Real-Time Aligned Reward Model beyond Semantics

Reinforcement Learning from Human Feedback (RLHF) is a pivotal technique for aligning large language models (LLMs) with human preferences, yet it is susceptible to reward overoptimization, in which policy models overfit to the reward model, exploit spurious reward patterns instead of faithfully capturing human intent. Prior mitigations primarily relies on surface semantic information and fails to efficiently address the misalignment between the reward model (RM) and the policy model caused by continuous policy distribution shifts. This inevitably leads to an increasing reward discrepancy, exacerbating reward overoptimization. To address these limitations, we introduce R2M (Real-Time Aligned Reward Model), a novel lightweight RLHF framework. R2M goes beyond vanilla reward models that solely depend on the semantic representations of a pretrained LLM. Instead, it leverages the evolving hidden states of the policy (namely policy feedback) to align with the real-time distribution shift of the policy during the RL process. This work points to a promising new direction for improving the performance of reward models through real-time utilization of feedback from policy models.

RAG-GFM: Overcoming In-Memory Bottlenecks in Graph Foundation Models via Retrieval-Augmented Generation

Graph Foundation Models (GFMs) have emerged as a frontier in graph learning, which are expected to deliver transferable representations across diverse tasks. However, GFMs remain constrained by in-memory bottlenecks: they attempt to encode knowledge into model parameters, which limits semantic capacity, introduces heavy lossy compression with conflicts, and entangles graph representation with the knowledge in ways that hinder efficient adaptation, undermining scalability and interpretability. In this work,we propose RAG-GFM, a Retrieval-Augmented Generation aided Graph Foundation Model that offloads knowledge from parameters and complements parameterized learning. To externalize graph knowledge, we build a dual-modal unified retrieval module, where a semantic store from prefix-structured text and a structural store from centrality-based motif. To preserve heterogeneous information, we design a dual-view alignment objective that contrasts both modalities to capture both content and relational patterns. To enable efficient downstream adaptation, we perform in-context augmentation to enrich supporting instances with retrieved texts and motifs as contextual evidence. Extensive experiments on five benchmark graph datasets demonstrate that RAG-GFM consistently outperforms 13 state-of-the-art baselines in both cross-domain node and graph classification, achieving superior effectiveness and efficiency.

Overcoming In-Memory Bottlenecks in Graph Foundation Models via Retrieval-Augmented Generation

Graph Foundation Models (GFMs) have emerged as a frontier in graph learning, which are expected to deliver transferable representations across diverse tasks. However, GFMs remain constrained by in-memory bottlenecks: they attempt to encode knowledge into model parameters, which limits semantic capacity, introduces heavy lossy compression with conflicts, and entangles graph representation with the knowledge in ways that hinder efficient adaptation, undermining scalability and interpretability. In this work,we propose RAG-GFM, a Retrieval-Augmented Generation aided Graph Foundation Model that offloads knowledge from parameters and complements parameterized learning. To externalize graph knowledge, we build a dual-modal unified retrieval module, where a semantic store from prefix-structured text and a structural store from centrality-based motif. To preserve heterogeneous information, we design a dual-view alignment objective that contrasts both modalities to capture both content and relational patterns. To enable efficient downstream adaptation, we perform in-context augmentation to enrich supporting instances with retrieved texts and motifs as contextual evidence. Extensive experiments on five benchmark graph datasets demonstrate that RAG-GFM consistently outperforms 13 state-of-the-art baselines in both cross-domain node and graph classification, achieving superior effectiveness and efficiency.

