Yuhang Han

Yuhang Han, Wenzheng Yang, Yujie Chen et al.

Vision-language-model-based graphical user interface (GUI) agents have shown broad automation capabilities, yet deployment is bottlenecked by a key-value (KV) cache that grows linearly with interaction steps. For instance, UI-TARS-1.5-7B consumes 76 GB of GPU memory on merely five screenshots, approaching the capacity of mainstream 80 GB accelerators. Existing KV compression methods share two structural assumptions: aggregating visual-token importance into a single shared saliency map, and applying a fixed top-B cutoff to the fused score distribution. Pilot measurements refute both: spatial specialization lives at the attention-subspace level and migrates across layers, while the score distribution drifts in shape along a trajectory. We propose STaR-KV (Spatio-Temporal Adaptive Re-weighting), a training-free KV cache compression framework that calibrates token importance along three axes: (i) subspace-aware scoring driven by online spatial mutual information; (ii) a temporal stability discount that suppresses redundant cache entries from persistently attended subspaces; and (iii) an entropy-derived temperature that adaptively reshapes the score distribution. Across four GUI benchmarks, STaR-KV achieves the strongest average accuracy among state-of-the-art KV compression methods (e.g., GUIKV, SnapKV) at matched budgets, with no compression-stage FLOPs overhead (-0.07%) and cutting peak GPU memory by nearly 40% at a 20% KV-cache budget. Code is available at https://github.com/kawhiiiileo/STaR-KV.

Pei Yang, Wanyi Chen, Tongyun Yang et al.

LLM routing matters most in long-horizon applications such as coding agents, deep research systems, and computer-use agents, where a single user request triggers many model calls. Routing each call to the cheapest sufficient model can cut costs without sacrificing quality, yet existing router benchmarks evaluate routers only on one-shot prompts. They never expose the router-visible prefix at an intermediate agent step, never test whether a cheaper replacement preserves downstream task success, and often rely on online LLM judges at evaluation time. We introduce TwinRouterBench, a step-level routing benchmark with two tracks. The static track provides 970 router-visible prefixes from 520 instances across SWE-bench, BFCL, mtRAG, QMSum, and PinchBench, each paired with an execution-verified target tier estimated under a released downgrade-and-cascade protocol; scoring is deterministic arithmetic over tier labels, trajectory membership, and token costs, with no online evaluator-side LLM judge. The dynamic track supplies a harness that runs routers on the full 500-case SWE-bench Verified suite; in this paper we report a 100-case held-out evaluation disjoint from the static SWE supervision split. At each LLM call the router selects a concrete model from a locked pool, and success is measured by official task resolution and realized API spend. The two tracks support fast offline iteration followed by end-to-end validation under live agent execution. Code and data are available at https://github.com/CommonstackAI/TwinRouterBench.

Shuo Sun, Zekai Gu, Tianchen Sun et al.

Realistic and diverse traffic scenarios in large quantities are crucial for the development and validation of autonomous driving systems. However, owing to numerous difficulties in the data collection process and the reliance on intensive annotations, real-world datasets lack sufficient quantity and diversity to support the increasing demand for data. This work introduces DriveSceneGen, a data-driven driving scenario generation method that learns from the real-world driving dataset and generates entire dynamic driving scenarios from scratch. DriveSceneGen is able to generate novel driving scenarios that align with real-world data distributions with high fidelity and diversity. Experimental results on 5k generated scenarios highlight the generation quality, diversity, and scalability compared to real-world datasets. To the best of our knowledge, DriveSceneGen is the first method that generates novel driving scenarios involving both static map elements and dynamic traffic participants from scratch.

Zichen Wen, Boxue Yang, Shuang Chen et al.

We present Innovator-VL, a scientific multimodal large language model designed to advance understanding and reasoning across diverse scientific domains while maintaining excellent performance on general vision tasks. Contrary to the trend of relying on massive domain-specific pretraining and opaque pipelines, our work demonstrates that principled training design and transparent methodology can yield strong scientific intelligence with substantially reduced data requirements. (i) First, we provide a fully transparent, end-to-end reproducible training pipeline, covering data collection, cleaning, preprocessing, supervised fine-tuning, reinforcement learning, and evaluation, along with detailed optimization recipes. This facilitates systematic extension by the community. (ii) Second, Innovator-VL exhibits remarkable data efficiency, achieving competitive performance on various scientific tasks using fewer than five million curated samples without large-scale pretraining. These results highlight that effective reasoning can be achieved through principled data selection rather than indiscriminate scaling. (iii) Third, Innovator-VL demonstrates strong generalization, achieving competitive performance on general vision, multimodal reasoning, and scientific benchmarks. This indicates that scientific alignment can be integrated into a unified model without compromising general-purpose capabilities. Our practices suggest that efficient, reproducible, and high-performing scientific multimodal models can be built even without large-scale data, providing a practical foundation for future research.

