Latest AI Papers: RAG, Agents & LLMs

Hey guys! Here's a breakdown of the latest research papers, focusing on some seriously cool areas like RAG, Agents, SFT, RLHF, LLMs, Function Calling, and LLM Tabular data. I've organized everything for you, making it easier to dive into the papers that pique your interest. For a better reading experience and more papers, be sure to check out the Github page.

RAG

Let's kick things off with Retrieval-Augmented Generation (RAG). RAG is all about making language models smarter by letting them pull information from external sources. It's like giving your LLM a super-powered memory and the ability to look things up. This is super important because it can help LLMs give more accurate, detailed, and up-to-date responses. Here's what's new:

• An Ontology-Driven Graph RAG for Legal Norms: A Structural, Temporal, and Deterministic Approach: This paper focuses on using a graph-based RAG system for understanding legal rules. It leverages an ontology to structure and organize legal information, making it easier to retrieve relevant data. They've also focused on clarifying the terminology and refining the core contributions of their work.
• Inteligencia Artificial jurídica y el desafío de la veracidad: análisis de alucinaciones, optimización de RAG y principios para una integración responsable: This one is in Spanish and English. It dives into the challenges of truthfulness in legal AI, looking at hallucinations (when AI makes stuff up), how to optimize RAG systems, and how to responsibly integrate these technologies.
• MetaRAG: Metamorphic Testing for Hallucination Detection in RAG Systems: This paper explores using metamorphic testing to detect when RAG systems are hallucinating or making things up. It's a clever way to check the reliability of these systems.
• Automated Evidence Extraction and Scoring for Corporate Climate Policy Engagement: A Multilingual RAG Approach: Uses a multilingual RAG approach to automatically extract and score evidence related to corporate climate policy engagement. Pretty useful for companies trying to understand the climate impact.
• Speaking at the Right Level: Literacy-Controlled Counterspeech Generation with RAG-RL: This one focuses on generating counterspeech (responses to harmful speech) that's tailored to the audience's literacy level. It uses a combination of RAG and Reinforcement Learning (RL).
• LLM Ensemble for RAG: Role of Context Length in Zero-Shot Question Answering for BioASQ Challenge: Explores using an ensemble of LLMs within a RAG framework for zero-shot question answering, with a focus on the role of context length. It was presented at CEUR-WS, CLEF2025.
• All for law and law for all: Adaptive RAG Pipeline for Legal Research: This paper looks at creating an adaptive RAG pipeline specifically designed for legal research, helping legal professionals find the information they need more efficiently.
• TextlessRAG: End-to-End Visual Document RAG by Speech Without Text: An interesting paper exploring how to do RAG using speech input directly, without needing to transcribe the speech to text first. It's all about visual documents and cutting out the middleman.
• CyberRAG: An Agentic RAG cyber attack classification and reporting tool: This paper focuses on building a RAG system that acts as an agent to classify and report cyber attacks. Useful for cybersecurity!.
• An Iterative LLM Framework for SIBT utilizing RAG-based Adaptive Weight Optimization: It uses RAG to help optimize the weights in a framework.
• A Multimodal RAG Framework for Housing Damage Assessment: Collaborative Optimization of Image Encoding and Policy Vector Retrieval: This paper looks at a multimodal RAG framework for assessing housing damage, combining image encoding and policy vector retrieval for better assessment.
• Aligning LLMs for the Classroom with Knowledge-Based Retrieval -- A Comparative RAG Study: Examines how to align LLMs for use in the classroom, using knowledge-based retrieval as part of the process. It compares different approaches and was submitted to the IEEE.
• FinRAGBench-V: A Benchmark for Multimodal RAG with Visual Citation in the Financial Domain: This paper introduces a benchmark for evaluating RAG systems in the financial domain that includes visual citations. This helps with accuracy and reliability when dealing with financial data.
• Towards End-to-End Model-Agnostic Explanations for RAG Systems: This paper aims to create explanations for how RAG systems work. It helps us to understand why these systems give the answers they do. It was accepted at a workshop at SIGIR 2025.
• Domain-Aware RAG: MoL-Enhanced RL for Efficient Training and Scalable Retrieval: Focuses on making RAG systems more efficient and scalable by using domain-aware techniques and a Mixture of LoRA (MoL)-enhanced Reinforcement Learning (RL) approach. Great stuff for improving performance.

