Papers

AIThis paper introduces 'The AI Scientist', a pipeline that automates the entire scientific research lifecycle. The system generates ideas, writes code, runs experiments, analyzes data, writes the full manuscript, and performs its own peer review, with the generated manuscript passing the first round of peer review for a top-tier machine learning conference workshop.

AIThis paper systematically organizes the latest techniques in LLM across four dimensions: pre-training, post-training, utilization, and evaluation. It also identifies key research issues and challenges such as theoretical foundations, efficient scaling, alignment, and agentic capabilities.

AIAxiom is an RNN that constructs a unitary transition matrix via a product of Householder reflections, enabling lossless processing of sequences of arbitrary length. It achieves 76.5-99.9% accuracy on the delayed copy task with 13x fewer parameters than LSTM, and when attached to GPT-2, it retrieves facts from 7 chunks back with 62.3% accuracy.

AIWe prove in four stages that the token embedding layer is the geometric foundation of transformer attention. Fourier-based training techniques (PFFT, FGP) and No-Q attention preserve the embedding's geometric authority and achieve lightweight performance.

AIWe propose Cognitive Impedance Matching Theory (CIMT), a compiler theory that enhances the capabilities of a fixed LLM model through world-side interface, validation, repair, and audit design. Human evaluators or LLM judges are not treated as privileged evaluators but modeled as fallible measurement channels, providing a framework to prove system reliability using only observable atomic data.

AICo-Scientist is a multi-agent AI system built on Gemini that assists scientists by generating and refining novel research hypotheses based on their objectives and existing evidence. It uses an asynchronous task execution framework and a tournament evolution process to improve hypothesis quality, demonstrating practical value in biomedical applications like drug repurposing.

AIA field experiment with 758 knowledge workers at a global consulting firm examined GPT-4's impact on task performance. The study found that while AI significantly boosted productivity and quality on most tasks, it hindered performance on a complex managerial task, revealing a 'jagged technological frontier'.

AIThis paper presents an interface theory that distinguishes genuine improvement from illusion in AI systems that change their own evaluation criteria. It develops a mathematical framework ensuring the stability of self-improvement loops through replayable observable interfaces and certified stable gains.

AIThis study quantitatively evaluates two specialized clinical AI tools against three frontier general-purpose LLMs across three medical benchmarks. The frontier LLMs outperformed the clinical AI tools in all evaluations, which performed comparably to an automated Google Search AI Overview on the real clinical queries benchmark.

AIProposes the Spiral Time framework, which represents time as 2D spiral coordinates (radius=trend, angle=seasonality). In LSTM and Transformer experiments, it improves MAPE by 83% over scalar time, with monotonic performance gains across all experiments.

AIThis study analyzes the practical effectiveness of differential privacy (DP) guarantees when adapting large language models (LLM) for sensitive applications. The researchers benchmark privacy risks using state-of-the-art attacks while systematically varying the adaptation data distribution, identifying key factors for achieving practical privacy protection.

AIThis paper provides a comprehensive survey of Multimodal Large Language Models (MLLMs) focused on vision-language tasks such as image captioning and visual question answering. It examines MLLM architectures, training pipelines, and practical applications, while highlighting fundamental constraints like information bottlenecks and data-processing limits.

AIQwen-Scope open-sources 14 groups of sparse autoencoders (SAEs) for Qwen3/3.5 series models and demonstrates their use in inference-time steering, evaluation analysis, data-centric workflows, and post-training optimization, showing that SAEs can serve as practical interfaces for model development beyond post-hoc analysis.

AIThis study analyzed 7.3 million academic articles from 2020 to 2025 to track the widespread adoption of large language models (LLMs) in scholarly research. The analysis found that by 2025, an estimated 57% of published articles showed evidence of LLM influence, with significant variation in adoption rates across regions, institutional ranks, and academic disciplines.

AIThis research identifies necessary fixes to scale Schedule-Free Learning for large language models. The proposed ScheduleFree+ method significantly outperforms Warmup-Stable-Decay schedules, especially in long-duration training.