Yuhang Han, Xuyang Liu, Zihan Zhang et al.

The quadratic complexity of Multimodal Large Language Models (MLLMs) with respect to context length poses significant computational and memory challenges, hindering their real-world deployment. In the paper, we devise a ''filter-correlate-compress'' framework to accelerate the MLLM by systematically optimizing multimodal context length during prefilling. The framework first implements FiCoCo-V, a training-free method operating within the vision encoder. It employs a redundancy-based token discard mechanism that uses a novel integrated metric to accurately filter out redundant visual tokens. To mitigate information loss, the framework introduces a correlation-based information recycling mechanism that allows preserved tokens to selectively recycle information from correlated discarded tokens with a self-preserving compression, thereby preventing the dilution of their own core content. The framework's FiCoCo-L variant further leverages task-aware textual priors to perform token reduction directly within the LLM decoder. Extensive experiments demonstrate that the FiCoCo series effectively accelerates a range of MLLMs, achieves up to 14.7x FLOPs reduction with 93.6% performance retention. Our methods consistently outperform state-of-the-art training-free approaches, showcasing effectiveness and generalizability across model architectures, sizes, and tasks without requiring retraining. Code: https://github.com/kawhiiiileo/FiCoCo

Yaojie Zhang, Jianuo Huang, Junlong Ke et al.

Speculative decoding accelerates memory-bound LLM inference without quality degradation by using a fast drafter to propose multiple candidate tokens and the target model to verify them in parallel. However, conventional sequential speculative decoding suffers from mutual waiting between drafting and verification, and repeated exchange of intermediate states further increases memory access overhead. Parallel speculative decoding addresses this limitation by performing drafting and verification within a single target forward pass, allowing future drafts to be prepared while current candidates are being verified. Although effective at small batch sizes, existing parallel speculative decoding methods either require costly continual pretraining with quality degradation or suffer from low acceptance rates. More importantly, this paradigm inherently suffers from uncertainty in both the bonus token and the accepted length, leading to draft verification mismatch and causing throughput gains to collapse at large batch sizes. To address these limitations, we introduce FlexDraft, a lossless speculative decoding framework that flexibly adapts to varying batch sizes through three key designs. (1) Attention Tuning enables block diffusion drafting by tuning only the attention projectors of the final few layers on mask tokens, while keeping the autoregressive path frozen to preserve the target distribution and produce high quality drafts with minimal trainable parameters. (2) Bonus-guided Calibration uses a lightweight MLP conditioned on the resolved bonus token to calibrate draft logits, mitigating draft verification mismatch caused by bonus token uncertainty. (3) Flex Decoding dynamically switches between parallel draft and verify at small batch sizes and sequential draft then verify at large batch sizes, and adjusts verification length based on draft confidence to eliminate redundant computation.

Xuyang Liu, Ziming Wang, Junjie Chen et al.

Large vision-language models (LVLMs) excel at visual understanding, but face efficiency challenges due to quadratic complexity in processing long multi-modal contexts. While token compression can reduce computational costs, existing approaches are designed for single-view LVLMs and fail to consider the unique multi-view characteristics of high-resolution LVLMs with dynamic cropping. Existing methods treat all tokens uniformly, but our analysis reveals that global thumbnails can naturally guide the compression of local crops by providing holistic context for informativeness evaluation. In this paper, we first analyze dynamic cropping strategy, revealing both the complementary nature between thumbnails and crops, and the distinctive characteristics across different crops. Based on our observations, we propose "Global Compression Commander" (GlobalCom$^2$), a novel plug-and-play token compression framework for HR-LVLMs. GlobalCom$^2$ leverages thumbnail as the "commander" to guide the compression of local crops, adaptively preserving informative details while eliminating redundancy. Extensive experiments show that GlobalCom$^2$ maintains over 90% performance while compressing 90% visual tokens, reducing FLOPs and peak memory to 9.1% and 60%. Our code is available at https://github.com/xuyang-liu16/GlobalCom2.

