Yoonjeon Kim

Unified multimodal foundation models have enabled both visual understanding and generation within a single framework. Building on this foundation, supervised fine-tuning and reinforcement learning have been employed to facilitate interleaved visual and textual thinking, tightly coupling image generation with textual reasoning. Existing works predominantly rely on autoregressive unified models as the backbone, which incurs substantial computational overhead due to full regeneration of image token sequences at every rollout step. In this work, we instead leverage multimodal discrete diffusion models to develop a reinforcement learning framework for interleaved reasoning. By exploiting bidirectional context modeling, our approach enables localized visual editing, allowing targeted modifications and reducing rollout computation during GRPO by 26.9% compared to full-image editing baselines, with only a minimal performance drop. However, bidirectional multimodal decoding in discrete diffusion models introduces a non-trivial challenge: rewards become implicitly coupled across interleaved image and text tokens, leading to spurious cross-modal credit assignment. To address this, we propose factorized reward assignment across text and vision streams, assigning rewards to their corresponding token segments for stable credit propagation. This yields 11.2% gains over joint reward assignment baselines and 38.04% improvements over the supervised fine-tuned model.

Recent research on reasoning models explores the meta-awareness of language models, including their ability to determine optimal thinking duration, recognize knowledge boundaries, and structure concept-level thinking. While current large reasoning models depend solely on answer-based verification, we show that adding meta-awareness objectives leads to significant performance gains over models without such meta-knowledge. MAPR utilizes a self-generated task of predicting rollout statistics — specifically length, pass-rate, and concepts used — allowing for verification against the actual statistics. Furthermore, by leveraging this self-predictive capability, the model can regulate its reasoning behavior by (i) filtering out trivial or unsolvable prompts, (ii) reducing lengthy generations that tend to be incorrect, and (iii) generating hints relevant to the problem. MAPR yields significant improvements in both accuracy and training efficiency on various reasoning benchmarks: a 1.28× GRPO training speedup, 83.18% gain in accuracy on AIME25, and 13.04% average gain over six mathematics benchmarks.

The development of vision-language and generative models has significantly advanced text-guided image editing, which seeks the preservation of core elements in the source image while implementing modifications based on the target text. However, existing metrics have a context-blindness problem, indiscriminately applying the same evaluation criteria on completely different pairs of source image and target text, biasing towards either modification or preservation. Directional CLIP similarity, the only metric that considers both source image and target text, is also biased towards modification aspects. We propose AugCLIP, a context-aware metric that adaptively coordinates preservation and modification aspects depending on the specific context. This is done by deriving the CLIP representation of an ideally edited image using a multi-modal large language model to augment textual descriptions, then calculating a modification vector through a hyperplane that separates source and target attributes in CLIP space. Extensive experiments on five benchmark datasets show that AugCLIP aligns remarkably well with human evaluation standards, outperforming existing metrics.