to dealing with unstructured evidence (e.g., natural language sentences and documents, news, etc), while verification under structured evidence, such as tables, graphs, and databases, remains under-explored. This paper specifically aims to study the fact verification given semi-structured data as evidence. To this end, we construct a large-scale dataset called TabFact with 16k Wikipedia tables as the evidence for 118k human-annotated natural language statements, which are labeled as either ENTAILED or REFUTED. TabFact is challenging since it involves both soft linguistic reasoning and hard symbolic reasoning. To address these reasoning challenges, we design two different models: Table-BERT and Latent Program Algorithm (LPA). Table-BERT leverages the state-of-the-art pre-trained language model to encode the linearized tables and statements into continuous vectors for verification. LPA parses statements into programs and executes them against the tables to obtain the returned binary value for verification. Both methods achieve similar accuracy but still lag far behind human performance. We also perform a comprehensive analysis to demonstrate great future opportunities. The data and code of the dataset are provided in \url{https://github.com/wenhuchen/Table-Fact-Checking}.

Shiyang Li, Zijian Zhang, Guangyan Sun et al.

Large language models show promise for automated CUDA programming, however even the strongest coding models (e.g., Claude-Opus-4.6) may still fall short of expert-level, architecture-aware optimization. We introduce CUDAHercules, a benchmark that evaluates generated CUDA against end-to-end human-expert SOTA systems. It spans single kernels, module-level operators, full applications, and unsolved challenge tasks across Ampere, Hopper, and Blackwell GPUs, with end-to-end tasks gated by domain-specific semantic validators. Evaluating models such as Claude-Opus-4.6 and GPT-5.4 shows a large gap between runnable CUDA and expert CUDA engineering: models often compile and pass tests, but rarely recover the optimization strategies needed to match expert performance. Application semantics further reduce success, and iterative or tool-augmented feedback can improve correctness while drifting toward slow fallback implementations. These results show that automated CUDA programming remains far from fully solved and requires stronger hardware reasoning, better tool use, and training objectives that connect code understanding to hardware architecture-grounded intelligence.

Shiyang Li, Haoyang Chen, Mattia Fazzini et al.

Debugging CUDA programs has long been challenging because failures often arise from subtle interactions among hardware behavior, compiler decisions, memory hierarchy, and asynchronous execution. More importantly, with the rapid expansion of GPU usage across scientific computing, machine learning, graphics, and systems workloads, CUDA debugging has become more challenging than ever. Current evaluations of LLM-based CUDA programming largely miss this setting: a model can pass correctness tests with repair by degeneration, simplifying the CUDA code into a safer but slower program that abandons the original optimization structure. We introduce CUDABEAVER, a benchmark for CUDA debugging from real failing workspaces produced during LLM-based CUDA generation. Each task provides the broken candidate, native build/test commands, raw error evidence, and a single editable file. CUDABEAVER evaluates whether a fixer truly repairs the failing CUDA code or merely finds a slower test-passing replacement, reporting results by failure category, debugging trajectory, stagnation mode, and performance preservation. We further propose pass@k(M,C,A), a protocol-conditional CUDA debugging metric by making the fixer M, corpus C, and protocol axes Aexplicit. Using this metric across 213 tasks and seven frontier LLMs, we show that protocol-aware evaluation gives a more faithful view of CUDA debugging ability: when performance-loss tolerance is high, fixers appear much stronger, but even a minor stricter performance requirement can sharply reduce measured success, shifting scores by up to 40 percentage points.

Yuebo Luo, Shiyang Li, Yifei Feng et al.

Graph Neural Networks (GNNs) show strong promise for circuit analysis, but scaling to modern large-scale circuit graphs is limited by GPU memory and training cost, especially for deep models. We revisit deep GNNs for circuit graphs and show that, when trainable, they significantly outperform shallow architectures, motivating an efficient, domain-specific training framework. We propose Grouped-Sparse-Reversible GNN (GSR-GNN), which enables training GNNs with up to hundreds of layers while reducing both compute and memory overhead. GSR-GNN integrates reversible residual modules with a group-wise sparse nonlinear operator that compresses node embeddings without sacrificing task-relevant information, and employs an optimized execution pipeline to eliminate fragmented activation storage and reduce data movement. On sampled circuit graphs, GSR-GNN achieves up to 87.2\% peak memory reduction and over 30$\times$ training speedup with negligible degradation in correlation-based quality metrics, making deep GNNs practical for large-scale EDA workloads.

