Toward Trustworthy Foundation AI: Systems That See, Know, and Think as We Do

Abstract

Recent foundation models exhibit a “Paradox of Competence”: they can achieve silver medals in the International Math Olympiad yet frequently fail at trivial tasks like counting objects or interpreting basic charts. This disparity reveals that current “black-box” models often rely on statistical shortcuts rather than genuine understanding, making them brittle and difficult to trust in high-stakes collaboration.

 

In this talk, I present a research roadmap for bridging this gap to build Trustworthy Intelligence: systems that are perceptually grounded, reasoning-justified, and architecturally modular. The talk is structured around three key inquiries:

 

1. See As We See: I will first diagnose the “Grounding Gap” using Chimera, a semiotic benchmark that reveals how Vision-Language Models (VLMs) frequently cheat via visual, linguistic, and knowledge shortcuts rather than processing the actual visual signal.

 

2. Know As We Know: I will demonstrate that possessing knowledge is insufficient if it cannot be flexibly applied. Through work on Visual Math and Modality Fusion, we show that models suffer from a “Compositional Bottleneck”, failing to integrate basic knowledge/skills such as robust visual recognition with symbolic reasoning.

 

3. Think As We Think: Finally, I will present a technical deep-dive into the mechanisms of In-Context Learning (ICL). By testing the Structured Task Hypothesis, we provide empirical evidence that Large Models naturally generalize by composing latent primitives. I will argue that leveraging this latent compositionality towards explicit Modular Architectures is the key to creating AI partners that are not only powerful but transparent, verifiable, and safe.

 

Speaker: Mr. Yifan HOU
Date: 5 Feb 2026 (Thursday)
Time: 9:30am – 10:30am
Venue: LAU 6-209
Poster: Click here