Ziyu Yao

Ziyu Yao, Xuxin Cheng, Yuexian Zou · pku

This paper presents a significant contribution to the field of repetitive action counting through the introduction of a new approach called Pose Saliency Representation. The proposed method efficiently represents each action using only two salient poses instead of redundant frames, which significantly reduces the computational cost while improving the performance. Moreover, we introduce a pose-level method, PoseRAC, which is based on this representation and achieves state-of-the-art performance on two new version datasets by using Pose Saliency Annotation to annotate salient poses for training. Our lightweight model is highly efficient, requiring only 20 minutes for training on a GPU, and infers nearly 10x faster compared to previous methods. In addition, our approach achieves a substantial improvement over the previous state-of-the-art TransRAC, achieving an OBO metric of 0.56 compared to 0.29 of TransRAC. The code and new dataset are available at https://github.com/MiracleDance/PoseRAC for further research and experimentation, making our proposed approach highly accessible to the research community.

Atish Kumar Dipongkor, Ziyu Yao, Kevin Moran

The study of Code Stylometry, and in particular Code Authorship Attribution (CAA), aims to analyze coding styles to identify the authors of code samples. CAA has been illustrated to be an important component of automating software engineering (SE) tasks such as bug triaging, fault localization, and test prioritization. In addition, CAA is also important in cybersecurity and software forensics for addressing copyright disputes and detecting plagiarism. Past techniques for CAA tend to leverage hand-crafted code-related features typically carry limitations that prevent proper authorship characterization and lead to sensitivities to adversarial attacks. Recently, transformer-based Language Models (LMs) have shown remarkable efficacy across a range of SE tasks, and in authorship attribution for natural language in the NLP domain. However, their effectiveness in CAA is not well understood. As such, we conduct the first extensive empirical study applying two larger state-of-the-art code LMs, and five smaller code LMs to the task of CAA on six diverse datasets that encompass 12k code snippets written by 463 developers. Furthermore, we perform an in-depth quantitative and qualitative analysis of our studied models' performance on CAA using established interpretability techniques. Our results illustrate important aspects of the behavior of LMs in understanding stylometric code patterns.

Saurabh Srivastava, Chengyue Huang, Weiguo Fan et al.

Large language models (LLMs) have revolutionized zero-shot task performance, mitigating the need for task-specific annotations while enhancing task generalizability. Despite its advancements, current methods using trigger phrases such as "Let's think step by step" remain limited. This study introduces PRomPTed, an approach that optimizes the zero-shot prompts for individual task instances following an innovative manner of "LLMs in the loop". Our comprehensive evaluation across 13 datasets and 10 task types based on GPT-4 reveals that PRomPTed significantly outperforms both the naive zero-shot approaches and a strong baseline (i.e., "Output Refinement") which refines the task output instead of the input prompt. Our experimental results also confirmed the generalization of this advantage to the relatively weaker GPT-3.5. Even more intriguingly, we found that leveraging GPT-3.5 to rewrite prompts for the stronger GPT-4 not only matches but occasionally exceeds the efficacy of using GPT-4 as the prompt rewriter. Our research thus presents a huge value in not only enhancing zero-shot LLM performance but also potentially enabling supervising LLMs with their weaker counterparts, a capability attracting much interest recently. Finally, our additional experiments confirm the generalization of the advantages to open-source LLMs such as Mistral 7B and Mixtral 8x7B.

Mohamed Aghzal, Erion Plaku, Ziyu Yao

Large language models (LLMs) have achieved remarkable success across a wide spectrum of tasks; however, they still face limitations in scenarios that demand long-term planning and spatial reasoning. To facilitate this line of research, in this work, we propose a new benchmark, termed $\textbf{P}$ath $\textbf{P}$lanning from $\textbf{N}$atural $\textbf{L}$anguage ($\textbf{PPNL}$). Our benchmark evaluates LLMs' spatial-temporal reasoning by formulating ''path planning'' tasks that require an LLM to navigate to target locations while avoiding obstacles and adhering to constraints. Leveraging this benchmark, we systematically investigate LLMs including GPT-4 via different few-shot prompting methodologies as well as BART and T5 of various sizes via fine-tuning. Our experimental results show the promise of few-shot GPT-4 in spatial reasoning, when it is prompted to reason and act interleavedly, although it still fails to perform long-term temporal reasoning. In contrast, while fine-tuned LLMs achieved impressive results on in-distribution reasoning tasks, they struggled to generalize to larger environments or environments with more obstacles.

Murong Yue, Jie Zhao, Min Zhang et al.

Large language models (LLMs) such as GPT-4 have exhibited remarkable performance in a variety of tasks, but this strong performance often comes with the high expense of using paid API services. In this paper, we are motivated to study building an LLM cascade to save the cost of using LLMs, particularly for performing reasoning (e.g., mathematical, causal) tasks. Our cascade pipeline follows the intuition that simpler questions can be addressed by a weaker but more affordable LLM, whereas only the challenging questions necessitate the stronger and more expensive LLM. To realize this decision-making, we consider the "answer consistency" of the weaker LLM as a signal of the question difficulty and propose several methods for the answer sampling and consistency checking, including one leveraging a mixture of two thought representations (i.e., Chain-of-Thought and Program-of-Thought). Through experiments on six reasoning benchmark datasets, with GPT-3.5-turbo and GPT-4 being the weaker and stronger LLMs, respectively, we demonstrate that our proposed LLM cascades can achieve performance comparable to using solely the stronger LLM but require only 40% of its cost.

Daking Rai, Yilun Zhou, Shi Feng et al.

