News

  1. Co-authored SwitchMem, which learns when GUI agents should remember in text or images.

    Across three GUI benchmarks, choosing between text and screenshots at each step worked better than always showing the agent the same kind of memory.

  2. Released CVRR, a study of how vision-language models use latent visual reasoning.

    CVRR repeatedly revisits the image, then answers using the final recurrent state as its only source of visual information. This lets us test whether the model relies on that computation when it answers.

  3. Started my Ph.D. in Computer Science at Korea University’s DMIS Lab.

    I joined Prof. Jaewoo Kang’s group to study representation learning and mechanistic interpretability.

  4. Released SaMer, which merges visual tokens while preserving evidence for retrieval.

    The method reduces the number of image tokens a retrieval system needs to store and score while retaining object-level evidence that a text query can match.