Wenkai Wang

Wenkai Wang, Tao Xiong, Jingchen Ni et al.

Real-world professional desktop workflows in specialized creative and engineering software unfold over long horizons and often require human-in-the-loop coordination, where agents proactively seek necessary information and users provide additional instructions, clarifications, feedback, or corrections as the task progresses. Yet existing desktop GUI benchmarks mostly reduce this setting to short, simplified tasks with all user instructions provided upfront. To address this issue, we introduce DeskCraft, a desktop GUI benchmark targeting long horizon creative and engineering workflows and proactive human-agent collaboration. DeskCraft organizes tasks into a multilevel difficulty taxonomy, with long horizon tasks requiring over 50 execution steps, and covers professional creative software across design, video, audio, and 3D creation. Furthermore, DeskCraft formalizes human-agent collaboration into an interaction protocol covering mid-turn and post-turn exchanges. Mid-turn interaction captures both agent-initiated clarification under uncertainty and user-initiated interruption during execution, while post-turn interaction accommodates user-driven feedback after the agent signals completion, together spanning the full space of realistic collaboration patterns. We evaluate 18 proprietary and open source agents on 538 tasks and find that GPT-5.4 reaches 31.6% on standard tasks and 27.6% on interactive tasks. Further analyses reveal persistent failures in long horizon workflow delivery and proactive clarification. We will open-source all evaluation codes, tasks, and data at https://github.com/mrwwk/DeskCraft.

Wenkai Wang, Xiyun Li, Hongcan Guo et al. · tencent-ai

Graphical User Interface (GUI) grounding requires mapping natural language instructions to precise pixel coordinates. However, due to visually homogeneous elements and dense layouts, models typically grasp semantic intent yet struggle with achieving precise localization. While scaling sampling attempts (Pass@k) reveals potential gains, static self-consistency strategies derived from geometric clustering often yield limited improvements, as the model's predictions tend to be spatially dispersed. In this paper, we propose replacing static consistency strategies with a learnable selection mechanism that selects the optimal target by critiquing its own proposals rendered on the screenshot. Given the significant disparity between the model's grounding and critiquing capabilities, we propose a co-evolving Propose-then-Critic framework. To jointly optimize these, we introduce a maturity-aware adaptive co-evolutionary reinforcement learning paradigm. This approach dynamically balances the training objectives of proposer and critic, where the diversity of the proposer's outputs enhances critic robustness, while the critic's maturing discrimination capability conversely unlocks the proposer's potential for extensive spatial exploration, fostering the mutual reinforcement and co-evolution of both capabilities, thereby ensuring generalizability to adapt to diverse and complex interface layouts. Extensive experiments over 6 benchmarks show that our method significantly enhances both grounding accuracy and critic reliability.

Yixuan Wang, Yue Huang, Hong Qian et al.

Creativity has become a core competence in the era of LLMs and human-AI collaboration, underpinning innovation in real-world problem solving. Crucially, the systematic improvement of creativity necessitates scientifically valid assessment instruments. Psychometric research recognizes context-based assessment as an effective way to measure creative thinking. However, high-quality expert-designed contexts remain scarce. Existing LLM-based generators often struggle with insufficient assessment cues, weak narrative coherence, limited stylistic diversity, and poor support for creative thinking. To address these challenges, we propose AlphaContext, an evolutionary tree-based psychometric context generator for creativity assessment. First, the HyperTree Outline Planner formalizes expert-designed outlining as a rule-guided hypertree and performs top-down hierarchical planning. The MCTS-based Context Generator fills the outline via MCTS to balance global structure and local quality. Then, the Evolutionary Context Optimizer evolves contexts with MAP-Elites by repeatedly updating niche elites to jointly improve diversity and quality. Finally, the Assessment-Guided Evolution Refiner simulates virtual participants with diverse styles and recycles weak contexts for further evolution. Experiments show that AlphaContext yields an average improvement of 8% over competitive methods across 6 quality metrics.