Mechanistic interpretability (MI) is an emerging sub-field of interpretability that seeks to understand a neural network model by reverse-engineering its internal computations. Recently, MI has garnered significant attention for interpreting transformer-based language models (LMs), resulting in many novel insights yet introducing new challenges. However, there has not been work that comprehensively reviews these insights and challenges, particularly as a guide for newcomers to this field. To fill this gap, we provide a comprehensive survey from a task-centric perspective, organizing the taxonomy of MI research around specific research questions or tasks. We outline the fundamental objects of study in MI, along with the techniques, evaluation methods, and key findings for each task in the taxonomy. In particular, we present a task-centric taxonomy as a roadmap for beginners to navigate the field by helping them quickly identify impactful problems in which they are most interested and leverage MI for their benefit. Finally, we discuss the current gaps in the field and suggest potential future directions for MI research.

Yuqing Zhou, Ruixiang Tang, Ziyu Yao et al.

Language models (LMs), despite their advances, often depend on spurious correlations, undermining their accuracy and generalizability. This study addresses the overlooked impact of subtler, more complex shortcuts that compromise model reliability beyond oversimplified shortcuts. We introduce a comprehensive benchmark that categorizes shortcuts into occurrence, style, and concept, aiming to explore the nuanced ways in which these shortcuts influence the performance of LMs. Through extensive experiments across traditional LMs, large language models, and state-of-the-art robust models, our research systematically investigates models' resilience and susceptibilities to sophisticated shortcuts. Our benchmark and code can be found at: https://github.com/yuqing-zhou/shortcut-learning-in-text-classification.

Daking Rai, Yilun Zhou, Bailin Wang et al.

While large language models (LLMs) have demonstrated strong capability in structured prediction tasks such as semantic parsing, few amounts of research have explored the underlying mechanisms of their success. Our work studies different methods for explaining an LLM-based semantic parser and qualitatively discusses the explained model behaviors, hoping to inspire future research toward better understanding them.

Xiang Yue, Ziyu Yao, Huan Sun

Synthesizing QA pairs with a question generator (QG) on the target domain has become a popular approach for domain adaptation of question answering (QA) models. Since synthetic questions are often noisy in practice, existing work adapts scores from a pretrained QA (or QG) model as criteria to select high-quality questions. However, these scores do not directly serve the ultimate goal of improving QA performance on the target domain. In this paper, we introduce a novel idea of training a question value estimator (QVE) that directly estimates the usefulness of synthetic questions for improving the target-domain QA performance. By conducting comprehensive experiments, we show that the synthetic questions selected by QVE can help achieve better target-domain QA performance, in comparison with existing techniques. We additionally show that by using such questions and only around 15% of the human annotations on the target domain, we can achieve comparable performance to the fully-supervised baselines.

Yangzhe Kong, Daeun Song, Jing Liang et al.

We present a novel method, AutoSpatial, an efficient approach with structured spatial grounding to enhance VLMs' spatial reasoning. By combining minimal manual supervision with large-scale Visual Question-Answering (VQA) pairs auto-labeling, our approach tackles the challenge of VLMs' limited spatial understanding in social navigation tasks. By applying a hierarchical two-round VQA strategy during training, AutoSpatial achieves both global and detailed understanding of scenarios, demonstrating more accurate spatial perception, movement prediction, Chain of Thought (CoT) reasoning, final action, and explanation compared to other SOTA approaches. These five components are essential for comprehensive social navigation reasoning. Our approach was evaluated using both expert systems (GPT-4o, Gemini 2.0 Flash, and Claude 3.5 Sonnet) that provided cross-validation scores and human evaluators who assigned relative rankings to compare model performances across four key aspects. Augmented by the enhanced spatial reasoning capabilities, AutoSpatial demonstrates substantial improvements by averaged cross-validation score from expert systems in: perception & prediction (up to 10.71%), reasoning (up to 16.26%), action (up to 20.50%), and explanation (up to 18.73%) compared to baseline models trained only on manually annotated data.

Binfeng Xu, Xukun Liu, Hua Shen et al.

Augmented Language Models (ALMs) empower large language models with the ability to use tools, transforming them into intelligent agents for real-world interactions. However, most existing frameworks for ALMs, to varying degrees, are deficient in the following critical features: flexible customization, collaborative democratization, and holistic evaluation. We present gentopia, an ALM framework enabling flexible customization of agents through simple configurations, seamlessly integrating various language models, task formats, prompting modules, and plugins into a unified paradigm. Furthermore, we establish gentpool, a public platform enabling the registration and sharing of user-customized agents. Agents registered in gentpool are composable such that they can be assembled together for agent collaboration, advancing the democratization of artificial intelligence. To ensure high-quality agents, gentbench, an integral component of gentpool, is designed to thoroughly evaluate user-customized agents across diverse aspects such as safety, robustness, efficiency, etc. We release gentopia on Github and will continuously move forward.

Ziyu Yao, Xuxin Cheng, Zhiqi Huang · pku

Talking head generation is a significant research topic that still faces numerous challenges. Previous works often adopt generative adversarial networks or regression models, which are plagued by generation quality and average facial shape problem. Although diffusion models show impressive generative ability, their exploration in talking head generation remains unsatisfactory. This is because they either solely use the diffusion model to obtain an intermediate representation and then employ another pre-trained renderer, or they overlook the feature decoupling of complex facial details, such as expressions, head poses and appearance textures. Therefore, we propose a Facial Decoupled Diffusion model for Talking head generation called FD2Talk, which fully leverages the advantages of diffusion models and decouples the complex facial details through multi-stages. Specifically, we separate facial details into motion and appearance. In the initial phase, we design the Diffusion Transformer to accurately predict motion coefficients from raw audio. These motions are highly decoupled from appearance, making them easier for the network to learn compared to high-dimensional RGB images. Subsequently, in the second phase, we encode the reference image to capture appearance textures. The predicted facial and head motions and encoded appearance then serve as the conditions for the Diffusion UNet, guiding the frame generation. Benefiting from decoupling facial details and fully leveraging diffusion models, extensive experiments substantiate that our approach excels in enhancing image quality and generating more accurate and diverse results compared to previous state-of-the-art methods.

