Xiang Cheng, Yulan Hu, Xiangwen Zhang, Lu Xu, Lide Tan, Zheng Pan, Xin Li, Yong Liu. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Zhicong Li, Lingjie Jiang, Yulan Hu, Xingchen Zeng, Yixia Li, Xiangwen Zhang, Guanhua Chen, Zheng Pan, Xin Li, Yong Liu. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

In classical Reinforcement Learning from Human Feedback (RLHF), Reward Models (RMs) serve as the fundamental signal provider for model alignment.As Large Language Models evolve into agentic systems capable of autonomous tool invocation and complex reasoning, reward modeling faces a key challenge: the lack of benchmarks specifically designed to assess RM capabilities in toolintegrated environments.To address this gap, we present Plan-RewardBench, a trajectorylevel preference benchmark for evaluating how well judges distinguish preferred versus distractor agent trajectories in complex tool-using scenarios.Plan-RewardBench covers four representative task families-(i) Safety Refusal, (ii) Tool-Irrelevance / Unavailability, (iii) Complex Planning, and (iv) Robust Error Recoverywith validated positive trajectories and confusable hard negatives constructed via multimodel natural rollouts, rule-based perturbations, and minimal-edit LLM perturbations.We benchmark representative RMs (generative, discriminative, and LLM-as-Judge) under a unified pairwise protocol, reporting accuracy trends across trajectory lengths and task categories, and providing diagnostic analyses of prevalent failure modes.Results show that all three evaluator families face substantial challenges, with performance degrading sharply on long-horizon trajectories, underscoring the need for specialized training in agentic, trajectory-level reward modeling.Ultimately, Plan-RewardBench serves as both a practical evaluation suite and a reusable blueprint for constructing agentic planning preference data.

Rewards serve as proxies for human preferences and play a crucial role in Reinforcement Learning from Human Feedback (RLHF).However, if these rewards are inherently imperfect, exhibiting various biases, they can adversely affect the alignment of large language models (LLMs).In this paper, we collectively define the various biases present in rewards as the problem of reward unfairness.We propose a bias-agnostic method to address the issue of reward fairness from a resource allocation perspective, without specifically designing for each type of bias, yet effectively mitigating them.Specifically, we model preference learning as a resource allocation problem, treating rewards as resources to be allocated while considering the trade-off between utility and fairness in their distribution.We propose two methods, Fairness Regularization and Fairness Coefficient, to achieve fairness in rewards.We apply our methods in both verification and reinforcement learning scenarios to obtain a fairness reward model and a policy model, respectively.Experiments conducted in these scenarios demonstrate that our approach aligns LLMs with human preferences in a more fair manner.Our data and code are available at https://github.com/ shoyua/Towards-Reward-Fairness.

In this work, we propose FoRA-UA, a novel method that, using only 1-5% of the standard LoRA's parameters, achieves state-ofthe-art performance across a wide range of tasks.Specifically, we explore scenarios with extremely limited parameter budgets and derive two key insights: (1) fix-sized sparse frequency representations approximate small matrices more accurately; and (2) with a fixed number of trainable parameters, introducing a smaller intermediate representation to approximate larger matrices results in lower construction error.These findings form the foundation of our FoRA-UA method.By inserting a small intermediate parameter set, we achieve greater model compression without sacrificing performance.We evaluate FoRA-UA across diverse tasks, including natural language understanding (NLU), natural language generation (NLG), instruction tuning, and image classification, demonstrating strong generalisation and robustness under extreme compression. 1

In recent years, the field of heterogeneous graph learning has garnered significant interest. Various efforts have been made towards learning heterogeneous graph representations, such as designing meta-paths to mine implicit graph knowledge or directly applying Graph Neural Networks (GNNs) for graph representation. However, these methods fail to fully capture available graph knowledge while ensuring scalability across diverse graph settings. In this paper, we address these challenges by introducing IEGraph, a heterogeneous Graph learning approach that capitalizes on both implicit and explicit graph knowledge. This encompasses two training stages: the implicit label-free stage and the explicit label-based stage, fostering comprehensive utilization of graph information. The label-free stage extracts implicit graph knowledge by constructing local and global training samples for contrastive pre-training, while the label-based stage further employs explicit labeled data to fine-tune the model. We carry out experiments on diverse heterogeneous graphs, and the results show that IEGraph achieves commendable performance compared to other state-of-the-art baselines.

