- LLMs Research
- Posts
- Creative way of using LLMs
Creative way of using LLMs
Exploring interesting applications of LLMs from research papers published between February 15th-28th, 2025
In partnership with
Highlights
LEDD: Find your data lake treasures with LLM-powered semantic search!
LLMs in Mobile Apps: Discover the secrets & struggles of adding LLMs to your Android apps!
BayesGenie: Edit like a pro with AI that combines LLMs & Bayesian optimization!
Talking Like P&IDs: Chat with your P&IDs using natural language and AI-powered knowledge graphs!
ArtInsight: Bridge vision gaps with AI that describes children's art for BLV families!
Learning Code-Edit Embedding: Level up your debugging with AI that learns from your coding edits!
Creative ways to use LLMs!!
The AI That Called 📞 2,739 Strangers (And Got Them Talking)
Paper: Telephone Surveys Meet Conversational AI: Evaluating a LLM-Based Telephone Survey System at Scale
Traditional telephone surveys require substantial human resources for recruiting and training interviewers. This research introduces an autonomous AI system that dramatically scales survey deployment without compromising data quality.
In a pilot study with 75 calls in the United States and a larger study with 2,739 calls in Peru, the system was deployed an STT-LLM-TTS pipeline that achieved 89% response parity with human interviews across Peru. But here's the twist: When respondents hesitated, the AI used "cognitive mirroring" - subtly matching regional speech patterns, reducing hang-ups by 22% compared to rigid IVR systems.
At $0.83 per completed survey vs. $25+ for human teams, this could democratize global policy research. The AI even detected vocal stress patterns correlating with response accuracy (r=0.67, p<0.01)
Stanford's code-editing model analyzed 1.4M student submissions
Providing timely, personalized feedback in programming education is challenging. By modeling how students debug their code, educators can design targeted support tools that improve learning outcomes.
This paper presents an encoder-decoder model that learns "code-edit embeddings" from consecutive student code submissions. By analyzing the patterns in how students edit their code after receiving test case feedback, the system can generate personalized next-step coding suggestions, identify common debugging patterns through clustering, and better understand the learning process. Educators can provide more targeted assistance based on individual problem-solving approaches.
Optimize global IT operations with our World at Work Guide
Explore this ready-to-go guide to support your IT operations in 130+ countries. Discover how:
Standardizing global IT operations enhances efficiency and reduces overhead
Ensuring compliance with local IT legislation to safeguard your operations
Integrating Deel IT with EOR, global payroll, and contractor management optimizes your tech stack
Leverage Deel IT to manage your global operations with ease.
ArtInsight: Making Children's Art Accessible to All
Art is meant to be shared and appreciated, but how do blind or low-vision family members engage with children's artwork? ArtInsight offers a touching solution to this challenge using LLMs.
The Innovation: The ArtInsight system uses large language models to generate rich, descriptive content about children's artwork. It goes beyond basic visual descriptions to create detailed artistic interpretations, audio recordings of children explaining their art, and AI-generated questions to prompt meaningful discussions
This paper shows how LLMs can bridge experiential gaps and create more inclusive family interactions around visual creativity.
Engineering Diagrams That Speak Your Language
Process engineers are familiar with the complex schematics known as Piping and Instrumentation Diagrams (P&IDs). Now, researchers have enabled natural language communication with these diagrams through a three-component system: transforming P&IDs into graph representations via the DEXPI data model, creating P&ID knowledge graphs, and integrating these with LLMs through graph-based retrieval augmented generation.
This innovation allows engineers to query complex diagrams using plain language, potentially reducing errors and improving efficiency in industrial settings where these diagrams are essential.
BayesGenie: Precision Image Editing Through Language
Image editing requires precise control and semantic accuracy—critical for both creative professionals and everyday users. This research introduces a novel method that marries LLMs with Bayesian optimization to refine image edits using natural language instructions.
The model, dubbed BayesGenie, uses LLM-generated edits and applies Bayesian optimization to fine-tune parameters, thereby ensuring that the edited image retains both accuracy and semantic consistency. Users can specify desired changes in natural language without manually selecting image areas.
Discovering Hidden Data in Vast Data Lakes
Many organizations store vast amounts of unstructured data in data lakes, and finding useful information within them can be overwhelming.
The LEDD system uses an LLM to automatically generate hierarchical catalogs that organize data semantically. It also supports natural language queries, which make it easier to search for specific information across millions of data points. Though exact numerical improvements in search speed or accuracy are not detailed in the abstract, the paper highlights that LEDD can significantly reduce the time and effort needed to locate valuable data.
A detailed survey paper on LLMs in mobile apps
Why?: The research is significant because it explores how LLMs can transform mobile application development, making apps more intelligent and personalized while addressing the unique challenges developers face.
How?: The research methodology involved constructing a dataset of 149 LLM-enabled Android apps and conducting an exploratory analysis to examine the deployment and usage of LLMs within these applications, focusing on integration strategies and challenges.
Results: The analysis revealed key characteristics of the dataset, prevalent integration strategies, and common challenges faced by developers in the context of mobile app development.
Reply