AI
Fragility of Value under Imperfect Alignment
Researchers have proposed a model to address concerns about AI systems being aligned with human values. They argue that optimizing too heavily for imp...
AI
AI
Hybrid LLM-Augmented Reinforcement Learning Agents for Complex Sequential Decision Tasks
Researchers have proposed a new approach to creating autonomous agents that can tackle complex decision-making tasks. The method combines the strength...
AI
BrainBench: Benchmarking Large Language Models for Comprehensive EEG Understanding
Researchers have developed a new benchmark called BrainBench to evaluate the ability of large language models (LLMs) to understand electroencephalogra...
AI
Mechanist: AI as a Scientific Instrument for Discovering the Mechanisms of Intelligence
Researchers have developed Mechanist, an AI system designed to use machine learning models as scientific instruments for discovering the underlying me...
AI
Reconstruction: A Blind Benchmark for Recovering Research Ideas from Pre-Publication Bibliographies
Researchers have created a benchmark to test whether language models can recover the original research idea behind a published paper based on its pre-...
AI
GRIP: Grounded Reasoning via Information-Restricted Premises
Researchers have developed a new approach called GRIP, which aims to improve the performance of retrieval-augmented generation (RAG) models. These mod...
AI
Automated Computational Energy Minimization of ML Algorithms using Constrained Bayesian Optimization
Researchers have developed a method to optimize machine learning algorithms for low energy consumption while maintaining their performance. The approa...
AI
`From Prompt to Perturbation': An Adaptive Framework for Voice-Based Jailbreaks on Audio LLMs
Researchers have developed a new framework for testing the vulnerability of large language models (LLMs) to audio-based attacks. The framework, called...
AI
Iterative Flow Matching: Path Correction and Gradual Refinement for Enhanced Generative Modeling
Researchers propose an iterative process to improve generative modeling for image generation. They use flow matching, a technique that can lead to hal...
AI
Sleeping Kelly
Researchers have revisited the 'Sleeping Beauty problem', a thought experiment in decision-making under uncertainty. The problem involves Sleeping Bea...
AI
Jailbreaking in the Haystack
Researchers have developed a method called NINJA to 'jailbreak' large language models by appending harmless content to user goals. This allows attacke...
AI
CausalProfiler: Generating Synthetic Benchmarks for Rigorous and Transparent Evaluation of Causal Machine Learning
Researchers have created CausalProfiler, a tool that generates synthetic benchmarks for evaluating causal machine learning methods. These benchmarks a...