Generative graph self-supervised learning (SSL), represented by masked graph autoencoders (GAEs), has shown great potential in graph representation learning. Existing masked GAEs typically rely on reconstruction criteria, such as mean squared error, to measure the discrepancy between the input graph and the reconstructed output. However, this learning paradigm struggles with perturbed graph characteristics, hindering the learning of robust graph representations. To address this, we introduce AMGAE -- an Adversarial Masked Graph AutoEncoder, which enhances the robustness of masked GAEs by integrating an adversarial learning strategy. Specifically, we design AMGAE to comprise a generator and a discriminator, optimized alternately and interconnected by a binary discrimination task (BDT). We treat the entire masked GAE as the generator, which produces a reconstructed output using the visible graph features. Then, we synthesize the reconstructed output by substituting the visible node features with the corresponding raw input features. Finally, we employ an additional GNN layer as the discriminator to determine the authenticity of the node-level features synthesized by BDT. By introducing the adversarial strategy, AMGAE reformulates masked GAE learning into a min-max game, which facilitates the learning of robust graph representations. We conduct extensive experiments on three graph tasks, demonstrating that AMGAE performs favorably against diverse baselines.

In the existing spectral GNNs, polynomial-based methods occupy the mainstream in designing a filter through the Laplacian matrix. However, polynomial combinations factored by the Laplacian matrix naturally have limitations in message passing (e.g., over-smoothing). Furthermore, most existing spectral GNNs are based on polynomial bases, which struggle to capture the high-frequency parts of the graph spectral signal. Additionally, we also find that even increasing the polynomial order does not change this situation, which means polynomial-based models have a natural deficiency when facing high-frequency signals. To tackle these problems, we propose WaveNet, which aims to effectively capture the high-frequency part of the graph spectral signal from the perspective of wavelet bases through reconstructing the message propagation matrix. We utilize Multi-Resolution Analysis (MRA) to model this question, and our proposed method can reconstruct arbitrary filters theoretically. We also conduct node classification experiments on real-world graph benchmarks and achieve superior performance on most datasets. Our code is available at https://github.com/Bufordyang/WaveNet

Generative self-supervised learning, represented by graph autoencoders (GAEs), has begun to exhibit significant potential in addressing graph tasks. However, GAEs often rely on Graph Neural Networks (GNNs) for encoding and decoding, this can pose a computation challenge due to the inherent complexities of the aggregation mechanism in GNNs. Furthermore, the bipartite structure of GAEs introduces additional computational burdens. In contrast, Multi-Layer Perceptrons (MLPs) have no graph dependency and can train much faster than GNNs. Motivated by this, in this work, we introduce a simple yet effective alternative: the GNN-Free Masked AutoEncoder (GFMAE), which employs MLPs rather than GNNs to serve as the backbone model to speed up training. Additionally, we devise comprehensive decoding strategies to compensate for the inability of MLPs in characterizing the graph. Our comprehensive experiments conducted on eight datasets demonstrate that GFMAE achieves performance comparable to GNNs while also enhancing the training efficiency of generative models with GNNs as the backbone.

Recent progress has propelled the development of realistic talking-face videos for avatars. Yet, animating 3D cartoon avatars remains intricate due to the imprecise nature of facial-driven data. This often manifests as inconsistent mouth configurations and rigid facial expressions, curbing the animation’s realism. Addressing these issues, we introduce a conformer-based framework that derives expression coefficients directly from phonemes, thereby elevating prediction precision and minimizing manual oversight. Furthermore, by harnessing a pre-trained emotion blending module coupled with the keyframe of the target emotional character, we employ a zero-shot adaptation technique. This serves to amplify emotional expressions and bolster the authenticity of lip dynamics. Our methodology adeptly registers nuanced expression shifts in avatars, leading to remarkably lifelike animations, as substantiated by our experimental findings.

Abstract The problem of finding the maximum independent sets (or maximum cliques) of a given graph is fundamental in graph theory and is also one of the most important in terms of the application of graph theory. Let $A(n,d,w)$ be the size of the maximum independent set of $Q_{n}^{(d-1,w)}$, which is the induced subgraph of points of weight $w$ of the $d-1^{th}$-power of $n$-dimensional hypercubes. In order to further understand and study the dependent set of $Q_{n}^{(d-1,w)}$, we explore its clique number and the structure of the maximum clique. This paper obtains the clique number and the structure of the maximum clique of $Q_{n}^{(d-1,w)}$ for $5\leq d\leq 6$. Moreover, the characterizations for $A(n,d,w)=2$ and $3$ are also given.