Dongwoo Kim

Training data attribution on 32B-parameter models no longer requires intractable gradient storage: extreme 1-bit eigenbasis projection achieves competitive influence estimation accuracy at over 100× lower per-example footprint than standard baselines.

Harness-aware post-training can drastically improve LLM agent performance, especially when facing shifting tool environments, revealing a key design dimension often overlooked in AI systems.

Machine learning resolves 20,000 ambiguous X-ray source matches, revealing the limitations of traditional spatial cross-matching methods.

Ortho-ReID achieves a remarkable boost in clothes-changing person re-identification accuracy by leveraging instance-adaptive low-rank subspaces, outperforming existing methods significantly.

Current concept unlearning methods for diffusion models are easily bypassed by paraphrasing prompts, but PURE leverages cross-attention activations to achieve significantly better erasure.

Optimal transport provides a surprisingly tight and efficiently computable bound on transductive generalization in graph node classification, revealing how GNN depth impacts representation geometry.

Skip the retraining: compose pre-trained GFlowNets at inference time to tackle multi-objective generation, even with complex non-linear reward functions.