A comprehensive index of artificial intelligence and machine-learning research with AI-generated summaries, citation metrics, and direct links to papers and code.
Mohammad Tavakoli, Alireza Salemi, Carrie Ye, et al.
Introduces BEAM, a benchmark for long-term memory in LLMs with up to 10M tokens, and LIGHT, a memory framework inspired by human cognition that improves performance by 3.5%-12.69%.
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.
Tom Everitt, Cristina Garbacea, Alexis Bellot, et al.
This paper introduces a method to evaluate the goal-directedness of LLMs by measuring how effectively they use their capabilities toward given goals, showing it is consistent across tasks and distinct from task performance.
Zhuoshi Pan, Qizhi Pei, Yu Li, et al.
Introduces REST, a stress-testing framework that evaluates large reasoning models by presenting multiple problems simultaneously, revealing performance degradation and stronger discriminative power than single-question benchmarks.
Ryan Sze-Yin Chan, Federico Nanni, Tomas Lazauskas, et al.
This paper presents a retrieval-augmented reasoning system using a lean Qwen2.5-Instruct model fine-tuned with synthetic reasoning traces, achieving near-frontier performance on domain-specific queries while remaining locally deployable.
Wenrui Bao, Zhiben Chen, Dan Xu, et al.
Learn2PD trains a lightweight filter to adaptively unmask tokens in diffusion LLMs, achieving up to 22.58x speedup without performance loss.
Nour Shaheen, Junwei Ma, Alex Labach, et al.
This paper investigates why tabular foundation models generalize surprisingly well in in-context learning, proposing a theoretical framework and empirical analysis to explain their performance.
Yizhou Fan, Luzhen Tang, Huixiao Le, et al.
A randomized experiment shows ChatGPT improves short-term writing performance but may induce metacognitive laziness, with no gains in intrinsic motivation or knowledge transfer.
Zeyu Ren, Ling Yue, Ran Li, et al.
FlowEvo is a training-free framework where workflows and executable skills co-evolve at inference time, improving agent performance across benchmarks while reducing token usage.
Adam Zweiger, Jyothish Pari, Han Guo, et al.
SEAL enables LLMs to self-adapt by generating their own finetuning data and update directives, using reinforcement learning to optimize downstream performance.
Andrew Birrell, Bruce Jay Nelson
This paper describes the design and implementation of a remote procedure call (RPC) package, covering binding, transport protocol, optimizations, and performance measurements.
Tzu‐Mao Li, Michaël Gharbi, Andrew Adams, et al.
This paper extends the Halide image processing language with reverse-mode automatic differentiation and automatic scheduling, enabling high-performance gradient computation for arbitrary programs.