Model releases, architecture work and prompting research on large language models — from frontier-lab announcements to the arXiv papers behind them.
arXiv:2603.19896v2 Announce Type: replace Abstract: Tool-using large language model (LLM) agents often face a fundamental tension between answer quality and execution cost. Fixed workflows are stable...
arXiv:2609.14796v1 Announce Type: new Abstract: The threat that AI persuasion poses to human control has been acknowledged in the literature, but not yet systematically studied. Now that persuasion a...
arXiv:2609.14819v1 Announce Type: cross Abstract: Large language models (LLMs) have a growing range of applications in medicine, and their evaluation is critical for ensuring they provide benefit and...
The paper introduces Vocalizer, a mobile app that lets users submit spoken online reviews enhanced by a large language model. A longitudinal study shows that users often use the AI to add detail and that interactive AI features boost confidence and willingness to share reviews. The authors also outline the benefits and challenges of embedding AI assistance in review-writing systems.
The paper introduces Counteraction-Aware Multi-Teacher On-Policy Distillation (CaMOPD), a method designed to recover general capabilities in large language models while preserving domain-specific behavior. CaMOPD tackles two failure modes of standard Multi-Teacher On-Policy Distillation—conflicting recovery and preservation gradients, and weak correction signals—by using decoupled alternating training and selecting samples with large teacher‑student log‑probability gaps. Experiments on role‑play dialogue and medical reasoning QA show that CaMOPD outperforms baselines in general capability recovery while maintaining domain specialization, and gradient coherence analyses confirm more coherent correction signals.
arXiv:2603.18740v3 Announce Type: replace-cross Abstract: Automated Code Review (ACR) systems integrating Large Language Models (LLMs) are increasingly adopted in software development workflows, rang...
HGSQ is a real‑time aerial small‑object detector that uses a heatmap‑guided sparse query strategy to focus computation on foreground regions. It introduces a lightweight Heatmap Budget Predictor to generate a foreground budget map, and then employs Heatmap‑Guided Sparse Query Selection, Heatmap‑Gated Lite Snake Convolution, and Adaptive Query‑Decoder Budgeting to efficiently process only small‑object areas. On NWPU VHR‑10 and VisDrone2019, HGSQ achieves 95.10 mAP50 and 54.8 mAP50 respectively while running at 96 FPS with only 48.6 GFLOPs on an RTX 4070.
CHAI for LLMs is a framework that improves large language models’ performance on code‑mixed translation tasks by using LLMs as annotators to create preference data, applying reinforcement learning from AI feedback, incorporating LLM‑generated domain knowledge for iterative refinement, and evaluating on real‑world datasets. The approach yields a 68.45% average win rate over state‑of‑the‑art open‑source models in human‑adjudicated tests. It demonstrates a scalable method to enhance code‑mixed language understanding in open‑source LLMs.
The paper argues that diffusion-based action policies can use a frozen, observation‑free backbone as a reusable trajectory prior, with task adaptation handled entirely by the conditioning pathway. By pretraining a general action head on forward‑kinematics data and then freezing it, the authors show that a single backbone can match or outperform normally trained models on MimicGen and LIBERO. Their experiments reveal that a small 5 M‑parameter MLP backbone can rival large U‑Net and transformer backbones, indicating that action backbones are often over‑parameterized and that image‑style architectures may not be the best fit for low‑dimensional action generation.
The paper introduces Semantic Confusion to assess how consistently large language models refuse similar prompts. It presents ParaGuard, a 10k‑prompt corpus of controlled paraphrase clusters, and proposes three token‑level metrics—Confusion Index, Confusion Rate, and Confusion Depth—to measure contradictory refusal decisions across meaning‑preserving paraphrases. Experiments show that global false rejection rates can mask local inconsistencies, revealing that refusal evaluation must consider both frequency and consistency across nearby paraphrases.
The paper introduces an enhanced end‑to‑end circuit analysis framework built on Gemini 2.5 Pro, targeting engineering education. It addresses two key failure modes—circuit‑recognition hallucinations and reasoning‑process hallucinations—by adding a YOLO detector for source polarity re‑identification and an ngspice verification loop for iterative refinement. The resulting pipeline achieves 97.59 % accuracy on 83 undergraduate problems, markedly outperforming the baseline Gemini model and demonstrating significant gains across varied diagram styles and textbooks.
