Kangyu Wang

Weixun Wang, XiaoXiao Xu, Wanhe An et al.

Agentic crafting requires LLMs to operate in real-world environments over multiple turns by taking actions, observing outcomes, and iteratively refining artifacts. Despite its importance, the open-source community lacks a principled, end-to-end ecosystem to streamline agent development. We introduce the Agentic Learning Ecosystem (ALE), a foundational infrastructure that optimizes the production pipeline for agentic model. ALE consists of three components: ROLL, a post-training framework for weight optimization; ROCK, a sandbox environment manager for trajectory generation; and iFlow CLI, an agent framework for efficient context engineering. We release ROME, an open-source agent grounded by ALE and trained on over one million trajectories. Our approach includes data composition protocols for synthesizing complex behaviors and a novel policy optimization algorithm, Interaction-Perceptive Agentic Policy Optimization (IPA), which assigns credit over semantic interaction chunks rather than individual tokens to improve long-horizon training stability. Empirically, we evaluate ROME within a structured setting and introduce Terminal Bench Pro, a benchmark with improved scale and contamination control. ROME demonstrates strong performance across benchmarks like SWE-bench Verified and Terminal Bench, proving the effectiveness of ALE.

Yuxin Ma, Lun Du, Lanning Wei et al.

Diffusion-based large language models (dLLMs) have emerged as a promising alternative to autoregressive (AR) LLMs, leveraging denoising-based generation to enable inherent parallelism. Even more and more open-sourced dLLM models emerge, yet their widespread adoption remains constrained by the lack of a standardized and efficient inference framework. We present dInfer, an efficient and extensible framework for dLLM inference. dInfer decomposes the inference pipeline into four modular components--model, diffusion iteration manager, decoding strategy, and KV-cache manager--and integrates novel algorithms for each component alongside system-level optimizations. Through this combination of algorithmic innovations and system enhancements, dInfer achieves substantial efficiency gains without compromising output quality on LLaDA-MoE. At batch size 1, it surpasses 1,100 tokens per second on HumanEval and averages over 800 tokens per second across six benchmarks on $8\times$ H800 GPUs. Compared to prior systems, dInfer delivers a $10\times$ speedup over Fast-dLLM while maintaining similar model performance. Even compared to the AR model (with a comparable number of activation parameters and performance) QWen2.5-3B, which is highly optimized with the latest vLLM inference engine, dInfer still delivers a $2$-$3\times$ speedup. The implementation of dInfer is open-sourced at https://github.com/inclusionAI/dInfer.

Zhihao He, Tieyuan Chen, Kangyu Wang et al.

Standard Autoregressive Video LLMs inevitably suffer from causal masking biases that hinder global spatiotemporal modeling, leading to suboptimal understanding efficiency. We propose VidLaDA, a Video LLM based on Diffusion Language Model utilizing bidirectional attention to capture bidirectional dependencies. To further tackle the inference bottleneck of diffusion decoding on massive video tokens, we introduce MARS-Cache. This framework accelerates inference by combining asynchronous visual cache refreshing with frame-wise chunk attention, effectively pruning redundancy while preserving global connectivity via anchor tokens. Extensive experiments show VidLaDA outperforms diffusion baselines and rivals state-of-the-art autoregressive models (e.g., Qwen2.5-VL and LLaVA-Video), with MARS-Cache delivering over 12x speedup without compromising reasoning accuracy. Code and checkpoints are open-sourced at https://github.com/ziHoHe/VidLaDA.

Zicheng Zhao, Kangyu Wang, Shijie Li et al.

Despite advancements in Video Large Language Models (Vid-LLMs) improving multimodal understanding, challenges persist in streaming video reasoning due to its reliance on contextual information. Existing paradigms feed all available historical contextual information into Vid-LLMs, resulting in a significant computational burden for visual data processing. Furthermore, the inclusion of irrelevant context distracts models from key details. This paper introduces a challenging task called Context-guided Streaming Video Reasoning (CogStream), which simulates real-world streaming video scenarios, requiring models to identify the most relevant historical contextual information to deduce answers for questions about the current stream. To support CogStream, we present a densely annotated dataset featuring extensive and hierarchical question-answer pairs, generated by a semi-automatic pipeline. Additionally, we present CogReasoner as a baseline model. It efficiently tackles this task by leveraging visual stream compression and historical dialogue retrieval. Extensive experiments prove the effectiveness of this method. The project is released on https://github.com/LiamZhao326/CogStream.

Fengqi Zhu, Zebin You, Yipeng Xing et al.