Agent

Next up, we've got Agents. These are AI systems designed to act autonomously, make decisions, and interact with their environment to achieve goals. Think of them as smart little robots that can explore, learn, and solve problems on their own. The latest research covers a wide range of applications:

• DeepDive: Advancing Deep Search Agents with Knowledge Graphs and Multi-Turn RL: This paper focuses on improving deep search agents by using knowledge graphs and multi-turn Reinforcement Learning (RL). It is useful for improving search capabilities.
• DECAMP: Towards Scene-Consistent Multi-Agent Motion Prediction with Disentangled Context-Aware Pre-Training: This focuses on predicting the movements of multiple agents in a scene, ensuring the predictions are consistent with the environment. This has applications in autonomous driving and robotics.
• Abduct, Act, Predict: Scaffolding Causal Inference for Automated Failure Attribution in Multi-Agent Systems: This paper focuses on figuring out why things fail in multi-agent systems. The authors used causal inference to do so.
• RecoWorld: Building Simulated Environments for Agentic Recommender Systems: This paper is all about creating simulated environments where agents can interact with and learn from recommender systems. It's like a training ground for improving how recommendations are made.
• Web3 x AI Agents: Landscape, Integrations, and Foundational Challenges: Explores the intersection of Web3 and AI agents, looking at the current landscape, how they can be integrated, and the main challenges to be solved.
• Robot guide with multi-agent control and automatic scenario generation with LLM: Focuses on the design of a robot guide system that uses multi-agent control and LLMs to generate scenarios automatically. The guide is a great way to show how the technologies can be used together.
• We Need a New Ethics for a World of AI Agents: This paper delves into the ethical considerations of AI agents and the need for new ethical frameworks as these agents become more prevalent in our world.
• A Holistic Architecture for Monitoring and Optimization of Robust Multi-Agent Path Finding Plan Execution: This paper is about creating a complete system for monitoring and optimizing the execution of multi-agent pathfinding plans.
• Compartmentalised Agentic Reasoning for Clinical NLI: Explores using agentic reasoning for clinical Natural Language Inference (NLI). It's about helping AI understand and reason about medical information.
• Towards Fully Automated Molecular Simulations: Multi-Agent Framework for Simulation Setup and Force Field Extraction: This paper aims to fully automate molecular simulations using a multi-agent framework. It covers the setup of simulations and the extraction of force fields, which is pretty complex.
• Federated Multi-Agent Reinforcement Learning for Privacy-Preserving and Energy-Aware Resource Management in 6G Edge Networks: Uses Federated Multi-Agent Reinforcement Learning in 6G edge networks to manage resources while preserving privacy and conserving energy. Great for communications!.
• Virtual Agent Economies: Explores the idea of virtual agent economies, likely focusing on how AI agents can interact, trade, and create economic systems within virtual environments.
• Tangential Action Spaces: Geometry, Memory and Cost in Holonomic and Nonholonomic Agents: This paper looks at the geometry, memory, and cost aspects of agents in holonomic (can move in any direction) and nonholonomic (restricted movement) scenarios.
• XAgents: A Unified Framework for Multi-Agent Cooperation via IF-THEN Rules and Multipolar Task Processing Graph: It is about creating a framework for multi-agent cooperation, using IF-THEN rules and a special type of graph for processing tasks.
• Exploring a Gamified Personality Assessment Method through Interaction with LLM Agents Embodying Different Personalities: Focuses on a gamified approach to personality assessment, using LLM agents that embody different personalities. It offers a fresh perspective on human-computer interaction.

SFT

Now, let's move on to Supervised Fine-Tuning (SFT). This is a crucial step in training language models. It involves taking a pre-trained model and then