A comprehensive index of artificial intelligence and machine-learning research with AI-generated summaries, citation metrics, and direct links to papers and code.
Unknown
Proposes stratified per-question probability evaluation and step-wise mitigation to detect and reduce benchmark contamination in LLM pretraining.
Unknown
The paper argues benchmark contamination is not a single failure but at least five distinct types, and proposes a taxonomy organized by which mitigation each type defeats.
Unknown
This paper identifies and addresses gradient pathologies in physics-informed neural networks, where loss terms of different nature cause training difficulties, and proposes mitigation strategies.
Unknown
This paper characterizes possible failure modes in physics-informed neural networks (PINNs), identifying causes and proposing mitigation strategies.
Unknown
This paper proposes a flexible definition of hallucination in foundation models for decision-making and reviews state-of-the-art detection and mitigation methods, including those requiring internal model access.
Unknown
This paper investigates hallucination detection and mitigation in language models, highlighting the importance of context and proposing an iterative retraining approach.
Unknown
This paper rigorously examines mitigation strategies for LLM benchmark data contamination, proposing a framework to evaluate their effectiveness and highlighting the inadequacy of current methods.
Unknown
This paper investigates whether model editing harms the general abilities of large language models and proposes regularization as a mitigation strategy.
Unknown
Agentrim introduces a tool risk mitigation framework for agentic AI systems, combining LLM-based reasoning with reinforcement learning to dynamically restrict tool permissions and reduce security risks.
M. Ferrag, N. Tihanyi, Djallel Hamouda, et al.
A unified threat model for LLM-agent ecosystems covering host-to-tool and agent-to-agent attacks, with over 30 techniques and mitigation strategies.
Manuel Cossio
This paper provides a comprehensive taxonomy of LLM hallucinations, arguing their theoretical inevitability and emphasizing the need for robust detection, mitigation, and human oversight.
Yung-Sung Chuang, Linlu Qiu, Cheng-Yu Hsieh, et al.
Proposes Lookback Lens, a simple hallucination detector using attention weight ratios, effective across tasks and models, and reduces hallucinations via classifier-guided decoding.