Ziyu Yao · pku

Federated Learning (FL) is a distributed machine learning paradigm that enables collaboration among multiple clients to train a shared model without sharing raw data. However, a major challenge in FL is the label imbalance, where clients may exclusively possess certain classes while having numerous minority and missing classes. Previous works focus on optimizing local updates or global aggregation but ignore the underlying imbalanced label distribution across clients. In this paper, we propose a novel approach ReGL to address this challenge, whose key idea is to Recover the Global data distribution Locally. Specifically, each client uses generative models to synthesize images that complement the minority and missing classes, thereby alleviating label imbalance. Moreover, we adaptively fine-tune the image generation process using local real data, which makes the synthetic images align more closely with the global distribution. Importantly, both the generation and fine-tuning processes are conducted at the client-side without leaking data privacy. Through comprehensive experiments on various image classification datasets, we demonstrate the remarkable superiority of our approach over existing state-of-the-art works in fundamentally tackling label imbalance in FL.

Yunxiang Peng, Mengmeng Ma, Ziyu Yao et al.

Reliable generalization metrics are fundamental to the evaluation of machine learning models. Especially in high-stakes applications where labeled target data are scarce, evaluation of models' generalization performance under distribution shift is a pressing need. We focus on two practical scenarios: (1) Before deployment, how to select the best model for unlabeled target data? (2) After deployment, how to monitor model performance under distribution shift? The central need in both cases is a reliable and label-free proxy metric. Yet existing proxy metrics, such as model confidence or accuracy-on-the-line, are often unreliable as they only assess model output while ignoring the internal mechanisms that produce them. We address this limitation by introducing a new perspective: using the inner workings of a model, i.e., circuits, as a predictive metric of generalization performance. Leveraging circuit discovery, we extract the causal interactions between internal representations as a circuit, from which we derive two metrics tailored to the two practical scenarios. (1) Before deployment, we introduce Dependency Depth Bias, which measures different models' generalization capability on target data. (2) After deployment, we propose Circuit Shift Score, which predicts a model's generalization under different distribution shifts. Across various tasks, both metrics demonstrate significantly improved correlation with generalization performance, outperforming existing proxies by an average of 13.4\% and 34.1\%, respectively. Our code is available at https://github.com/deep-real/GenCircuit.

Jinhao Pan, Chahat Raj, Ziyu Yao et al.

Large Language Models (LLMs) often exhibit social biases inherited from their training data. While existing benchmarks evaluate bias by term-based mode through direct term associations between demographic terms and bias terms, LLMs have become increasingly adept at avoiding biased responses, leading to seemingly low levels of bias. However, biases persist in subtler, contextually hidden forms that traditional benchmarks fail to capture. We introduce the Description-based Bias Benchmark (DBB), a novel dataset designed to assess bias at the semantic level that bias concepts are hidden within naturalistic, subtly framed contexts in real-world scenarios rather than superficial terms. We analyze six state-of-the-art LLMs, revealing that while models reduce bias in response at the term level, they continue to reinforce biases in nuanced settings. Data, code, and results are available at https://github.com/JP-25/Description-based-Bias-Benchmark.

Mohamed Aghzal, Gregory J. Stein, Ziyu Yao

Large language model (LLM) web agents are increasingly used for web navigation but remain far from human reliability on realistic, long-horizon tasks. Existing evaluations focus primarily on end-to-end success, offering limited insight into where failures arise. We propose a hierarchical planning framework to analyze web agents across three layers (i.e., high-level planning, low-level execution, and replanning), enabling process-based evaluation of reasoning, grounding, and recovery. Our experiments show that structured Planning Domain Definition Language (PDDL) plans produce more concise and goal-directed strategies than natural language (NL) plans, but low-level execution remains the dominant bottleneck. These results indicate that improving perceptual grounding and adaptive control, not only high-level reasoning, is critical for achieving human-level reliability. This hierarchical perspective provides a principled foundation for diagnosing and advancing LLM web agents.

Tianbao Xie, Chen Henry Wu, Peng Shi et al.

Structured knowledge grounding (SKG) leverages structured knowledge to complete user requests, such as semantic parsing over databases and question answering over knowledge bases. Since the inputs and outputs of SKG tasks are heterogeneous, they have been studied separately by different communities, which limits systematic and compatible research on SKG. In this paper, we overcome this limitation by proposing the UnifiedSKG framework, which unifies 21 SKG tasks into a text-to-text format, aiming to promote systematic SKG research, instead of being exclusive to a single task, domain, or dataset. We use UnifiedSKG to benchmark T5 with different sizes and show that T5, with simple modifications when necessary, achieves state-of-the-art performance on almost all of the 21 tasks. We further demonstrate that multi-task prefix-tuning improves the performance on most tasks, largely improving the overall performance. UnifiedSKG also facilitates the investigation of zero-shot and few-shot learning, and we show that T0, GPT-3, and Codex struggle in zero-shot and few-shot learning for SKG. We also use UnifiedSKG to conduct a series of controlled experiments on structured knowledge encoding variants across SKG tasks. UnifiedSKG is easily extensible to more tasks, and it is open-sourced at https://github.com/hkunlp/unifiedskg.

