Multimodal Faithful Decoder Mitigates Hallucinations
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Multimodal generative models often produce output text that is unfaithful to the input context, leading to hallucinations where the model generates untrue facts associated with the input data.
Innovation Solution
The system employs a multimodal faithful decoder that mitigates hallucinations by encoding input data, generating intermediate partial sentences, forming complete sentences, encoding these sentences, and calculating a faithfulness score to re-rank the partial sentences based on their fidelity to the input.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If multimodal generative models are used to generate output text, then the model can process multiple types of input data (audio, image, text), but the generated text becomes unfaithful to the input context and produces hallucinations
Solution Approach 1:
The patent implements a feedback mechanism by computing a faithfulness score that compares the encoded representation of the input data with the encoded representation of the generated text. This feedback loop allows the system to evaluate and adjust the generated output to ensure it remains faithful to the input context, directly addressing the reliability issue while maintaining multimodal versatility
Solution Approach 2:
The patent introduces an intermediary faithfulness scoring mechanism that acts as a mediator between the input data and the generated text. This intermediary component encodes both the input data and generated text, compares their representations, and provides a faithfulness score that guides the generation process to maintain reliability without sacrificing multimodal adaptability
2Loss of information
If the model generates creative and detailed captions, then the output is more informative, but the likelihood of hallucinations increases
Solution Approach 1:
The faithfulness score computation provides continuous feedback during the caption generation process, allowing the model to maintain information completeness while checking against the input context. This feedback mechanism ensures that creative and detailed captions remain grounded in actual input data, reducing hallucinations while preserving information richness
Solution Approach 2:
The patent generates multiple candidate captions and selectively applies faithfulness scoring to the most promising candidates. This partial action approach allows the model to explore creative possibilities while maintaining accuracy by focusing verification efforts on the most likely final outputs, balancing information completeness with factual accuracy
Data Source
AI summary
Systems and techniques are described for generating and using unimodal/multimodal generative models that mitigate hallucinations. For example, a computing device can encode input data to generate encoded representations of the input data. The computing device can obtain intermediate data including a plurality of partial sentences associated with the input data and can generate, based on the intermediate data, at least one complete sentence associated with the input data. The computing device can encode the at least one complete sentence to generate at least one encoded representation of the at least one complete sentence. The computing device can generate a faithfulness score based on a comparison of the encoded representations of the input data and the at least one encoded representation of the at least one complete sentence. The computing device can re-rank the plurality of partial sentences of the intermediate data based on the faithfulness score to generate re-ranked data.