Kunyu Shi

Jintao Tong, Shilin Yan, Hongwei Xue et al.

Multimodal Large Language Models (MLLMs) have made remarkable progress in multimodal perception and reasoning by bridging vision and language. However, most existing MLLMs perform reasoning primarily with textual CoT, which limits their effectiveness on vision-intensive tasks. Recent approaches inject a fixed number of continuous hidden states as "visual thoughts" into the reasoning process and improve visual performance, but often at the cost of degraded text-based logical reasoning. We argue that the core limitation lies in a rigid, pre-defined reasoning pattern that cannot adaptively choose the most suitable thinking modality for different user queries. We introduce SwimBird, a reasoning-switchable MLLM that dynamically switches among three reasoning modes conditioned on the input: (1) text-only reasoning, (2) vision-only reasoning (continuous hidden states as visual thoughts), and (3) interleaved vision-text reasoning. To enable this capability, we adopt a hybrid autoregressive formulation that unifies next-token prediction for textual thoughts with next-embedding prediction for visual thoughts, and design a systematic reasoning-mode curation strategy to construct SwimBird-SFT-92K, a diverse supervised fine-tuning dataset covering all three reasoning patterns. By enabling flexible, query-adaptive mode selection, SwimBird preserves strong textual logic while substantially improving performance on vision-dense tasks. Experiments across diverse benchmarks covering textual reasoning and challenging visual understanding demonstrate that SwimBird achieves state-of-the-art results and robust gains over prior fixed-pattern multimodal reasoning methods.

Peiran Xu, Zhuohao Li, Xiaoying Xing et al.

Large Language Models (LLMs) increasingly rely on external tools such as search engines to solve complex agentic tasks that require reasoning and external knowledge retrieval. Recently, reinforcement learning with verifiable rewards (RLVR) has demonstrated its effectiveness in advancing capabilities of LLMs by rewarding the final answers via outcome rewards. While straightforward to supervise, outcome rewards only provide sparse signals and delayed feedback, which limits their effectiveness on long trajectories. Process rewards address this by evaluating intermediate steps, providing fine-grained supervision and encouraging grounded problem solving. However, it is notoriously hard to annotate step-wise labels, especially in non-verifiable process without "golden" answers. Furthermore, step-wise judgment requires the balance between local quality with contribution to the final outcome, as optimizing towards higher process reward may not always align with better final outcomes. To address the above challenges, we introduce Principle Process Reward (PPR), an RL approach that unifies principled step-level assessment and outcome verification. We train a principle-based reward model to improve the transparency and reliability of process evaluation, and further introduce a Reward Normalization (ReNorm) strategy to calibrate outcome and process rewards. Experiment results show that PPR achieves state-of-the-art performance across a wide range of benchmarks, demonstrating its impressive robustness and generalization. Our code and model collection is available in this link.

Shilin Yan, Jintao Tong, Hongwei Xue et al.

The advent of agentic multimodal models has empowered systems to actively interact with external environments. However, current agents suffer from a profound meta-cognitive deficit: they struggle to arbitrate between leveraging internal knowledge and querying external utilities. Consequently, they frequently fall prey to blind tool invocation, resorting to reflexive tool execution even when queries are resolvable from the raw visual context. This pathological behavior precipitates severe latency bottlenecks and injects extraneous noise that derails sound reasoning. Existing reinforcement learning protocols attempt to mitigate this via a scalarized reward that penalizes tool usage. Yet, this coupled formulation creates an irreconcilable optimization dilemma: an aggressive penalty suppresses essential tool use, whereas a mild penalty is entirely subsumed by the variance of the accuracy reward during advantage normalization, rendering it impotent against tool overuse. To transcend this bottleneck, we propose HDPO, a framework that reframes tool efficiency from a competing scalar objective to a strictly conditional one. By eschewing reward scalarization, HDPO maintains two orthogonal optimization channels: an accuracy channel that maximizes task correctness, and an efficiency channel that enforces execution economy exclusively within accurate trajectories via conditional advantage estimation. This decoupled architecture naturally induces a cognitive curriculum-compelling the agent to first master task resolution before refining its self-reliance. Extensive evaluations demonstrate that our resulting model, Metis, reduces tool invocations by orders of magnitude while simultaneously elevating reasoning accuracy.