Ziyu Yao, Yiqi Tang, Wen-tau Yih et al.

Despite the widely successful applications, bootstrapping and fine-tuning semantic parsers are still a tedious process with challenges such as costly data annotation and privacy risks. In this paper, we suggest an alternative, human-in-the-loop methodology for learning semantic parsers directly from users. A semantic parser should be introspective of its uncertainties and prompt for user demonstration when uncertain. In doing so it also gets to imitate the user behavior and continue improving itself autonomously with the hope that eventually it may become as good as the user in interpreting their questions. To combat the sparsity of demonstration, we propose a novel annotation-efficient imitation learning algorithm, which iteratively collects new datasets by mixing demonstrated states and confident predictions and re-trains the semantic parser in a Dataset Aggregation fashion (Ross et al., 2011). We provide a theoretical analysis of its cost bound and also empirically demonstrate its promising performance on the text-to-SQL problem. Code will be available at https://github.com/sunlab-osu/MISP.

Daking Rai, Mor Geva, Ziyu Yao

Circuit discovery aims to explain how language models (LMs) implement a specific task by localizing and interpreting a circuit, a computational subgraph responsible for the LM's behavior. Existing circuit discovery methods are hypothesis-driven; they first informally define a task with a dataset, and then apply a circuit discovery algorithm over that dataset to obtain a single circuit. This imposes two strong assumptions: that the LM implements the task with a single circuit, and that the dataset adequately represents the task as humans understand it. We systematically test these assumptions across four previously studied tasks and find that even minor dataset variations that preserve task semantics can produce circuits with low edge overlap and cross-dataset faithfulness. More strikingly, when applied to a mixed dataset with two distinct tasks whose separately discovered circuits have near-zero cross-faithfulness, existing methods still return a single circuit with high faithfulness across both tasks. This indicates that current methods discover dataset-specific circuits, rather than general task circuits. We propose Data-driven Circuit Discovery (DCD), a new discovery framework that drops both assumptions: instead of returning a single circuit for a dataset, DCD first clusters examples in the dataset by how similarly the model processes them and discovers a separate circuit for each group. This allows distinct mechanisms to appear separately rather than merged into a single circuit; each circuit explains its group, not the full task. Experiments show that DCD discovers multiple circuits per dataset, each more faithful to its group than a single circuit discovered by existing methods. Broadly, DCD lets the data reveal mechanistic structure within LMs, rather than relying on human-defined task boundaries that may not align with how models organize their computation.

Murong Yue, Wenhan Lyu, Jennifer Suh et al.

Collaborative problem solving (CPS) is essential in mathematics education, fostering deeper learning through the exchange of ideas. Yet, classrooms often lack the resources, time, and peer dynamics needed to sustain productive CPS. Recent advancements in Large Language Models (LLMs) offer a promising avenue to enhance CPS in mathematical education. We designed and developed MathVC, a multi-persona LLM simulated virtual classroom platform to facilitate CPS in mathematics. MathVC combines a meta planning controller that monitors CPS stages-sense-making, team organization, planning, execution, validation, and predicts the next speaker, with a persona simulation stack that encodes mathematical thinking via a task schema and error-injected persona schemas seeded from teacher-specified misconceptions. We evaluated MathVC with 14 U.S. middle schoolers. Students reported constructive interaction and reaching shared solutions, describing gains in engagement, motivation, and confidence through diverse perspectives, immediate scaffolding, and human-like fallibility. Our findings also provide insights into simulating peers via LLM-based technologies for collaboration to support learning.

Dong Shu, Xuansheng Wu, Haiyan Zhao et al.

Large Language Models (LLMs) have transformed natural language processing, yet their internal mechanisms remain largely opaque. Recently, mechanistic interpretability has attracted significant attention from the research community as a means to understand the inner workings of LLMs. Among various mechanistic interpretability approaches, Sparse Autoencoders (SAEs) have emerged as a promising method due to their ability to disentangle the complex, superimposed features within LLMs into more interpretable components. This paper presents a comprehensive survey of SAEs for interpreting and understanding the internal workings of LLMs. Our major contributions include: (1) exploring the technical framework of SAEs, covering basic architecture, design improvements, and effective training strategies; (2) examining different approaches to explaining SAE features, categorized into input-based and output-based explanation methods; (3) discussing evaluation methods for assessing SAE performance, covering both structural and functional metrics; and (4) investigating real-world applications of SAEs in understanding and manipulating LLM behaviors.

Qineng Wang, Chen Qian, Xiaochang Li et al.

Network traffic refers to the amount of data being sent and received over the internet or any system that connects computers. Analyzing and understanding network traffic is vital for improving network security and management. However, the analysis of network traffic is challenging due to the diverse nature of data packets, which often feature heterogeneous headers and encrypted payloads lacking semantics. To capture the latent semantics of traffic, a few studies have adopted pre-training techniques based on the Transformer encoder or decoder to learn the representations from massive traffic data. However, these methods typically excel in traffic understanding (classification) or traffic generation tasks. To address this issue, we develop Lens, a foundation model for network traffic that leverages the T5 architecture to learn the pre-trained representations from large-scale unlabeled data. Harnessing the strength of the encoder-decoder framework, which captures the global information while preserving the generative ability, our model can better learn the representations from raw data. To further enhance pre-training effectiveness, we design a novel loss that combines three distinct tasks: Masked Span Prediction (MSP), Packet Order Prediction (POP), and Homologous Traffic Prediction (HTP). Evaluation results across various benchmark datasets demonstrate that the proposed Lens outperforms the baselines in most downstream tasks related to both traffic understanding and generation. Notably, it also requires much less labeled data for fine-tuning compared to current methods.

