Reward models (RMs) are a core component in the post-training of large language models (LLMs), serving as proxies for human preference evaluation and guiding model alignment. However, training reliable RMs under limited resources remains challenging due to the reliance on large-scale preference annotations and the high cost of fine-tuning LLMs. To address this, we propose SparseRM, which leverages Sparse Autoencoder (SAE) to extract preference-relevant information encoded in model representations, enabling the construction of a lightweight and interpretable reward model. SparseRM first employs SAE to decompose LLM representations into interpretable directions that capture preference-relevant features. The representations are then projected onto these directions to compute alignment scores, which quantify the strength of each preference feature in the representations. A simple reward head aggregates these scores to predict preference scores. Experiments on three preference modeling tasks show that SparseRM achieves superior performance over most mainstream RMs while using less than 1% of trainable parameters. Moreover, it integrates seamlessly into downstream alignment pipelines, highlighting its potential for efficient alignment.

Training Large Language Models (LLMs) for chain-of-thought reasoning presents a significant challenge: supervised fine-tuning on a single "golden" rationale hurts generalization as it penalizes equally valid alternatives, whereas reinforcement learning with verifiable rewards struggles with credit assignment and prohibitive computational cost. To tackle these limitations, we introduce InTRO (In-Token Rationality Optimization), a new framework that enables both token-level exploration and self-feedback for accurate and concise reasoning. Instead of directly optimizing an intractable objective over all valid reasoning paths, InTRO leverages correction factors—token-wise importance weights estimated by the information discrepancy between the generative policy and its answer-conditioned counterpart, for informative next-token selection. This approach allows the model to perform token-level exploration and receive self-generated feedback within a single forward pass, ultimately encouraging accurate and concise rationales. Across six math-reasoning benchmarks, InTRO consistently outperforms other baselines, raising solution accuracy by up to 20% relative to the base model. Its chains of thought are also notably more concise, exhibiting reduced verbosity. Beyond this, InTRO enables cross-domain transfer, successfully adapting to out-of-domain reasoning tasks that extend beyond the realm of mathematics, demonstrating robust generalization.

Visual Question Answering (VQA) aims to answer questions utilizing information from both textual and visual modalities. New data categories and novel combinations of the two modalities will continuously emerge in practical applications, necessitating continual learning. For this unique compositional generalization challenge, existing work adopts prototype learning to separately capture sample-invariant features of question types and image objects, retrieving the most similar prototype from two modalities as generalizable representations. However, for semantic fuzzy samples that deviate from the corresponding prototype, relying on a single prototype may lead to error-prone representation learning. Additionally, for the catastrophic forgetting challenge, existing methods employ memory to store past samples but lack an effective selection of samples to be stored, merely updating the memory through random sampling. In this paper, we propose ProtoGroup, a multi-prototype grouping continual learning framework for VQA, which groups prototypes based on their similarity to obtain more accurate and stable sample-invariant features. We also devise a memory sample selection method based on the correlation with prototype groups from old tasks. Experimental results demonstrate our framework outperforms previous methods significantly across multiple datasets and settings.

Preference alignment methods are increasingly critical for steering large language models (LLMs) to generate outputs consistent with human values. While recent approaches often rely on synthetic data generated by LLMs for scalability and cost-efficiency reasons, this reliance can introduce distributional shifts that undermine the nuanced representation of human preferences needed for desirable outputs. In this paper, we propose a novel distribution-aware optimization framework that improves preference alignment despite such shifts. Our approach first leverages well-learned classifiers to assign a calibration value to each training sample, quantifying its alignment with the target human-preferred distribution. These values are then incorporated into a robust optimization objective that minimizes the worst-case loss over regions of the data space most relevant to human preferences. By explicitly focusing optimization on the target distribution, our approach mitigates the impact of distributional mismatch and improves the generation of responses that better reflect intended values.

Multimodal Large Language Models (MLLMs) exhibit impressive performance across vision-language tasks, but still face the hallucination challenges, where generated texts are factually inconsistent with visual input. Existing mitigation methods focus on surface symptoms of hallucination and heavily rely on post-hoc corrections, extensive data curation, or costly inference schemes. In this work, we identify two key factors of MLLM hallucination: Insufficient Visual Context, where ambiguous visual contexts lead to language speculation, and Progressive Textual Drift, where model attention strays from visual inputs in longer responses. To address these problems, we propose a novel Complementary Visual Grounding (CVG) framework. CVG exploits the intrinsic architecture of MLLMs, without requiring any external tools, models, or additional data. CVG first disentangles visual context into two complementary branches based on query relevance, then maintains steadfast visual grounding during the auto-regressive generation. Finally, it contrasts the output distributions of two branches to produce a faithful response. Extensive experiments on various hallucination and general benchmarks demonstrate that CVG achieves state-of-the-art performances across MLLM architectures and scales.

