The widespread adoption of reinforcement learning-based alignment highlights the growing importance of reward models. Various benchmarks have been built to evaluate reward models in various domains and scenarios. However, a significant gap remains in assessing reward models for long-form generation, despite its critical role in real-world applications. To bridge this, we introduce Long-form RewardBench, the first reward modeling testbed specifically designed for long-form generation. Our benchmark encompasses five key subtasks: QA, RAG, Chat, Writing, and Reasoning. We collected instruction and preference data through a meticulously designed multi-stage data collection process, and conducted extensive experiments on 20+ mainstream reward models, including both classifiers and generative models. Our findings reveal that current models still lack long-form reward modeling capabilities. Furthermore, we designed a novel Long-form Needle-in-a-Haystack Test, which revealed a correlation between reward modeling performance and the error's position within a response, as well as the overall response length, with distinct characteristics observed between classification and generative models. Finally, we demonstrate that classifier exhibit better generalizability compared to generative models trained on the same data. As the first benchmark for long-form reward modeling, this work aims to offer a robust platform for visualizing progress in this crucial area.
论文检索
输入标题、作者或关键词，从 7,876 篇学术成果中精准定位
The emergence of large language models (LLMs) marks a transformative era in artificial intelligence~(AI). However, systematically evaluating the capability of LLMs is challenging due to the necessity of a large number of labeled test data. To tackle this problem, in the conventional AI field, AutoEval has been proposed to estimate the capability of AI models without data labeling effort. Unfortunately, even though multiple AutoEval methods have been proposed, most are constructed for classification tasks and evaluated only on image datasets. As a result, their effectiveness for LLMs is unclear, as LLMs often target generation tasks. In this work, we introduce the first AutoEval benchmark specifically designed to estimate the capability of LLMs using unlabeled test data, AEBench. Besides existing AutoEval methods, AEBench also supports our designed method, which utilizes the correlation between data uncertainty and model ability for the capability estimation. In total, AEBench covers 12 AutoEval methods and 120 method combinations. Based on AEBench, we conducted a comprehensive study to explore the usefulness of AutoEval on LLMs. Experimental results on 10 datasets demonstrated that our designed uncertainty features-based methods perform the best in achieving the lowest estimation errors.
Deploying language models often requires navigating accuracy vs. performance trade-offs to meet latency constraints while preserving utility. Traditional model distillation reduces size but incurs substantial costs through training separate models. We introduce ModularStarEncoder (MoSE), a 1-billion-parameter multi-exit encoder for code retrieval and classification that employs a novel Self-Distillation mechanism. This approach significantly enhances lower-layer representations, enabling flexible deployment of different model portions with favorable performance trade-offs. Our architecture improves text-to-code and code-to-code search by targeting specific encoder layers as exit heads, where higher layers guide earlier ones during training, thereby improving intermediate representations at minimal additional cost. We further enhance MoSE with a repository-level contextual loss that maximizes training context window utilization. Additionally, we release a new dataset created through code translation that extends text-to code benchmarks with cross-language code-to-code pairs. Evaluations demonstrate the effectiveness of Self-Distillation as a principled approach to trading inference cost for accuracy across various code understanding tasks.
Graph Neural Networks (GNNs) offer superior modeling capabilities for text classification by capturing complex spatial features within semantic representations. However, existing graph-based approaches often suffer from computational inefficiency and limited ability to model both fine-grained local structures and the sequential nature of text. To address these challenges, we propose HC2-GNN, a Hierarchical Clustering and Coarsening Graph Neural Network, which introduces a novel lightweight graph clustering algorithm called Compromise Conductance Graph Clustering (C2GC). C2GC enables efficient graph clustering while simultaneously preserving both the textual order and the topological coherence of subgraphs. Furthermore, it incorporates a virtue cluster mechanism that expands each subgraph with semantically relevant neighbors, explicitly enabling cross-cluster information propagation without compromising local structural integrity. HC2-GNN aggregates local and global features by combining subgraph-level and full-graph representations, enhancing semantic discriminability for classification. Extensive experiments on benchmark datasets demonstrate that HC2-GNN consistently outperforms existing state-of-the-art text classification methods.