Mohamed Aghzal, Erion Plaku, Gregory J. Stein et al.

The planning ability of Large Language Models (LLMs) has garnered increasing attention in recent years due to their remarkable capacity for multi-step reasoning and their ability to generalize across a wide range of domains. While some researchers emphasize the potential of LLMs to perform complex planning tasks, others highlight significant limitations in their performance, particularly when these models are tasked with handling the intricacies of long-horizon reasoning. In this survey, we critically investigate existing research on the use of LLMs in automated planning, examining both their successes and shortcomings in detail. We illustrate that although LLMs are not well-suited to serve as standalone planners because of these limitations, they nonetheless present an enormous opportunity to enhance planning applications when combined with other approaches. Thus, we advocate for a balanced methodology that leverages the inherent flexibility and generalized knowledge of LLMs alongside the rigor and cost-effectiveness of traditional planning methods.

Zihao Li, Xu Wang, Yuzhe Yang et al.

Large Language Models (LLMs) demonstrate the ability to solve reasoning and mathematical problems using the Chain-of-Thought (CoT) technique. Expanding CoT length, as seen in models such as DeepSeek-R1, significantly enhances this reasoning for complex problems, but requires costly and high-quality long CoT data and fine-tuning. This work, inspired by the deep thinking paradigm of DeepSeek-R1, utilizes a steering technique to enhance the reasoning ability of an LLM without external datasets. Our method first employs Sparse Autoencoders (SAEs) to extract interpretable features from vanilla CoT. These features are then used to steer the LLM's internal states during generation. Recognizing that many LLMs do not have corresponding pre-trained SAEs, we further introduce a novel SAE-free steering algorithm, which directly computes steering directions from the residual activations of an LLM, obviating the need for an explicit SAE. Experimental results demonstrate that both our SAE-based and subsequent SAE-free steering algorithms significantly enhance the reasoning capabilities of LLMs.

Saurabh Srivastava, Sweta Pati, Ziyu Yao

In this work, we study the effect of annotation guidelines -- textual descriptions of event types and arguments, when instruction-tuning large language models for event extraction. We conducted a series of experiments with both human-provided and machine-generated guidelines in both full- and low-data settings. Our results demonstrate the promise of annotation guidelines when there is a decent amount of training data and highlight its effectiveness in improving cross-schema generalization and low-frequency event-type performance.

Saurabh Srivastava, Ziyu Yao

Large Reasoning Models (LRMs) such as DeepSeek-R1 and OpenAI o1 have demonstrated remarkable capabilities in various reasoning tasks. Their strong capability to generate and reason over intermediate thoughts has also led to arguments that they may no longer require extensive prompt engineering or optimization to interpret human instructions and produce accurate outputs. In this work, we aim to systematically study this open question, using the structured task of event extraction for a case study. We experimented with two LRMs (DeepSeek-R1 and o1) and two general-purpose Large Language Models (LLMs) (GPT-4o and GPT-4.5), when they were used as task models or prompt optimizers. Our results show that on tasks as complicated as event extraction, LRMs as task models still benefit from prompt optimization, and that using LRMs as prompt optimizers yields more effective prompts. Our finding also generalizes to tasks beyond event extraction. Finally, we provide an error analysis of common errors made by LRMs and highlight the stability and consistency of LRMs in refining task instructions and event guidelines.

Ziyu Yao, Xuxin Cheng, Zhiqi Huang et al. · pku

Repetitive action counting, which aims to count periodic movements in a video, is valuable for video analysis applications such as fitness monitoring. However, existing methods largely rely on regression networks with limited representational capacity, which hampers their ability to accurately capture variable periodic patterns. Additionally, their supervised learning on narrow, limited training sets leads to overfitting and restricts their ability to generalize across diverse scenarios. To address these challenges, we propose CountLLM, the first large language model (LLM)-based framework that takes video data and periodic text prompts as inputs and outputs the desired counting value. CountLLM leverages the rich clues from explicit textual instructions and the powerful representational capabilities of pre-trained LLMs for repetitive action counting. To effectively guide CountLLM, we develop a periodicity-based structured template for instructions that describes the properties of periodicity and implements a standardized answer format to ensure consistency. Additionally, we propose a progressive multimodal training paradigm to enhance the periodicity-awareness of the LLM. Empirical evaluations on widely recognized benchmarks demonstrate CountLLM's superior performance and generalization, particularly in handling novel and out-of-domain actions that deviate significantly from the training data, offering a promising avenue for repetitive action counting.

Sai Adith Senthil Kumar, Hao Yan, Saipavan Perepa et al.

Large Language Models (LLMs) are now increasingly widely used to simulate personas in virtual environments, leveraging their instruction-following capability. However, we discovered that even state-of-the-art LLMs cannot simulate personas with reversed performance (e.g., student personas with low proficiency in educational settings), which impairs the simulation diversity and limits the practical applications of the simulated environments. In this work, using mathematical reasoning as a representative scenario, we propose the first benchmark dataset for evaluating LLMs on simulating personas with reversed performance, a capability that we dub "counterfactual instruction following". We evaluate both open-weight and closed-source LLMs on this task and find that LLMs, including the OpenAI o1 reasoning model, all struggle to follow counterfactual instructions for simulating reversedly performing personas. Intersectionally simulating both the performance level and the race population of a persona worsens the effect even further. These results highlight the challenges of counterfactual instruction following and the need for further research.

Mohamed Aghzal, Xiang Yue, Erion Plaku et al.

