Not All Tokens Matter Equally: Dynamic In-context Vector Distillation with Decisive-Token Supervision for Long-form Medical Report Generation

Distilling demonstration effects into hidden-space interventions offers a lightweight alternative to full finetuning. However, existing multimodal variants are mostly evaluated on short-form tasks, where outputs end after a few tokens. Extending these methods to long-form generation exposes a fundamental yet underexamined limitation: token-level distillation implicitly treats all output tokens as equally informative, but long-form outputs are dominated by high-frequency template and grammatical tokens, while the tokens that actually determine output quality are sparsely distributed. In medical report generation (MRG), two such decisive tokens stand out: pathology-related tokens that determine diagnostic content, and the end-of-sequence (EOS) event that determines termination. Both receive insufficient supervision under uniform cross-entropy, and autoregressive decoding further compounds the problem by drifting away from teacher-forced trajectories. We propose DIVE, a frozen-backbone distillation framework that addresses long-form report generation through two complementary mechanisms matched to these failures. Decisive-token supervision restores supervision balance by upweighting the cross-entropy contribution of pathology-related tokens and the EOS event, ensuring that content fidelity and termination are learned during training rather than imposed at decoding time. State-conditioned dynamic steering replaces fixed open-loop residuals with hidden-state-dependent adapters, allowing the injected signal to adapt as decoding drifts. Experiments on MIMIC-CXR and CheXpert Plus with two medical VLM backbones show that DIVE consistently ranks among the strongest methods across lexical and clinical-proxy metrics. Our method achieves the best BLEU-4, ROUGE-L, and RadGraph F1 in all dataset--backbone settings, while remaining competitive on coarse label-level CheXbert F1.

Beyond Rejection Sampling: Trajectory Fusion for Scaling Mathematical Reasoning

Large language models (LLMs) have made impressive strides in mathematical reasoning, often fine-tuned using rejection sampling that retains only correct reasoning trajectories. While effective, this paradigm treats supervision as a binary filter that systematically excludes teacher-generated errors, leaving a gap in how reasoning failures are modeled during training. In this paper, we propose TrajFusion, a fine-tuning strategy that reframes rejection sampling as a structured supervision construction process. Specifically, TrajFusion forms fused trajectories that explicitly model trial-and-error reasoning by interleaving selected incorrect trajectories with reflection prompts and correct trajectories. The length of each fused sample is adaptively controlled based on the frequency and diversity of teacher errors, providing richer supervision for challenging problems while safely reducing to vanilla rejection sampling fine-tuning (RFT) when error signals are uninformative. TrajFusion requires no changes to the architecture or training objective. Extensive experiments across multiple math benchmarks demonstrate that TrajFusion consistently outperforms RFT, particularly on challenging and long-form reasoning problems.

Quantifying the Robustness of Retrieval-Augmented Language Models Against Spurious Features in Grounding Data

Robustness has become a critical attribute for the deployment of RAG systems in real-world applications. Existing research focuses on robustness to explicit noise (e.g., document semantics) but overlooks spurious features (a.k.a. implicit noise). While previous works have explored spurious features in LLMs, they are limited to specific features (e.g., formats) and narrow scenarios (e.g., ICL). In this work, we statistically confirm the presence of spurious features in the RAG paradigm, a robustness problem caused by the sensitivity of LLMs to semantic-agnostic features. Moreover, we provide a comprehensive taxonomy of spurious features and empirically quantify their impact through controlled experiments. Further analysis reveals that not all spurious features are harmful and they can even be beneficial sometimes. Extensive evaluation results across multiple LLMs suggest that spurious features are a widespread and challenging problem in the field of RAG. The code and dataset will be released to facilitate future research. We release all codes and data at: $\\\href{https://github.com/maybenotime/RAG-SpuriousFeatures}{https://github.com/maybenotime/RAG-SpuriousFeatures}$.

MuDAF: Long-Context Multi-Document Attention Focusing through Contrastive Learning on Attention Heads

Large Language Models (LLMs) frequently show distracted attention due to irrelevant information in the input, which severely impairs their long-context capabilities. Inspired by recent studies on the effectiveness of retrieval heads in long-context factutality, we aim at addressing this distraction issue through improving such retrieval heads directly. We propose Multi-Document Attention Focusing (MuDAF), a novel method that explicitly optimizes the attention distribution at the head level through contrastive learning. According to the experimental results, MuDAF can significantly improve the long-context question answering performance of LLMs, especially in multi-document question answering. Extensive evaluations on retrieval scores and attention visualizations show that MuDAF possesses great potential in making attention heads more focused on relevant information and reducing attention distractions.

Zero-Shot NAS via the Suppression of Local Entropy Decrease

Architecture performance evaluation is the most time-consuming part of neural architecture search (NAS). Zero-Shot NAS accelerates the evaluation by utilizing zero-cost proxies instead of training. Though effective, existing zero-cost proxies require invoking backpropagations or running networks on input data, making it difficult to further accelerate the computation of proxies. To alleviate this issue, architecture topologies are used to evaluate the performance of networks in this study. We prove that particular architectural topologies decrease the local entropy of feature maps, which degrades specific features to a bias, thereby reducing network performance. Based on this proof, architectural topologies are utilized to quantify the suppression of local entropy decrease (SED) as a data-free and running-free proxy. Experimental results show that SED outperforms most state-of-the-art proxies in terms of architecture selection on five benchmarks, with computation time reduced by three orders of magnitude. We further compare the SED-based NAS with state-of-the-art proxies. SED-based NAS selects the architecture with higher accuracy and fewer parameters in only one second. The theoretical analyses of local entropy and experimental results demonstrate that the suppression of local entropy decrease facilitates selecting optimal architectures in Zero-Shot NAS.