The paper investigates the energy costs of multilingual large language model (LLM) inference, revealing significant disparities across languages. Using the ML.Energy framework, the authors find that energy consumption per output token can differ by up to 8.3×, and total energy for a fixed request set can vary up to 179×, with English being the cheapest and Pashto the most expensive. The study attributes these differences to higher per-token costs for complex or rare scripts and longer outputs for low‑resource languages, and notes that high‑energy languages also tend to have lower task accuracy.
MoARa introduces a module-aware rank allocation strategy and a block-wise magnitude-direction decomposition to improve low-rank gradient projection for large language model pre‑training. By profiling Transformer modules and tailoring projection ranks, it reduces the number of steps and wall‑clock time needed to reach target perplexity. Experiments on Llama, Qwen, and DeepSeek models show up to 41.7% fewer steps and 37.1% less training time with minimal memory overhead.
EventVL introduces the first generative event-based multimodal large language model (MLLM) designed for explicit semantic understanding of event streams. The framework leverages a newly annotated dataset of nearly 1.4 million event–image/video–text pairs and incorporates an Event Spatiotemporal Representation to capture comprehensive event information, along with Dynamic Semantic Alignment to refine sparse semantic spaces. Experiments demonstrate that EventVL outperforms existing MLLM baselines in event captioning and scene description generation tasks, advancing the field of event vision.
The paper investigates how AI‑generated data, treated as anomalies linked to main data points, influences batch decompositions and undersampling in random datasets. Using redundancy graphs and iterative methods, it derives bounds for the minimum size of a strongly dissimilar decomposition and shows a phase transition: when anomalies are few, the minimum size depends on main data, but beyond a threshold it is dominated by anomalies. Additionally, the authors present a size criticality result for strong similarity in randomly undersampled datasets, illustrated with categorical data examples where the overall space far exceeds the dataset size.
TwinICL is a procedurally generated benchmark that pairs matched text and image versions of tasks to enable controlled comparison of in‑context learning (ICL) across modalities. Experiments on six open‑weight models and 38 tasks show that multimodal ICL consistently underperforms text‑only ICL, with varying gaps by task family. Interventions targeting visual access, task framing, and reasoning can recover strong multimodal performance on a diagnostic subset, yet a modality gap remains even when explicit task instructions are provided, highlighting the dual role of demonstrations as context and evidence.
SAILOR is a proof‑of‑concept system that helps language models translate natural‑language optimization problem descriptions into executable code by detecting missing numerical values. It asks users targeted follow‑up questions, prioritizing them based on uncertainty and solver estimates of impact, and updates the model before returning a solution. In tests on 1,723 benchmark instances, SAILOR achieved exact objective‑value agreement between 27.0% and 87.6% while asking an average of 1.4–5.7 questions per instance.
The paper introduces scKITE, a single-cell foundation model that incorporates biological knowledge—cell-level text annotations and gene-level regulatory information—into a shared Transformer encoder via lightweight auxiliary decoders used only during pretraining. This approach provides a new scaling dimension beyond merely increasing data size, enabling the model to achieve superior performance on diverse downstream tasks with only 179,067 pretraining samples, less than 0.5% of the data used by previous strong scFMs. The study demonstrates that knowledge-enhanced pretraining can yield significant gains while reducing computational cost.
The paper introduces FlockMTL, an extension for database management systems that deeply integrates large language models and retrieval‑augmented generation into DuckDB. It provides model‑driven scalar and aggregate functions, cost‑based optimizations like batching and caching, and new SQL DDL abstractions (PROMPT and MODEL) to treat LLMs as first‑class schema objects. These features aim to simplify the development of knowledge‑intensive analytical applications by reducing the effort required to orchestrate heterogeneous data systems and manage LLM context.
SynGhost is a novel task‑agnostic backdoor attack that injects invisible syntactic backdoors into pre‑training corpora of language models. It uses an entropy‑based poisoning filter, contrastive learning to select optimal targets, and an awareness module to reduce interference between backdoors, thereby preserving the model’s pre‑training performance. Experiments demonstrate that SynGhost can transfer to multiple downstream tasks and withstand several defense mechanisms such as perplexity checks, fine‑pruning, and the maxEntropy filter.