arXiv subjects

Xinghan Liu

Publications and source records attributed to Xinghan Liu.

12 recordsLinked to original sources

arXiv2025

MobileRL: Online Agentic Reinforcement Learning for Mobile GUI Agents

Building general-purpose graphical user interface (GUI) agents has become increasingly promising with the progress in vision language models. However, developing effective mobile GUI agents with reinforcement learning (RL) remains challenging due to the heavy-tailed distribution of task difficulty and the inefficiency of large-scale environment sampling. We present an online agentic reinforcement learning framework MobileRL to enhance GUI agents in mobile environments. Its core component is the Difficulty-ADAptive GRPO (ADAGRPO) algorithm. In ADAGRPO, we design difficulty-adaptive positive replay and failure curriculum filtering to adapt the model to different task difficulties. We introduce the shortest-path reward adjustment strategy to reshape rewards concerning the task length in multi-turn agentic tasks. Those strategies jointly stabilize RL training, improve sample efficiency, and generate strong performance across diverse mobile apps and tasks. We apply MOBILERL to two open models (Qwen2.5-VL-7B-Instruct and GLM-4.1V-9B-Base). The resultant MOBILERL-9B model achieves state-of-the-art results in terms of success rates on both AndroidWorld (80.2%) and AndroidLab (53.6%). The MOBILERL framework is open-sourced at: https://github.com/THUDM/MobileRL.

cs.LG↗

arXiv2025

Why this and not that? A Logic-based Framework for Contrastive Explanations

We define several canonical problems related to contrastive explanations, each answering a question of the form ''Why P but not Q?''. The problems compute causes for both P and Q, explicitly comparing their differences. We investigate the basic properties of our definitions in the setting of propositional logic. We show, inter alia, that our framework captures a cardinality-minimal version of existing contrastive explanations in the literature. Furthermore, we provide an extensive analysis of the computational complexities of the problems. We also implement the problems for CNF-formulas using answer set programming and present several examples demonstrating how they work in practice.

cs.AI↗

arXiv2025

A Non-Interventionist Approach to Causal Reasoning based on Lewisian Counterfactuals

We present a computationally grounded semantics for counterfactual conditionals in which i) the state in a model is decomposed into two elements: a propositional valuation and a causal base in propositional form that represents the causal information available at the state; and ii) the comparative similarity relation between states is computed from the states' two components. We show that, by means of our semantics, we can elegantly formalize the notion of actual cause without recurring to the primitive notion of intervention. Furthermore, we provide a succinct formulation of the model checking problem for a language of counterfactual conditionals in our semantics. We show that this problem is PSPACE-complete and provide a reduction of it into QBF that can be used for automatic verification of causal properties.

cs.LO↗

arXiv2024

AutoGLM: Autonomous Foundation Agents for GUIs

We present AutoGLM, a new series in the ChatGLM family, designed to serve as foundation agents for autonomous control of digital devices through Graphical User Interfaces (GUIs). While foundation models excel at acquiring human knowledge, they often struggle with decision-making in dynamic real-world environments, limiting their progress toward artificial general intelligence. This limitation underscores the importance of developing foundation agents capable of learning through autonomous environmental interactions by reinforcing existing models. Focusing on Web Browser and Phone as representative GUI scenarios, we have developed AutoGLM as a practical foundation agent system for real-world GUI interactions. Our approach integrates a comprehensive suite of techniques and infrastructures to create deployable agent systems suitable for user delivery. Through this development, we have derived two key insights: First, the design of an appropriate "intermediate interface" for GUI control is crucial, enabling the separation of planning and grounding behaviors, which require distinct optimization for flexibility and accuracy respectively. Second, we have developed a novel progressive training framework that enables self-evolving online curriculum reinforcement learning for AutoGLM. Our evaluations demonstrate AutoGLM's effectiveness across multiple domains. For web browsing, AutoGLM achieves a 55.2% success rate on VAB-WebArena-Lite (improving to 59.1% with a second attempt) and 96.2% on OpenTable evaluation tasks. In Android device control, AutoGLM attains a 36.2% success rate on AndroidLab (VAB-Mobile) and 89.7% on common tasks in popular Chinese APPs.