Despite their promise to perform complex reasoning, large language models (LLMs) have been shown to have limited effectiveness in end-to-end planning. This has inspired an intriguing question: if these models cannot plan well, can they still contribute to the planning framework as a helpful plan evaluator? In this work, we generalize this question to consider LLMs augmented with visual understanding, i.e., Vision-Language Models (VLMs). We introduce PathEval, a novel benchmark evaluating VLMs as plan evaluators in complex path-planning scenarios. Succeeding in the benchmark requires a VLM to be able to abstract traits of optimal paths from the scenario description, demonstrate precise low-level perception on each path, and integrate this information to decide the better path. Our analysis of state-of-the-art VLMs reveals that these models face significant challenges on the benchmark. We observe that the VLMs can precisely abstract given scenarios to identify the desired traits and exhibit mixed performance in integrating the provided information. Yet, their vision component presents a critical bottleneck, with models struggling to perceive low-level details about a path. Our experimental results show that this issue cannot be trivially addressed via end-to-end fine-tuning; rather, task-specific discriminative adaptation of these vision encoders is needed for these VLMs to become effective path evaluators.

Siddarth Mamidanna, Daking Rai, Ziyu Yao et al.

Large language models (LLMs) demonstrate proficiency across numerous computational tasks, yet their inner workings remain unclear. In theory, the combination of causal self-attention and multilayer perceptron layers allows every token to access and compute information based on all preceding tokens. In practice, to what extent are such operations present? In this paper, on mental math tasks (i.e., direct math calculation via next-token prediction without explicit reasoning), we investigate this question in three steps: inhibiting input-specific token computations in the initial layers, restricting the routes of information transfer across token positions in the next few layers, and forcing all computation to happen at the last token in the remaining layers. With two proposed techniques, Context-Aware Mean Ablation (CAMA) and Attention-Based Peeking (ABP), we identify an All-for-One subgraph (AF1) with high accuracy on a wide variety of mental math tasks, where meaningful computation occurs very late (in terms of layer depth) and only at the last token, which receives information of other tokens in few specific middle layers. Experiments on a variety of models and arithmetic expressions show that this subgraph is sufficient and necessary for high model performance, transfers across different models, and works on a variety of input styles. Ablations on different CAMA and ABP alternatives reveal their unique advantages over other methods, which may be of independent interest.

Murong Yue, Zhiwei Liu, Liangwei Yang et al.

Large Language Models (LLMs) equipped with external tools have demonstrated enhanced performance on complex reasoning tasks. The widespread adoption of this tool-augmented reasoning is hindered by the scarcity of domain-specific tools. For instance, in domains such as physics question answering, suitable and specialized tools are often missing. Recent work has explored automating tool creation by extracting reusable functions from Chain-of-Thought (CoT) reasoning traces; however, these approaches face a critical scalability bottleneck. As the number of generated tools grows, storing them in an unstructured collection leads to significant retrieval challenges, including an expanding search space and ambiguity between function-related tools. To address this, we propose a systematic approach to automatically refactor an unstructured collection of tools into a structured tool library. Our system first generates discrete, task-specific tools and clusters them into semantically coherent topics. Within each cluster, we introduce a multi-agent framework to consolidate scattered functionalities: a code agent refactors code to extract shared logic and creates versatile, aggregated tools, while a reviewing agent ensures that these aggregated tools maintain the complete functional capabilities of the original set. This process transforms numerous question-specific tools into a smaller set of powerful, aggregated tools without loss of functionality. Experimental results demonstrate that our approach significantly improves tool retrieval accuracy and overall reasoning performance across multiple reasoning tasks. Furthermore, our method shows enhanced scalability compared with baselines as the number of question-specific increases.

Daking Rai, Samuel Miller, Kevin Moran et al.

Despite remarkable advances in coding capabilities, language models (LMs) still struggle with simple syntactic tasks such as generating balanced parentheses. In this study, we investigate the underlying mechanisms behind the persistence of these errors across LMs of varying sizes (124M-7B) to both understand and mitigate the errors. Our study reveals that LMs rely on a number of components (attention heads and FF neurons) that independently make their own predictions. While some components reliably promote correct answers across a generalized range of inputs (i.e., implementing "sound mechanisms''), others are less reliable and introduce noise by promoting incorrect tokens (i.e., implementing "faulty mechanisms''). Errors occur when the faulty mechanisms overshadow the sound ones and dominantly affect the predictions. Motivated by this insight, we introduce RASteer, a steering method to systematically identify and increase the contribution of reliable components for improving model performance. RASteer substantially improves performance on balanced parentheses tasks, boosting accuracy of some models from $0$% to around $100$% without impairing the models' general coding ability. We further demonstrate its broader applicability in arithmetic reasoning tasks, achieving performance gains of up to around $20$%.

Murong Yue, Ziyu Yao

Batch prompting, which combines a batch of multiple queries sharing the same context in one inference, has emerged as a promising solution to reduce inference costs. However, our study reveals a significant security vulnerability in batch prompting: malicious users can inject attack instructions into a batch, leading to unwanted interference across all queries, which can result in the inclusion of harmful content, such as phishing links, or the disruption of logical reasoning. In this paper, we construct BATCHSAFEBENCH, a comprehensive benchmark comprising 150 attack instructions of two types and 8k batch instances, to study the batch prompting vulnerability systematically. Our evaluation of both closed-source and open-weight LLMs demonstrates that all LLMs are susceptible to batch-prompting attacks. We then explore multiple defending approaches. While the prompting-based defense shows limited effectiveness for smaller LLMs, the probing-based approach achieves about 95% accuracy in detecting attacks. Additionally, we perform a mechanistic analysis to understand the attack and identify attention heads that are responsible for it.

