filmov
tv
AriGraph: Learning Knowledge Graph World Models with Episodic Memory for LLM Agents
Показать описание
Excited to have Petr and Nikita present their work on AriGraph! It is a way to use episodic and semantic memory jointly to aid in retrieval of key information for an agent to perform decision making on. It leverages the power of Knowledge Graphs to sufficiently constrain the context and make it relevant for the agent. It also features a dynamically updatable way of modifying the knowledge graph based on most recent observations, thereby enabling the agent to have the most recent information.
I think knowledge graphs could form a semantically meaningful abstraction space for memory, on which association can be done. It would likely be better than naive RAG over data chunks. This, combined with other abstraction spaces, like summary of environment or reflection, could likely help an agent perform better over arbitrary environments.
~~~
Speaker Profile:
1. Petr Anokhin, PhD in neuroscience, senior researcher at the Artificial Intelligence Research Institute (AIRI) in Moscow, Russia.
2. Nikita Semenov, research engineer at the Skolkovo Center for Applied AI, Moscow, Russia. Fourth-year student at the HSE University, Faculty of Mathematics.
AIRI group overview:
Our group specializes in the study of agentic behavior, with a specific emphasis on the implementation of long-term memory systems. This involves exploring and creating techniques that allow agents to retain and utilize information over extended periods, enhancing their learning and decision-making capabilities.
Abstract:
Advancements in generative AI, particularly Large Language Models (LLMs), have significantly expanded the capabilities of autonomous agents. Achieving true autonomy requires accumulating and updating knowledge gained from interactions with the environment and effectively utilizing it. Traditional agentic memory approaches often rely on unstructured memory systems using methods like full history observation, summarization, or retrieval augmentation. These methods, however, fall short in supporting the complex reasoning and planning needed for sophisticated decision-making. In our research, we introduce "AriGraph," an innovative memory graph approach that seamlessly integrates semantic and episodic memories during environmental interaction. This structured memory aid not only facilitates efficient associative retrieval of relevant concepts but also enhances the agent's planning and decision-making processes. Through our experiments, we demonstrate that our Ariadne LLM agent, equipped with AriGraph, significantly surpasses conventional methods in complex scenarios within the TextWorld environment. Notable improvements are observed in diverse tasks including the cooking challenge from the First TextWorld Problems competition and novel tasks such as house cleaning and Treasure Hunting puzzles, showcasing its robust zero-shot task handling capabilities.
~~~
Related Resources:
Paper at 1:23:52 in which we talked about human priors in Atari Games, and how human priors are important for humans to solve games.
~~~
0:00 Introduction
1:19 Foreword by authors
5:35 Introduction to Agents
8:57 Memory Types: RAG vs Large Context
17:55 TextWorld Introduction
21:55 Is Graph Traversal the same as reasoning?
28:15 Environment Tasks
32:30 LLM Baselines
34:36 How to imbue semantic memory and episodic memory as a Knowledge Graph
46:55 Agent Workflow
52:00 AriGraph Structure
52:52 Extracting Semantic Memory
54:29 Episodic Memory
54:50 Memory Retrieval Process
59:04 Illustrative walkthrough of memory retrieval
1:01:10 Navigation Capabilities of LLM as a function of required actions
1:14:44 Results
1:31:20 Segway to Discussion
1:32:31 Discussion (including Emotions)
1:52:26 Conclusion
~~~
AI and ML enthusiast. Likes to think about the essences behind breakthroughs of AI and explain it in a simple and relatable way. Also, I am an avid game creator.
I think knowledge graphs could form a semantically meaningful abstraction space for memory, on which association can be done. It would likely be better than naive RAG over data chunks. This, combined with other abstraction spaces, like summary of environment or reflection, could likely help an agent perform better over arbitrary environments.
~~~
Speaker Profile:
1. Petr Anokhin, PhD in neuroscience, senior researcher at the Artificial Intelligence Research Institute (AIRI) in Moscow, Russia.
2. Nikita Semenov, research engineer at the Skolkovo Center for Applied AI, Moscow, Russia. Fourth-year student at the HSE University, Faculty of Mathematics.
AIRI group overview:
Our group specializes in the study of agentic behavior, with a specific emphasis on the implementation of long-term memory systems. This involves exploring and creating techniques that allow agents to retain and utilize information over extended periods, enhancing their learning and decision-making capabilities.
Abstract:
Advancements in generative AI, particularly Large Language Models (LLMs), have significantly expanded the capabilities of autonomous agents. Achieving true autonomy requires accumulating and updating knowledge gained from interactions with the environment and effectively utilizing it. Traditional agentic memory approaches often rely on unstructured memory systems using methods like full history observation, summarization, or retrieval augmentation. These methods, however, fall short in supporting the complex reasoning and planning needed for sophisticated decision-making. In our research, we introduce "AriGraph," an innovative memory graph approach that seamlessly integrates semantic and episodic memories during environmental interaction. This structured memory aid not only facilitates efficient associative retrieval of relevant concepts but also enhances the agent's planning and decision-making processes. Through our experiments, we demonstrate that our Ariadne LLM agent, equipped with AriGraph, significantly surpasses conventional methods in complex scenarios within the TextWorld environment. Notable improvements are observed in diverse tasks including the cooking challenge from the First TextWorld Problems competition and novel tasks such as house cleaning and Treasure Hunting puzzles, showcasing its robust zero-shot task handling capabilities.
~~~
Related Resources:
Paper at 1:23:52 in which we talked about human priors in Atari Games, and how human priors are important for humans to solve games.
~~~
0:00 Introduction
1:19 Foreword by authors
5:35 Introduction to Agents
8:57 Memory Types: RAG vs Large Context
17:55 TextWorld Introduction
21:55 Is Graph Traversal the same as reasoning?
28:15 Environment Tasks
32:30 LLM Baselines
34:36 How to imbue semantic memory and episodic memory as a Knowledge Graph
46:55 Agent Workflow
52:00 AriGraph Structure
52:52 Extracting Semantic Memory
54:29 Episodic Memory
54:50 Memory Retrieval Process
59:04 Illustrative walkthrough of memory retrieval
1:01:10 Navigation Capabilities of LLM as a function of required actions
1:14:44 Results
1:31:20 Segway to Discussion
1:32:31 Discussion (including Emotions)
1:52:26 Conclusion
~~~
AI and ML enthusiast. Likes to think about the essences behind breakthroughs of AI and explain it in a simple and relatable way. Also, I am an avid game creator.
1:55:09
AriGraph: Learning Knowledge Graph World Models with Episodic Memory for LLM Agents
1:02:12
AriGraph (Part 2) - Knowledge Graph Construction and Retrieval Details
0:23:34
Thoth - a recommendation engine for Python applications
0:23:34
Thoth a recommendation engine for Python applications
Комментарии