Towards Truthful Multilingual Large Language Models: Benchmarking and Alignment Strategies

In the era of large language models (LLMs), building multilingual large language models (MLLMs) that can serve users worldwide holds great significance. However, existing research seldom focuses on the truthfulness of MLLMs. Meanwhile, contemporary multilingual aligning technologies struggle to balance massive languages and often exhibit serious truthfulness gaps across different languages, especially those that differ greatly from English. In our work, we construct a benchmark for truthfulness evaluation in multilingual scenarios and explore the ways to align facts across languages to enhance the truthfulness of MLLMs. Furthermore, we propose Fact-aware Multilingual Selective Synergy (FaMSS) to optimize the data allocation across a large number of languages and different data types. Experimental results demonstrate that our approach can effectively reduce the multilingual representation disparity and enhance the multilingual capabilities of LLMs.

Finetuning Large Language Model for Personalized Ranking

Large Language Models (LLMs) have demonstrated remarkable performance across various domains, motivating researchers to investigate their potential use in recommendation systems. However, directly applying LLMs to recommendation tasks has proven challenging due to the significant disparity between the data used for pre-training LLMs and the specific requirements of recommendation tasks. In this study, we introduce Direct Multi-Preference Optimization (DMPO), a streamlined framework designed to bridge the gap and enhance the alignment of LLMs for recommendation tasks. DMPO enhances the performance of LLM-based recommenders by simultaneously maximizing the probability of positive samples and minimizing the probability of multiple negative samples. We conducted experimental evaluations to compare DMPO against traditional recommendation methods and other LLM-based recommendation approaches. The results demonstrate that DMPO significantly improves the recommendation capabilities of LLMs across three real-world public datasets in few-shot scenarios. Additionally, the experiments indicate that DMPO exhibits superior generalization ability in cross-domain recommendations. A case study elucidates the reasons behind these consistent improvements and also underscores DMPO's potential as an explainable recommendation system.

Beyond Surface: Probing LLaMA Across Scales and Layers

This paper presents an in-depth analysis of Large Language Models (LLMs), focusing on LLaMA, a prominent open-source foundational model in natural language processing. Instead of assessing LLaMA through its generative output, we design multiple-choice tasks to probe its intrinsic understanding in high-order tasks such as reasoning and computation. We examine the model horizontally, comparing different sizes, and vertically, assessing different layers. We unveil several key and uncommon findings based on the designed probing tasks: (1) Horizontally, enlarging model sizes almost could not automatically impart additional knowledge or computational prowess. Instead, it can enhance reasoning abilities, especially in math problem solving, and helps reduce hallucinations, but only beyond certain size thresholds; (2) In vertical analysis, the lower layers of LLaMA lack substantial arithmetic and factual knowledge, showcasing logical thinking, multilingual and recognitive abilities, with top layers housing most computational power and real-world knowledge.

Breaking Language Barriers in Multilingual Mathematical Reasoning: Insights and Observations

Existing research predominantly focuses on developing powerful language learning models (LLMs) for mathematical reasoning within monolingual languages, with few explorations in preserving efficacy in a multilingual context. To bridge this gap, this paper pioneers exploring and training powerful Multilingual Math Reasoning (xMR) LLMs. Firstly, by utilizing translation, we construct the first multilingual math reasoning instruction dataset, MGSM8KInstruct, encompassing ten distinct languages, thus addressing the issue of training data scarcity in xMR tasks. Based on the collected dataset, we propose different training strategies to build powerful xMR LLMs, named MathOctopus, notably outperform conventional open-source LLMs and exhibit superiority over ChatGPT in few-shot scenarios. Notably, MathOctopus-13B reaches 47.6% accuracy which exceeds ChatGPT 46.3% on MGSM testset. Beyond remarkable results, we unearth several pivotal observations and insights from extensive experiments: (1) When extending the rejection sampling strategy to the multilingual context, it proves effective for model performances, albeit limited. (2) Employing parallel corpora for math Supervised Fine-Tuning (SFT) across multiple languages not only significantly enhances model performance multilingually but also elevates their monolingual performance. This indicates that crafting multilingual corpora can be regarded as a vital strategy for enhancing model performance in a specific language, especially in mathematical reasoning tasks. For instance, MathOctopus-7B improves its counterparts that trained on English from 42.2% to 50.8% on GSM8K testset.

RUEL: Retrieval-Augmented User Representation with Edge Browser Logs for Sequential Recommendation

Online recommender systems (RS) aim to match user needs with the vast amount of resources available on various platforms. A key challenge is to model user preferences accurately under the condition of data sparsity. To address this challenge, some methods have leveraged external user behavior data from multiple platforms to enrich user representation. However, all of these methods require a consistent user ID across platforms and ignore the information from similar users. In this study, we propose RUEL, a novel retrieval-based sequential recommender that can effectively incorporate external anonymous user behavior data from Edge browser logs to enhance recommendation. We first collect and preprocess a large volume of Edge browser logs over a one-year period and link them to target entities that correspond to candidate items in recommendation datasets. We then design a contrastive learning framework with a momentum encoder and a memory bank to retrieve the most relevant and diverse browsing sequences from the full browsing log based on the semantic similarity between user representations. After retrieval, we apply an item-level attentive selector to filter out noisy items and generate refined sequence embeddings for the final predictor. RUEL is the first method that connects user browsing data with typical recommendation datasets and can be generalized to various recommendation scenarios and datasets. We conduct extensive experiments on four real datasets for sequential recommendation tasks and demonstrate that RUEL significantly outperforms state-of-the-art baselines. We also conduct ablation studies and qualitative analysis to validate the effectiveness of each component of RUEL and provide additional insights into our method.