Reinforcement Learning with Verifiable Rewards (RLVR) has proven effective for enhancing Large Language Models (LLMs) on complex reasoning tasks. However, existing methods suffer from an exploration dilemma: the sharply peaked initial policies of pre-trained LLMs confine standard RL algorithms to a narrow set of solutions, boosting single-solution accuracy (pass@1) but suppressing solution diversity and multi-solution performance (pass@k). As a result, RLVR often distills existing capabilities rather than discovering new reasoning strategies. To overcome this, we introduce a Risk-Sensitive Reinforcement Learning framework. Our approach employs a risk-seeking objective that interpolates between mean and maximum rewards, leading to a novel algorithm, Risk-Sensitive GRPO (RS-GRPO), which drives deeper exploration by amplifying learning from challenging prompts. Remarkably, RS-GRPO is simple to implement, requiring only minor code modifications. On six mathematical reasoning benchmarks and with five different LLMs, RS-GRPO consistently improves pass@k performance while maintaining or enhancing pass@1 accuracy.

The pretrain-transfer paradigm, which underpins the success of large language models (LLMs), has demonstrated the immense power of creating foundation models that learn generalizable representations from vast datasets. However, extending this paradigm to Operations Research (OR) problems on graph structures remains challenging due to the fundamental conflict between the statistical flexibility of language and the strict combinatorial constraints of graphs. To bridge this gap, we introduce the Graph Foundation Model (GFM), the first framework capable of solving all distance-based optimization problems on graph structures. By introducing the LLM-like self-supervised pre-training paradigm on the paths generated from random walks in the graph, GFM is compelled to internalize the graph's complex topological and combinatorial rules, where the connectivity of the structure itself can be treated as the supervisory signal. Unlike existing neural methods that learn complex and task-specific solving policies, our approach leverages the pre-trained GFM as a foundational model of the graph's intrinsic structure, which in turn enables a simple generative heuristic to tackle a diverse range of optimization challenges effectively. Comprehensive experiments on networks ranging from 20 to 893 nodes demonstrate that GFM achieves competitive performance against specialized solvers across a variety of distinct optimization task classes, while maintaining significantly faster inference times. Our work establishes a new paradigm of adapting the pretrain-transfer framework to graph optimization, opening the door for applying foundation model innovations to OR.

The pipeline illustrates the full workflow from data collection in VR, through self-annotation and human baseline evaluation, to modeling and analysis of cognitive states. As virtual reality (VR) and augmented reality (AR) continue to gain popularity, head and hand motion data captured by consumer VR systems have become ubiquitous. Prior work shows such telemetry can be highly identifying and reflect broad user traits, often aligning with intuitive "folk theories" of body language. However, it remains unclear to what extent motion kinematics encode more nuanced cognitive states, such as confusion, hesitation, and readiness, which lack clear correlates with motion. To investigate this, we introduce a novel dataset of head and hand motion with frame-level annotations of these states collected during structured decision-making tasks. Our findings suggest that deep temporal models can infer subtle cognitive states from motion alone, achieving comparable performance with human observers. This work demonstrates that standard VR telemetry contains strong patterns related to users' internal cognitive processes, which opens the door for a new gener- To enhance reproducibility and support future work, we will make our dataset and modeling framework publicly available. Virtual Reality (VR) is rapidly evolving from a specialized tool for simulation and entertainment into a mainstream computing platform for work, education, and social interaction. As users spend more time in these immersive environments, the quality of human-computer interaction becomes paramount. The next generation of VR systems must move beyond explicit, command-based interfaces and develop the capacity for implicit, nuanced understanding. This requires an ability to perceive and adapt to a user's cognitive state in real-time, creating experiences that are more intuitive, supportive, and effective. The key to unlocking this capability lies in decoding the rich, continuous, and often subconscious stream of motion data generated by every user.