Yining Lu, Zilong Wang, Shiyang Li et al.

Prior works in multi-objective reinforcement learning typically use linear reward scalarization with fixed weights, which provably fail to capture non-convex Pareto fronts and thus yield suboptimal results. This limitation becomes especially critical in online preference alignment for large language models. Here, stochastic trajectories generated by parameterized policies create highly non-linear and non-convex mappings from parameters to objectives that no single static weighting scheme can find optimal trade-offs. We address this limitation by introducing dynamic reward weighting, which adaptively adjusts reward weights during the online reinforcement learning process. Unlike existing approaches that rely on fixed-weight interpolation, our dynamic weighting continuously balances and prioritizes objectives in training, facilitating effective exploration of Pareto fronts in objective space. We introduce two approaches of increasing sophistication and generalizability: (1) hypervolume-guided weight adaptation and (2) gradient-based weight optimization, offering a versatile toolkit for online multi-objective alignment. Our extensive experiments demonstrate their compatibility with commonly used online reinforcement learning algorithms (including GRPO, REINFORCE, and RLOO), effectiveness across multiple mathematical reasoning datasets, and applicability to different model families, consistently achieving Pareto dominant solutions with fewer training steps than fixed-weight linear scalarization baselines.

Yuebo Luo, Shiyang Li, Junran Tao et al.

The increasing scale and complexity of integrated circuit design have led to increased challenges in Electronic Design Automation (EDA). Graph Neural Networks (GNNs) have emerged as a promising approach to assist EDA design as circuits can be naturally represented as graphs. While GNNs offer a foundation for circuit analysis, they often fail to capture the full complexity of EDA designs. Heterogeneous Graph Neural Networks (HGNNs) can better interpret EDA circuit graphs as they capture both topological relationships and geometric features. However, the improved representation capability comes at the cost of even higher computational complexity and processing cost due to their serial module-wise message-passing scheme, creating a significant performance bottleneck. In this paper, we propose DR-CircuitGNN, a fast GPU kernel design by leveraging row-wise sparsity-aware Dynamic-ReLU and optimizing SpMM kernels during heterogeneous message-passing to accelerate HGNNs training on EDA-related circuit graph datasets. To further enhance performance, we propose a parallel optimization strategy that maximizes CPU-GPU concurrency by concurrently processing independent subgraphs using multi-threaded CPU initialization and GPU kernel execution via multiple cudaStreams. Our experiments show that on three representative CircuitNet designs (small, medium, large), the proposed method can achieve up to 3.51x and 4.09x speedup compared to the SOTA for forward and backward propagation, respectively. On full-size CircuitNet and sampled Mini-CircuitNet, our parallel design enables up to 2.71x speed up over the official DGL implementation cuSPARSE with negligible impact on correlation scores and error rates.

Qingru Zhang, Liang Qiu, Ilgee Hong et al. · gatech

Supervised fine-tuning (SFT) has emerged as a crucial method for aligning large language models (LLMs) with human-annotated demonstrations. However, SFT, being an off-policy approach similar to behavior cloning, often struggles with overfitting and poor out-of-domain generalization, especially in limited-data scenarios. To address these limitations, we propose Self-Rewarding PPO, a novel fine-tuning method that leverages on-policy techniques to enhance generalization performance. Our approach combines the strengths of SFT and proximal policy optimization (PPO) to achieve more effective alignment from demonstration data. At its core is a reward function designed as the log policy ratio between the SFT model and the pretrained base model. This function serves as an implicit reward signal, using the pretrained policy as a baseline and the SFT policy as a target. By doing so, it enables on-policy fine-tuning without relying on human preference annotations. The integration of this self-rewarding mechanism with PPO addresses key limitations of SFT, improving generalization, data efficiency, and robustness. Our empirical evaluation across a range of natural language processing tasks demonstrates that Self-Rewarding PPO consistently outperforms traditional SFT methods. The results highlight the effectiveness of our approach in aligning LLMs using demonstration data, particularly in scenarios where high-quality annotated data is scarce.

