Soo Gon Kim

Papers

Beyond Family Labels

Computer Vision

Potential-Guided Particle Steering

Robotics

Beyond Family Labels

A Taxonomic Ornstein–Uhlenbeck Prior for Avian 3D Shape Recovery

NeurIPS 2026 · Oral

10,000 species

17 with 3D

Family: 1 or 0

SupCon

1pull

0push

Ornstein–Uhlenbeck

OU Covariance

$${\Sigma }_{ij}=\frac{{\sigma }^2}{2\alpha }{e}^{-\alpha d_{ij}}$$

Shape-code Regularization

$${\mathcal{L}}_{\mathrm{OU}}=\frac{D}{2}\log |\Sigma |+\frac{1}{2}\sum _{d=1}^D({\beta }_{\cdot ,d}-\mu {)}^{\top }{\Sigma }^{-1}({\beta }_{\cdot ,d}-\mu )$$

Unseen ↑

	CUBseen	CowBirdunseen
AniMer+	59.40	16.40
Ours	57.84	18.86

PCK@0.1

Ablation

PCK@0.1

Potential-Guided Particle Steering

for Negation-Constrained Dexterous Grasping

Under Review

“Grasp the handle, not the body.”

57.9% violation

Training Data

5/ 49,433

avoid-part instructions · DexGYSNet

Split

Inference-time Steering

57.9 → 17.2%

CSR 61.5%

Projects

QR-Bloom

Diffusion

Da Vinci Code AI

Reinforcement Learning

QR-Bloom

Conv2D = im2col

input

kernel

im2col

patches

×

kernel

=

output

2×2

fast

Conv3D = loop

Conv3D = Σ Conv2D

Depth = batch

slices

W_k

im2col

depth = batch

×

=

plane_k

×3 → Σ

Browser Inference

≈ 9× faster

Da Vinci Code Agent

Model Architecture

53.6%

win rate vs. Transformer baseline

Zhang et al., Mastering Da Vinci Code, arXiv 2506.12801 (2025) · 10,000 games

Attention Analysis

Open Source

MariaDB MCP

LLM Agent · MCP

MariaDB MCP

#7 merged

Optional
embeddings

#13 merged

Foreign keys
in schema

Thank you.