Cross-modal alignment aims to build a bridge connecting vision and language. It is an important multi-modal task that efficiently learns the semantic similarities be-tween images and texts. Traditional finegrained alignment methods heavily rely on pre-trained object detectors to extract region features for subsequent region-word alignment, thereby incurring substantial computational costs for region detection and error propagation issues for two-stage training. In this paper, we focus on the mainstream vision transformer, incorporating patch features for patch-word alignment, while addressing the resultant issue of visual patch redundancy and patch ambiguity for semantic alignment. We propose a novel Linguistic-Aware Patch Slim-ming (LAPS) framework for fine-grained alignment, which explicitly identifies redundant visual patches with language supervision and rectifies their semantic and spatial information to facilitate more effective and consistent patchword alignment. Extensive experiments on various evaluation benchmarks and model backbones show LAPS out-performs the state-of-the-art fine-grained alignment methods by 5%-15% rSum. Our code is available at https://github.com/CrossmodalGroup/LAPS.

Image-text matching, a bridge connecting image and language, is an important task, which generally learns a holistic cross-modal embedding to achieve a high-quality semantic alignment between the two modalities. However, previous studies only focus on capturing fragment-level relation within a sample from a particular modality, e.g., salient regions in an image or text words in a sentence, where they usually pay less attention to capturing instance-level interactions among samples and modalities, e.g., multiple images and texts. In this paper, we argue that sample relations could help learn subtle differences for hard negative instances, and thus transfer shared knowledge for infrequent samples should be promising in obtaining better holistic embeddings. Therefore, we propose a novel hierarchical relation modeling framework (HREM), which explicitly capture both fragment-and instance-level relations to learn discriminative and robust cross-modal embeddings. Extensive experiments on Flickr30K and MS-COCO show our proposed method out-performs the state-of-the-art ones by 4%-10% in terms of rSum. Our code is available at https://github.com/CrossmodalGroup/HREM.

Deep metric learning aims to learn an embedding space, where semantically similar samples are close together and dissimilar ones are repelled against. To explore more hard and informative training signals for augmentation and generalization, recent methods focus on generating synthetic samples to boost metric learning losses. However, these methods just use the deterministic and class-independent generations (e.g., simple linear interpolation), which only can cover the limited part of distribution spaces around original samples. They have overlooked the wide characteristic changes of different classes and can not model abundant intra-class variations for generations. Therefore, generated samples not only lack rich semantics within the certain class, but also might be noisy signals to disturb training. In this paper, we propose a novel intra-class adaptive augmentation (IAA) framework for deep metric learning. We reasonably estimate intra-class variations for every class and generate adaptive synthetic samples to support hard samples mining and boost metric learning losses. Further, for most datasets that have a few samples within the class, we propose the neighbor correction to revise the inaccurate estimations, according to our correlation discovery where similar classes generally have similar variation distributions. Extensive experiments on five benchmarks show our method significantly improves and outperforms the state-of-the-art methods on retrieval performances by 3%-6%.

The core purpose of deep metric learning is to construct an embedding space, where objects belonging to the same class are gathered together and the ones from different classes are pushed apart. Most existing approaches typically insist to inter-class characteristics,e.g., class-level information or instance-level similarity, to obtain semantic relevance of data points and get a large margin between different classes in the embedding space. However, the intra-class characteristics,e.g., local manifold structure or relative relationship within the same class, are usually overlooked in the learning process. Hence the output embeddings have limitation in retrieving a good ranking result if existing multiple positive samples. And the local data structure of embedding space cannot be fully exploited since lack of relative ranking information. As a result, the model is prone to overfitting on a train set and get low generalization on the test set (unseen classes) when losing sight of intra-class variance. This paper presents a novel self-supervised synthesis ranking auxiliary framework, which captures intra-class characteristics as well as inter-class characteristics for better metric learning. Our method designs a synthetic samples generation of polar coordinates to generate measurable intra-class variance with different strength and diversity in the latent space, which can simulate the various local structure change of intra-class in the initial data domain. And then formulates a self-supervised learning procedure to fully exploit this property and preserve it in the embedding space. As a result, the learned embedding space not only keeps inter-class discrimination but also owns subtle intra-class diversity, leading to better global and local embedding structures. Extensive experiments on five benchmarks show that our method significantly improves and outperforms the state-of-the-art methods on the performances of both retrieval and ranking by 2%-4% (personal use of this material is permitted. However, permission to use this material for any other purposes must be obtained from the IEEE by sending an email to [email protected]).

Deep metric learning aims to learn a deep embedding space, where similar objects are pushed towards together and different objects are repelled against. Existing approaches typically use inter-class characteristics, e.g. class-level information or instance-level similarity, to obtain semantic relevance of data points and get a large margin between different classes in the embedding space. However, the intra-class characteristics, e.g. local manifold structure or relative relationship within the same class, are usually overlooked in the learning process. Hence the data structure cannot be fully exploited and the output embeddings have limitation in retrieval. More importantly, retrieval results lack in a good ranking. This paper presents a novel self-supervised ranking auxiliary framework, which captures intra-class characteristics as well as inter-class characteristics for better metric learning. Our method defines specific transform functions to simulates the local structure change of intra-class in the initial image domain, and formulates a self-supervised learning procedure to fully exploit this property and preserve it in the embedding space. Extensive experiments on three standard benchmarks show that our method significantly improves and outperforms the state-of-the-art methods on the performances of both retrieval and ranking by 2%-4%.