Breaking the Adversarial Robustness-Performance Trade-off in Text Classification via Manifold Purification
PDF ↗A persistent challenge in text classification (TC) is that enhancing model robustness against adversarial attacks typically degrades performance on clean data. We argue that this challenge can be resolved by modeling the distribution of clean samples in the encoder’s embedding manifold. To this end, we propose the Manifold-Correcting Causal Flow (MC²F), a two-module system that operates directly on sentence embeddings. A Stratified Riemannian Continuous Normalizing Flow (SR-CNF) learns the density of the clean data manifold. It identifies out-of-distribution embeddings, which are then corrected by a Geodesic Purification Solver. This solver projects adversarial points back onto the learned manifold via the shortest path, restoring a clean, semantically coherent representation. We conducted extensive evaluations on text classification (TC) across three datasets and multiple adversarial attacks. The results demonstrate that our method, MC²F, not only establishes a new state-of-the-art in adversarial robustness but also fully preserves performance on clean data, even yielding modest gains in Accuracy.
ERank: Fusing Supervised Fine-Tuning and Reinforcement Learning for Effective and Efficient Text Reranking
PDF ↗Text reranking models are a crucial component in modern systems like Retrieval-Augmented Generation, tasked with selecting the most relevant documents prior to generation. However, current Large Language Models (LLMs) powered rerankers often face a fundamental trade-off. On one hand, Supervised Fine-Tuning based pointwise methods that frame relevance as a binary classification task lack the necessary scoring discrimination, particularly for those built on reasoning LLMs. On the other hand, approaches designed for complex reasoning often employ powerful yet inefficient listwise formulations, rendering them impractical for low latency applications. To resolve this dilemma, we introduce ERank, a highly Effective and Efficient pointwise reranker built from a reasoning LLM that excels across diverse relevance scenarios. We propose a novel two-stage training pipeline that begins with Supervised Fine-Tuning (SFT). In this stage, we move beyond binary labels and train the model generatively to output fine grained integer scores, which significantly enhances relevance discrimination. The model is then further refined using Reinforcement Learning (RL) with a novel, listwise derived reward. This technique instills global ranking awareness into the efficient pointwise architecture. We evaluate the ERank reranker on the BRIGHT, FollowIR, TREC DL, and BEIR benchmarks, demonstrating superior effectiveness and robustness compared to existing approaches. On the reasoning-intensive BRIGHT benchmark, our ERank-4B achieves an nDCG@10 of 38.7, while a larger 32B variant reaches a state of the art nDCG@10 of 40.2.
SEVADE: Self-Evolving Multi-Agent Analysis with Decoupled Evaluation for Hallucination-Resistant Sarcasm Detection
PDF ↗Sarcasm detection is a crucial yet challenging Natural Language Processing task. Existing Large Language Model methods are often limited by single-perspective analysis, static reasoning pathways, and a susceptibility to hallucination when processing complex ironic rhetoric, which impacts their accuracy and reliability. To address these challenges, we propose SEVADE, a novel Self-Evolving multi-agent Analysis framework with Decoupled Evaluation for hallucination-resistant sarcasm detection. The core of our framework is a Dynamic Agentive Reasoning Engine (DARE), which utilizes a team of specialized agents grounded in linguistic theory to perform a multifaceted deconstruction of the text and generate a structured reasoning chain. Subsequently, a separate lightweight rationale adjudicator (RA) performs the final classification based solely on this reasoning chain. This decoupled architecture is designed to mitigate the risk of hallucination by separating complex reasoning from the final judgment. Extensive experiments on four benchmark datasets demonstrate that our framework achieves state-of-the-art performance, with average improvements of 7.01% in Accuracy and 6.55% in Macro-F1 score.
Large language models (LLMs) have revolutionized AI applications, yet their high computational and memory demands hinder their widespread deployment. Existing compression techniques focus on intra-block optimizations (e.g., low-rank approximation or attention head pruning), while the repetitive layered structure of transformers implies significant inter-block redundancy - a dimension largely unexplored beyond key-value (KV) caching. Inspired by dictionary learning in convolutional networks, we propose a framework for structured weight sharing across transformer layers. Our approach decomposes attention projection matrices (Q, K, V, O) into shared dictionary atoms, reducing the attention module's parameters by 66.7% (e.g., 226.5M -> 75M in a 700M-parameter model) while achieving on-par performance. Unlike complex methods requiring distillation or architectural changes, MASA (Matrix Atom Sharing in Attention) operates as a drop-in replacement - trained with standard optimizers - and represents each layer's weights as linear combinations of shared matrix atoms. Experiments across scales (100M-700M parameters) show that MASA achieves better benchmark accuracy and perplexity than grouped-query attention (GQA), low-rank baselines and recently proposed Repeat-all-over/Sequential sharing at comparable parameter budgets. Ablation studies confirm robustness to the dictionary size and the efficacy of shared representations in capturing cross-layer statistical regularities. Extending to Vision Transformers (ViT), MASA matches performance metrics on image classification tasks with 66.7% fewer attention parameters. By combining dictionary learning strategies with transformer efficiency, MASA offers a scalable blueprint for parameter-efficient models without sacrificing performance. Finally, we investigate the possibility of employing MASA on large pretrained models to reduce their number of parameters without experiencing any significant drop in their performance.