cs.HC↗

arXiv2024

ChatGLM: A Family of Large Language Models from GLM-130B to GLM-4 All Tools

We introduce ChatGLM, an evolving family of large language models that we have been developing over time. This report primarily focuses on the GLM-4 language series, which includes GLM-4, GLM-4-Air, and GLM-4-9B. They represent our most capable models that are trained with all the insights and lessons gained from the preceding three generations of ChatGLM. To date, the GLM-4 models are pre-trained on ten trillions of tokens mostly in Chinese and English, along with a small set of corpus from 24 languages, and aligned primarily for Chinese and English usage. The high-quality alignment is achieved via a multi-stage post-training process, which involves supervised fine-tuning and learning from human feedback. Evaluations show that GLM-4 1) closely rivals or outperforms GPT-4 in terms of general metrics such as MMLU, GSM8K, MATH, BBH, GPQA, and HumanEval, 2) gets close to GPT-4-Turbo in instruction following as measured by IFEval, 3) matches GPT-4 Turbo (128K) and Claude 3 for long context tasks, and 4) outperforms GPT-4 in Chinese alignments as measured by AlignBench. The GLM-4 All Tools model is further aligned to understand user intent and autonomously decide when and which tool(s) touse -- including web browser, Python interpreter, text-to-image model, and user-defined functions -- to effectively complete complex tasks. In practical applications, it matches and even surpasses GPT-4 All Tools in tasks like accessing online information via web browsing and solving math problems using Python interpreter. Over the course, we have open-sourced a series of models, including ChatGLM-6B (three generations), GLM-4-9B (128K, 1M), GLM-4V-9B, WebGLM, and CodeGeeX, attracting over 10 million downloads on Hugging face in the year 2023 alone. The open models can be accessed through https://github.com/THUDM and https://huggingface.co/THUDM.

cs.CL↗

arXiv2024

ChatGLM-Math: Improving Math Problem-Solving in Large Language Models with a Self-Critique Pipeline

Large language models (LLMs) have shown excellent mastering of human language, but still struggle in real-world applications that require mathematical problem-solving. While many strategies and datasets to enhance LLMs' mathematics are developed, it remains a challenge to simultaneously maintain and improve both language and mathematical capabilities in deployed LLM systems.In this work, we tailor the Self-Critique pipeline, which addresses the challenge in the feedback learning stage of LLM alignment. We first train a general Math-Critique model from the LLM itself to provide feedback signals. Then, we sequentially employ rejective fine-tuning and direct preference optimization over the LLM's own generations for data collection. Based on ChatGLM3-32B, we conduct a series of experiments on both academic and our newly created challenging dataset, MathUserEval. Results show that our pipeline significantly enhances the LLM's mathematical problem-solving while still improving its language ability, outperforming LLMs that could be two times larger. Related techniques have been deployed to ChatGLM\footnote{\url{https://chatglm.cn}}, an online serving LLM. Related evaluation dataset and scripts are released at \url{https://github.com/THUDM/ChatGLM-Math}.

cs.CL↗

arXiv2024

A Logic of "Black Box" Classifier Systems

Binary classifiers are traditionally studied by propositional logic (PL). PL can only represent them as white boxes, under the assumption that the underlying Boolean function is fully known. Binary classifiers used in practical applications and trained by machine learning are however opaque. They are usually described as black boxes. In this paper, we provide a product modal logic called PLC (Product modal Logic for binary input Classifier) in which the notion of "black box" is interpreted as the uncertainty over a set of classifiers. We give results about axiomatics and complexity of satisfiability checking for our logic. Moreover, we present a dynamic extension in which the process of acquiring new information about the actual classifier can be represented.

cs.LO↗

arXiv2023

SemProtector: A Unified Framework for Semantic Protection in Deep Learning-based Semantic Communication Systems

Recently proliferated semantic communications (SC) aim at effectively transmitting the semantics conveyed by the source and accurately interpreting the meaning at the destination. While such a paradigm holds the promise of making wireless communications more intelligent, it also suffers from severe semantic security issues, such as eavesdropping, privacy leaking, and spoofing, due to the open nature of wireless channels and the fragility of neural modules. Previous works focus more on the robustness of SC via offline adversarial training of the whole system, while online semantic protection, a more practical setting in the real world, is still largely under-explored. To this end, we present SemProtector, a unified framework that aims to secure an online SC system with three hot-pluggable semantic protection modules. Specifically, these protection modules are able to encrypt semantics to be transmitted by an encryption method, mitigate privacy risks from wireless channels by a perturbation mechanism, and calibrate distorted semantics at the destination by a semantic signature generation method. Our framework enables an existing online SC system to dynamically assemble the above three pluggable modules to meet customized semantic protection requirements, facilitating the practical deployment in real-world SC systems. Experiments on two public datasets show the effectiveness of our proposed SemProtector, offering some insights of how we reach the goal of secrecy, privacy and integrity of an SC system. Finally, we discuss some future directions for the semantic protection.

