Small large language models (LLMs) often face difficulties in aligning output to human preferences, particularly when operating under severe performance gaps. In this work, we propose two lightweight DPO-based variants -- Adaptive Margin-Sigmoid Loss and APO-hinge-zero -- to better address underperformance scenarios by introducing margin-based objectives and selective update mechanisms. Our APO-hinge-zero method, which combines hinge-induced hard-example mining with the chosen-focused optimization of APO-zero, achieves strong results. In AlpacaEval, APO-hinge-zero improves the win rate by +2.0 points and the length-controlled win rate by +1.4 points compared to the APO-zero baseline. In MT-Bench, our methods maintain competitive performance in diverse categories, particularly excelling in STEM and Humanities tasks. These results demonstrate that simple modifications to preference-based objectives can significantly enhance small LLM alignment under resource constraints, offering a practical path toward more efficient deployment.

Large language models (LLMs) excel across diverse tasks but face significant deployment challenges due to high inference costs. LLM inference comprises prefill (compute-bound) and decode (memory-bound) stages, with decode dominating latency particularly for long sequences. Current decoder-only models handle both stages uniformly, despite their distinct computational profiles. We propose OverFill, which decouples these stages to optimize accuracy-efficiency tradeoffs. OverFill begins with a full model for prefill, processing system and user inputs in parallel. It then switches to a dense pruned model, while generating tokens sequentially. Leveraging more compute during prefill, OverFill improves generation quality with minimal latency overhead. Our 3B-to-1B OverFill configuration outperforms 1B pruned models by 83.2%, while the 8B-to-3B configuration improves over 3B pruned models by 79.2% on average across standard benchmarks. OverFill matches the performance of same-sized models trained from scratch, while using significantly less training data. Our code is available at https://github.com/friendshipkim/overfill.

Structure elucidation is a fundamental technique for understanding the microscopic composition of matter and is widely applied across various disciplines in the natural sciences and engineering. However, existing methods often rely heavily on prior databases or known structural information, making it difficult to resolve unknown structures. In addition, complex structures typically require the joint analysis of multiple spectroscopic modalities. This process heavily depends on expert domain knowledge and is often accompanied by high costs in terms of both time and instrumentation. To address these challenges, we propose SpectraLLM, the first large language model designed to support multi-modal spectroscopic joint reasoning. SpectraLLM is capable of processing either single or multiple spectroscopic inputs and performing end-to-end structure elucidation. By integrating continuous and discrete spectroscopic modalities into a shared semantic space, SpectraLLM learns to uncover substructural patterns that are consistent and complementary across spectra, enabling precise molecular structure elucidation. We pretrain and fine-tune SpectraLLM in the domain of small molecules, and evaluate it on six standardized, publicly available chemical datasets. The model achieves state-of-the-art performance, significantly outperforming existing approaches trained on single modalities. Notably, SpectraLLM demonstrates strong robustness and generalization even for single-spectrum inference, while its multi-modal reasoning capability further improves the accuracy of structural prediction.

