Wang Yang

Wang Yang, Debargha Ganguly, Xinpeng Li et al.

Hybrid reasoning language models are commonly controlled through high-level Think/No-think instructions to regulate reasoning behavior, yet we found that such mode switching is largely driven by a small set of trigger tokens rather than the instructions themselves. Through attention analysis and controlled prompting experiments, we show that a leading ``Okay'' token induces reasoning behavior, while the newline pattern following ``</think>'' suppresses it. Based on this observation, we propose Mid-Think, a simple training-free prompting format that combines these triggers to achieve intermediate-budget reasoning, consistently outperforming fixed-token and prompt-based baselines in terms of the accuracy-length trade-off. Furthermore, applying Mid-Think to RL training after SFT reduces training time by approximately 15% while improving final performance of Qwen3-8B on AIME from 69.8% to 72.4% and on GPQA from 58.5% to 61.1%, demonstrating its effectiveness for both inference-time control and RL-based reasoning training.

Chuang Ma, Qianying Liu, Tomoyuki Obuchi et al.

Multimodal large language models (MLLMs) remain unreliable on spatial multiple-choice questions, and their failures are often attributed to poorly attended visual information. In this work, we identify a complementary failure mode, spatial lexical bias: adding a spatial relation word to the answer options can attract the model's decision and make the newly added option likely to be selected. Using nine open-weight MLLMs, we show that this phenomenon is widely observed. In particular, models can answer a binary spatial question correctly, yet consistently select an incorrect third spatial option once it is added to the answer set. We isolate such binary-stable but ternary-fragile cases as diagnostic examples and leverage mechanistic interpretability tools, revealing that a substantial part of the failure instead originates on the language side rather than the visual side: visual attention analyses and residual-stream probes show the correct spatial relation remains internally available on these failures, while irrelevant-option controls, activation patching, and sparse component interventions trace the bias to specific LLM-side channels and neurons. Based on this finding, we show that a lightweight LLM-only DPO update on tiny single-object-pair synthetic data mitigates the bias, lifting four-way robust accuracy by up to 100 points on synthetic data, and by 68.0, 32.6, and 20.1 points on broader evaluation datasets WhatsUp, SpatialMQA-Direct, and VSR.

Candi Zheng, Wang Yang, Shiyi Chen

The maximal entropy moment method (MEM) is systematic solution of the challenging problem: generating extended hydrodynamic equations valid for both dense and rarefied gases. However, simulating MEM suffers from a computational expensive and ill-conditioned maximal entropy problem. It causes numerical overflow and breakdown when the numerical precision is insufficient, especially for flows like high-speed shock waves. It also prevents modern GPUs from accelerating MEM with their enormous single floating-point precision computation power. This paper aims to stabilize MEM, making it possible to simulating very strong normal shock waves on modern GPUs at single precision. We improve the condition number of the maximal entropy problem by proposing gauge transformations, which moves not only flow fields but also hydrodynamic equations into a more optimal coordinate system. We addressed numerical overflow and breakdown in the maximal entropy problem by employing the canonical form of distribution and a modified Newton optimization method. Moreover, we discovered a counter-intuitive phenomenon that over-refined spatial mesh beyond mean free path degrades the stability of MEM. With these techniques, we accomplished single-precision GPU simulations of high speed shock wave up to Mach 10 utilizing 35 moments MEM, while previous methods only achieved Mach 4 on double-precision.

Wei Zhiwei, Xiao Yi, Tong Ying et al.

Building patterns are important urban structures that reflect the effect of the urban material and social-economic on a region. Previous researches are mostly based on the graph isomorphism method and use rules to recognize building patterns, which are not efficient. The knowledge graph uses the graph to model the relationship between entities, and specific subgraph patterns can be efficiently obtained by using relevant reasoning tools. Thus, we try to apply the knowledge graph to recognize linear building patterns. First, we use the property graph to express the spatial relations in proximity, similar and linear arrangement between buildings; secondly, the rules of linear pattern recognition are expressed as the rules of knowledge graph reasoning; finally, the linear building patterns are recognized by using the rule-based reasoning in the built knowledge graph. The experimental results on a dataset containing 1289 buildings show that the method in this paper can achieve the same precision and recall as the existing methods; meanwhile, the recognition efficiency is improved by 5.98 times.