Text-attributed heterogeneous graphs (TAHGs), characterized by nodes interconnected through diverse relationships and enriched with textual descriptions, are prevalent in numerous real-world applications. Recent advancements in integrating pre-trained language models (PLMs) and large language models (LLMs) with heterogeneous graph neural networks (HGNNs) have enhanced learning on TAHGs. However, the absence of standardized benchmark datasets tailored to TAHGs has impeded further progress. To bridge this gap, we propose the Text-attributed Heterogeneous Graphs Benchmark (THGB), a comprehensive collection of heterogeneous graphs from diverse domains, with each node enriched by relevant text attributes. Alongside dataset construction, we conduct extensive benchmark experiments using various graph learning methods, including GNN, PLM-GNN, and LLM-GNN approaches, for node classification and link prediction tasks. We evaluated model performance across supervised, few-shot, and zero-shot learning scenarios to assess their ability to leverage limited and unseen data. Our experiments highlight THGB's potential to improve the integration of heterogeneous structural and textual information. By providing curated datasets, robust evaluation protocols, and baseline implementations, THGB introduces a standardized benchmark and solid groundwork for TAHGs research.
Accurate muscle-mass assessment is crucial for staging and managing sarcopenia, yet existing methods suffer from modality-specific limitations and weak integration of muscle function indicators. To solve these limitations, we propose a Dual-source Features Graph for Sarcopenia Evaluation (DFGSE) to synergize high- and low-energy whole-body Dual-energy X-ray Absorptiometry (DXA) images, local high-energy DXA images, and blood-borne biochemical markers. Specifically, the feature extraction module employs dual-energy feature extraction to disentangle soft-tissue and skeletal cues from low-energy images, while skeleton-aware detection extracts joint features from high-energy images. It yields global and local DXA embeddings, complemented by blood-test representations. In the relevance exploration module, inter- and intra-modality correlations are computed via bilinear transformations to form adjacency matrices for the global, local, and blood modality representations. These matrices seed the Multi-type Multi-relation Graph Convolutional Network (MMGCN) – the core of the relation learning module – which captures both direct and indirect interactions among modalities through relation-aware message passing. Finally, the graph-fused representations are used by a muscle-mass prediction head trained with cross-entropy loss. Experiments on the public MURA dataset and two independent sarcopenia cohorts demonstrate that DFGSE consistently outperforms machine learning and state-of-the-art graph-based methods, in terms of four evaluation metrics for classification task.
Shared & Domain Self-Adaptive Experts with Frequency-Aware Discrimination for Continual Test-Time Adaptation
PDF ↗This paper focuses on the Continual Test-Time Adaptation (CTTA) task, aiming to enable an agent to continuously adapt to evolving target domains while retaining previously acquired domain knowledge for effective reuse when those domains reappear. Existing shared-parameter paradigms struggle to balance adaptation and forgetting, leading to decreased efficiency and stability. To address this, we propose a frequency-aware shared and self-adaptive expert framework, consisting of two key components: (i) a dual-branch expert architecture that extracts general features and dynamically models domain-specific representations, effectively reducing cross-domain interference and repetitive learning cost; and (ii) an online Frequency-aware Domain Discriminator (FDD), which leverages the robustness of low-frequency image signals for online domain shift detection, guiding dynamic allocation of expert resources for more stable and realistic adaptation. Additionally, we introduce a Continual Repeated Shifts (CRS) benchmark to simulate periodic domain changes for more realistic evaluation. Experimental results show that our method consistently outperforms existing approaches on both classification and segmentation CTTA tasks under standard and CRS settings, with ablations and visualizations confirming its effectiveness and robustness.
Ordinal classification has been widely applied in many high-stakes applications, e.g., medical imaging and diagnosis, where reliable uncertainty quantification (UQ) is essential for decision making. Conformal prediction (CP) is a general UQ framework that provides statistically valid guarantees, which is especially useful in practice. However, prior ordinal CP methods mainly focus on heuristic algorithms or restrictively require the underlying model to predict a unimodal distribution over ordinal labels. Consequently, they provide limited insight into coverage–efficiency trade-offs, or a model-agnostic and distribution-free nature favored by CP methods. To this end, we fill this gap by propose an ordinal-CP method that is model-agnostic and provides instance-level optimal prediction intervals. Specifically, we formulate conformal ordinal classification as a minimum-length covering problem at the instance level. To solve this problem, we develop a sliding-window algorithm that is optimal on each calibration data, with only a linear time complexity in K, the # of label candidates. The local optimality per instance further also improves predictive efficiency in expectation. Moreover, we propose a length-regularized variant that shrinks prediction set size while preserving coverage. Experiments on four benchmark datasets from diverse domains are conducted to demonstrate the significantly improved predictive efficiency of the proposed methods over baselines (by 15%↓ on average over four datasets).