Wenkai Wang, Hongcan Guo, Zheqi Lv et al.

Self-evaluation, a model's ability to assess the correctness of its own output, is crucial for Large Multimodal Models (LMMs) to achieve self-improvement in multi-turn conversations, yet largely absent in foundation models. Recent work has employed reinforcement learning (RL) to enhance self-evaluation; however, its fixed reward mechanism suffers from reward hacking when optimizing multiple training objectives, leading to model collapse. In this paper we propose AdaPO, an online reinforcement learning framework capable of adaptively adjusting training objective in real time according to the current training state for each task. Specifically, to mitigate reward hacking , AdaPO introduces an Adaptive Reward Model (ARM) and a Reward Aware Dynamic KL Regularization mechanism. ARM assesses the task's training state from the distribution of model generated multi-turn trajectories' performance. Reward Aware Dynamic KL replaces a fixed penalty with dynamic coefficients which is modulated by the reward gap between different multi-turn situations. Notably, our method automatically and smoothly adjusts its learning focus based on sub-tasks' training progress without manual intervention. Extensive experiments over 8 benchmarks and various models show that our method significantly enhances both direct reasoning and self-evaluation capability. We will release our code to contribute to the community.

Wenkai Wang, Vincent Lee, Yizhen Zheng

Large language model hallucination represents a critical challenge where outputs deviate from factual accuracy due to distributional biases in training data. While recent investigations establish that specific hidden layers exhibit differences between hallucinatory and factual content, the precise localization of hallucination signals within layers remains unclear, limiting the development of efficient detection methods. We propose a dual-model architecture integrating a Projected Fusion (PF) block for adaptive inter-layer feature weighting and a Differential Feature Learning (DFL) mechanism that identifies discriminative features by computing differences between parallel encoders learning complementary representations from identical inputs. Through systematic experiments across HaluEval's question answering, dialogue, and summarization datasets, we demonstrate that hallucination signals concentrate in highly sparse feature subsets, achieving significant accuracy improvements on question answering and dialogue tasks. Notably, our analysis reveals a hierarchical "funnel pattern" where shallow layers exhibit high feature diversity while deep layers demonstrate concentrated usage, enabling detection performance to be maintained with minimal degradation using only 1\% of feature dimensions. These findings suggest that hallucination signals are more concentrated than previously assumed, offering a pathway toward computationally efficient detection systems that could reduce inference costs while maintaining accuracy.

Zheqi Lv, Wenkai Wang, Jiawei Wang et al.

Efficient Multimodal Large Language Models (EMLLMs) can improve performance through Chain-of-Thought (CoT) reasoning, but they have poor self-evaluation capabilities during the CoT reasoning process. This is due to their tendency to simplify the reasoning process and the degradation of self-evaluation ability during downstream task fine-tuning. To address this, we intuitively propose \textit{Self-Evaluation Augmented Training (SEAT)}, which uses more powerful EMLLMs to evaluate CoT reasoning data. The evaluation data is then used to train EMLLMs. However, due to the difficulties EMLLMs face with processing long token input-output sequences, and the degradation of self-evaluation ability as a basis for CoT reasoning, the SEAT method is not fully adapted. Therefore, we further propose \textit{Cascaded Self-Evaluation Augmented Training (Cas-SEAT)}, which converts long prompts into cascaded short prompts, each focusing on a specific task. Additionally, we mix CoT reasoning and self-evaluation data to preserve its CoT reasoning ability while enhancing the self-evaluation capability of EMLLMs. We also conduct \textit{Double-level Data Filtering (DDF)}, which includes source data filtering and labeled data filtering, using both manual selection and MLLMs for filtering. Cas-SEAT and DDF work together to improve the performance of EMLLMs. Experiments show that Cas-SEAT achieves an average improvement of 22.16% across multiple datasets, and DDF significantly reduces the resource consumption of training