Yipu Dou, Wang Yang

We study how to allocate a fixed supervised fine-tuning budget when three objectives must be balanced at once: multi-turn safety alignment, low over-refusal on benign boundary queries, and instruction following under verifiable constraints. We propose MOSAIC (Multi-Objective Slice-Aware Iterative Curation for Alignment), a multi-objective framework for closed-loop data mixture search built on a unified L1-L3 evaluation interface. MOSAIC turns slice-level failure profiles into executable data actions, including dataset-level mixture ratios, bucket-level weights, and focus criteria. Under a fixed 1M-token budget and five rounds of independent fine-tuning from the same base model, MOSAIC improves internal XGuard from 2.76 to 4.67 while keeping OrBench at 4.41 and IFEval at 3.65. The final Pareto solution also generalizes better than a random static LoRA baseline on independent attack, over-refusal, and capability tests, suggesting that structured failure diagnosis can serve as a practical control signal for budgeted data construction. Code is available at https://github.com/douyipu/mosaic.

Yipu Dou, Wang Yang

As Large Language Models (LLMs) evolve from static chatbots into autonomous agents capable of tool execution, the landscape of AI safety is shifting from content moderation to action security. However, existing red-teaming frameworks remain bifurcated: they either focus on rigid, script-based text attacks or lack the architectural modularity to simulate complex, multi-turn agentic exploitations. In this paper, we introduce AJAR (Adaptive Jailbreak Architecture for Red-teaming), a proof-of-concept framework designed to bridge this gap through Protocol-driven Cognitive Orchestration. Built upon the robust runtime of Petri, AJAR leverages the Model Context Protocol (MCP) to decouple adversarial logic from the execution loop, encapsulating state-of-the-art algorithms like X-Teaming as standardized, plug-and-play services. We validate the architectural feasibility of AJAR through a controlled qualitative case study, demonstrating its ability to perform stateful backtracking within a tool-use environment. Furthermore, our preliminary exploration of the "Agentic Gap" reveals a complex safety dynamic: while tool usage introduces new injection vectors via code execution, the cognitive load of parameter formatting can inadvertently disrupt persona-based attacks. AJAR is open-sourced to facilitate the standardized, environment-aware evaluation of this emerging attack surface. The code and data are available at https://github.com/douyipu/ajar.

Phat Thanh Dang, Saahil Thoppay, Wang Yang et al.

Large language models suffer issues when operated on long contexts that are larger than their training context length due to the standard position encoding for tokens in the attention layer. Tokens a long distance apart will rarely have an effect on each other and long prompts yield unexpected results. To solve this problem, we propose SELF (Self-Extend the Context Length With Logistic Growth Function): a solution of grouping consecutive tokens at varying group sizes using a logistic capacity equation combined with a constant group size at smaller relative distances. Our model had an increase in performance of up to 12% compared to the LongLM extension method in LEval (specifically on the Qwen model). On summarization related tasks in LongBench, our model performed up to 6.4% better than LongLM (specifically on the Llama-2-7b model). On reading comprehension tasks from LEval, our model performed up to 5.4% better than the LongLM. Our code is available at https://github.com/alexeipc/SELF-LLM.

Dou Goodman, Hao Xin, Wang Yang et al.

In recent years, neural networks have been extensively deployed for computer vision tasks, particularly visual classification problems, where new algorithms reported to achieve or even surpass the human performance. Recent studies have shown that they are all vulnerable to the attack of adversarial examples. Small and often imperceptible perturbations to the input images are sufficient to fool the most powerful neural networks. \emph{Advbox} is a toolbox to generate adversarial examples that fool neural networks in PaddlePaddle, PyTorch, Caffe2, MxNet, Keras, TensorFlow and it can benchmark the robustness of machine learning models. Compared to previous work, our platform supports black box attacks on Machine-Learning-as-a-service, as well as more attack scenarios, such as Face Recognition Attack, Stealth T-shirt, and DeepFake Face Detect. The code is licensed under the Apache 2.0 and is openly available at https://github.com/advboxes/AdvBox. Advbox now supports Python 3.

Xiaoran Cai, Wang Yang, Xiyu Ren et al.

Sustainability or ESG rating agencies use company disclosures and external data to produce scores or ratings that assess the environmental, social, and governance performance of a company. However, sustainability ratings across agencies for a single company vary widely, limiting their comparability, credibility, and relevance to decision-making. To harmonize the rating results, we propose adopting a universal human-AI collaboration framework to generate trustworthy benchmark datasets for evaluating sustainability rating methodologies. The framework comprises two complementary parts: STRIDE (Sustainability Trust Rating & Integrity Data Equation) provides principled criteria and a scoring system that guide the construction of firm-level benchmark datasets using large language models (LLMs), and SR-Delta, a discrepancy-analysis procedural framework that surfaces insights for potential adjustments. The framework enables scalable and comparable assessment of sustainability rating methodologies. We call on the broader AI community to adopt AI-powered approaches to strengthen and advance sustainability rating methodologies that support and enforce urgent sustainability agendas.