Samuel Miller, Daking Rai, Ziyu Yao

Recently, language models (LMs) have shown impressive proficiency in code generation tasks, especially when fine-tuned on code-specific datasets, commonly known as Code LMs. However, our understanding of the internal decision-making processes of Code LMs, such as how they use their (syntactic or semantic) knowledge, remains limited, which could lead to unintended harm as they are increasingly used in real life. This motivates us to conduct one of the first Mechanistic Interpretability works to understand how Code LMs perform a syntactic completion task, specifically the closing parenthesis task, on the CodeLlama-7b model (Roziere et al. 2023). Our findings reveal that the model requires middle-later layers until it can confidently predict the correct label for the closing parenthesis task. Additionally, we identify that while both multi-head attention (MHA) and feed-forward (FF) sub-layers play essential roles, MHA is particularly crucial. Furthermore, we also discover attention heads that keep track of the number of already closed parentheses precisely but may or may not promote a correct number of closing parentheses that are still missing, leading to a positive or negative impact on the model's performance.

Daking Rai, Ziyu Yao

Large language models (LLMs) have shown strong arithmetic reasoning capabilities when prompted with Chain-of-Thought (CoT) prompts. However, we have only a limited understanding of how they are processed by LLMs. To demystify it, prior work has primarily focused on ablating different components in the CoT prompt and empirically observing their resulting LLM performance change. Yet, the reason why these components are important to LLM reasoning is not explored. To fill this gap, in this work, we investigate ``neuron activation'' as a lens to provide a unified explanation to observations made by prior work. Specifically, we look into neurons within the feed-forward layers of LLMs that may have activated their arithmetic reasoning capabilities, using Llama2 as an example. To facilitate this investigation, we also propose an approach based on GPT-4 to automatically identify neurons that imply arithmetic reasoning. Our analyses revealed that the activation of reasoning neurons in the feed-forward layers of an LLM can explain the importance of various components in a CoT prompt, and future research can extend it for a more complete understanding.

Mohamed Aghzal, Erion Plaku, Ziyu Yao

Large Language Models (LLMs) have shown impressive capabilities across a wide variety of tasks. However, they still face challenges with long-horizon planning. To study this, we propose path planning tasks as a platform to evaluate LLMs' ability to navigate long trajectories under geometric constraints. Our proposed benchmark systematically tests path-planning skills in complex settings. Using this, we examined GPT-4's planning abilities using various task representations and prompting approaches. We found that framing prompts as Python code and decomposing long trajectory tasks improve GPT-4's path planning effectiveness. However, while these approaches show some promise toward improving the planning ability of the model, they do not obtain optimal paths and fail at generalizing over extended horizons.

Daking Rai, Bailin Wang, Yilun Zhou et al.

Compositional and domain generalization present significant challenges in semantic parsing, even for state-of-the-art semantic parsers based on pre-trained language models (LMs). In this study, we empirically investigate improving an LM's generalization in semantic parsing with two simple techniques: at the token level, we introduce a token preprocessing method to preserve the semantic boundaries of tokens produced by LM tokenizers; at the sequence level, we propose to use special tokens to mark the boundaries of components aligned between input and output. Our experimental results on two text-to-SQL semantic parsing datasets show that our token preprocessing, although simple, can substantially improve the LM performance on both types of generalization, and our component boundary marking method is particularly helpful for compositional generalization.

Saurabh Srivastava, Gaurav Singh, Shou Matsumoto et al.

In this work, we present the first dataset, MailEx, for performing event extraction from conversational email threads. To this end, we first proposed a new taxonomy covering 10 event types and 76 arguments in the email domain. Our final dataset includes 1.5K email threads and ~4K emails, which are annotated with totally ~8K event instances. To understand the task challenges, we conducted a series of experiments comparing three types of approaches, i.e., fine-tuned sequence labeling, fine-tuned generative extraction, and few-shot in-context learning. Our results showed that the task of email event extraction is far from being addressed, due to challenges lying in, e.g., extracting non-continuous, shared trigger spans, extracting non-named entity arguments, and modeling the email conversational history. Our work thus suggests more future investigations in this domain-specific event extraction task.

Hao Yan, Saurabh Srivastava, Yintao Tai et al.

Interactive semantic parsing based on natural language (NL) feedback, where users provide feedback to correct the parser mistakes, has emerged as a more practical scenario than the traditional one-shot semantic parsing. However, prior work has heavily relied on human-annotated feedback data to train the interactive semantic parser, which is prohibitively expensive and not scalable. In this work, we propose a new task of simulating NL feedback for interactive semantic parsing. We accompany the task with a novel feedback evaluator. The evaluator is specifically designed to assess the quality of the simulated feedback, based on which we decide the best feedback simulator from our proposed variants. On a text-to-SQL dataset, we show that our feedback simulator can generate high-quality NL feedback to boost the error correction ability of a specific parser. In low-data settings, our feedback simulator can help achieve comparable error correction performance as trained using the costly, full set of human annotations.

Ziyu Yao, Frank F. Xu, Pengcheng Yin et al.

While most neural generative models generate outputs in a single pass, the human creative process is usually one of iterative building and refinement. Recent work has proposed models of editing processes, but these mostly focus on editing sequential data and/or only model a single editing pass. In this paper, we present a generic model for incremental editing of structured data (i.e., "structural edits"). Particularly, we focus on tree-structured data, taking abstract syntax trees of computer programs as our canonical example. Our editor learns to iteratively generate tree edits (e.g., deleting or adding a subtree) and applies them to the partially edited data, thereby the entire editing process can be formulated as consecutive, incremental tree transformations. To show the unique benefits of modeling tree edits directly, we further propose a novel edit encoder for learning to represent edits, as well as an imitation learning method that allows the editor to be more robust. We evaluate our proposed editor on two source code edit datasets, where results show that, with the proposed edit encoder, our editor significantly improves accuracy over previous approaches that generate the edited program directly in one pass. Finally, we demonstrate that training our editor to imitate experts and correct its mistakes dynamically can further improve its performance.