--Global KV-cache sharing has emerged as a key optimization for accelerating large language model (LLM) inference. However, it exposes a new class of timing side-channel attacks, enabling adversaries to infer sensitive user inputs via shared cache entries. Existing defenses, such as per-user isolation, eliminate leakage but degrade performance by up to 38.9% in time-to-first-token (TTFT), making them impractical for high-throughput deployment. T o address this gap, we introduce SafeKV (Secure and Flexible KV Cache Sharing), a privacy-aware KV-cache management framework that selectively shares non-sensitive entries while confining sensitive content to private caches. SafeKV comprises three components: (i) a hybrid, multi-tier detection pipeline that integrates rule-based pattern matching, a general-purpose privacy detector, and context-aware validation; (ii) a unified radix-tree index that manages public and private entries across heterogeneous memory tiers (HBM, DRAM, SSD); and (iii) entropy-based access monitoring to detect and mitigate residual information leakage. Our evaluation shows that SafeKV mitigates 94%-97% of timing-based side-cahnnel attacks. Compare to per-user isolation method, SafeKV improves TTFT by up to 40.58% and throughput by up to 2.66 across diverse LLMs and workloads. By combining fine-grained privacy control with high cache reuse efficiency, SafeKV reclaims the performance advantages of global sharing while providing robust runtime privacy guarantees for LLM inference. Large language models (LLMs) now underpin applications from dialogue to complex reasoning. To meet time-sensitive inference demands, key-value (KV) caching stores intermediate attention states ("keys" and "values") to eliminate redundant computation for sequential or similar prompts, thereby accelerating generation [70]. This efficiency gain is amplified through KV cache sharing across multiple requests. In particular, prompts with common prefixes, such as shared dialogue history or structured prompting patterns, enable substantial throughput improvements and latency reduction. Consequently, KV -cache sharing has become a critical mechanism for boosting throughput and reducing response latency in large-scale, multi-user LLM deployments. Empirical studies confirm that a substantial portion of real-world prompts exhibit prefix-level or structural overlap [42], [74], making shared KV reuse both practical and highly beneficial. Despite these performance benefits, KV cache sharing raises serious privacy and security concerns in shared or multi-tenant deployments. Specifically, KV -cache sharing across mutually untrusted users can lead to unintended information leakage.

Large language models (LLMs), especially Explicit Long Chain-of-Thought (CoT) reasoning models like DeepSeek-R1 and QWQ, have demonstrated powerful reasoning capabilities, achieving impressive performance in commonsense reasoning and mathematical inference. Despite their effectiveness, Long-CoT reasoning models are often criticized for their limited ability and low efficiency in knowledge-intensive domains such as molecule discovery. Success in this field requires a precise understanding of domain knowledge, including molecular structures and chemical principles, which is challenging due to the inherent complexity of molecular data and the scarcity of high-quality expert annotations. To bridge this gap, we introduce Mol-R1, a novel framework designed to improve explainability and reasoning performance of R1-like Explicit Long-CoT reasoning LLMs in text-based molecule generation. Our approach begins with a high-quality reasoning dataset curated through Prior Regulation via In-context D istillation ( PRID), a dedicated distillation strategy to effectively generate paired reasoning traces guided by prior regulations. Building upon this, we introduce MoIA, Molecular I terative Adaptation, a sophisticated training strategy that iteratively combines Supervised Fine-tuning (SFT) with Reinforced Policy Optimization (RPO), tailored to boost the reasoning performance of R1-like reasoning models for molecule discovery. Finally, we examine the performance of Mol-R1 in the text-based molecule reasoning generation task, showing superior performance against existing baselines.

Large Language Models (LLMs) are increasingly employed as AI tutors due to their scalability and potential for personalized instruction. However, off-the-shelf LLMs often underperform in educational settings: they frequently reveal answers too readily, fail to adapt their responses to student uncertainty, and remain vulnerable to emotionally manipulative prompts. To address these challenges, we introduce CoDAE, a framework that adapts LLMs for educational use through Chain-of-Thought (CoT) data augmentation. We collect real-world dialogues between students and a ChatGPT-based tutor and enrich them using CoT prompting to promote step-by-step reasoning and pedagogically aligned guidance. Furthermore, we design targeted dialogue cases to explicitly mitigate three key limitations: over-compliance, low response adaptivity, and threat vulnerability. We fine-tune four open-source LLMs on different variants of the augmented datasets and evaluate them in simulated educational scenarios using both automatic metrics and LLM-as-a-judge assessments. Our results show that models fine-tuned with CoDAE deliver more pedagogically appropriate guidance, better support reasoning processes, and effectively resist premature answer disclosure.

