A comprehensive index of artificial intelligence and machine-learning research with AI-generated summaries, citation metrics, and direct links to papers and code.
Claude E. Shannon
Shannon extends his mathematical theory of communication to continuous signals and messages via a limiting process from the discrete case, revealing new effects and specializations.
Unknown
The paper analyzes uncertainty calibration of process reward models, extending evaluation to a 72B parameter variant on MATH500 and AIME24-25 datasets.
Unknown
This paper shows that Process Reward Models can generalize from mathematical reasoning to code, enhancing LLM reasoning in test-time scaling across domains.
Giorgio Roffo
This tutorial explores advanced LLM techniques including RAG, PAL, ReAct, LangChain, fine-tuning strategies, and transformer architectures to address limitations like temporal cutoffs and mathematical inaccuracies.
Aviral Kumar, Vincent Zhuang, Rishabh Agarwal, et al.
SCoRe is a multi-turn online RL method that significantly improves LLM self-correction using only self-generated data, achieving state-of-the-art gains on MATH and HumanEval.
Arian Hosseini, Alessandro Sordoni, Daniel Toyama, et al.
This paper reveals a significant reasoning gap in LLMs when solving compositional math problems, showing that performance on standard benchmarks masks systematic differences in reasoning abilities.
Bogdan Georgiev, J. Gómez-Serrano, Terence Tao, et al.
This paper demonstrates AlphaEvolve, an LLM-guided evolutionary coding agent, as a tool for autonomously discovering novel mathematical constructions, rediscovering best-known solutions and improving several across 67 problems.
Yue Wang, Qiuzhi Liu, Jiahao Xu, et al.
Identifies 'underthinking' in o1-like LLMs, where frequent thought switching harms reasoning, and proposes a decoding penalty (TIP) to improve accuracy on math problems.
Unknown
A survey of the mathematical foundations of geometric deep learning, focusing on group equivariant and gauge equivariant neural networks.
Unknown
This paper proposes a mathematical philosophy of mechanistic interpretability, defining MI and identifying preconditions for its success.
Chia-Hsuan Lee, Sihui Dai, Mingyang Zhou, et al.
DASH assigns segment-level credit based on drift from correctness to reduce overthinking in reasoning language models, improving accuracy on competition math benchmarks.
Unknown
CoLD introduces a counterfactually-guided length debiasing method for Process Reward Models to mitigate length bias in mathematical reasoning evaluation.