Xiang Yue, Xinliang Frederick Zhang, Ziyu Yao et al.

Clinical question answering (QA) aims to automatically answer questions from medical professionals based on clinical texts. Studies show that neural QA models trained on one corpus may not generalize well to new clinical texts from a different institute or a different patient group, where large-scale QA pairs are not readily available for model retraining. To address this challenge, we propose a simple yet effective framework, CliniQG4QA, which leverages question generation (QG) to synthesize QA pairs on new clinical contexts and boosts QA models without requiring manual annotations. In order to generate diverse types of questions that are essential for training QA models, we further introduce a seq2seq-based question phrase prediction (QPP) module that can be used together with most existing QG models to diversify the generation. Our comprehensive experiment results show that the QA corpus generated by our framework can improve QA models on the new contexts (up to 8% absolute gain in terms of Exact Match), and that the QPP module plays a crucial role in achieving the gain.

Ziyu Yao, Yu Su, Huan Sun et al.

As a promising paradigm, interactive semantic parsing has shown to improve both semantic parsing accuracy and user confidence in the results. In this paper, we propose a new, unified formulation of the interactive semantic parsing problem, where the goal is to design a model-based intelligent agent. The agent maintains its own state as the current predicted semantic parse, decides whether and where human intervention is needed, and generates a clarification question in natural language. A key part of the agent is a world model: it takes a percept (either an initial question or subsequent feedback from the user) and transitions to a new state. We then propose a simple yet remarkably effective instantiation of our framework, demonstrated on two text-to-SQL datasets (WikiSQL and Spider) with different state-of-the-art base semantic parsers. Compared to an existing interactive semantic parsing approach that treats the base parser as a black box, our approach solicits less user feedback but yields higher run-time accuracy.

Boyuan Pan, Hao Li, Ziyu Yao et al.

This paper investigates a new task named Conversational Question Generation (CQG) which is to generate a question based on a passage and a conversation history (i.e., previous turns of question-answer pairs). CQG is a crucial task for developing intelligent agents that can drive question-answering style conversations or test user understanding of a given passage. Towards that end, we propose a new approach named Reinforced Dynamic Reasoning (ReDR) network, which is based on the general encoder-decoder framework but incorporates a reasoning procedure in a dynamic manner to better understand what has been asked and what to ask next about the passage. To encourage producing meaningful questions, we leverage a popular question answering (QA) model to provide feedback and fine-tune the question generator using a reinforcement learning mechanism. Empirical results on the recently released CoQA dataset demonstrate the effectiveness of our method in comparison with various baselines and model variants. Moreover, to show the applicability of our method, we also apply it to create multi-turn question-answering conversations for passages in SQuAD.

Ziyu Yao, Jayavardhan Reddy Peddamail, Huan Sun

To accelerate software development, much research has been performed to help people understand and reuse the huge amount of available code resources. Two important tasks have been widely studied: code retrieval, which aims to retrieve code snippets relevant to a given natural language query from a code base, and code annotation, where the goal is to annotate a code snippet with a natural language description. Despite their advancement in recent years, the two tasks are mostly explored separately. In this work, we investigate a novel perspective of Code annotation for Code retrieval (hence called `CoaCor'), where a code annotation model is trained to generate a natural language annotation that can represent the semantic meaning of a given code snippet and can be leveraged by a code retrieval model to better distinguish relevant code snippets from others. To this end, we propose an effective framework based on reinforcement learning, which explicitly encourages the code annotation model to generate annotations that can be used for the retrieval task. Through extensive experiments, we show that code annotations generated by our framework are much more detailed and more useful for code retrieval, and they can further improve the performance of existing code retrieval models significantly.

Ziyu Yao, Xiujun Li, Jianfeng Gao et al.

Given a text description, most existing semantic parsers synthesize a program in one shot. However, it is quite challenging to produce a correct program solely based on the description, which in reality is often ambiguous or incomplete. In this paper, we investigate interactive semantic parsing, where the agent can ask the user clarification questions to resolve ambiguities via a multi-turn dialogue, on an important type of programs called "If-Then recipes." We develop a hierarchical reinforcement learning (HRL) based agent that significantly improves the parsing performance with minimal questions to the user. Results under both simulation and human evaluation show that our agent substantially outperforms non-interactive semantic parsers and rule-based agents.

Ziyu Yao, Daniel S. Weld, Wei-Peng Chen et al.

Stack Overflow (SO) has been a great source of natural language questions and their code solutions (i.e., question-code pairs), which are critical for many tasks including code retrieval and annotation. In most existing research, question-code pairs were collected heuristically and tend to have low quality. In this paper, we investigate a new problem of systematically mining question-code pairs from Stack Overflow (in contrast to heuristically collecting them). It is formulated as predicting whether or not a code snippet is a standalone solution to a question. We propose a novel Bi-View Hierarchical Neural Network which can capture both the programming content and the textual context of a code snippet (i.e., two views) to make a prediction. On two manually annotated datasets in Python and SQL domain, our framework substantially outperforms heuristic methods with at least 15% higher F1 and accuracy. Furthermore, we present StaQC (Stack Overflow Question-Code pairs), the largest dataset to date of ~148K Python and ~120K SQL question-code pairs, automatically mined from SO using our framework. Under various case studies, we demonstrate that StaQC can greatly help develop data-hungry models for associating natural language with programming language.