Context and background:
SingularityNET Foundation, in collaboration with other partners such as the OpenCog Foundation and TrueAGI, is working toward a scalable implementation of the Hyperon AGI framework running on decentralized infrastructure, and toward implementation of the PRIMUS cognitive architecture within this framework.
Hyperon and PRIMUS are complex systems involving multiple components, which need to demonstrate appropriate functionalities both individually and in combination. This RFP aims to address a portion of this overall need, via funding the initial iteration of one significant component of PRIMUS within Hyperon.
The goal of this RFP is to explore how Probabilistic Logic Networks (PLN) can be applied to enhance LLMs. By leveraging PLN’s probabilistic reasoning capabilities, the aim is to improve memory systems, reasoning capabilities, and information retrieval processes. While a comparison to existing methods, such as retrieval-augmented generation (RAG), is of interest, this research encourages proposals that consider alternative approaches to integrating PLN with LLMs.
There are two suggested avenues for exploration, but researchers are welcome to propose other innovative methods where PLN could enhance LLM performance.
- Batch Mode PLN on Atomspace Knowledge Graphs:
This approach applies PLN in batch mode on an Atomspace knowledge graph to enrich the graph with more informative nodes and probabilistic links. The enhanced knowledge graph would then support LLM memory by improving the structure of the graph to facilitate more accurate and relevant information retrieval. The output of the PLN-enriched knowledge graph can be used to develop a probabilistically weighted graphRAG, improving retrieval without the need for real-time inference. This method operates in the background, refining the knowledge graph over time and offering a probabilistic edge over static graphs commonly used in graphRAG systems.
- Real-Time PLN Queries on Atomspace:
Another potential method involves using PLN to perform real-time queries on the knowledge graph. Instead of relying solely on pre-built graphs, this approach would allow LLMs to engage dynamically with the knowledge graph. By running live PLN queries, LLMs could reason probabilistically in real time, enabling context-sensitive decisions based on the updated graph. While this approach offers more nuanced reasoning capabilities, it may be limited by the current speed of MeTTa interpreters, making real-time inference less practical at this time. However, future improvements in interpreter performance could make this a highly adaptive solution for LLM memory and reasoning.
Expected Outcomes:
- PLN-enhanced Knowledge Graphs: Demonstrate how PLN can be used to enrich knowledge graphs, improving their structure and utility for LLM memory and reasoning tasks. The graph should include useful nodes and probabilistically weighted links that enhance retrieval processes.
- Improved LLM Memory and Reasoning: Develop and demonstrate systems where PLN improves LLM memory retrieval and reasoning, either through batch-mode knowledge graph enrichment or real-time queries, compared to existing methods like graphRAG.
- Performance Evaluation: Conduct a thorough evaluation of the proposed approaches (batch mode, real-time PLN queries, or alternative methods), including performance metrics, improvements in reasoning, and memory accuracy.
- Comparative Study: Compare the effectiveness of PLN-enhanced systems against traditional RAG or graphRAG approaches, demonstrating the unique benefits PLN offers for improving LLM memory, retrieval, and reasoning.
|
Join the Discussion (0)
Please create account or login to post comments.