This paper studies confidence calibration in LLMs across nine models and three factual QA datasets, comparing standard free-generation with distractor-augmented prompting. It shows that explicitly adding distractors greatly reduces overconfidence, yielding up to 460% relative accuracy gains and 90% ECE reductions. While large RLHF-tuned models are generally better calibrated, they can become miscalibrated on easy questions. Smaller models benefit more from distractors but remain imperfect. The study highlights persistent failures, especially on person-based queries, and offers practical guidance for more reliable LLM deployment.

Mem0 is a scalable memory-centric architecture that dynamically extracts, consolidates, and retrieves salient conversational information to extend LLM context. Its graph-based variant captures relational structures. On the LOCOMO benchmark, both outperform six baselines—boosting accuracy by up to 26%, while cutting latency by 91% and token cost by over 90%.

This work investigates Multimodal Large Language Models' (MLLMs) ability to perceive small versus large visual details in question answering tasks. The study shows that MLLMs' accuracy is sensitive to subject size and can be improved using visual cropping methods. These findings suggest caution and potential improvements for detail-sensitive applications.

This paper introduces FIRE, a novel multimodal methodology for generating recipes from food images, contributing to the growing field of food computing. FIRE effectively produces food titles, ingredients, and cooking instructions using the BLIP model, a Vision Transformer with a decoder, and the T5 model. The paper also explores practical applications like recipe customization and recipe-to-code generation for automated cooking.

This paper introduces a knowledge-injection framework to enhance the functional grounding of agents in text-based games, addressing existing limitations in coherence, contextual awareness, and learning. It incorporates domain knowledge through memory of past actions and object affordances, aiding two types of agents: reinforcement learning and language model agents. The framework employs strategies like knowledge graphs and input encoding augmentations. Tested on 10 tasks in the ScienceWorld environment, the study reveals how task properties, model architectures, and domain knowledge interact in interactive contexts.