Test-time compute has emerged as a key paradigm for enhancing LLM reasoning, yet prevailing approaches like Best-of-N and majority voting apply uniform depth across inputs, wasting computation on simple queries while potentially under-thinking complex ones. We present FR-Ponder, a single-graph, backbone-training-free framework that allocates instance-adaptive reasoning compute via latent steering. A less than 1M-param controller observes hidden states and decides to halt or apply a small ponder step by adding a pre-computed steering vector to frozen representations. Our method extracts the latent steering vector associated with deeper reasoning outputs and direct IO from LLM and re-applies it through a tunable scaling factor, allowing the model to adapt its reasoning depth to the complexity of each input. To balance performance and computational cost, we employ Group Relative Policy Optimization (GRPO) as a reward signal to adaptively regulate reasoning depth, achieving task accuracy while mitigating overreasoning. Through curriculum learning and careful reward engineering, FR-Ponder learns calibrated compute allocation correlated with problem difficulty. On GSM8K and MA TH500, FR-Ponder improves the compute-accuracy frontier, delivering lower FLOPs with better matched accuracy and comparing favorably to early-exit baselines, without modifying backbone weights. Analyses visualize interpretable steering directions and show learned compute allocation correlates with problem difficulty.

This paper presents the opdainlp team's solution for the GPU track of the CPDC 2025 challenge. The challenge consists of three tasks, aiming to build an in-game conversational AI that adheres to character personas, aligns with the game's worldview, and supports function calling. Considering both effectiveness and resource/time constraints during inference, we synthesized data for some of the tasks based on the datasets provided by the competition organizers. We employed Qwen3-14B with LoRA fine-tuning and model fusion, and utilized a base model integrated with multiple LoRA adapters during inference. Specifically, in the competition, we used three distinct LoRA adapters to handle tool calling, response generation with tool call results, and response generation without tool call results, respectively. MultiLoRA inference was implemented using vLLM. Our solution achieved the first place in Task 1 and Task 3, and the second place in Task 2 of the GPU track.

Positive-unlabeled (PU) learning is a weakly supervised binary classification problem, in which the goal is to learn a binary classifier from only positive and unlabeled data, without access to negative data. In recent years, many PU learning algorithms have been developed to improve model performance. However, experimental settings are highly inconsistent, making it difficult to identify which algorithm performs better. In this paper, we propose the first PU learning benchmark to systematically compare PU learning algorithms. During our implementation, we identify subtle yet critical factors that affect the realistic and fair evaluation of PU learning algorithms. On the one hand, many PU learning algorithms rely on a validation set that includes negative data for model selection. This is unrealistic in traditional PU learning settings, where no negative data are available. To handle this problem, we systematically investigate model selection criteria for PU learning. On the other hand, the problem settings and solutions of PU learning have different families, i.e., the one-sample and two-sample settings. However, existing evaluation protocols are heavily biased towards the one-sample setting and neglect the significant difference between them. We identify the internal label shift problem of unlabeled training data for the one-sample setting and propose a simple yet effective calibration approach to ensure fair comparisons within and across families. We hope our framework will provide an accessible, realistic, and fair environment for evaluating PU learning algorithms in the future.

Prostate cancer (PCa) is the most frequently diagnosed malignancy in men and the eighth leading cause of cancer death worldwide. Multiparametric MRI (mpMRI) has become central to the diagnostic pathway for men at intermediate risk, improving de-tection of clinically significant PCa (csPCa) while reducing unnecessary biopsies and over-diagnosis. However, mpMRI remains limited by false positives, false negatives, and moderate to substantial interobserver agreement. Time-dependent diffusion (TDD) MRI, a novel sequence that enables tissue microstructure characterization, has shown encouraging preclinical performance in distinguishing clinically significant from insignificant PCa. Combining TDD-derived metrics with machine learning may provide robust, zone-specific risk prediction with less dependence on reader training and improved accuracy compared to current standard-of-care. This study protocol out-lines the rationale and describes the prospective evaluation of a home-developed AI-enhanced TDD-MRI software (PROSTDAI) in routine diagnostic care, assessing its added value against PI-RADS v2.1 and validating results against MRI-guided prostate biopsy.

