In training every agent we use a distributed framework for simulation and training. For simulation, we run 6400 Hanabi environments in parallel and the trajectories are batched together for efficient GPU computation. This is done efficiently as every thread can hold many environments in which many agents interact. Every agent chooses actions based on neural network calls, which are more intensive and done by GPUs. By doing these calls asynchronously it allows a thread to support multiple environments while waiting for prior agents' actions to be computed.

Output: deterministic upper-bound for the frequency of z in the data set: ˆfCMSup (z). The CMS-CU algorithm Algorithm A2 CMS-CU Input: Data set Z1,...,Zm. Output: deterministic upper-bound for the frequency of z in the data set: ˆfCMS CUup (z). Input: A (trainable) rule for computing nested intervals [ˆLm,α(; t), ˆUm,α(; t)], t T. Input: Number of data points mtrain0

Two 2025 publications, "AI 2027" (Kokotajlo et al., 2025) and "If Anyone Builds It, Everyone Dies" (Yudkowsky & Soares, 2025), assert that superintelligent artificial intelligence will almost certainly destroy or render humanity obsolete within the next decade. Both rest on the classic chain formulated by Good (1965) and Bostrom (2014): intelligence explosion, superintelligence, lethal misalignment. This article subjects each link to the empirical record of 2023-2025. Sixty years after Good's speculation, none of the required phenomena (sustained recursive self-improvement, autonomous strategic awareness, or intractable lethal misalignment) have been observed. Current generative models remain narrow, statistically trained artefacts: powerful, opaque, and imperfect, but devoid of the properties that would make the catastrophic scenarios plausible. Following Whittaker (2025a, 2025b, 2025c) and Zuboff (2019, 2025), we argue that the existential-risk thesis functions primarily as an ideological distraction from the ongoing consolidation of surveillance capitalism and extreme concentration of computational power. The thesis is further inflated by the 2025 AI speculative bubble, where trillions in investments in rapidly depreciating "digital lettuce" hardware (McWilliams, 2025) mask lagging revenues and jobless growth rather than heralding superintelligence. The thesis remains, in November 2025, a speculative hypothesis amplified by a speculative financial bubble rather than a demonstrated probability.

While Multimodal Large Language Models (MLLMs) excel on benchmarks, their processing paradigm differs from the human ability to integrate visual information. Unlike humans who naturally bridge details and high-level concepts, models tend to treat these elements in isolation. Prevailing evaluation protocols often decouple low-level perception from high-level reasoning, overlooking their semantic and causal dependencies, which yields non-diagnostic results and obscures performance bottlenecks. We present VCU-Bridge, a framework that operationalizes a human-like hierarchy of visual connotation understanding: multi-level reasoning that advances from foundational perception through semantic bridging to abstract connotation, with an explicit evidence-to-inference trace from concrete cues to abstract conclusions. Building on this framework, we construct HVCU-Bench, a benchmark for hierarchical visual connotation understanding with explicit, level-wise diagnostics. Comprehensive experiments demonstrate a consistent decline in performance as reasoning progresses to higher levels. We further develop a data generation pipeline for instruction tuning guided by Monte Carlo Tree Search (MCTS) and show that strengthening low-level capabilities yields measurable gains at higher levels. Interestingly, it not only improves on HVCU-Bench but also brings benefits on general benchmarks (average +2.53%), especially with substantial gains on MMStar (+7.26%), demonstrating the significance of the hierarchical thinking pattern and its effectiveness in enhancing MLLM capabilities. The project page is at https://vcu-bridge.github.io .

Large Language Models (LLMs) are increasingly tasked with creative generation, including the simulation of fictional characters. However, their ability to portray non-prosocial, antagonistic personas remains largely unexamined. We hypothesize that the safety alignment of modern LLMs creates a fundamental conflict with the task of authentically role-playing morally ambiguous or villainous characters. To investigate this, we introduce the Moral RolePlay benchmark, a new dataset featuring a four-level moral alignment scale and a balanced test set for rigorous evaluation. We task state-of-the-art LLMs with role-playing characters from moral paragons to pure villains. Our large-scale evaluation reveals a consistent, monotonic decline in role-playing fidelity as character morality decreases. We find that models struggle most with traits directly antithetical to safety principles, such as ``Deceitful'' and ``Manipulative'', often substituting nuanced malevolence with superficial aggression. Furthermore, we demonstrate that general chatbot proficiency is a poor predictor of villain role-playing ability, with highly safety-aligned models performing particularly poorly. Our work provides the first systematic evidence of this critical limitation, highlighting a key tension between model safety and creative fidelity. Our benchmark and findings pave the way for developing more nuanced, context-aware alignment methods.