Yuhang Han, Yuyang Wu, Zhengbo Jiao et al.

Reinforcement Learning from Verifiable Rewards (RLVR) has substantially enhanced the reasoning capabilities of large language models in abstract reasoning tasks. However, its application to Large Vision-Language Models (LVLMs) remains constrained by a structural representational bottleneck. Existing approaches generally lack explicit modeling and effective utilization of visual information, preventing visual representations from being tightly coupled with the reinforcement learning optimization process and thereby limiting further improvements in multimodal reasoning performance. To address this limitation, we propose KAWHI (Key-Region Aligned Weighted Harmonic Incentive), a plug-and-play reward reweighting mechanism that explicitly incorporates structured visual information into uniform reward policy optimization methods (e.g., GRPO and GSPO). The method adaptively localizes semantically salient regions through hierarchical geometric aggregation, identifies vision-critical attention heads via structured attribution, and performs paragraph-level credit reallocation to align spatial visual evidence with semantically decisive reasoning steps. Extensive empirical evaluations on diverse reasoning benchmarks substantiate KAWHI as a general-purpose enhancement module, consistently improving the performance of various uniform reward optimization methods. Project page: KAWHI (https://kawhiiiileo.github.io/KAWHI_PAGE/)

Xinying Guo, Chenxi Jiang, Hyun Bin Kim et al.

Robotic manipulation often requires memory: occlusion and state changes can make decision-time observations perceptually aliased, making action selection non-Markovian at the observation level because the same observation may arise from different interaction histories. Most embodied agents implement memory via semantically compressed traces and similarity-based retrieval, which discards disambiguating fine-grained perceptual cues and can return perceptually similar but decision-irrelevant episodes. Inspired by human episodic memory, we propose Chameleon, which writes geometry-grounded multimodal tokens to preserve disambiguating context and produces goal-directed recall through a differentiable memory stack. We also introduce Camo-Dataset, a real-robot UR5e dataset spanning episodic recall, spatial tracking, and sequential manipulation under perceptual aliasing. Across tasks, Chameleon consistently improves decision reliability and long-horizon control over strong baselines in perceptually confusable settings.

Jiahui Li, Tianle Shen, Zekai Gu et al.

Motion prediction is a challenging problem in autonomous driving as it demands the system to comprehend stochastic dynamics and the multi-modal nature of real-world agent interactions. Diffusion models have recently risen to prominence, and have proven particularly effective in pedestrian motion prediction tasks. However, the significant time consumption and sensitivity to noise have limited the real-time predictive capability of diffusion models. In response to these impediments, we propose a novel diffusion-based, acceleratable framework that adeptly predicts future trajectories of agents with enhanced resistance to noise. The core idea of our model is to learn a coarse-grained prior distribution of trajectory, which can skip a large number of denoise steps. This advancement not only boosts sampling efficiency but also maintains the fidelity of prediction accuracy. Our method meets the rigorous real-time operational standards essential for autonomous vehicles, enabling prompt trajectory generation that is vital for secure and efficient navigation. Through extensive experiments, our method speeds up the inference time to 136ms compared to standard diffusion model, and achieves significant improvement in multi-agent motion prediction on the Argoverse 1 motion forecasting dataset.

Yichen Han, Yuhang Han, Bojun Liu et al.

Prompt engineering is crucial for fully leveraging large language models (LLMs), yet most existing optimization methods follow a single trajectory, resulting in limited adaptability, gradient conflicts, and high computational overhead. We propose MAPGD (Multi-Agent Prompt Gradient Descent), a novel framework that reconceptualizes prompt optimization as a collaborative process among specialized agents. Each agent focuses on a distinct refinement dimension, such as instruction clarity, example selection, format structure, or stylistic adaptation, and their contributions are coordinated through semantic gradient embedding, conflict detection, and fusion. To further enhance robustness and stability, MAPGD introduces two new mechanisms: Hypersphere Constrained Gradient Clustering (HCGC), which enforces angular margin constraints for compact and well-separated clusters, and Channel Adaptive Agent Weighting (CAAW), which dynamically reweights agent contributions based on validation performance. Experiments on classification and reasoning benchmarks show that MAPGD consistently surpasses single-agent and random baselines in both accuracy and efficiency. Ablation studies confirm the effectiveness of gradient fusion, agent specialization, and conflict resolution. Together, these components establish MAPGD as a unified, gradient-based, and interpretable framework for robust prompt optimization with theoretical convergence guarantees.