cs.CR↗

arXiv2023

A unified logical framework for explanations in classifier systems

Recent years have witnessed a renewed interest in Boolean function in explaining binary classifiers in the field of explainable AI (XAI). The standard approach of Boolean function is propositional logic. We present a modal language of a ceteris paribus nature which supports reasoning about binary input classifiers and their properties. We study a family of classifier models, axiomatize it as two proof systems regarding the cardinality of the language and show completeness of our axiomatics. Moreover, we prove that satisfiability checking problem for our modal language is NEXPTIME-complete in the infinite-variable case, while it becomes polynomial in the finite-variable case. We furthermore identify an interesting NP fragment of our language in the infinite-variable case. We leverage the language to formalize counterfactual conditional as well as a variety of notions of explanation including abductive, contrastive and counterfactual explanations, and biases. Finally, we present two extensions of our language: a dynamic extension by the notion of assignment enabling classifier change and an epistemic extension in which the classifier's uncertainty about the actual input can be represented.

cs.LO↗

arXiv2023

Boosting Physical Layer Black-Box Attacks with Semantic Adversaries in Semantic Communications

End-to-end semantic communication (ESC) system is able to improve communication efficiency by only transmitting the semantics of the input rather than raw bits. Although promising, ESC has also been shown susceptible to the crafted physical layer adversarial perturbations due to the openness of wireless channels and the sensitivity of neural models. Previous works focus more on the physical layer white-box attacks, while the challenging black-box ones, as more practical adversaries in real-world cases, are still largely under-explored. To this end, we present SemBLK, a novel method that can learn to generate destructive physical layer semantic attacks for an ESC system under the black-box setting, where the adversaries are imperceptible to humans. Specifically, 1) we first introduce a surrogate semantic encoder and train its parameters by exploring a limited number of queries to an existing ESC system. 2) Equipped with such a surrogate encoder, we then propose a novel semantic perturbation generation method to learn to boost the physical layer attacks with semantic adversaries. Experiments on two public datasets show the effectiveness of our proposed SemBLK in attacking the ESC system under the black-box setting. Finally, we provide case studies to visually justify the superiority of our physical layer semantic perturbations.

eess.SP↗

arXiv2022

Modelling and Explaining Legal Case-based Reasoners through Classifiers

This paper brings together two lines of research: factor-based models of case-based reasoning (CBR) and the logical specification of classifiers. Logical approaches to classifiers capture the connection between features and outcomes in classifier systems. Factor-based reasoning is a popular approach to reasoning by precedent in AI & Law. Horty (2011) has developed the factor-based models of precedent into a theory of precedential constraint. In this paper we combine the modal logic approach (binary-input classifier, BLC) to classifiers and their explanations given by Liu & Lorini (2021) with Horty's account of factor-based CBR, since both a classifier and CBR map sets of features to decisions or classifications. We reformulate case bases of Horty in the language of BCL, and give several representation results. Furthermore, we show how notions of CBR, e.g. reason, preference between reasons, can be analyzed by notions of classifier system.

cs.AI↗

arXiv2022

CogVideo: Large-scale Pretraining for Text-to-Video Generation via Transformers

Large-scale pretrained transformers have created milestones in text (GPT-3) and text-to-image (DALL-E and CogView) generation. Its application to video generation is still facing many challenges: The potential huge computation cost makes the training from scratch unaffordable; The scarcity and weak relevance of text-video datasets hinder the model understanding complex movement semantics. In this work, we present 9B-parameter transformer CogVideo, trained by inheriting a pretrained text-to-image model, CogView2. We also propose multi-frame-rate hierarchical training strategy to better align text and video clips. As (probably) the first open-source large-scale pretrained text-to-video model, CogVideo outperforms all publicly available models at a large margin in machine and human evaluations.

cs.CV↗