Ask or search anything...

This survey paper from researchers at Tongji University and Fudan University offers a comprehensive, systematic synthesis of Retrieval-Augmented Generation (RAG) for Large Language Models. It structures the field by delineating three evolutionary paradigms—Naive, Advanced, and Modular RAG—and details advancements across retrieval, generation, and augmentation components, while also providing a thorough framework for evaluating RAG systems.

G-Memory introduces a hierarchical, graph-based memory system designed for Large Language Model-based Multi-Agent Systems, enabling them to learn from complex collaborative histories. The system consistently improves MAS performance, achieving gains up to 20.89% in embodied action and 10.12% in knowledge question-answering tasks, while maintaining resource efficiency.

The ReSearch framework enables Large Language Models to integrate multi-step reasoning with external search, learning interactively via reinforcement learning without supervised intermediate steps. It yields substantial performance gains on complex multi-hop question answering benchmarks and reveals emergent self-correction capabilities.

MemOS, a memory operating system for AI systems, redefines memory as a first-class system resource to address current Large Language Model limitations in long-context reasoning, continuous personalization, and knowledge evolution. This framework unifies heterogeneous memory types (plaintext, activation, parameter) using a standardized MemCube unit, achieving superior performance on benchmarks like LoCoMo and PreFEval, and demonstrating robust, low-latency memory operations.

Researchers introduced LIBERO-Plus, a diagnostic benchmark for vision-language-action (VLA) models, revealing that current models exhibit substantial fragility to environmental perturbations and frequently ignore linguistic instructions. Fine-tuning with a generalized dataset significantly enhances their robustness.

This survey establishes "Agentic Science" as a paradigm for autonomous scientific discovery, offering a unified framework that integrates agent capabilities, scientific workflows, and domain-specific applications across natural sciences. It charts the evolution of AI from computational tools to autonomous research partners, highlighting over 20 validated scientific discoveries made by AI agents.

MaAS introduces an 'agentic supernet' to enable dynamic, query-dependent allocation of multi-agent system resources and architectures. The framework achieves superior performance across diverse tasks while significantly reducing inference costs and demonstrating strong adaptability compared to prior fixed-architecture methods.

G-Designer introduces a framework using Graph Neural Networks to dynamically generate task-aware communication topologies for LLM-based multi-agent systems. This approach achieved superior performance on various benchmarks while significantly reducing token consumption by up to 95.33% and demonstrating high adversarial robustness.

This survey offers a comprehensive review of how Reinforcement Learning (RL) is applied across the entire lifecycle of Large Language Models, from pre-training to alignment and reinforced reasoning. It particularly emphasizes the role of Reinforcement Learning with Verifiable Rewards (RLVR) in advancing LLM reasoning capabilities and compiles key datasets, benchmarks, and open-source tools for the field.

SRPO (Self-Referential Policy Optimization) enhances Vision-Language-Action (VLA) models for robotic manipulation by addressing reward sparsity, generating dense, progress-wise rewards using the model's own successful trajectories and latent world representations from V-JEPA 2. The method achieved a 99.2% success rate on the LIBERO benchmark, a 103% relative improvement over its one-shot SFT baseline, and demonstrated strong generalization on the LIBERO-Plus benchmark.

SRPO enhances multimodal large language models by integrating explicit self-reflection and self-correction capabilities through a two-stage training framework. The approach achieves state-of-the-art performance among open-source models, scoring 78.5% on MathVista with SRPO-32B, and showing competitive results against leading closed-source models across diverse reasoning benchmarks.

The RISEBench benchmark evaluates Large Multi-modality Models (LMMs) on Reasoning-Informed Visual Editing (RISE), assessing their capacity for visual modifications guided by complex instructions and logical inference across temporal, causal, spatial, and logical domains. Initial evaluations show leading proprietary models like GPT-4o-Image achieve only 28.9% overall accuracy, indicating substantial challenges, particularly in logical reasoning tasks.

R&D-Agent-Quant (R&D-Agent(Q)) is a multi-agent framework that automates the entire quantitative financial research and development pipeline by jointly optimizing financial factors and predictive models. It achieves superior investment strategy performance with up to 2 times higher annualized returns and 70% fewer factors compared to state-of-the-art baselines, demonstrating robustness across diverse financial markets.

A comprehensive survey proposes a novel two-dimensional taxonomy to organize research on integrating Large Language Models and Knowledge Graphs for Question Answering, detailing methods by complex QA categories and the specific functions KGs serve. The work reviews state-of-the-art techniques, highlights their utility in mitigating LLM weaknesses such as factual inaccuracy and limited reasoning, and identifies critical future research directions.

NEMORI, a self-organizing agent memory system, was developed by researchers from Tongji University, Shanghai University of Finance and Economics, Beihang University, and Tanka AI, drawing inspiration from cognitive science to address the 'amnesia' of large language models. The system established new state-of-the-art performance on the LoCoMo dataset with an LLM score of 0.744 using gpt-4o-mini, while simultaneously reducing token usage by 88% compared to full context baselines.

This work uncovers that architectural decoupling in unified multimodal models (UMMs) improves performance by inducing task-specific attention patterns, rather than eliminating task conflicts. Researchers from CUHK MMLab and Meituan introduce an Attention Interaction Alignment (AIA) loss, a regularization technique that guides UMMs' attention toward optimal task-specific behaviors without architectural changes, enhancing both understanding and generation performance for models like Emu3 and Janus-Pro.

Robotic visuomotor policies achieve dramatically improved spatial generalization by removing proprioceptive state inputs. This approach leverages a relative end-effector action space and comprehensive egocentric vision to enable robust task performance across varied spatial configurations, while also enhancing data efficiency and cross-embodiment adaptation.

Researchers from a global consortium, including Tianjin University and Huawei Noah’s Ark Lab, developed Embodied Arena, a comprehensive platform for evaluating Embodied AI agents, featuring a systematic capability taxonomy and an automated, LLM-driven data generation pipeline. This platform integrates over 22 benchmarks and 30 models, revealing that specialized embodied models often outperform general models on targeted tasks and identifying object and spatial perception as key performance bottlenecks.

MasRouter addresses the Multi-Agent System Routing (MASR) problem by dynamically configuring multi-agent systems to balance performance and cost. It uses a cascaded controller network to determine collaboration modes, allocate agent roles, and route agents to specific LLMs based on query characteristics, achieving Pareto optimality and improved cost-efficiency on benchmarks like HumanEval and MBPP.