Developing a universal graph model capable of generalizing across diverse graph domains has consistently been a key objective in graph learning. Recently, many studies have focused on achieving in-context learning (ICL) on graphs, which can generalize to novel tasks without the need for fine-tuning, similar to large language models (LLMs) such as GPT-3. These researches can be primarily divided into graph-based methods and LLM-based methods. However, the generalization performance of the former is limited by the representation capability of GNNs, while the latter faces the challenge of LLMs understanding graph structures. Therefore, we propose CAGML, a context-aware graph meta-learning model, which learns to generalize to cross-domain and cross-granularity graph tasks using a meta-trained Transformer. Firstly, we formulate graph few-shot learning tasks as a structure-aware sequence modeling problem to unify cross-domain and cross-granularity tasks. Then, a structure-aware Transformer (SAT) is introduced as a graph in-context learner to make predictions with a few labels and the task-specific structural context. Finally, we pre-train SAT in a meta-optimization manner on large-scale citation network and knowledge graph. Experiments on 6 cross-domain graph datasets show that, without fine-tuning, CAGML can achieve state-of-the-art (SOTA) performance in terms of average performance across cross-granularity tasks on adopted datasets.

While Knowledge Graphs (KGs) can ground Large Language Models (LLMs) in factual knowledge, existing LLM-KG integration methods for complex reasoning are plagued by computational inefficiency and semantic inconsistency. The tight coupling of LLM inference and KG traversal leads to prohibitive costs, while spurious reasoning paths often misguide learning-based agents, causing reward hacking. To this end, we propose EAT (Exploration-and-Thinking), a novel agentic framework that synergizes LLMs with Reinforcement Learning (RL) for efficient and faithful reasoning on KGs. EAT's core innovations are twofold: (1) an adaptive retrieval-augmented generation mechanism that decouples language comprehension from structured exploration, dramatically improving efficiency; and (2) an LLM-guided reward shaping strategy that explicitly penalizes semantically inconsistent paths and promotes logically valid trajectories grounded in the KG. Extensive experiments on benchmarks like WebQSP, CWQ, and GrailQA show that EAT achieves state-of-the-art performance. Notably, it surpasses the reasoning capability of GPT-4o while utilizing a significantly smaller 8B-parameter LLM.

The goal of stance detection is to automatically recognize the author's expressed attitude in text towards a given target. However, social media users often express themselves briefly and implicitly, which leads to a significant number of comments lacking explicit reference information to the target, posing a challenge for stance detection. To address the missing relationship between text and target, existing studies primarily focus on incorporating external knowledge, which inevitably introduces noise information. In contrast to their work, we are dedicated to mining implicit relational information within data. Typically, users tend to emphasize their attitudes towards a relevant topic or aspect of the target while concealing others when expressing opinions. Motivated by this phenomenon, we suggest that the potential correlation between text and target can be learned from instances with similar topics. Therefore, we design a pretext task to mine the topic associations between samples and model this topic association as a dynamic weight introduced into contrastive learning. In this way, we can selectively cluster samples that have similar topics and consistent stances, while enlarging the gap between samples with different stances in the feature space. Additionally, we propose a nearest-neighbor prediction mechanism for stance classification to better utilize the features we constructed. Our experiments on two datasets demonstrate the advanced and generalization ability of our method, yielding the state-of-the-art results.

As the intelligent software, the development and application of large language models are extremely hot topics recently, bringing tremendous changes to general AI and software industry. Nonetheless, large language models, especially open source ones, incontrollably suffer from some potential software quality issues such as instability, inaccuracy, and insecurity, making software testing necessary. In this paper, we propose the first solution for functional testing of open large language models to check full-scene availability and conclude empirical principles for better steering large language models, particularly considering their black box and intelligence properties. Specifically, we focus on the model’s causal reasoning ability, which is the core of artificial intelligence but almost ignored by most previous work. First, for comprehensive evaluation, we deconstruct the causal reasoning capability into five dimensions and summary the forms of causal reasoning task as causality identification and causality matching. Then, rich datasets are introduced and further modified to generate test cases along with different ability dimensions and task forms to improve the testing integrity. Moreover, we explore the ability boundary of open large language models in two usage modes: prompting and lightweight fine-tuning. Our work conducts comprehensive functional testing on the causal reasoning ability of open large language models, establishes benchmarks, and derives empirical insights for practical usage. The proposed testing solution can be transferred to other similar evaluation tasks as a general framework for large language models or their derivations.