Unlocking the Potentials of Retrieval-Augmented Generation for Diffusion Language Models

Diffusion Language Models (DLMs) have recently demonstrated remarkable capabilities in natural language processing tasks. However, the potential of Retrieval-Augmented Generation (RAG), which shows great successes for enhancing large language models (LLMs), has not been well explored, due to the fundamental difference between LLM and DLM decoding. To fill this critical gap, we systematically test the performance of DLMs within the RAG framework. Our findings reveal that DLMs coupled with RAG show promising potentials with stronger dependency on contextual information, but suffer from limited generation precision. We identify a key underlying issue: Response Semantic Drift (RSD), where the generated answer progressively deviates from the query's original semantics, leading to low precision content. We trace this problem to the denoising strategies in DLMs, which fail to maintain semantic alignment with the query throughout the iterative denoising process. To address this, we propose Semantic-Preserving REtrieval-Augmented Diffusion (SPREAD), a novel framework that introduces a query-relevance-guided denoising strategy. By actively guiding the denoising trajectory, SPREAD ensures the generation remains anchored to the query's semantics and effectively suppresses drift. Experimental results demonstrate that SPREAD significantly enhances the precision and effectively mitigates RSD of generated answers within the RAG framework.

Your Group-Relative Advantage Is Biased

Reinforcement Learning from Verifier Rewards (RLVR) has emerged as a widely used approach for post-training large language models on reasoning tasks, with group-based methods such as GRPO and its variants gaining broad adoption. These methods rely on group-relative advantage estimation to avoid learned critics, yet its theoretical properties remain poorly understood. In this work, we uncover a fundamental issue of group-based RL: the group-relative advantage estimator is inherently biased relative to the true (expected) advantage. We provide the first theoretical analysis showing that it systematically underestimates advantages for hard prompts and overestimates them for easy prompts, leading to imbalanced exploration and exploitation. To address this issue, we propose History-Aware Adaptive Difficulty Weighting (HA-DW), an adaptive reweighting scheme that adjusts advantage estimates based on an evolving difficulty anchor and training dynamics. Both theoretical analysis and experiments on five mathematical reasoning benchmarks demonstrate that HA-DW consistently improves performance when integrated into GRPO and its variants. Our results suggest that correcting biased advantage estimation is critical for robust and efficient RLVR training.

LLMBoost: Make Large Language Models Stronger with Boosting

Ensemble learning of LLMs has emerged as a promising alternative to enhance performance, but existing approaches typically treat models as black boxes, combining the inputs or final outputs while overlooking the rich internal representations and interactions across models.In this work, we introduce LLMBoost, a novel ensemble fine-tuning framework that breaks this barrier by explicitly leveraging intermediate states of LLMs. Inspired by the boosting paradigm, LLMBoost incorporates three key innovations. First, a cross-model attention mechanism enables successor models to access and fuse hidden states from predecessors, facilitating hierarchical error correction and knowledge transfer. Second, a chain training paradigm progressively fine-tunes connected models with an error-suppression objective, ensuring that each model rectifies the mispredictions of its predecessor with minimal additional computation. Third, a near-parallel inference paradigm design pipelines hidden states across models layer by layer, achieving inference efficiency approaching single-model decoding. We further establish the theoretical foundations of LLMBoost, proving that sequential integration guarantees monotonic improvements under bounded correction assumptions. Extensive experiments on commonsense reasoning and arithmetic reasoning tasks demonstrate that LLMBoost consistently boosts accuracy while reducing inference latency.

Towards Long-window Anchoring in Vision-Language Model Distillation

While large vision-language models (VLMs) demonstrate strong long-context understanding, their prevalent small branches fail on linguistics-photography alignment for a limited window size. We discover that knowledge distillation improves students' capability as a complement to Rotary Position Embeddings (RoPE) on window sizes (anchored from large models). Building on this insight, we propose LAid, which directly aims at the transfer of long-range attention mechanisms through two complementary components: (1) a progressive distance-weighted attention matching that dynamically emphasizes longer position differences during training, and (2) a learnable RoPE response gain modulation that selectively amplifies position sensitivity where needed. Extensive experiments across multiple model families demonstrate that LAid-distilled models achieve up to 3.2 times longer effective context windows compared to baseline small models, while maintaining or improving performance on standard VL benchmarks. Spectral analysis also suggests that LAid successfully preserves crucial low-frequency attention components that conventional methods fail to transfer. Our work not only provides practical techniques for building more efficient long-context VLMs but also offers theoretical insights into how positional understanding emerges and transfers during distillation.

