Current Position
Assistant Professor of MS&E and, by courtesy, of Computer Science and Electrical Engineering
Intersection with the project
Overview
Vasilis Syrgkanis’s research underpins the adaptability and analytical intelligence pillars of the agricultural chatbot, especially regarding learning user preferences and understanding the cognitive capabilities of artificial intelligence. The researcher’s work focuses on the intersection between machine learning and causal inference, providing the technical basis for the system to not only process information, but “learn” dynamically from the history of interactions and execute complex reasoning needed for the financial management of the rural property.
”Personalized Adaptation via In-Context Preference Learning” (2024)
This research introduces the Preference Pretrained Transformer (PPT), a framework that allows personalization of language models without the need for constant retraining. The central term is In-Context Learning (ICL), where the model adapts its behavior by appending the history of recent interactions to its prompt. Another relevant term is the Contextual Bandit, a scenario where the AI makes decisions (such as suggesting a response or action) based on context to maximize utility for the user.
Dynamic Adaptation to the Producer’s Routine
- “PPT leverages the in-context learning capabilities of transformers to dynamically adapt to individual preferences.” “PPT leverages the in-context learning capabilities of transformers to dynamically adapt to individual preferences.”
- Intersection: This technology is what supports the chatbot’s “Continuous Learning Engine.” Since each farm has specific suppliers and routes, the bot uses the WhatsApp conversation history to learn what is a priority for that producer without requiring bureaucratic manual configurations.
Continuous Improvement through Interaction
- “The model interacts with the user, appends the interaction history to its context and generates more personalized responses.” “The pretrained model interacts with the user, appends the interaction history to its context and generates more personalized responses.”
- Intersection: In the project, the user sends audios about urgent demands. The chatbot uses this technique so that, over time, the semantic interpretation and alerts become more accurate, reducing failures due to forgetfulness by up to 80% as it “gets to know” the manager’s habits better.
”Discovering Hierarchical Latent Capabilities of Language Models via Causal Representation Learning” (2024)
This work investigates how the abilities of a language model (like Claude, used in the project) are structured hierarchically. The term Causal Representation Learning refers to the identification of hidden factors that explain the AI’s performance. The research demonstrates that complex abilities, such as Mathematical Reasoning, depend on fundamental abilities, such as Instruction Following.
Causal Dependence for Financial Management
- “we reveal a clear causal direction starting from general problem-solving capabilities, advancing through instruction-following proficiency, and culminating in mathematical reasoning ability.” “we reveal a clear causal direction starting from general problem-solving capabilities, advancing through instruction-following proficiency, and culminating in mathematical reasoning ability.”
- Intersection: For the chatbot to provide cash flow reports with predictive analyses, it must first be excellent at following the instructions contained in the producer’s audios. Syrgkanis’s research validates that the success of the bot’s financial automation depends on the AI’s robustness in interpreting simple field commands.
Accuracy in Interpreting Operational Commands
- Mathematical tasks demand precise adherence to instructions for correct formatting and problem interpretation, where misunderstandings severely impact accuracy.” “mathematical tasks demand precise adherence to instructions for correct formatting and problem interpretation, where misunderstandings severely impact accuracy.”
- Intersection: The chatbot extracts tax data from receipt images and audios. Syrgkanis’s finding about the causal link between “following instructions” and “mathematical accuracy” explains why Advanced Audio Processing is the project’s critical feature: if the AI does not correctly interpret the vocal instruction, the calculations of input and fuel waste will be compromised.
Full papers
Vasilis Syrgkanis - Personalized Adaptation via In-Context Preference Learning.pdf Vasilis Syrgkanis - Discovering Hierarchical Latent Capabilities of Language Models via Causal Representation Learning.pdf