Human Cognition Inspired RAG with Knowledge Graph for Complex Problem Solving

Oct 13, 2025 by leolavanderia.com.br 94 views

Latest 15 Papers - March 13, 2025

Introduction

The field of artificial intelligence (AI) has witnessed tremendous growth in recent years, with advancements in various areas such as natural language processing (NLP), computer vision, and machine learning. In this article, we will discuss the latest 15 papers in the field of AI, covering topics such as retrieval-augmented generation (RAG), agent systems, self-supervised learning from text (SFT), reinforcement learning from human feedback (RLHF), and large language models (LLM).

RAG

RAG is a technique that combines the strengths of retrieval and generation to produce high-quality text. The following papers discuss various aspects of RAG:

RAG-Adapter: A Plug-and-Play RAG-enhanced Framework for Long Video Understanding: This paper proposes a RAG-enhanced framework for long video understanding, which can be used for various applications such as video summarization and question answering.
RAAD-LLM: Adaptive Anomaly Detection Using LLMs and RAG Integration: This paper presents a method for adaptive anomaly detection using LLMs and RAG integration, which can be used for various applications such as fraud detection and network intrusion detection.
OpenRAG: Optimizing RAG End-to-End via In-Context Retrieval Learning: This paper proposes an optimization method for RAG end-to-end via in-context retrieval learning, which can be used for various applications such as text classification and sentiment analysis.
OCR Hinders RAG: Evaluating the Cascading Impact of OCR on Retrieval-Augmented Generation: This paper evaluates the cascading impact of OCR on retrieval-augmented generation, which can be used for various applications such as document analysis and text recognition.
Machine Against the RAG: Jamming Retrieval-Augmented Generation with Blocker Documents: This paper proposes a method for jamming retrieval-augmented generation with blocker documents, which can be used for various applications such as text classification and sentiment analysis.
RAG-Enhanced Collaborative LLM Agents for Drug Discovery: This paper presents a method for RAG-enhanced collaborative LLM agents for drug discovery, which can be used for various applications such as drug discovery and personalized medicine.
SmartRAG: Jointly Learn RAG-Related Tasks From the Environment Feedback: This paper proposes a method for jointly learning RAG-related tasks from the environment feedback, which can be used for various applications such as text classification and sentiment analysis.
Human Cognition Inspired RAG with Knowledge Graph for Complex Problem Solving: This paper presents a method for human cognition-inspired RAG with knowledge graph for complex problem solving, which can be used for various applications such as text classification and sentiment analysis.
MM-PoisonRAG: Disrupting Multimodal RAG with Local and Global Poisoning Attacks: This paper proposes a method for disrupting multimodal RAG with local and global poisoning attacks, which can be used for various applications such as text classification and sentiment analysis.
FinTMMBench: Benchmarking Temporal-Aware Multi-Modal RAG in Finance: This paper presents a benchmarking method for temporal-aware multi-modal RAG in finance, which can be used for various applications such as financial analysis and risk management.
Beyond RAG: Task-Aware KV Cache Compression for Comprehensive Knowledge Reasoning: This paper proposes a method for task-aware KV cache compression for comprehensive knowledge reasoning, which can be used for various applications such as text classification and sentiment analysis.
More Documents, Same Length: Isolating the Challenge of Multiple Documents in RAG: This paper evaluates the challenge of multiple documents in RAG, which can be used for various applications such as text classification and sentiment analysis.
In-depth Analysis of Graph-based RAG in a Unified Framework: This paper presents an in-depth analysis of graph-based RAG in a unified framework, which can be used for various applications such as text classification and sentiment analysis.
LaRA: Benchmarking Retrieval-Augmented Generation and Long-Context LLMs -- No Silver Bullet for LC or RAG Routing: This paper presents a benchmarking method for retrieval-augmented generation and long-context LLMs, which can be used for various applications such as text classification and sentiment analysis.
Measuring and Enhancing Trustworthiness of LLMs in RAG through Grounded Attributions and Learning to Refuse: This paper proposes a method for measuring and enhancing trustworthiness of LLMs in RAG through grounded attributions and learning to refuse, which can be used for various applications such as text classification and sentiment analysis.

