Retrieval-augmented generation (RAG) enhances large language models (LLMs) with external knowledge but still suffers from long contexts and disjoint retrieval-generation optimization. In this work, we propose CLaRa (Continuous Latent Reasoning), a unified framework that performs embedding-based compression and joint optimization in a shared continuous space. To obtain semantically rich and retrievable compressed vectors, we introduce SCP, a key-preserving data synthesis framework using QA and paraphrase supervision. CLaRa then trains the reranker and generator end-to-end via a single language modeling loss, with gradients flowing through both modules using a differentiable top-k estimator. Theoretically, this unified optimization aligns retrieval relevance with answer quality. Experiments across multiple QA benchmarks show that CLaRa achieves state-of-the-art compression and reranking performance, often surpassing text-based fine-tuned baselines.

Most applications of generative AI involve a sequential interaction in which a person inputs a prompt and waits for a response, and where reaction time and adaptiv-ity are not important factors. In contrast, live jamming is a collaborative interaction that requires real-time coordination and adaptation without access to the other player's future moves, while preserving diversity to sustain a creative flow. Reinforcement learning post-training enables effective adaptation through on-policy interaction, yet it often reduces output diversity by exploiting coherence-based rewards. This collapse, known as "reward hacking", affects many RL post-training pipelines, but is especially harmful in live jamming, where musical creativity relies on dynamic variation and mutual responsiveness. In this paper, we propose a novel adversarial training method on policy-generated trajectories to mitigate reward hacking in RL post-training for melody-to-chord accompaniment. A co-evolving discriminator separates policy trajectories from the data distribution, while the policy maximizes the discriminator output in addition to coherence rewards to prevent collapse to trivial outputs. We evaluate accompaniment quality and output diversity in simulation with both fixed test melodies and learned melody agents, and we conduct a user study with the model deployed in a real-time interactive system with expert musicians. Quantitative evaluation and user feedback demonstrate improved output diversity, harmonic coherence, adaptation speed and user agency. Our results demonstrate a simple yet effective method to mitigate reward hacking in RL post-training of generative sequence models. The combination of large-scale transformer-based models and reinforcement learning (RL) post-training has revolutionized AI, with over 1 billion people now using large language models (LLMs) trained with these techniques (OpenAI, 2025; Perez, 2025). However, most applications of generative AI still involve a slow back-and-forth interaction, where the user inputs a request, and then waits--sometimes several minutes--for a response.

The growth of Massive Open Online Courses (MOOCs) presents significant challenges for personalized learning, where concept recommendation is crucial. Existing approaches typically rely on heterogeneous information networks or knowledge graphs to capture conceptual relationships, combined with knowledge tracing models to assess learners' cognitive states. However, these methods face significant limitations due to their dependence on high-quality structured knowledge graphs, which are often scarce in real-world educational scenarios. To address this fundamental challenge, this paper proposes CLLMRec, a novel framework that leverages Large Language Models through two synergistic technical pillars: Semantic Alignment and Prerequisite Knowledge Distillation. The Semantic Alignment component constructs a unified representation space by encoding unstructured textual descriptions of learners and concepts. The Prerequisite Knowledge Distillation paradigm employs a teacher-student architecture, where a large teacher LLM (implemented as the Prior Knowledge Aware Component) extracts conceptual prerequisite relationships from its internalized world knowledge and distills them into soft labels to train an efficient student ranker. Building upon these foundations, our framework incorporates a fine-ranking mechanism that explicitly models learners' real-time cognitive states through deep knowledge tracing, ensuring recommendations are both structurally sound and cognitively appropriate. Extensive experiments on two real-world MOOC datasets demonstrate that CLLMRec significantly outperforms existing baseline methods across multiple evaluation metrics, validating its effectiveness in generating truly cognitive-aware and personalized concept recommendations without relying on explicit structural priors.