Weizhi Wang, Rongmei Lin, Shiyang Li et al. · amazon-science

The Multimodal Large Language Models (MLLMs) are continually pre-trained on a mixture of image-text caption data and interleaved document data, while the high-quality data filtering towards image-text interleaved document data is under-explored. We propose to train an efficient MLLM as a Unified Mulitmodal Data Quality Classifier to Filter both high-quality image-text caption and interleaved data (UniFilter). To address the challenge of collecting diverse labeled multimodal data, we introduce a semi-synthetic approach that leverages readily available raw images and generates corresponding text across four quality levels. This method enables efficient creation of sample-score pairs for both caption and interleaved document data to train UniFilter. We apply UniFilter to curate high-quality caption data from DataComp caption dataset and interleaved data from the OBELICS image-text interleaved dataset. MLLMs pre-trained on the filtered data demonstrate significantly enhanced capabilities compared to those trained on baseline-filtered data, achieving stronger zero-shot reasoning and in-context learning capabilities. After visual supervised fine-tuning, these UniFilter-induced MLLMs achieve stronger performance on various benchmarks, highlighting the downstream benefits of high-quality multimodal pre-training. We release the synthetic training data used for training UniFilter, the UniFilter model checkpoints, and the high-quality interleaved document subset OBELICS-HQ, curated by UniFilter, to the community for reproduction and further development.

Zhenyu Shu, Jiawei Wen, Shiyang Li et al.

The task of 3D shape captioning occupies a significant place within the domain of computer graphics and has garnered considerable interest in recent years. Traditional approaches to this challenge frequently depend on the utilization of costly voxel representations or object detection techniques, yet often fail to deliver satisfactory outcomes. To address the above challenges, in this paper, we introduce Diff-3DCap, which employs a sequence of projected views to represent a 3D object and a continuous diffusion model to facilitate the captioning process. More precisely, our approach utilizes the continuous diffusion model to perturb the embedded captions during the forward phase by introducing Gaussian noise and then predicts the reconstructed annotation during the reverse phase. Embedded within the diffusion framework is a commitment to leveraging a visual embedding obtained from a pre-trained visual-language model, which naturally allows the embedding to serve as a guiding signal, eliminating the need for an additional classifier. Extensive results of our experiments indicate that Diff-3DCap can achieve performance comparable to that of the current state-of-the-art methods.

Xianjun Yang, Liqiang Xiao, Shiyang Li et al.

Current jailbreaking work on large language models (LLMs) aims to elicit unsafe outputs from given prompts. However, it only focuses on single-turn jailbreaking targeting one specific query. On the contrary, the advanced LLMs are designed to handle extremely long contexts and can thus conduct multi-turn conversations. So, we propose exploring multi-turn jailbreaking, in which the jailbroken LLMs are continuously tested on more than the first-turn conversation or a single target query. This is an even more serious threat because 1) it is common for users to continue asking relevant follow-up questions to clarify certain jailbroken details, and 2) it is also possible that the initial round of jailbreaking causes the LLMs to respond to additional irrelevant questions consistently. As the first step (First draft done at June 2024) in exploring multi-turn jailbreaking, we construct a Multi-Turn Jailbreak Benchmark (MTJ-Bench) for benchmarking this setting on a series of open- and closed-source models and provide novel insights into this new safety threat. By revealing this new vulnerability, we aim to call for community efforts to build safer LLMs and pave the way for a more in-depth understanding of jailbreaking LLMs.

Shiyang Li

To address the limitations of medium- and long-term four-dimensional (4D) trajectory prediction models, this paper proposes a hybrid CNN-LSTM-attention-adaboost neural network model incorporating a multi-strategy improved snake-herd optimization (SO) algorithm. The model applies the Adaboost algorithm to divide multiple weak learners, and each submodel utilizes CNN to extract spatial features, LSTM to capture temporal features, and attention mechanism to capture global features comprehensively. The strong learner model, combined with multiple sub-models, then optimizes the hyperparameters of the prediction model through the natural selection behavior pattern simulated by SO. In this study, based on the real ADS-B data from Xi'an to Tianjin, the comparison experiments and ablation studies of multiple optimizers are carried out, and a comprehensive test and evaluation analysis is carried out. The results show that SO-CLA-adaboost outperforms traditional optimizers such as particle swarm, whale, and gray wolf in handling large-scale high-dimensional trajectory data. In addition, introducing the full-strategy collaborative improvement SO algorithm improves the model's prediction accuracy by 39.89%.

