LLM Ambiguity Detection via Perturbed Queries
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Large language models (LLMs) often produce incorrect outputs due to ambiguity in user queries, as they are not specifically trained to recognize ambiguities and may misinterpret small changes in input, leading to significant changes in responses without human knowledge of the incorrect output.
Innovation Solution
A method is introduced that generates an ambiguity query by perturbing the original query with multiple perturbations and sending it to the LLM via an API, which includes training the model to recognize ambiguities, and using a prompt manager to detect ambiguities and request clarification from users to generate revised queries, thereby improving the accuracy of responses.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If LLMs are trained for general tasks without specific task training, then adaptability is improved, but output accuracy deteriorates due to inability to recognize ambiguities
Solution Approach 1:
An ambiguity detection module is introduced as an intermediary component between the user interface and the LLM. This module receives the original query, generates perturbed queries, and detects ambiguities before the query is processed by the LLM, thereby improving output accuracy without compromising the model's general adaptability
Solution Approach 2:
The system performs preliminary ambiguity detection and query perturbation generation before the LLM processes the query. By pre-processing the query to identify and handle ambiguities in advance, the system ensures more accurate LLM responses while maintaining its general task capabilities
2Speed
If LLMs process original queries directly, then response speed is improved, but measurement precision deteriorates due to undetected ambiguities
Solution Approach 1:
The system applies partial action by performing ambiguity detection only when necessary - it generates perturbed queries and detects ambiguities selectively rather than processing every query through the complete ambiguity detection pipeline, thus balancing response speed with query interpretation accuracy
3Measurement precision
If multiple perturbations are generated to detect ambiguities, then detection precision is improved, but device complexity increases
Solution Approach 1:
The ambiguity detection process is segmented into distinct functional components: a perturbation generation module that creates multiple perturbed queries, and an ambiguity detection module that analyzes these queries. This segmentation allows the system to achieve high detection precision through multiple perturbations while managing complexity through modular architecture
Data Source
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.