The Contrastive Language-Image Pre-Training (CLIP) model excels in few-shot learning by aligning visual and textual representations. Our study shows that template-sample similarity (TSS), defined as the resemblance between a text template and an image sample, introduces bias. This bias leads the model to rely on template proximity rather than true sample-to-category alignment, reducing both accuracy and robustness in classification. We present a framework that uses empty prompts, textual inputs that convey the idea of “emptiness” without category information. These prompts capture unbiased template features and offset TSS bias. The framework employs two stages. During pre-training, empty prompts reveal and reduce template-induced bias within the CLIP encoder. During few-shot fine-tuning, a bias calibration loss enforces correct alignment between images and their categories, ensuring the model focuses on relevant visual cues. Experiments across multiple benchmarks demonstrate that our template correction method significantly reduces performance fluctuations caused by TSS, yielding higher classification accuracy and stronger robustness.
The success of large language models (LLMs) in cognitive tasks prompts the question of whether their next-token prediction (NTP) paradigm can be adapted to model physiological signals from wearable devices. A key target for this adaptation is photoplethysmography (PPG), the most prevalent sensing modality in consumer wearables for non-invasive monitoring of diverse physiological conditions. Unlike in NLP, where NTP aligns with generative objectives, physiological signal analysis involves fundamentally different tasks, such as continuous parameter estimation (regression) and discrete state recognition (classification). This disparity creates a semantic mismatch between the pre-training paradigm and the downstream tasks. To bridge this gap, we propose PPGPT, the first foundation model that reformulates NTP into next-feature token prediction (NFTP), learning hierarchical feature transition probabilities to unify pre-training and downstream objectives. PPGPT features a novel dual-stream encoder that generates feature tokens by jointly modeling temporal dynamics and local-global morphological patterns. The model is developed using a two-stage training framework: it is first pre-trained on a large-scale mixed dataset of 1.6 billion data points and then validated on our newly released BioMTL benchmark, which includes data from 172 subjects over 285 days across seven different tasks. Extensive experiments show that PPGPT significantly outperforms competing methods, achieving a 16.5% improvement in F1-score and a 25.9% reduction in Mean Absolute Error (MAE). Furthermore, the model demonstrates robust few-shot learning capabilities.
Embedding-based generalized zero-shot learning (GZSL) models often first forge robust latent semantic correlations between visual and attribute features so that knowledge can generalize to unseen categories. Despite leveraging attributes as priors and learning a shared embedding space, current methods exhibit two critical flaws. First, attributes with heterogeneous granularity are treated uniformly, leading to semantic ambiguity. Second, the source of class-level misclassification seldom aligns with attribute-level errors, preventing models from targeting the specific attributes responsible. To overcome these limitations, we introduce Structured Attribute-Guided Enhancement (SAGE), a unified framework for GZSL. Consensus-aware bidirectional attention first synchronizes visual–semantic focus regions via a mutual-distillation scheme. Next, we partition all attributes into pairwise-disjoint subsets—Global, Context, and Local—and couple them with visual features extracted at matching spatial scales. Finally, we design a cross-sample, subset-aware distillation mechanism—when a sample is misclassified, SAGE identifies the culpable attribute subset, retrieves high-confidence prototypes from a memory bank, and applies a Kullback–Leibler (KL) divergence constraint to the corresponding feature branch. Comprehensive experiments and ablations on the challenging AwA2, CUB, and SUN benchmarks demonstrate the contribution of each component, with SAGE achieving a new state-of-the-art throughout. These findings underscore SAGE’s robustness and versatility, marking a substantial advance in generalized zero-shot learning and paving the way for broader zero-resource recognition.
Knowledge distillation from Artificial Neural Networks (ANNs) to Spiking Neural Networks (SNNs) is a prominent training paradigm. However, its efficacy is fundamentally limited by a spectral mismatch: SNNs, with their intrinsic low-pass filtering characteristics, struggle to learn high-frequency details from their ANN teachers, creating a bottleneck in knowledge transfer at both the feature and logit levels. To address this, we propose Bi-Spectrum Distillation (BSD), a novel framework that mitigates the mismatch from two complementary perspectives. First, at the feature level, our Spectral Residual Distillation (SRD) enhances the student SNN's features with a parameter-efficient, learnable filter that adaptively compensates for high-frequency information loss, which transforms the student's output to better match the teacher's rich spectral target. Second, at the logits level, our Spectral Semantic Distillation (SSD) enhances fine-grained classification by distilling high-frequency components from teacher-ordered logits. Extensive experiments on CIFAR-10/100, ImageNet, and CIFAR10-DVS demonstrate that BSD achieves new state-of-the-art performance across both CNN and Transformer-based SNNs, validating its effectiveness and broad applicability.