Yuan Gao, Kunyu Shi, Pengkai Zhu et al.

We propose Strongly Supervised pre-training with ScreenShots (S4) - a novel pre-training paradigm for Vision-Language Models using data from large-scale web screenshot rendering. Using web screenshots unlocks a treasure trove of visual and textual cues that are not present in using image-text pairs. In S4, we leverage the inherent tree-structured hierarchy of HTML elements and the spatial localization to carefully design 10 pre-training tasks with large scale annotated data. These tasks resemble downstream tasks across different domains and the annotations are cheap to obtain. We demonstrate that, compared to current screenshot pre-training objectives, our innovative pre-training method significantly enhances performance of image-to-text model in nine varied and popular downstream tasks - up to 76.1% improvements on Table Detection, and at least 1% on Widget Captioning.

Kunyu Shi, Qi Dong, Luis Goncalves et al.

Sequence-to-sequence vision-language models are showing promise, but their applicability is limited by their inference latency due to their autoregressive way of generating predictions. We propose a parallel decoding sequence-to-sequence vision-language model, trained with a Query-CTC loss, that marginalizes over multiple inference paths in the decoder. This allows us to model the joint distribution of tokens, rather than restricting to conditional distribution as in an autoregressive model. The resulting model, NARVL, achieves performance on-par with its state-of-the-art autoregressive counterpart, but is faster at inference time, reducing from the linear complexity associated with the sequential generation of tokens to a paradigm of constant time joint inference.

Pengxiang Li, Joey Tsai, Hongwei Xue et al.

Diffusion Language Models (DLMs) promise highly parallel text generation, yet their practical inference speed is often bottlenecked by suboptimal decoding schedulers. Standard approaches rely on 'scattered acceptance'-committing high confidence tokens at disjoint positions throughout the sequence. This approach inadvertently fractures the Key-Value (KV) cache, destroys memory locality, and forces the model into costly, repeated repairs across unstable token boundaries. To resolve this, we present the Longest Stable Prefix (LSP) scheduler, a training-free and model-agnostic inference paradigm based on monolithic prefix absorption. In each denoising step, LSP evaluates token stability via a single forward pass, dynamically identifies a contiguous left-aligned block of stable predictions, and snaps its boundary to natural linguistic or structural delimiters before an atomic commitment. This prefix-first topology yields dual benefits: systemically, it converts fragmented KV cache updates into efficient, contiguous appends; algorithmically, it preserves bidirectional lookahead over a geometrically shrinking active suffix, drastically reducing token flip rates and denoiser calls. Extensive evaluations on LLaDA-8B and Dream-7B demonstrate that LSP accelerates inference by up to 3.4x across rigorous benchmarks including mathematical reasoning, code generation, multilingual (CJK) tasks, and creative writing while matching or slightly improving output quality. By fundamentally restructuring the commitment topology, LSP bridges the gap between the theoretical parallelism of DLMs and practical hardware efficiency.

Zhaoyang Zhang, Yantao Shen, Kunyu Shi et al.

We present a vision-language model whose parameters are jointly trained on all tasks and fully shared among multiple heterogeneous tasks which may interfere with each other, resulting in a single model which we named Musketeer. The integration of knowledge across heterogeneous tasks is enabled by a novel feature called Task Explanation Prompt (TEP). With rich and structured information such as task input/output format, TEP reduces interference among tasks, allowing the model to focus on their shared structure. With a single model, Musketeer achieves results comparable to or better than strong baselines trained on single tasks, almost uniformly across multiple tasks.

Justin Lazarow, Kwonjoon Lee, Kunyu Shi et al.

Panoptic segmentation requires segments of both "things" (countable object instances) and "stuff" (uncountable and amorphous regions) within a single output. A common approach involves the fusion of instance segmentation (for "things") and semantic segmentation (for "stuff") into a non-overlapping placement of segments, and resolves overlaps. However, instance ordering with detection confidence do not correlate well with natural occlusion relationship. To resolve this issue, we propose a branch that is tasked with modeling how two instance masks should overlap one another as a binary relation. Our method, named OCFusion, is lightweight but particularly effective in the instance fusion process. OCFusion is trained with the ground truth relation derived automatically from the existing dataset annotations. We obtain state-of-the-art results on COCO and show competitive results on the Cityscapes panoptic segmentation benchmark.