Research

Demonstrates that distractor-augmented prompting substantially mitigates overconfidence in LLMs—yielding up to 460% relative accuracy gains and 90% reductions in expected calibration error—and derives principled guidelines for reliable, well-calibrated deployment.

Proposes a scalable, memory-centric architecture that dynamically extracts, consolidates, and retrieves salient conversational information to augment long-term LLM context. On the LOCOMO benchmark it improves accuracy by up to 26% while reducing latency by 91% and token expenditure by over 90%.

Establishes that the visual question-answering accuracy of MLLMs is highly sensitive to subject scale, and introduces training-free visual cropping techniques that enhance the perception of fine-grained details.

Introduces FIRE, a multimodal framework that synthesizes food titles, ingredients, and cooking instructions from images by integrating BLIP, a Vision Transformer decoder, and T5—facilitating downstream applications such as recipe customization and recipe-to-code generation for automated cooking.

Presents a knowledge-injection framework that strengthens the functional grounding of reinforcement learning and language model agents in text-based games through knowledge graphs and memory of prior actions and object affordances, evaluated systematically across 10 ScienceWorld tasks.