Wang Yang, Xiang Yue, Vipin Chaudhary et al.

Recent advances leverage post-training to enhance model reasoning performance, which typically requires costly training pipelines and still suffers from inefficient, overly lengthy outputs. We introduce Speculative Thinking, a training-free framework that enables large reasoning models to guide smaller ones during inference at the reasoning level, distinct from speculative decoding, which operates at the token level. Our approach is based on two observations: (1) reasoning-supportive tokens such as "wait" frequently appear after structural delimiters like "\n\n", serving as signals for reflection or continuation; and (2) larger models exhibit stronger control over reflective behavior, reducing unnecessary backtracking while improving reasoning quality. By strategically delegating reflective steps to a more capable model, our method significantly boosts the reasoning accuracy of reasoning models while shortening their output. With the assistance of the 32B reasoning model, the 1.5B model's accuracy on MATH500 increases from 83.2% to 89.4%, marking a substantial improvement of 6.2%. Simultaneously, the average output length is reduced from 5439 tokens to 4583 tokens, representing a 15.7% decrease. Moreover, when applied to a non-reasoning model (Qwen-2.5-7B-Instruct), our framework boosts its accuracy from 74.0% to 81.8% on the same benchmark, achieving a relative improvement of 7.8%.

Shouren Wang, Wang Yang, Chuang Ma et al.

Hybrid-thinking language models expose explicit think and no-think modes, but current designs do not separate them cleanly. Even in no-think mode, models often emit long and self-reflective responses, causing reasoning leakage. Existing work reduces this issue through better data curation and multi-stage training, yet leakage remains because both modes are still encoded in the same feed-forward parameters. We propose Path-Lock Expert (PLE), an architecture-level solution that replaces the single MLP in each decoder layer with two semantically locked experts, one for think and one for no-think, while keeping attention, embeddings, normalization, and the language-model head shared. A deterministic control-token router selects exactly one expert path for the entire sequence, so inference preserves the dense model's per-token computation pattern and each expert receives mode-pure updates during supervised fine-tuning. Across math and science reasoning benchmarks, PLE maintains strong think performance while producing a substantially stronger no-think mode that is more accurate, more concise, and far less prone to reasoning leakage. On Qwen3-4B, for example, PLE reduces no-think reflective tokens on AIME24 from 2.54 to 0.39 and improves no-think accuracy from 20.67% to 40.00%, all while preserving think-mode performance. These results suggest that controllable hybrid thinking is fundamentally an architectural problem, and separating mode-specific feed-forward pathways is a simple and effective solution.

Wang Yang, Hongye Jin, Jingfeng Yang et al.

Supervised Fine-Tuning (SFT) has been a go-to and effective method for enhancing long chain-of-thought (CoT) reasoning in relatively small LLMs by fine-tuning them with long CoT responses from larger LLMs. To continually improve reasoning abilities, we can either collect new high-quality long CoT reasoning SFT data or repeatedly train on existing SFT datasets. However, acquiring new long CoT SFT data is costly and limited, while repeated training often results in a performance plateau or decline. To further boost the performance with the SFT data, we propose Thinking Preference Optimization (ThinkPO), a simple yet effective post-SFT method that enhances long CoT reasoning without requiring new long CoT responses. Instead, ThinkPO utilizes readily available or easily obtainable short CoT reasoning responses as rejected answers and long CoT responses as chosen answers for the same question. It then applies direct preference optimization to encourage the model to favor longer reasoning outputs. Experiments show that ThinkPO further improves the reasoning performance of SFT-ed models, e.g. it increases math reasoning accuracy of SFT-ed models by 8.6% and output length by 25.9%. Notably, ThinkPO is capable of continually boosting the performance of the publicly distilled SFT model, e.g., increasing the official DeepSeek-R1-Distill-Qwen-7B's performance on MATH500 from 87.4% to 91.2%.

Wang Yang, Zirui Liu, Hongye Jin et al.

Recent language models exhibit strong reasoning capabilities, yet the influence of long-context capacity on reasoning remains underexplored. In this work, we hypothesize that current limitations in reasoning stem, in part, from insufficient long-context capacity, motivated by empirical observations such as (1) higher context window length often leads to stronger reasoning performance, and (2) failed reasoning cases resemble failed long-context cases. To test this hypothesis, we examine whether enhancing a model's long-context ability before Supervised Fine-Tuning (SFT) leads to improved reasoning performance. Specifically, we compared models with identical architectures and fine-tuning data but varying levels of long-context capacity. Our results reveal a consistent trend: models with stronger long-context capacity achieve significantly higher accuracy on reasoning benchmarks after SFT. Notably, these gains persist even on tasks with short input lengths, indicating that long-context training offers generalizable benefits for reasoning performance. These findings suggest that long-context modeling is not just essential for processing lengthy inputs, but also serves as a critical foundation for reasoning. We advocate for treating long-context capacity as a first-class objective in the design of future language models.