Zeyu Ling, Bo Han, Shiyang Li et al.

Large language models (LLMs) are, by design, inherently capable of multi-task learning: through a unified next-token prediction paradigm, they can naturally address a wide variety of downstream tasks. Prior work in the motion domain has demonstrated some generality by adapting LLMs via a Motion Tokenizer coupled with an autoregressive Transformer to generate and understand human motion. However, this generality remains limited in scope and yields only modest performance gains. We introduce VersatileMotion, a unified multimodal motion LLM that combines a novel motion tokenizer, integrating VQ-VAE with flow matching, and an autoregressive transformer backbone to seamlessly support at least nine distinct motion-related tasks. VersatileMotion is the first method to handle single-agent and multi-agent motions in a single framework and enable cross-modal conversion between motion, text, music, and speech, achieving state-of-the-art performance on seven of these tasks. Each sequence in MotionHub may include one or more of the following annotations: natural-language captions, music or audio clips, speech transcripts, and multi-agent interaction data. To facilitate evaluation, we define and release benchmark splits covering nine core tasks. Extensive experiments demonstrate the superior performance, versatility, and potential of VersatileMotion as a foundational model for future understanding and generation of motion.

Lichang Chen, Shiyang Li, Jun Yan et al.

Large language models (LLMs) strengthen instruction-following capability through instruction-finetuning (IFT) on supervised instruction/response data. However, widely used IFT datasets (e.g., Alpaca's 52k data) surprisingly contain many low-quality instances with incorrect or irrelevant responses, which are misleading and detrimental to IFT. In this paper, we propose a simple and effective data selection strategy that automatically identifies and filters out low-quality data using a strong LLM (e.g., ChatGPT). To this end, we introduce AlpaGasus, which is finetuned on only 9k high-quality data filtered from the 52k Alpaca data. AlpaGasus significantly outperforms the original Alpaca as evaluated by GPT-4 on multiple test sets and the controlled human evaluation. Its 13B variant matches $>90\%$ performance of its teacher LLM (i.e., Text-Davinci-003 generating the 52k data) on test tasks. It also provides 5.7x faster training, reducing the training time for a 7B variant from 80 minutes (for Alpaca) to 14 minutes. Moreover, the experiments prove the efficacy of our method across diverse datasets, base models, and LLM filters. Overall, AlpaGasus demonstrates a novel data-centric IFT paradigm that can be generally applied to instruction-tuning data, leading to faster training and better instruction-following models. Our project page is available at: https://lichang-chen.github.io/AlpaGasus/

Shiyang Li, Yifan Gao, Haoming Jiang et al.

Answering complex questions often requires reasoning over knowledge graphs (KGs). State-of-the-art methods often utilize entities in questions to retrieve local subgraphs, which are then fed into KG encoder, e.g. graph neural networks (GNNs), to model their local structures and integrated into language models for question answering. However, this paradigm constrains retrieved knowledge in local subgraphs and discards more diverse triplets buried in KGs that are disconnected but useful for question answering. In this paper, we propose a simple yet effective method to first retrieve the most relevant triplets from KGs and then rerank them, which are then concatenated with questions to be fed into language models. Extensive results on both CommonsenseQA and OpenbookQA datasets show that our method can outperform state-of-the-art up to 4.6% absolute accuracy.

Shiyang Li, Semih Yavuz, Wenhu Chen et al.

Task-adaptive pre-training (TAPT) and Self-training (ST) have emerged as the major semi-supervised approaches to improve natural language understanding (NLU) tasks with massive amount of unlabeled data. However, it's unclear whether they learn similar representations or they can be effectively combined. In this paper, we show that TAPT and ST can be complementary with simple TFS protocol by following TAPT -> Finetuning -> Self-training (TFS) process. Experimental results show that TFS protocol can effectively utilize unlabeled data to achieve strong combined gains consistently across six datasets covering sentiment classification, paraphrase identification, natural language inference, named entity recognition and dialogue slot classification. We investigate various semi-supervised settings and consistently show that gains from TAPT and ST can be strongly additive by following TFS procedure. We hope that TFS could serve as an important semi-supervised baseline for future NLP studies.