Is the Information Bottleneck Robust Enough? Towards Label-Noise Resistant Information Bottleneck Learning

The Information Bottleneck (IB) principle facilitates effective representation learning by preserving label-relevant information while compressing irrelevant information. However, its strong reliance on accurate labels makes it inherently vulnerable to label noise, prevalent in real-world scenarios, resulting in significant performance degradation and overfitting. To address this issue, we propose LaT-IB, a novel Label-Noise ResistanT Information Bottleneck method which introduces a "Minimal-Sufficient-Clean" (MSC) criterion. Instantiated as a mutual information regularizer to retain task-relevant information while discarding noise, MSC addresses standard IB's vulnerability to noisy label supervision. To achieve this, LaT-IB employs a noise-aware latent disentanglement that decomposes the latent representation into components aligned with to the clean label space and the noise space. Theoretically, we first derive mutual information bounds for each component of our objective including prediction, compression, and disentanglement, and moreover prove that optimizing it encourages representations invariant to input noise and separates clean and noisy label information. Furthermore, we design a three-phase training framework: Warmup, Knowledge Injection and Robust Training, to progressively guide the model toward noise-resistant representations. Extensive experiments demonstrate that LaT-IB achieves superior robustness and efficiency under label noise, significantly enhancing robustness and applicability in real-world scenarios with label noise.

Fine-Tuned LLMs Know They Don't Know: A Parameter-Efficient Approach to Recovering Honesty

The honesty of Large Language Models (LLMs) is increasingly important for safe deployment in high-stakes domains. However, this crucial trait is severely undermined by supervised fine-tuning (SFT), a common technique for model specialization. Existing recovery methods rely on data-intensive global parameter adjustments, implicitly assuming that SFT deeply corrupts the models' ability to recognize their knowledge boundaries. However, we observe that fine-tuned LLMs still preserve this ability; what is damaged is their capacity to faithfully express that awareness. Building on this, we propose Honesty-Critical Neurons Restoration (HCNR) to surgically repair this suppressed capacity. HCNR identifies and restores key expression-governing neurons to their pre-trained state while harmonizing them with task-oriented neurons via Hessian-guided compensation. Experiments on four QA tasks and five LLM families demonstrate that HCNR effectively recovers 33.25% of the compromised honesty while achieving at least 2.23x speedup with over 10x less data compared to baseline methods, offering a practical solution for trustworthy LLM deployment.

USF-Net: A Unified Spatiotemporal Fusion Network for Ground-Based Remote Sensing Cloud Image Sequence Extrapolation

Ground-based remote sensing cloud image sequence extrapolation is a key research area in the development of photovoltaic power systems. However, existing approaches exhibit several limitations:(1)they primarily rely on static kernels to augment feature information, lacking adaptive mechanisms to extract features at varying resolutions dynamically;(2)temporal guidance is insufficient, leading to suboptimal modeling of long-range spatiotemporal dependencies; and(3)the quadratic computational cost of attention mechanisms is often overlooked, limiting efficiency in practical deployment. To address these challenges, we propose USF-Net, a Unified Spatiotemporal Fusion Network that integrates adaptive large-kernel convolutions and a low-complexity attention mechanism, combining temporal flow information within an encoder-decoder framework. Specifically, the encoder employs three basic layers to extract features. Followed by the USTM, which comprises:(1)a SiB equipped with a SSM that dynamically captures multi-scale contextual information, and(2)a TiB featuring a TAM that effectively models long-range temporal dependencies while maintaining computational efficiency. In addition, a DSM with a TGM is introduced to enable unified modeling of temporally guided spatiotemporal dependencies. On the decoder side, a DUM is employed to address the common "ghosting effect." It utilizes the initial temporal state as an attention operator to preserve critical motion signatures. As a key contribution, we also introduce and release the ASI-CIS dataset. Extensive experiments on ASI-CIS demonstrate that USF-Net significantly outperforms state-of-the-art methods, establishing a superior balance between prediction accuracy and computational efficiency for ground-based cloud extrapolation. The dataset and source code will be available at https://github.com/she1110/ASI-CIS.