Wang Yang, Chaoda Song, Xinpeng Li et al.

Existing Agent benchmarks suffer from two critical limitations: high environment interaction overhead (up to 41\% of total evaluation time) and imbalanced task horizon and difficulty distributions that make aggregate scores unreliable. To address these issues, we propose ACE-Bench built around a unified grid-based planning task, where agents must fill hidden slots in a partially completed schedule subject to both local slot constraints and global constraints. Our benchmark offers fine-grained control through two orthogonal axes: Scalable Horizons, controlled by the number of hidden slots $H$, and Controllable Difficulty, governed by a decoy budget $B$ that determines the number of globally misleading decoy candidates. Crucially, all tool calls are resolved via static JSON files under a Lightweight Environment design, eliminating setup overhead and enabling fast, reproducible evaluation suitable for training-time validation. We first validate that H and B provide reliable control over task horizon and difficulty, and that ACE-Bench exhibits strong domain consistency and model discriminability. We then conduct comprehensive experiments across 13 models of diverse sizes and families over 6 domains, revealing significant cross-model performance variation and confirming that ACE-Bench provides interpretable and controllable evaluation of agent reasoning.

Wang Yang, Hongye Jin, Shaochen Zhong et al.

Long-context capability is considered one of the most important abilities of LLMs, as a truly long context-capable LLM enables users to effortlessly process many originally exhausting tasks -- e.g., digesting a long-form document to find answers vs. directly asking an LLM about it. However, existing real-task-based long-context evaluation benchmarks have two major shortcomings. First, benchmarks like LongBench often do not provide proper metrics to separate long-context performance from the model's baseline ability, making cross-model comparison unclear. Second, such benchmarks are usually constructed with fixed input lengths, which limits their applicability across different models and fails to reveal when a model begins to break down. To address these issues, we introduce a length-controllable long-context benchmark and a novel metric that disentangles baseline knowledge from true long-context capabilities. Experiments demonstrate the superiority of our approach in effectively evaluating LLMs.

Wang Yang, Shouren Wang, Chaoda Song et al.

Group Relative Policy Optimization (GRPO) has become a key technique for improving reasoning abilities in large language models, yet its behavior under different domain sequencing strategies is poorly understood. In particular, the impact of sequential (one domain at a time) versus mixed-domain (multiple domain at a time) training in GRPO has not been systematically studied. We provide the first systematic analysis of training-order effects across math, science, logic, and puzzle reasoning tasks. We found (1) single-domain generalization is highly asymmetric: training on other domains improves math reasoning by approximately 25\% accuracy, while yielding negligible transfer to logic and puzzle; (2) cross-domain interactions are highly order-dependent: training in the order math$\rightarrow$science achieves 83\% / 41\% accuracy on math / science, while reversing the order to science$\rightarrow$math degrades performance to 77\% / 25\%; (3) no single strategy is universally optimal in multi-domain training: sequential training favors math (up to 84\%), mixed training favors science and logic, and poor ordering can incur large performance gaps (from 70\% to 56\%). Overall, our findings demonstrate that GRPO under multi-domain settings exhibits pronounced asymmetry, order sensitivity, and strategy dependence, highlighting the necessity of domain-aware and order-aware training design.

Shouren Wang, Wang Yang, Xianxuan Long et al.

Hybrid thinking enables LLMs to switch between reasoning and direct answering, offering a balance between efficiency and reasoning capability. Yet our experiments reveal that current hybrid thinking LLMs only achieve partial mode separation: reasoning behaviors often leak into the no-think mode. To understand and mitigate this, we analyze the factors influencing controllability and identify four that matter most: (1) larger data scale, (2) using think and no-think answers from different questions rather than the same question, (3) a moderate increase in no-think data number, and (4) a two-phase strategy that first trains reasoning ability and then applies hybrid think training. Building on these findings, we propose a practical recipe that, compared to standard training, can maintain accuracy in both modes while significantly reducing no-think output length (from $1085$ to $585$ on MATH500) and occurrences of reasoning-supportive tokens such as ``\texttt{wait}'' (from $5917$ to $522$ on MATH500). Our findings highlight the limitations of current hybrid thinking and offer directions for strengthening its controllability.