🎉 Accepted at #NeurIPS2026 Main Track! We propose DiSPO, a credit-assignment method for RL in diffusion language models. Huge thanks to my co-authors @frt03_ @chokkanorg. Looking forward to sharing more soon!

Our paper has been accepted to Findings of #EMNLP2026. This work brings bias evaluation of LLMs into more realistic settings. Led by @rem_h4_ , with @MasahiroKaneko_, @Bollegala , and @chokkanorg. Details coming soon.

Seoul-bound for #ICML2026 🇰🇷 🎉 Excited to present three co-authored papers: * Budget-aligned test-time scaling in LLMs w/ @SoraMiyamo0831 @chokkanorg⏱️Tue, Jul 7. 10:30 * Human-like multi-image spatial reasoning in MLLMs w/ @stjohn2007 @silviasetitech @sponddd Prof. Inoue @chokkanorg⏱️Tue, Jul 7. 10:30 * Byte-treating ability in language models w/ @sangwhanmoon @Youmima1015 @7a7hi @chokkanorg ⏱️ Thu, Jul 9. 10:30 🎉 Also presenting some work on Diffusion LMs at the AI4Math workshop: * Credit assignment for RL post-training of masked diffusion LMs w/ @frt03_ @chokkanorg ⏱️ Sat, Jul 11. 8:00 Always up for a chat about LLM reasoning & Diffusion LMs — come find me!

1/ New preprint: Drifting Objectives for Refining Discrete Diffusion Language Models Can drifting be used beyond continuous generators? We study this in the setting of refining pretrained discrete diffusion language models (DDLMs). Our method, TokenDrift, provides a differentiable soft-token interface that lets feature-space drifting signals update categorical token logits. Main observation: Gen.-PPL improves throughout drifting training at fixed denoising budgets.