Research

NAIRL’s Professor Tae-Hyun Oh of KAIST to Present Three Studies at ICML 2026

June 1, 2026

Professor Tae-Hyun Oh, School of Computing, KAIST [Source=KAIST]

A research team led by Professor Tae-Hyun Oh of the School of Computing at KAIST, affiliated with the National AI Research Lab (NAIRL), will present three papers at ICML 2026, spanning efficient model tuning, reliable image restoration, and a hidden capability of modern AI models.

The first study, “A Language-Guided Bayesian Optimization for Efficient LoRA Hyperparameter Search,” cuts the cost of fine-tuning large models. Finding good LoRA settings normally means searching tens of thousands of combinations. By letting a pre-trained language model explain in plain language what each setting does, the team steered the search toward strong candidates and reached over 20 percent better performance in only about 30 rounds, versus the roughly 45,000 a standard search explores.

The second study, “Measurement-Consistent Langevin Corrector for Stabilizing Latent Diffusion Inverse Problem Solvers,” improves AI that restores degraded images in fields like medical and satellite imaging. Existing methods struggle to stay faithful to the real data while reflecting what the model has learned, leaving results unstable. The team’s method keeps generation consistent with actual measurements while correcting distortions along the way, making restoration more stable and trustworthy.

Overview of the Measurement-Consistent Langevin Corrector (MCLC), which keeps the generation process consistent with actual measurements while probabilistically correcting distortions [Source=Professor Tae-Hyun Oh’s research team]

The third study, “Zero-Shot Rankability,” reveals that multimodal large language models already grasp order, such as small to large or dark to bright, without being taught. Earlier vision-language models like CLIP struggled with this, but the team found MLLMs embed this sense of order from the start, tracing it to their conditional embeddings and a small gap between what they see and read.

NAIRL will continue to introduce research that advances the efficiency, reliability, and understanding of AI to the broader domestic and international AI community.

Comparison between conventional linear probing, which requires separate training, and the proposed agentic zero-shot approach, which probes the ordinal structure of MLLM embeddings through a language interface alone [Source=Professor Tae-Hyun Oh’s research team]

🔗 Language-Guided BO: https://arxiv.org/abs/2602.11171 / Project page: https://baekseongeun.github.io/lora-bo/
🔗 Measurement-Consistent Langevin Corrector: https://arxiv.org/abs/2601.04791