Large Language Agents: Architectures
Welcome to our research hub on Large Language Agents: Architectures. Here you will find papers that show how Large Language Agents be patterned and chained together to replicate useful workflows
Large Language Agents in Education and Knowledge Work
This paper introduces “Interactionalism,” a framework for learning design in the GenAI era, emphasizing interactional intelligence—meta-cognitive and meta-emotional skills critical for working with AI agents. It proposes actionable pedagogical strategies to foster human-AI collaboration in skill development.
This study evaluates GPT-4’s effectiveness in financial statement analysis, demonstrating superior directional earnings predictions compared to human analysts, especially in challenging scenarios. GPT-4 matches state-of-the-art ML models in accuracy and generates valuable narrative insights. Trading strategies based on GPT-4 yield higher Sharpe ratios and alphas, indicating its potential for central
This paper advocates for developing AI systems as collaborative “thought partners” that engage in shared reasoning, complement human cognition, and build human-compatible models. It proposes a Bayesian framework for designing such systems, emphasizing reliability, trust, and mutual understanding in human-AI partnerships.
This paper explores the transformative potential of generative AI in human learning, highlighting its ability to personalize education, diversify resources, and enhance assessments. Key challenges include ethical concerns, model limitations, and traditional assessment disruption. Rigorous research and AI literacy are essential for effective integration.
This study, conducted with Boston Consulting Group, investigates AI’s impact on knowledge-intensive tasks using GPT-4. Consultants with AI access completed tasks 25% faster and produced 40% higher quality outcomes. Two successful AI integration patterns emerged: “Centaurs,” who selectively delegated tasks to AI, and “Cyborgs,” who fully integrated AI into their
Large Language Agents: Architectures
The paper introduces Cognitive Architectures for Language Agents (CoALA), a framework that organizes and guides the development of language agents by integrating principles from cognitive science and symbolic AI. CoALA proposes modular memory, structured action spaces, and a generalized decision-making process. The framework aims to contextualize current advancements and direct
Building Agents that Help Students Learn
Papers showing how LLM’s and LLA’s can be used as dialogical learning agents for enhanced student learning experiences and outcomes
The paper addresses the challenge of leveraging generative AI (gen AI) for equitable, high-quality education. The authors identify difficulties in translating pedagogical principles into effective AI prompts and the lack of robust evaluation practices as key obstacles. They introduce LearnLM-Tutor, a fine-tuned model based on Gemini, guided by seven educational
The paper introduces an AI-driven solution using open-source LLaMA-2 models to improve scalability in answering student questions on online platforms. By integrating retrieval-augmented generation, fine-tuning, and preference optimization, the approach achieves a 30% improvement in answer quality on a Piazza dataset. This work lays the foundation for customizable AI teaching
Building Evaluation and Feedback Agents
This paper explores the enhancement of large language models (LLMs) for personalized education, specifically in dialog tutoring. Current LLMs struggle to accurately detect student errors and tailor feedback. The authors introduce error detection verifiers grounded in real-world teaching practices and demonstrate, through empirical evaluations, that these verifiers improve the accuracy
The paper introduces PROMETHEUS, an open-source large language model (LLM) designed to evaluate long-form responses with customizable criteria, addressing limitations of proprietary LLMs like GPT-4. Trained on a new dataset, FEEDBACK COLLECTION, PROMETHEUS achieves a Pearson correlation of 0.897 with human evaluators, matching GPT-4’s performance. The model also outperforms other