A comprehensive index of artificial intelligence and machine-learning research with AI-generated summaries, citation metrics, and direct links to papers and code.
Yuri Kuratov, A. Bulatov, Petr Anokhin, et al.
Recurrent memory augmentation enables GPT-2 to process sequences up to 11 million elements, far exceeding standard methods limited to 10,000 elements.
Hanzhuo Tan, Qi Luo, Jing Li, et al.
LLM4Decompile is the first open-source LLM series trained to decompile binary code, outperforming GPT-4o and Ghidra by over 100% in re-executability.
Zhuowan Li, Cheng Li, Mingyang Zhang, et al.
Compares RAG and long-context LLMs, finding long-context outperforms RAG but at higher cost, and proposes Self-Route to route queries based on model self-reflection.
Jinyang Wu, Feihu Che, Mingkuan Feng, et al.
Defines seven noise types in RAG, builds NoiserBench benchmark, and shows some noise can benefit LLMs while other noise harms them.
Annie Wong, Thomas H. W. Back, A. Plaat, et al.
Evaluates prompting strategies in dynamic environments, finding strategic prompting can close performance gaps but reveals persistent reasoning limitations in LLMs.
Samar Khanna, Siddhant Kharbanda, Shufan Li, et al.
Mercury introduces diffusion-based LLMs for coding, achieving up to 10x throughput over speed-optimized models while maintaining comparable quality.
Yuzhong Zhao, Yue Liu, Junpeng Liu, et al.
GMPO improves GRPO stability by replacing arithmetic mean with geometric mean of token rewards, reducing outlier sensitivity and boosting reasoning performance.
Kaiwen Wei, Rui Shan, Dongsheng Zou, et al.
MIRAGE enhances test-time scaling for medical QA by combining multi-path parallel inference with structured knowledge graph retrieval to reduce error accumulation and improve traceability.
Orion Weller, Michael Boratko, Iftekhar Naim, et al.
This paper proves that single-vector embedding models have a fundamental dimension-dependent limit on the number of distinct top-k subsets they can represent, and shows this limitation arises even in simple realistic queries.
Rachit Bansal, Aston Zhang, Rishabh Tiwari, et al.
Proposes test-time training via targeted gradient updates on context to overcome static self-attention limitations in long-context LLMs, outperforming inference-time scaling.
Yi Yu, Liuyi Yao, Yuexiang Xie, et al.
Proposes Agentic Memory (AgeMem), a unified framework integrating long-term and short-term memory management into LLM agent policy via tool-based actions and progressive reinforcement learning.
Yaoming Li, Guangxiang Zhao, Qilong Shi, et al.
First primer synthesizing over 150 studies on post-training reasoning data, organizing the field around four key questions.