Memory-Based Function Calling for LLM Reasoning Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Large language models (LLMs) struggle with tasks requiring reasoning, leading to incorrect results due to approaches like Chain-of-Thought (CoT) producing wrong answers with high confidence.
Innovation Solution
A memory-based function calling method for LLMs that converts reasoning problems into classification tasks, using logical-linguistic primitives to iteratively process sentences and constraints, storing them in a memory-based scratchpad for logical inference.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If Chain-of-Though (CoT) approach is used to break down reasoning tasks into smaller prompts, then the model can perform reasoning tasks as a series of intermediate steps, but the model produces wrong answers with high confidence and reasons stepwise to the wrong answer
Solution Approach 1:
The patent introduces an intermediary memory component that stores constraints and intermediate representations during the reasoning process. This memory acts as a mediator between the LLM's generative reasoning and the final answer, allowing the model to retrieve and verify constraints systematically rather than relying solely on sequential generation, thereby reducing hallucinations while maintaining reasoning capability
Solution Approach 2:
The patent implements feedback mechanisms where the model continuously checks its intermediate reasoning steps against stored constraints and previous answers. This feedback loop allows the model to detect and correct wrong answers before finalizing, improving reliability while maintaining the stepwise reasoning approach through iterative verification
2Adaptability or versatility
If LLMs are used for reasoning tasks, then the model can follow instructions and interact with tools, but the model still struggles with tasks requiring reasoning and produces incorrect results
Solution Approach 1:
The patent creates a universal reasoning framework where the same memory-based constraint storage and verification mechanism works across different reasoning tasks and tool interactions. This multi-functional approach allows the LLM to maintain its versatility in following instructions and using tools while systematically improving reasoning accuracy through consistent constraint management
Solution Approach 2:
The patent segments the reasoning process into distinct phases: constraint extraction, memory storage, intermediate reasoning, and verification. This segmentation allows the model to handle complex reasoning tasks by breaking them down into manageable steps while maintaining reliability through systematic verification at each stage
Data Source
AI summary
In one embodiment, a method comprises: parsing, using a large language model, a constraint satisfaction problem statement into a sequence of sentences corresponding to a plurality of individual problem sentences and a plurality of individual answer choices; invoking respective primitives from a set of logical-linguistic primitive functions to iteratively process i) each of the plurality of individual problem sentences to establish a respective problem constraint with accompanying arguments and ii) each of the plurality of individual answer choices to establish a respective answer choice constraint with accompanying arguments; storing, during invoking, the respective problem constraints into an aggregate list of problem constraints; and solving, after completion of the aggregate list of problem constraints, the constraint satisfaction problem statement as a validity output for each answer choice constraint given the aggregate list of problem constraints.