Training AI Co-Scientists Using Rubric Rewards
AI co-scientists are emerging as a tool to assist human researchers in achieving their research goals. This work studies how to leverage existing research papers to train language models for better research plan generation, demonstrating significant improvements across multiple domains.
Eliciting Behaviors in Multi-Turn Conversations
This work presents an analytical framework for behavior elicitation in multi-turn conversations, exploring methods for generating effective test cases. The findings underline the importance of adapting evaluation methods to dynamic conversational contexts.
Fine-Tuning LLMs with Fine-Grained Human Feedback on Text Spans
A new method for fine-tuning language models based on structured feedback is proposed. The findings emphasize the efficacy of fine-grained feedback in achieving superior preference tuning, which can facilitate better conversational AI.
Improving Reasoning for Diffusion Language Models via Group Diffusion Policy Optimization
This paper addresses the challenges in adapting reinforcement learning to Diffusion Language Models (DLMs) and introduces a new algorithm that achieves significant improvements in reasoning benchmarks.
Multilingual Hidden Prompt Injection Attacks on LLM-Based Academic Reviewing
This research explores the vulnerabilities of LLMs in academic peer reviews due to hidden prompt injection, revealing significant differences in model susceptibility across languages.
Generative Lecture: Making Lecture Videos Interactive with LLMs and AI Clone Instructors
The paper introduces a generative AI framework that makes existing lecture videos interactive, enhancing personalized student learning experiences.
A Review of Community-Centric Power Systems Resilience Assessment and Enhancement Strategies
This review synthesizes resilience metrics and strategies for power system enhancements and highlights the integration of AI, offering insights into regulatory impacts.
AI tutoring can safely and effectively support students: An exploratory RCT in UK classrooms
The study evaluates the effectiveness of a generative AI tutoring model, showing positive results in student performance and tutor feedback.
Enhancing Web Payload Classification Using WAMM: An AI-Based Framework for Dataset Refinement and Model Evaluation
WAMM is introduced as an AI-driven framework for web attack detection, significantly improving performance by reclassifying HTTP requests in web applications.