Background Major depressive disorder (MDD) is a leading cause of global disability, yet current diagnostic approaches often rely on subjective assessments and lack the ability to integrate multimodal clinical information. Large language models (LLMs) hold promise for enhancing diagnostic accuracy through advanced reasoning but face challenges in interpretability, hallucination, and reliance on synthetic data. Methods We developed MDD-Thinker, an LLM-based diagnostic framework that integrates supervised fine-tuning (SFT) with reinforcement learning (RL) to strengthen reasoning ability and interpretability. Using the UK Biobank dataset, we generated 40,000 reasoning samples, supplemented with 10,000 samples from publicly available mental health datasets. The model was fine-tuned on these reasoning corpora, and its diagnostic and reasoning performance was evaluated against machine learning, deep learning, and state-of-the-art LLM baselines. Findings MDD-Thinker achieved an accuracy of 0.8268 and F1-score of 0.8081, significantly outperforming traditional baselines such as SVM and MLP, as well as general-purpose LLMs. Incorporating both SFT and RL yielded the greatest improvements, with relative gains of 29.0% in accuracy, 38.1% in F1-score, and 34.8% in AUC. Moreover, the model demonstrated comparable reasoning performance compared to much larger LLMs, while maintaining computational efficiency. Interpretation This study presents the first reasoning-enhanced LLM framework for MDD diagnosis trained on large-scale real-world clinical data. By integrating SFT and RL, MDD-Thinker balances accuracy, interpretability, and efficiency, offering a scalable approach for intelligent psychiatric diagnostics. These findings suggest that reasoning-oriented LLMs can provide clinically reliable support for MDD detection and may inform broader applications in mental health care.

The rapid growth of audio-centric platforms and applications such as WhatsApp and Twitter has transformed the way people communicate and share audio content in modern society. However, these platforms are increasingly misused to disseminate harmful audio content, such as hate speech, deceptive advertisements, and explicit material, which can have significant negative consequences (e.g., detrimental effects on mental health). In response, researchers and practitioners have been actively developing and deploying audio content moderation tools to tackle this issue. Despite these efforts, malicious actors can bypass moderation systems by making subtle alterations to audio content, such as modifying pitch or inserting noise. Moreover, the effectiveness of modern audio moderation tools against such adversarial inputs remains insufficiently studied. To address these challenges, we propose MTAM, a Metamorphic Testing framework for Audio content Moderation software. Specifically, we conduct a pilot study on 2000 audio clips and define 14 metamorphic relations across two perturbation categories: Audio Features-Based and Heuristic perturbations. MTAM applies these metamorphic relations to toxic audio content to generate test cases that remain harmful while being more likely to evade detection. In our evaluation, we employ MTAM to test five commercial textual content moderation software and an academic model against three kinds of toxic content. The results show that MTAM achieves up to 38.6%, 18.3%, 35.1%, 16.7%, and 51.1% error finding rates (EFR) when testing commercial moderation software provided by Gladia, Assembly AI, Baidu, Nextdata, and Tencent, respectively, and it obtains up to 45.7% EFR when testing the state-of-the-art algorithms from the academy.

Evaluating language models fairly is becoming harder as static benchmarks available on the internet risk contamination by training data. This makes it unclear whether models are truly reasoning or just recalling answers. In this paper, we introduce BeyondBench, an evaluation framework that avoids this problem by using algorithmic problem generation. Unlike traditional benchmarks that risk contamination from internet-scale training data, BeyondBench creates mathematically grounded problems on the fly, ensuring each test remains fresh and uncontaminated. Our framework covers 44 algorithmic tasks with a total of 117 variations, grouped into three difficulty levels: the Easy Suite (29 tasks) for basic arithmetic and statistics, the Medium Suite (5 tasks, 49 variations) for sequence patterns and reasoning, and the Hard Suite (10 tasks, 68 variations) tackling NP-complete and constraint satisfaction problems. Each task generates problems from a combinatorial space larger than 10^15 unique instances, with solutions verified deterministically by mathematical proofs. We evaluated 101 language models, including 85 open-source and 16 closed-source models, spanning sizes from 0.5B to 141B parameters and multiple quantization schemes. Our results show consistent reasoning deficiencies across model families, with performance degrading sharply as problem complexity increases from polynomial to exponential. In our Hard Suite evaluations, models such as Gemini-2.5-pro, Llama-3.3-70B, and Qwen2.5-72B achieved average accuracies of 56.38%, 26.91%, and 33.60%, respectively. Moreover, we observe that performance drops drastically without tool usage, with GPT-5, GPT-5-mini, and GPT-5-nano showing a decline of 16.81%, 28.05%, and 47.59% accuracy on the hard suite. Our leaderboard is publicly available at https://ctrl-gaurav.github.io/BeyondBench/