Agent

Agent systems are a type of AI system that can interact with their environment and make decisions based on their perception of the world. The following papers discuss various aspects of agent systems:

AgentOrca: A Dual-System Framework to Evaluate Language Agents on Operational Routine and Constraint Adherence: This paper proposes a dual-system framework for evaluating language agents on operational routine and constraint adherence, which can be used for various applications such as text classification and sentiment analysis.
Proto Successor Measure: Representing the Behavior Space of an RL Agent: This paper presents a method for representing the behavior space of an RL agent, which can be used for various applications such as text classification and sentiment analysis.
Collaborative Uncertainty Benefits Multi-Agent Multi-Modal Trajectory Forecasting: This paper proposes a method for collaborative uncertainty benefits multi-agent multi-modal trajectory forecasting, which can be used for various applications such as traffic prediction and route planning.
GTR: Guided Thought Reinforcement Prevents Thought Collapse in RL-based VLM Agent Training: This paper presents a method for guided thought reinforcement prevents thought collapse in RL-based VLM agent training, which can be used for various applications such as text classification and sentiment analysis.
Hierarchical Multi Agent DRL for Soft Handovers Between Edge Clouds in Open RAN: This paper proposes a method for hierarchical multi-agent DRL for soft handovers between edge clouds in open RAN, which can be used for various applications such as network optimization and resource allocation.
ToolFuzz -- Automated Agent Tool Testing: This paper presents a method for automated agent tool testing, which can be used for various applications such as software testing and quality assurance.
An autonomous rl agent methodology for dynamic Web ui testing in a bdd framework: This paper proposes a method for an autonomous RL agent methodology for dynamic Web UI testing in a BDD framework, which can be used for various applications such as software testing and quality assurance.
Learning to Detect Objects from Multi-Agent LiDAR Scans without Manual Labels: This paper presents a method for learning to detect objects from multi-agent LiDAR scans without manual labels, which can be used for various applications such as object detection and tracking.
Agent-Oriented Planning in Multi-Agent Systems: This paper proposes a method for agent-oriented planning in multi-agent systems, which can be used for various applications such as text classification and sentiment analysis.
Seeing and Reasoning with Confidence: Supercharging Multimodal LLMs with an Uncertainty-Aware Agentic Framework: This paper presents a method for supercharging multimodal LLMs with an uncertainty-aware agentic framework, which can be used for various applications such as text classification and sentiment analysis.
Reasoning in visual navigation of end-to-end trained agents: a dynamical systems approach: This paper proposes a method for reasoning in visual navigation of end-to-end trained agents, which can be used for various applications such as text classification and sentiment analysis.
General-Purpose Aerial Intelligent Agents Empowered by Large Language Models: This paper presents a method for general-purpose aerial intelligent agents empowered by large language models, which can be used for various applications such as text classification and sentiment analysis.
HASARD: A Benchmark for Vision-Based Safe Reinforcement Learning in Embodied Agents: This paper proposes a benchmark for vision-based safe reinforcement learning in embodied agents, which can be used for various applications such as text classification and sentiment analysis.
Multi-Agent Obstacle Avoidance using Velocity Obstacles and Control Barrier Functions: This paper presents a method for multi-agent obstacle avoidance using velocity obstacles and control barrier functions, which can be used for various applications such as text classification and sentiment analysis.
Guardians of the Agentic System: Preventing Many Shots Jailbreak with Agentic System: This paper proposes a method for preventing many shots jailbreak with agentic system, which can be used for various applications such as text classification and sentiment analysis.

SFT

SFT is a type of self-supervised learning that can be used for various applications such as text classification and sentiment analysis. The following papers discuss various aspects of SFT:

Bring Remote Sensing Object Detect Into Nature Language Model: Using SFT Method: This paper proposes a method for bringing remote sensing object detect into nature language model using SFT method, which can be used for various applications such as text classification and sentiment analysis.
R1-Zero's "Aha Moment" in Visual Reasoning on a 2B Non-SFT Model: This paper presents a method for R1-Zero's "Aha moment" in visual reasoning on a 2B non-SFT model, which can be used for various applications such as text classification and sentiment analysis.
Simplify RLHF as Reward-Weighted SFT: A Variational Method: This paper proposes a method for simplifying RLHF as reward-weighted SFT, which can be used for various applications such as text classification and sentiment analysis.
SFT Memorizes, RL Generalizes: A Comparative Study of Foundation Model Post-training: This paper presents a comparative study of foundation model post-training, which can be used for various applications such as text classification and sentiment analysis.
**Provably Mitigating Overoptimization in RLHF: Your SFT Loss is Implicitly an Ad
Q&A: Latest 15 Papers - March 13, 2025

Q: What is RAG and how does it relate to the latest 15 papers?