Clinical antimicrobial therapy requires the dynamic integration of pathogen profiles, host factors, pharmacological properties of antimicrobials, and the severity of infection.This complexity imposes fundamental limitations on the applicability of Large Language Models (LLMs) in high-stakes clinical decision-making including knowledge gaps, data privacy concerns, high deployment costs, and limited reasoning capabilities. This is the first author footnote. A hierarchical evaluation employing diverse teacher-model proxies reduces assessment costs, while modular interface design facilitates seamless system updates. Experimental results demonstrate that KRAL significantly outperforms traditional Retrieval-Augmented Generation (RAG) and Supervised Fine-Tuning (SFT) methods. It improves knowledge question-answering capability (Accuracy@1 on the external open-source benchmark MEDQA increased by 1.8% vs. SFT and 3.6% vs. RAG) and reasoning capability (Pass@1 on the external benchmark PUMCH Antimicrobial increased by 27% vs. SFT and 27.2% vs. RAG), achieved at 20% of SFT's long-term training costs. This establishes KRAL as an effective solution for enhancing local LLMs' clinical diagnostic capabilities, enabling low-cost, high-safety deployment in complex medical decision support. Introduction Antimicrobial therapy constitutes a cornerstone of modern clinical practice. The formulation of an effective regimen necessitates the integration of pathogen-specific factors, host characteristics, pharmacokinetic pharma-codynamic (PK/PD) properties of antimicrobials, and infection severity, all of which are dynamic and interrelated. This places significant cognitive load on clinicians, especially for non-infectious disease specialists or in situations where pathogens are unknown and time is limited, which may result in sub-optimal prescribing decisions, thereby increasing the likelihood of therapeutic failure, antimicrobial toxicity, and the emergence of multidrug-resistant (MDR) pathogens. Large language models (LLMs) have recently emerged as promising tools for enhancing clinical decision support systems (CDSS), owing to their advanced natural language understanding and generation capabilities. Nevertheless, the direct deployment of general-purpose LLMs in high-stakes clinical domains such as antimicrobial therapy is fraught with limitations, including: Knowledge bias: Medical content constitutes <0.3 % of the pre-training corpora[1] in mainstream LLMs (e.g., GPT-3), resulting in limited coverage of rare or emerging pathogens[2-4], outdated guideline adherence[5-9](Appendix A1), and suboptimal performance on atypical presentations. 2 Data privacy and compliance risks: The use of closed-source, cloud-based LLMs (e.g., GPT-4) for processing unencrypted protected health information (PHI) may violate HIPAA/GDPR-equivalent regulations, even under private deployment scenarios if online guideline updates are required[10-12].

reasoning from minimal examples remains a core unsolved problem for frontier foundation models such as GPT-5 and Grok 4. These models still fail to infer structured transformation rules from a handful of examples, which is a key hallmark of human intelligence. The Abstraction and Reasoning Corpus for Artificial General Intelligence (ARC-AGI) provides a rigorous testbed for this capability, demanding conceptual rule induction and transfer to novel tasks. Most existing methods treat ARC-AGI as a purely textual reasoning task, overlooking the fact that humans rely heavily on visual abstraction when solving such puzzles. However, our pilot experiments reveal a paradox: naively rendering ARC-AGI grids as images degrades performance due to imprecise rule execution. This leads to our central hypothesis that vision and language possess complementary strengths across distinct reasoning stages: vision supports global pattern abstraction and verification, whereas language specializes in symbolic rule formulation and precise execution. Building on this insight, we introduce two synergistic strategies: (1) Vision-Language Synergy Reasoning (VLSR), which decomposes ARC-AGI into modality-aligned subtasks; and (2) Modality-Switch Self-Correction (MSSC), which leverages vision to verify text-based reasoning for intrinsic error correction. Extensive experiments demonstrate that our approach yields up to a 4.33% improvement over text-only baselines across diverse flagship models and multiple ARC-AGI tasks. Our findings suggest that unifying visual abstraction with linguistic reasoning is a crucial step toward achieving general-izable, human-like intelligence in future foundation models. Source code is released at https://github.com/

Although large language models (LLMs) are increasingly trained using human feedback for safety and alignment with human values, alignment decisions often overlook human social diversity. This study examines how incorporating pluralistic values affects LLM behavior by systematically evaluating demographic variation and design parameters in the alignment pipeline. We collect alignment data from US and German participants (N = 1,095 participants, 27,375 ratings) who rated LLM responses across five dimensions: Toxicity, Emotional Awareness (EA), Sensitivity, Stereotypical Bias, and Helpfulness. We fine-tuned multiple Large Language Models and Large Reasoning Models using preferences from different social groups while varying rating scales, disagreement handling methods, and optimization techniques. The results revealed systematic demographic effects: male participants rated responses 18% less toxic than female participants; conservative and Black participants rated responses 27.9% and 44% higher on EA than liberal and White participants, respectively. Models fine-tuned on group-specific preferences exhibited distinct behaviors. Technical design choices showed strong effects: the preservation of rater disagreement achieved roughly 53% greater toxicity reduction than majority voting, and 5-point scales yielded about 22% more reduction than binary formats; and Direct Preference Optimization (DPO) consistently outperformed Group Relative Policy Optimization (GRPO) in multi-value optimization. These findings represent a preliminary step in answering a critical question: How should alignment balance expert-driven and user-driven signals to ensure both safety and fair representation?