Shiyang Li, Semih Yavuz, Kazuma Hashimoto et al.

Dialogue state trackers have made significant progress on benchmark datasets, but their generalization capability to novel and realistic scenarios beyond the held-out conversations is less understood. We propose controllable counterfactuals (CoCo) to bridge this gap and evaluate dialogue state tracking (DST) models on novel scenarios, i.e., would the system successfully tackle the request if the user responded differently but still consistently with the dialogue flow? CoCo leverages turn-level belief states as counterfactual conditionals to produce novel conversation scenarios in two steps: (i) counterfactual goal generation at turn-level by dropping and adding slots followed by replacing slot values, (ii) counterfactual conversation generation that is conditioned on (i) and consistent with the dialogue flow. Evaluating state-of-the-art DST models on MultiWOZ dataset with CoCo-generated counterfactuals results in a significant performance drop of up to 30.8% (from 49.4% to 18.6%) in absolute joint goal accuracy. In comparison, widely used techniques like paraphrasing only affect the accuracy by at most 2%. Human evaluations show that COCO-generated conversations perfectly reflect the underlying user goal with more than 95% accuracy and are as human-like as the original conversations, further strengthening its reliability and promise to be adopted as part of the robustness evaluation of DST models.

Shiyang Li, Jianshu Chen, Dian Yu

Recently, pretrained language models (e.g., BERT) have achieved great success on many downstream natural language understanding tasks and exhibit a certain level of commonsense reasoning ability. However, their performance on commonsense tasks is still far from that of humans. As a preliminary attempt, we propose a simple yet effective method to teach pretrained models with commonsense reasoning by leveraging the structured knowledge in ConceptNet, the largest commonsense knowledge base (KB). Specifically, the structured knowledge in KB allows us to construct various logical forms, and then generate multiple-choice questions requiring commonsense logical reasoning. Experimental results demonstrate that, when refined on these training examples, the pretrained models consistently improve their performance on tasks that require commonsense reasoning, especially in the few-shot learning setting. Besides, we also perform analysis to understand which logical relations are more relevant to commonsense reasoning.

Shiyang Li, Xiaoyong Jin, Yao Xuan et al.

Time series forecasting is an important problem across many domains, including predictions of solar plant energy output, electricity consumption, and traffic jam situation. In this paper, we propose to tackle such forecasting problem with Transformer [1]. Although impressed by its performance in our preliminary study, we found its two major weaknesses: (1) locality-agnostics: the point-wise dot-product self-attention in canonical Transformer architecture is insensitive to local context, which can make the model prone to anomalies in time series; (2) memory bottleneck: space complexity of canonical Transformer grows quadratically with sequence length $L$, making directly modeling long time series infeasible. In order to solve these two issues, we first propose convolutional self-attention by producing queries and keys with causal convolution so that local context can be better incorporated into attention mechanism. Then, we propose LogSparse Transformer with only $O(L(\log L)^{2})$ memory cost, improving forecasting accuracy for time series with fine granularity and strong long-term dependencies under constrained memory budget. Our experiments on both synthetic data and real-world datasets show that it compares favorably to the state-of-the-art.

Tianyi Liu, Shiyang Li, Jianping Shi et al.

Asynchronous momentum stochastic gradient descent algorithms (Async-MSGD) is one of the most popular algorithms in distributed machine learning. However, its convergence properties for these complicated nonconvex problems is still largely unknown, because of the current technical limit. Therefore, in this paper, we propose to analyze the algorithm through a simpler but nontrivial nonconvex problem - streaming PCA, which helps us to understand Aync-MSGD better even for more general problems. Specifically, we establish the asymptotic rate of convergence of Async-MSGD for streaming PCA by diffusion approximation. Our results indicate a fundamental tradeoff between asynchrony and momentum: To ensure convergence and acceleration through asynchrony, we have to reduce the momentum (compared with Sync-MSGD). To the best of our knowledge, this is the first theoretical attempt on understanding Async-MSGD for distributed nonconvex stochastic optimization. Numerical experiments on both streaming PCA and training deep neural networks are provided to support our findings for Async-MSGD.