A: RAG stands for Retrieval-Augmented Generation, which is a technique that combines the strengths of retrieval and generation to produce high-quality text. The latest 15 papers discuss various aspects of RAG, including its applications in text classification, sentiment analysis, and more.

Q: What is the main contribution of the paper "RAG-Adapter: A Plug-and-Play RAG-enhanced Framework for Long Video Understanding"?

A: The main contribution of this paper is the proposal of a RAG-enhanced framework for long video understanding, which can be used for various applications such as video summarization and question answering.

Q: How does the paper "RAAD-LLM: Adaptive Anomaly Detection Using LLMs and RAG Integration" contribute to the field of anomaly detection?

A: This paper proposes a method for adaptive anomaly detection using LLMs and RAG integration, which can be used for various applications such as fraud detection and network intrusion detection.

Q: What is the main difference between the papers "OpenRAG: Optimizing RAG End-to-End via In-Context Retrieval Learning" and "OCR Hinders RAG: Evaluating the Cascading Impact of OCR on Retrieval-Augmented Generation"?

A: The main difference between these two papers is that the first paper proposes an optimization method for RAG end-to-end via in-context retrieval learning, while the second paper evaluates the cascading impact of OCR on retrieval-augmented generation.

Q: How does the paper "AgentOrca: A Dual-System Framework to Evaluate Language Agents on Operational Routine and Constraint Adherence" contribute to the field of agent systems?

A: This paper proposes a dual-system framework for evaluating language agents on operational routine and constraint adherence, which can be used for various applications such as text classification and sentiment analysis.

Q: What is the main contribution of the paper "Proto Successor Measure: Representing the Behavior Space of an RL Agent"?

A: The main contribution of this paper is the proposal of a method for representing the behavior space of an RL agent, which can be used for various applications such as text classification and sentiment analysis.

Q: How does the paper "Collaborative Uncertainty Benefits Multi-Agent Multi-Modal Trajectory Forecasting" contribute to the field of multi-agent systems?

A: This paper proposes a method for collaborative uncertainty benefits multi-agent multi-modal trajectory forecasting, which can be used for various applications such as traffic prediction and route planning.

Q: What is the main difference between the papers "GTR: Guided Thought Reinforcement Prevents Thought Collapse in RL-based VLM Agent Training" and "Hierarchical Multi Agent DRL for Soft Handovers Between Edge Clouds in Open RAN"?

A: The main difference between these two papers is that the first paper proposes a method for guided thought reinforcement prevents thought collapse in RL-based VLM agent training, while the second paper proposes a method for hierarchical multi-agent DRL for soft handovers between edge clouds in open RAN.

Q: How does the paper "Bring Remote Sensing Object Detect Into Nature Language Model: Using SFT Method" contribute to the field of SFT?

A: This paper proposes a method for bringing remote sensing object detect into nature language model using SFT method, which can be used for various applications such as text classification and sentiment analysis.

Q: What is the main contribution of the paper "R1-Zero's "Aha Moment" in Visual Reasoning on a 2B Non-SFT Model"?

A: The main contribution of this paper is the proposal of a method for R1-Zero's "Aha moment" in visual reasoning on a 2B non-SFT model, which can be used for various applications such as text classification and sentiment analysis.

Q: How does the paper "Simplify RLHF as Reward-Weighted SFT: A Variational Method" contribute to the field of RLHF?

A: This paper proposes a method for simplifying RLHF as reward-weighted SFT, which can be used for various applications such as text classification and sentiment analysis.

Q: What is the main difference between the papers "SFT Memorizes, RL Generalizes: A Comparative Study of Foundation Model Post-training" and "Provably Mitigating Overoptimization in RLHF: Your SFT Loss is Implicitly an Adversarial Regularizer"?

A: The main difference between these two papers is that the first paper presents a comparative study of foundation model post-training, while the second paper proposes a method for provably mitigating overoptimization in RLHF.