Developing efficient CUDA kernels is increasingly critical for AI applications such as large-scale LLM training. However, manual kernel design is both costly and time-consuming, motivating automatic approaches that leverage LLMs for code generation. Existing methods for automatic kernel generation, however, often produce low-efficiency kernels, incur high computational overhead, and fail to generalize across settings. In this work, we propose CudaForge, a training-free multi-agent workflow for CUDA kernel generation and optimization. Our workflow is inspired by the iterative workflow of human experts, which contains steps such as developing initial kernels, testing correctness, analyzing hardware feedback, and iterative improvement. More specifically, CudaForge employs two LLM agents: a Coder and a Judge, that iteratively generate, correct, and optimize CUDA kernels, while integrating hardware feedback such as Nsight Compute (NCU) metrics. In extensive evaluations, we show that CudaForge, by leveraging base models like OpenAI-o3, achieves 97.6\% correctness of generated kernels and an average 1.68$\times$ speedup over PyTorch baselines, substantially surpassing state-of-the-art models including OpenAI-o3 and Kevin on KernelBench.Beyond accuracy and speed, CudaForge demonstrates strong generalization across GPUs (A100, RTX 6000, 4090, 3090) and base models (OpenAI-o3, GPT-5, gpt-oss-120B, Claude-Sonnet-4, QwQ-32B), while maintaining high efficiency. In particular, generating an optimized kernel takes about 26.5 minutes on one RTX6000 and incurs about \$ 0.3 API cost, which is significantly cheaper than existing agentic work that costs 6 H100 hours and \$ 5 API cost per kernel. Our results highlight that multi-agent, training-free workflows can enable cost-effective, generalizable, and high-performance CUDA kernel optimization. Code available at https://github.com/OptimAI-Lab/CudaForge

With the rapid evolution of large language models (LLMs), the demand for automated, high-performance system kernels has emerged as a key enabler for accelerating development and deployment. We introduce TritonRL, a domain-specialized LLM for Triton kernel generation, trained with a novel training framework that enables robust and automated kernel synthesis. Unlike general-purpose programming languages, Triton kernel generation faces unique challenges due to data scarcity and incomplete evaluation criteria, vulnerable to reward hacking. Our approach addresses these challenges end-to-end by distilling Triton-specific knowledge through supervised fine-tuning on curated datasets, and further improving code quality via reinforcement learning (RL) with robust, verifiable rewards and hierarchical reward assignment. Our RL framework robustly detects reward hacking and guides both reasoning traces and code tokens through fine-grained verification and hierarchical reward decomposition, enabling the model to generate high-quality Triton kernels that can truly replace existing modules. With robust and fine-grained evaluation, our experiments on KernelBench demonstrate that TritonRL achieves state-of-the-art correctness and speedup, surpassing all other Triton-specific models and underscoring the effectiveness of our RL-based training paradigm.

Reinforcement Learning (RL) has emerged as a powerful paradigm for post-training Large Language Models (LLMs), significantly enhancing their capabilities on complex, multi-step reasoning tasks (Ouyang et al., 2022). Methods based on Reinforcement Learning from Human Feedback (RLHF) and Direct Preference Optimization (DPO) (Rafailov et al., 2023) have become standard practice for aligning LLMs. These paradigms are often powered by foundational algorithms like Proximal Policy Optimization (PPO) (Schulman et al., 2017), with state-of-the-art variants such as Group Relative Policy Optimization (GRPO) (Shao et al., 2024) pushing models to achieve remarkable performance on benchmarks like GSM8K (Cobbe et al., 2021) and MA TH (Hendrycks et al., 2021). These successes, often marked by state-of-the-art results (Lewkowycz et al., 2022; Lightman et al., 2023), are widely interpreted as a significant leap forward, suggesting that RL-based alignment is a key pathway toward developing more general and robust machine reasoning systems. Despite impressive reported gains, a key question is whether current benchmarks still meaningfully assess generalization. Our findings suggest that the traditional assumption underlying benchmark design, that a model's ability to perform well on unseen test examples is sufficient to measure generalization, no longer holds for RL. We find that RL-based reasoning models trained on the training split achieve nearly the same performance as those trained directly on the test split, indicating that "unseen-ness" alone is no longer the challenging or discriminative criterion. This calls for the rethinking of evaluation: rather than relying solely on disjoint train/test splits, future benchmarks must incorporate settings that remain sensitive to deeper forms of generalization and can reveal weaknesses that simple data separation fails to expose. To systematically investigate this phenomenon, we introduce a multi-faceted empirical framework designed not merely to measure performance, but to deconstruct it.

EHRNoteQA in two formats: open-ended and multi-choice question answering, and propose a reliable evaluation method for each. We evaluate 27 LLMs using EHRNoteQA and examine various factors affecting the model performance ( e.g., the length and number of discharge summaries).