LLM Ambiguity Detection via Perturbed Query Analysis
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Large language models (LLMs) often produce incorrect output due to ambiguities in user queries, which can lead to significant changes in response with minor input changes, and users may not recognize these errors.
Innovation Solution
A method is introduced that involves generating an ambiguity query from the original query, which includes training the LLM to recognize ambiguities, and using an application programming interface (API) to send this query to the LLM. The method receives a binary response and confidence value, allowing it to detect ambiguous queries and potentially revise them to improve accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If LLMs are used as general models for various tasks, then versatility is improved, but output accuracy deteriorates due to ambiguous queries
Solution Approach 1:
The system performs preliminary ambiguity detection and query refinement before the LLM generates its final response. By proactively identifying ambiguous queries through perturbation analysis and correcting them before processing, the system prevents accuracy degradation while preserving the LLM's general-purpose capabilities across multiple tasks.
2Ease of operation
If LLMs process general language queries, then ease of operation is improved, but reliability deteriorates because users cannot recognize incorrect output
Solution Approach 1:
The system implements a feedback mechanism where the LLM's own responses are analyzed for ambiguity indicators. When ambiguity is detected in the output, the system feeds this information back to refine the query or request clarification, creating a closed-loop system that improves reliability while maintaining ease of operation for users.
3Adaptability or versatility
If LLMs are trained for general tasks, then adaptability is improved, but ability to detect incorrect output deteriorates due to input ambiguity
Solution Approach 1:
The ambiguity detection process is segmented into distinct components: perturbation generation, multiple LLM evaluations of perturbed queries, confidence scoring, and ambiguity determination. This segmentation allows the system to maintain general task capability while systematically improving ambiguity detection through structured analysis of how query variations affect output consistency.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
Disambiguity in large language models (LLMs) includes receiving an original query in a user interface, generating an ambiguity query from the original query, and sending, via an application programming interface (API) of an LLM, the ambiguity query to the LLM. The ambiguity query includes the original query and training the LLM to recognize ambiguities. The method further includes receiving, via the API and responsive to the ambiguity query, a binary response and detecting, based at least in part on the binary response, the original query as ambiguous. Disambiguity may include detecting an ambiguity location in the original query using perturbed queries and the LLM.