Recent advances in vision-language models (VLMs) have enabled broad progress in the general medical field. However, pathology still remains a more challenging sub-domain, with current pathology-specific VLMs exhibiting limitations in both diagnostic accuracy and reasoning plausibility. Such shortcomings are largely attributable to the nature of current pathology datasets, which are primarily composed of image–description pairs that lack the depth and structured diagnostic paradigms employed by real-world pathologists. In this study, we leverage pathology textbooks and real-world pathology experts to construct high-quality, reasoning-oriented datasets. Building on this, we introduce Patho-R1, a multimodal RL-based pathology Reasoner, trained through a three-stage pipeline: (1) continued pretraining on 3.5 million image-text pairs for knowledge infusion; (2) supervised fine-tuning on 500k high-quality Chain-of-Thought samples for reasoning incentivizing; (3) reinforcement learning using Group Relative Policy Optimization and Decoupled Clip and Dynamic sAmpling Policy Optimization strategies for multimodal reasoning quality refinement. To further assess the alignment quality of our dataset, we propose Patho-CLIP, trained on the same figure-caption corpus used for continued pretraining. Comprehensive experimental results demonstrate that both Patho-CLIP and Patho-R1 achieve robust performance across a wide range of pathology-related tasks, including zero-shot classification, cross-modal retrieval, Visual Question Answering, and Multiple Choice Question.
HALoRA: Low-Rank Adaptation with Hierarchical Budget Allocation for Efficient Vision-Language Alignment
PDF ↗With the emergence of large multimodal models, dual-encoder alignment via contrastive learning has seen a resurgence. However, the escalating model size demands effective Parameter-Efficient Fine-Tuning (PEFT). While LoRA is a promising inference-free alternative to adapters, we find that its naive application to multimodal tasks causes a severe rank imbalance, favoring the text modality and FFN layers. To address this, we propose HALoRA (Hierarchical Allocation LoRA), which introduces a component-wise budget allocator to ensure balanced fine-tuning across both modalities and their internal components. This is complemented by a gradient-approximated initialization to accelerate convergence. With only half the parameters of adapters, HALoRA achieves superior or competitive performance in retrieval and zero-shot classification. Our work presents a more principled approach to multimodal LoRA, uncovering an intriguing asymmetry in vision-language alignment.
Grassmannian manifolds offer a powerful carrier for geometric representation learning by modelling high-dimensional data as low-dimensional subspaces. However, existing approaches predominantly rely on static single-subspace representations, neglecting the dynamic interplay between multiple subspaces critical for capturing complex geometric structures. To address this limitation, we propose a topology-driven multi-subspace fusion framework that enables adaptive subspace collaboration on the Grassmannian. Our solution introduces two key innovations: (1) an adaptive multi-subspace construction mechanism that dynamically selects and weights task-relevant subspaces via topological convergence analysis, and (2) a multi-subspace interaction block that fuses heterogeneous geometric representations through Fréchet mean optimisation on the manifold. Theoretically, we establish the convergence guarantees of adaptive subspaces under a projection metric topology, ensuring stable gradient-based optimisation. Practically, we integrate Riemannian batch normalisation and mutual information regularisation to enhance discriminability and robustness. Extensive experiments on 3D action recognition (HDM05, FPHA), EEG classification (MAMEM-SSVEPII), and graph tasks demonstrate state-of-the-art performance. Our work not only advances geometric deep learning but also successfully adapts the proven multi-channel interaction philosophy of Euclidean networks to non-Euclidean domains, achieving superior discriminability and interpretability.
In the context of the Classification and Regression Trees (CART) algorithm, the efficient splitting of categorical features using standard criteria like GINI and Entropy is well-established. However, using the Mean Absolute Error (MAE) criterion for categorical features has traditionally relied on various numerical encoding methods. This paper demonstrates that unsupervised numerical encoding methods are not viable for MAE criteria. Furthermore, we present a novel and efficient splitting algorithm that addresses the challenges of handling categorical features with the MAE criterion. Our findings underscore the limitations of existing approaches and offer a promising solution to enhance the handling of categorical data in CART algorithms.