Synthetic Aperture Radar (SAR) is a crucial imaging modality thanks to its all-weather capability. Although recent advances in self-supervised learning and masked image modeling (MIM) have enabled SAR foundation models, these methods largely emphasize low-level visual features and often overlook multimodal alignment and zero-shot target recognition in SAR imagery. T o address this, we construct SARVLM-1M, a large-scale vision-language dataset with over one million image-text pairs aggregated from existing datasets. W e further propose a domain transfer training strategy to mitigate the large gap between natural and SAR imagery. Building on this, we develop SARVLM, the first vision language foundation model (VLM) tailored to SAR, comprising SARCLIP and SARCap. SARVLM is trained with a vision-language contrastive objective under the proposed domain transfer strategy, bridging SAR imagery and textual descriptions. Extensive experiments on image text retrieval, zero-shot classification, semantic localization, and imagery captioning demonstrate that SARVLM delivers superior feature extraction and interpretation, outperforming state-of-the-art VLMs and advancing SAR semantic understanding. Code and datasets will be released soon.

Despite their remarkable capabilities, Large Language Models (LLMs) struggle to effectively leverage historical interaction information in dynamic and complex environments. Memory systems enable LLMs to move beyond stateless interactions by introducing persistent information storage, retrieval, and utilization mechanisms. However, existing memory systems often introduce substantial time and computational overhead. To this end, we introduce a new memory system called LightMem, which strikes a balance between the performance and efficiency of memory systems. Inspired by the Atkinson-Shiffrin model of human memory, LightMem organizes memory into three complementary stages. First, cognition-inspired sensory memory rapidly filters irrelevant information through lightweight compression and groups information according to their topics. Next, topic-aware short-term memory consolidates these topic-based groups, organizing and summarizing content for more structured access. Finally, long-term memory with sleep-time update employs an offline procedure that decouples consolidation from online inference. On LongMemEval and LoCoMo, using GPT and Qwen backbones, LightMem consistently surpasses strong baselines, improving QA accuracy by up to 7.7% / 29.3%, reducing total token usage by up to 38x / 20.9x and API calls by up to 30x / 55.5x, while purely online test-time costs are even lower, achieving up to 106x / 117x token reduction and 159x / 310x fewer API calls. The code is available at https://github.com/zjunlp/LightMem.

As an Indo-Aryan language with limited available data, Chakma remains largely underrepresented in language models. In this work, we introduce a novel corpus of contextually coherent Bangla-transliterated Chakma, curated from Chakma literature, and validated by native speakers. Using this dataset, we fine-tune six encoder-based transformer models, including multilingual (mBERT, XLM-RoBERTa, DistilBERT), regional (BanglaBERT, IndicBERT), and monolingual English (DeBERTaV3) variants on masked language modeling (MLM) tasks. Our experiments show that fine-tuned multilingual models outperform their pre-trained counterparts when adapted to Bangla-transliterated Chakma, achieving up to 73.54% token accuracy and a perplexity as low as 2.90. Our analysis further highlights the impact of data quality on model performance and shows the limitations of OCR pipelines for morphologically rich Indic scripts. Our research demonstrates that Bangla-transliterated Chakma can be very effective for transfer learning for Chakma language, and we release our dataset to encourage further research on multilingual language modeling for low-resource languages.

Large language models (LLM) have emerged as a promising avenue for time series forecasting, offering the potential to integrate multimodal data. However, existing LLM-based approaches face notable limitations-such as marginalized role in model architectures, reliance on coarse statistical text prompts, and lack of interpretability. In this work, we introduce Augur, a fully LLM driven time series forecasting framework that exploits LLM causal reasoning to discover and use directed causal associations among covariates. Augur uses a two stage teacher student architecture where a powerful teacher LLM infers a directed causal graph from time series using heuristic search together with pairwise causality testing. A lightweight student agent then refines the graph and fine tune on high confidence causal associations that are encoded as rich textual prompts to perform forecasting. This design improves predictive accuracy while yielding transparent, traceable reasoning about variable interactions. Extensive experiments on real-world datasets with 26 baselines demonstrate that Augur achieves competitive performance and robust zero-shot generalization.