Q: How does the paper "Continual SFT Matches Multimodal RLHF with Negative Supervision" contribute to the field of SFT?

A: This paper proposes a method for continual SFT matches multimodal RLHF with negative supervision, which can be used for various applications such as text classification and sentiment analysis.

Q: What is the main contribution of the paper "UFT: Unifying Fine-Tuning of SFT and RLHF/DPO/UNA through a Generalized Implicit Reward Function"?

A: The main contribution of this paper is the proposal of a method for unifying fine-tuning of SFT and RLHF/DPO/UNA through a generalized implicit reward function, which can be used for various applications such as text classification and sentiment analysis.

Q: How does the paper "Getting More Juice Out of the SFT Data: Reward Learning from Human Demonstration Improves SFT for LLM Alignment" contribute to the field of SFT?

A: This paper proposes a method for getting more juice out of the SFT data, which can be used for various applications such as text classification and sentiment analysis.

Q: What is the main difference between the papers "Minor SFT loss for LLM fine-tune to increase performance and reduce model deviation" and "On a Rice theorem for dynamical properties of SFTs on groups"?

A: The main difference between these two papers is that the first paper proposes a method for minor SFT loss for LLM fine-tune, while the second paper presents a method for on a Rice theorem for dynamical properties of SFTs on groups.

Q: How does the paper "Automatically Generating Numerous Context-Driven SFT Data for LLMs across Diverse Granularity" contribute to the field of SFT?

A: This paper proposes a method for automatically generating numerous context-driven SFT data for LLMs across diverse granularity, which can be used for various applications such as text classification and sentiment analysis.

Q: What is the main contribution of the paper "SteerLM: Attribute Conditioned SFT as an (User-Steerable) Alternative to RLHF"?

A: The main contribution of this paper is the proposal of a method for SteerLM: attribute conditioned SFT as an (user-steerable) alternative to RLHF, which can be used for various applications such as text classification and sentiment analysis.

Q: How does the paper "SFT-KD-Recon: Learning a Student-friendly Teacher for Knowledge Distillation in Magnetic Resonance Image Reconstruction" contribute to the field of SFT?

A: This paper proposes a method for SFT-KD-Recon: learning a student-friendly teacher for knowledge distillation in magnetic resonance image reconstruction, which can be used for various applications such as text classification and sentiment analysis.

Q: What is the main difference between the papers "Strongly Aperiodic SFTs on Generalized Baumslag-Solitar groups" and "Cantor-Bendixson ranks of countable SFTs"?

A: The main difference between these two papers is that the first paper presents a method for strongly aperiodic SFTs on generalized Baumslag-Solitar groups, while the second paper presents a method for Cantor-Bendixson ranks of countable SFTs.

Q: How does the paper "LLM Misalignment via Adversarial RLHF Platforms" contribute to the field of RLHF?

A: This paper proposes a method for LLM misalignment via adversarial RLHF platforms, which can be used for various applications such as text classification and sentiment analysis.

Q: What is the main contribution of the paper "Taming Overconfidence in LLMs: Reward Calibration in RLHF"?

A: The main contribution of this paper is the proposal of a method for taming overconfidence in LLMs: reward calibration in RLHF, which can be used for various applications such as text classification and sentiment analysis.

Q: How does the paper "Reward Shaping to Mitigate Reward Hacking in RLHF" contribute to the field of RLHF?

A: This paper proposes a method for reward shaping to mitigate reward hacking in RLHF, which can be used for various applications such as text classification and sentiment analysis.

Q: What is the main difference between the papers "Can RLHF be More Efficient with Imperfect Reward Models? A Policy Coverage Perspective" and "Evaluating Defences against Unsafe Feedback in RLHF"?

A: The main difference between these two papers is that the first paper presents a method for can RLHF be more efficient with imperfect reward models, while the second paper presents a method for evaluating defences against unsafe feedback in RLHF.

Q: How does the paper "Projection Optimization: A General Framework for Multi-Objective and Multi-Group RLHF" contribute to the field of RLHF?

A: This paper proposes a method for projection optimization: a general framework for multi-objective and multi-group RLHF, which can be used for various applications such as text classification and sentiment analysis.

**Q: What is the main contribution of the