Conference presentations

A single compact network handles two tasks the literature usually splits across two networks: completing the hidden portion of an occluded object's mask (amodal segmentation) and predicting which object is in front of which (pairwise occlusion ordering). Every visible mask is compressed into a 64-dimensional shape code, those codes exchange information through a relational graph whose edges carry a learned three-way occlusion-order distribution, and the updated code is decoded back into a completed mask. The same edges deliver the ordering.

The proposed research uses multiple verification stages to check generated responses before they are returned to users. Unlike traditional RAG systems that directly provide generated answers, the proposed system includes verification agents that examine whether the responses are supported by evidence, free from contradictions, and factually consistent, and that establish reliability through confidence scoring.

This paper proposes a conceptual data-governance framework that turns national AI policy principles and strategies into measurable indicators, and visualizes them through a prototype monitoring dashboard.

This study analyzes constraint-aware constrained quadratic modeling (CQM) for hyperparameter tuning as search-space breadth increases. With eight hyperparameters fixed, candidate values per hyperparameter grow from 6 to 128, expanding the space from 6 8 to 128 8 configurations. Experiments compare D-Wave's hybrid CQM solver with greedy, tabu, and SCIP methods.

A two-stage SFT and RL framework with GRPO for small-model Text-to-SQL. The SFT stage teaches the model to generate, diagnose, and repair SQL using correct and incorrect teacher traces. The GRPO stage then improves the model with rewards for execution correctness, schema consistency, SQL validity, and repair quality. This produces stronger small language models with fewer SQL errors and more stable RL training.

This work focuses on finding and segmenting objects in videos based on text descriptions of how they move, for example “the person walking toward the camera” or “the car that brakes”.

Current AI scientist systems such as Google Co-Scientist and Sakana AI Scientist can generate hypotheses and write research papers, but their verification happens inside the same pipeline that generated the idea, causing bias and overconfidence in incorrect results. As evidence, state-of-the-art models detect only 18.4% of errors on the SPOT benchmark. This research explores a reliability-aware multi-agent framework that separates generation from verification through independent verifier, adversarial reviewer, and replication agents, to improve the trustworthiness and reproducibility of AI-generated scientific research.