Digital Humanities (DH) is an interdisciplinary field that integrates computational methods with humanities scholarship to investigate innovative topics. Each academic discipline follows a unique developmental path shaped by the topics researchers investigate and the methods they employ. With the help of bibliometric analysis, most of previous studies have examined DH across multiple dimensions such as research hotspots, co-author networks, and institutional rankings. However, these studies have often been limited in their ability to provide deep insights into the current state of technological advancements and topic development in DH. As a result, their conclusions tend to remain superficial or lack interpretability in understanding how methods and topics interrelate in the field. To address this gap, this study introduced a new concept of Topic-Method Composition (TMC), which refers to a hybrid knowledge structure generated by the co-occurrence of specific research topics and the corresponding method. Especially by analyzing the interaction between TMCs, we can see more clearly the intersection and integration of digital technology and humanistic subjects in DH. Moreover, this study developed a TMC-based workflow combining bibliometric analysis, topic modeling, and network analysis to analyze the development characteristics and patterns of research disciplines. By applying this workflow to large-scale bibliometric data, it enables a detailed view of the knowledge structures, providing a tool adaptable to other fields.

Molecular representation learning, a cornerstone for downstream tasks like molecular captioning and molecular property prediction, heavily relies on Graph Neural Networks (GNN). However, GNN suffers from the over-smoothing problem, where node-level features collapse in deep GNN layers. While existing feature projection methods with cross-attention have been introduced to mitigate this issue, they still perform poorly in deep features. This motivated our exploration of using Mamba as an alternative projector for its ability to handle complex sequences. However, we observe that while Mamba excels at preserving global topological information from deep layers, it neglects fine-grained details in shallow layers. The capabilities of Mamba and cross-attention exhibit a global-local trade-off. To resolve this critical global-local trade-off, we propose Hierarchical and Structure-Aware Network (HSA-Net), a novel framework with two modules that enables a hierarchical feature projection and fusion. Firstly, a Hierarchical Adaptive Projector (HAP) module is introduced to process features from different graph layers. It learns to dynamically switch between a cross-attention projector for shallow layers and a structure-aware Graph-Mamba projector for deep layers, producing high-quality, multi-level features. Secondly, to adaptively merge these multi-level features, we design a Source-Aware Fusion (SAF) module, which flexibly selects fusion experts based on the characteristics of the aggregation features, ensuring a precise and effective final representation fusion. Extensive experiments demonstrate that our HSA-Net framework quantitatively and qualitatively outperforms current state-of-the-art (SOTA) methods.

For this week's Open Questions column, Cal Newport is filling in for Joshua Rothman. Much of the euphoria and dread swirling around today's artificial-intelligence technologies can be traced back to January, 2020, when a team of researchers at OpenAI published a thirty-page report titled "Scaling Laws for Neural Language Models." The team was led by the A.I. researcher Jared Kaplan, and included Dario Amodei, who is now the C.E.O. of Anthropic. They investigated a fairly nerdy question: What happens to the performance of language models when you increase their size and the intensity of their training? Back then, many machine-learning experts thought that, after they had reached a certain size, language models would effectively start memorizing the answers to their training questions, which would make them less useful once deployed.

Elon Musk has threatened legal action against Apple on behalf of his artificial intelligence startup xAI, accusing the iPhone maker of favoring OpenAI and breaching antitrust regulations in managing the rankings in its App Store. The posts elicited snide responses from Sam Altman, the OpenAI CEO, and began a spat between the two former business partners on X. "Apple is behaving in a manner that makes it impossible for any AI company besides OpenAI to reach #1 in the App Store, which is an unequivocal antitrust violation. In a post earlier that day, he wrote: "Hey @Apple App Store, why do you refuse to put either X or Grok in your'Must Have' section when X is the #1 news app in the world and Grok is #5 among all apps? OpenAI's ChatGPT currently holds the top spot in the App Store's "Top Free Apps" section in the US, while xAI's Grok ranks fifth. Apple has a partnership with OpenAI that integrates ChatGPT into iPhones, iPads and Macs.