We introduce LLaDA-MoE, a large language diffusion model with the Mixture-of-Experts (MoE) architecture, trained from scratch on approximately 20T tokens. LLaDA-MoE achieves competitive performance with significantly reduced computational overhead by maintaining a 7B-parameter capacity while activating only 1.4B parameters during inference. Our empirical evaluation reveals that LLaDA-MoE achieves state-of-the-art performance among diffusion language models with larger parameters, surpassing previous diffusion language models LLaDA, LLaDA 1.5, and Dream across multiple benchmarks. The instruct-tuned model LLaDA-MoE-7B-A1B-Instruct demonstrates capabilities comparable to Qwen2.5-3B-Instruct in knowledge understanding, code generation, mathematical reasoning, agent and alignment tasks, despite using fewer active parameters. Our results show that integrating a sparse MoE architecture into the training objective of masked diffusion language models still brings out MoE's strengths under efficient inference with few active parameters, and opens ample room for further exploration of diffusion language models. LLaDA-MoE models are available at Huggingface.

Kangyu Wang, Zhiyun Jiang, Haibo Feng et al.

Diffusion large language models (dLLMs) generate text through iterative denoising steps, achieving parallel decoding by denoising only high-confidence positions at each step. However, existing approaches often repetitively remask tokens due to initially low confidence scores, leading to redundant iterations and limiting overall acceleration. Through the analysis of dLLM decoding traces, we observe that the model often determines the final prediction for a token several steps before the decoding step. To leverage this historical information and avoid redundant steps, we introduce the concept of Trace Credit, which quantifies each token's convergence potential by accumulating historical logits. Furthermore, we propose CreditDecoding, a training-free parallel decoding algorithm that accelerates the confidence convergence of correct but underconfident tokens by fusing current logits with Trace Credit. This process significantly reduces redundant iterations and enhances decoding robustness. On eight benchmarks, CreditDecoding achieves a 5.48 times speedup and a 0.48 performance improvement over LLaDA-8B-Instruct, and a 4.11 times speedup with a 0.15 performance improvement over LLaDA-MoE-Instruct. Importantly, CreditDecoding scales effectively to long sequences and is orthogonal to mainstream inference optimizations, making it a readily integrable and versatile solution.

Kangyu Wang, Hongliang He, Lin Liu et al.

Large Language Models (LLMs) and Multimodal Large Language Models (MLLMs) have ushered in a new era of AI capabilities, demonstrating near-human-level performance across diverse scenarios. While numerous benchmarks (e.g., MMLU) and leaderboards (e.g., Chatbot Arena) have been proposed to help evolve the development of LLMs and MLLMs, most rely on static datasets or crowdsourced general-domain prompts, often falling short of reflecting performance in real-world applications. To bridge this critical gap, we present Inclusion Arena, a live leaderboard that ranks models based on human feedback collected directly from AI-powered applications. Our platform integrates pairwise model comparisons into natural user interactions, ensuring evaluations reflect practical usage scenarios. For robust model ranking, we employ the Bradley-Terry model augmented with two key innovations: (1) Placement Matches, a cold-start mechanism to quickly estimate initial ratings for newly integrated models, and (2) Proximity Sampling, an intelligent comparison strategy that prioritizes battles between models of similar capabilities to maximize information gain and enhance rating stability. Extensive empirical analyses and simulations demonstrate that Inclusion Arena yields reliable and stable rankings, exhibits higher data transitivity compared to general crowdsourced datasets, and significantly mitigates the risk of malicious manipulation. By fostering an open alliance between foundation models and real-world applications, Inclusion Arena aims to accelerate the development of LLMs and MLLMs truly optimized for practical, user-centric deployments. The platform is publicly accessible at https://www.tbox.cn/about/model-ranking.

Kangyu Wang

Machine Learning ML agents have been increasingly used in decision-making across a wide range of tasks and environments. These ML agents are typically designed to balance multiple objectives when making choices. Understanding how their decision-making processes align with or diverge from human reasoning is essential. Human agents often encounter hard choices, that is, situations where options are incommensurable; neither option is preferred, yet the agent is not indifferent between them. In such cases, human agents can identify hard choices and resolve them through deliberation. In contrast, current ML agents, due to fundamental limitations in Multi-Objective Optimisation or MOO methods, cannot identify hard choices, let alone resolve them. Neither Scalarised Optimisation nor Pareto Optimisation, the two principal MOO approaches, can capture incommensurability. This limitation generates three distinct alignment problems: the alienness of ML decision-making behaviour from a human perspective; the unreliability of preference-based alignment strategies for hard choices; and the blockage of alignment strategies pursuing multiple objectives. Evaluating two potential technical solutions, I recommend an ensemble solution that appears most promising for enabling ML agents to identify hard choices and mitigate alignment problems. However, no known technique allows ML agents to resolve hard choices through deliberation, as they cannot autonomously change their goals. This underscores the distinctiveness of human agency and urges ML researchers to reconceptualise machine autonomy and develop frameworks and methods that can better address this fundamental gap.