AIThis research proposes 'natural identifiers (NIDs)' to address the challenges of privacy auditing for large language models (LLMs). NIDs are structured random strings naturally occurring in training datasets, enabling post-hoc differential privacy auditing without retraining and dataset inference without needing a private held-out dataset.

AIThis study proposes a variational recurrent neural network (VRNN) that integrates latent random variables into the hidden state of a recurrent neural network (RNN). Experiments on speech and handwriting data demonstrate that VRNN better models the variability of structured sequential data than existing models.

AIWe consider a partial observability model in bandit problems where the learner can observe losses of some other actions in addition to its own. We propose the first algorithm that guarantees near-optimal regret without prior knowledge of the observation system, and show that the implicit exploration strategy is more efficient computationally and information-theoretically than existing methods.

AITo address the difficulty of model selection in deep unsupervised domain adaptation (Deep UDA), we propose Deep Embedded Validation (DEV), which embeds adapted feature representations into the validation procedure to provide an unbiased estimator of the target domain risk. Variance is reduced using the control variate technique, and effectiveness is demonstrated theoretically and empirically.

AIChatlaw is a multi-agent legal assistant specialized in the Chinese legal system, using a Role-Aligned Mixture-of-Experts (RA-MoE) architecture. It achieves a 7.73% improvement in accuracy over GPT-4 on the LawBench benchmark and an 11-point higher score on the legal professional exam.

AIThis paper systematically categorizes and analyzes graph neural network (GNN) models that address heterophily in graphs. It summarizes various approaches to overcome the limitations of existing GNNs that assume homophily and suggests future research directions for heterophilic graph learning.

AIThis paper presents MicroGrowAgents, an AI-driven, agent-based system that automates the design of optimized microbial growth media by integrating knowledge graphs, metabolic modeling, and optimal experimental design. The system uses specialized agents to query biological knowledge, mine literature, and generate statistically optimal experiments, aiming to reduce experimental burden and accelerate the discovery of growth-promoting conditions.

AIThis paper proposes that the universe operates as a necessity system, where organizational constraint drives structural configurations toward the golden ratio through a recursive mechanism identical to the Fibonacci sequence. The theory aims to explain the cosmological constant without fine-tuning and presents five independent empirical pillars of support, including nuclear morphometry and high-redshift galaxy observations.

AIReveals that Multi-modal Large Language Models (MLLMs) risk leaking sensitive information embedded in images, and presents a comprehensive dataset MM-Privacy for evaluating these risks. Experiments confirm that various MLLMs expose personal information across different tasks, and task inconsistency increases privacy risks.

AIThis study examines the impact of weakened patent protection following the Alice Corp. vs. CLS Bank decision on firms' innovation, competition, acquisitions, lawsuits, and employment agreements. It uses large language models (LLMs) to identify the potential exposure of firms' patent portfolios to the Alice decision and investigates the resulting unequal impacts.

AIThis survey addresses the multilayered challenges of deploying large language models on edge hardware, including compression, compiler behavior, and system-level trade-offs. It provides a deployment-centric taxonomy of compression strategies, analyzing their interaction with hardware toolchains and highlighting limitations in current benchmarking suites.

AIAccurate carbon emission forecasting in power distribution networks is a critical challenge due to the integration of large-scale electric vehicles and renewable energy sources. This paper proposes CarbonGPT, a model that uses a causal encoder and a meta causal graph dictionary to address spurious correlations and enhance LLM-based prediction.

AIThis survey systematically analyzes methods for integrating external knowledge into LLMs to address hallucinations and knowledge gaps. It discusses parametric and non-parametric approaches to improve reasoning and factual accuracy in domain-specific tasks.

AIThis paper proposes BRIDGE, a benchmark for evaluating large language models on understanding real-world clinical practice texts. The benchmark includes various clinical tasks such as disease diagnosis, treatment planning, and patient status summarization.

AIThis research analyzes which brands are recommended within specific categories and the concentration of their ownership across large language models. The authors propose three exploratory metrics—Category Ownership Index (COI), Competitive Vacuum Index (CVI), and Displacement Score (DS)—and conduct an empirical analysis across 3 models, 5 industries, and 250 category queries, finding results that challenge a strong winner-takes-all narrative.