Dr. Kernel: Reinforcement Learning Done Right for Triton Kernel Generations

High-quality kernel is critical for scalable AI systems, and enabling LLMs to generate such code would advance AI development. However, training LLMs for this task requires sufficient data, a robust environment, and the process is often vulnerable to reward hacking and lazy optimization. In these cases, models may hack training rewards and prioritize trivial correctness over meaningful speedup. In this paper, we systematically study reinforcement learning (RL) for kernel generation. We first design KernelGYM, a robust distributed GPU environment that supports reward hacking check, data collection from multi-turn interactions and long-term RL training. Building on KernelGYM, we investigate effective multi-turn RL methods and identify a biased policy gradient issue caused by self-inclusion in GRPO. To solve this, we propose Turn-level Reinforce-Leave-One-Out (TRLOO) to provide unbiased advantage estimation for multi-turn RL. To alleviate lazy optimization, we incorporate mismatch correction for training stability and introduce Profiling-based Rewards (PR) and Profiling-based Rejection Sampling (PRS) to overcome the issue. The trained model, Dr.Kernel-14B, reaches performance competitive with Claude-4.5-Sonnet in Kernelbench. Finally, we study sequential test-time scaling for Dr.Kernel-14B. On the KernelBench Level-2 subset, 31.6% of the generated kernels achieve at least a 1.2x speedup over the Torch reference, surpassing Claude-4.5-Sonnet (26.7%) and GPT-5 (28.6%). When selecting the best candidate across all turns, this 1.2x speedup rate further increases to 47.8%. All resources, including environment, training code, models, and dataset, are included in https://www.github.com/hkust-nlp/KernelGYM.

Towards a Science of Collective AI: LLM-based Multi-Agent Systems Need a Transition from Blind Trial-and-Error to Rigorous Science

Recent advancements in Large Language Models (LLMs) have greatly extended the capabilities of Multi-Agent Systems (MAS), demonstrating significant effectiveness across a wide range of complex and open-ended domains. However, despite this rapid progress, the field still relies heavily on empirical trial-and-error. It lacks a unified and principled scientific framework necessary for systematic optimization and improvement. This bottleneck stems from the ambiguity of attribution: first, the absence of a structured taxonomy of factors leaves researchers restricted to unguided adjustments; second, the lack of a unified metric fails to distinguish genuine collaboration gain from mere resource accumulation. In this paper, we advocate for a transition to design science through an integrated framework. We advocate to establish the collaboration gain metric ($Γ$) as the scientific standard to isolate intrinsic gains from increased budgets. Leveraging $Γ$, we propose a factor attribution paradigm to systematically identify collaboration-driving factors. To support this, we construct a systematic MAS factor library, structuring the design space into control-level presets and information-level dynamics. Ultimately, this framework facilitates the transition from blind experimentation to rigorous science, paving the way towards a true science of Collective AI.

Entropy-Gated Selective Policy Optimization:Token-Level Gradient Allocation for Hybrid Training of Large Language Models

Hybrid training methods for large language models combine supervised fine tuning (SFT) on expert demonstrations with reinforcement learning (RL) on model rollouts, typically at the sample level. We propose Entropy Gated Selective Policy Optimization (EGSPO), a three stage framework that extends sample level mixing with token level gradient modulation. Stage 1, SFT expert learning, establishes a reliable warm up policy using expert demonstrations with a pure SFT loss. Stage 2, RL rollout generation, samples trajectories from the current policy and computes per token predictive entropy. Stage 3, the EGSPO mechanism, applies entropy gated gradient allocation: a predictive entropy module routes high entropy tokens to full PPO updates to encourage exploration, and low entropy tokens to attenuated PPO updates to reduce variance and preserve knowledge. Critically, both branches incorporate the advantage function A_t, ensuring that incorrect trajectories receive consistent negative learning signals and preventing reinforcement of confident errors. EGSPO achieves consistent improvements on mathematical reasoning benchmarks, with gains of 3.8 percent on AIME and 2.9 percent on MATH over the CHORD phi baseline, while incurring only 3.4 percent additional computational overhead.

Seeing Is Believing? A Benchmark for Multimodal Large Language Models on Visual Illusions and Anomalies

Multimodal Large Language Models (MLLMs) have shown remarkable proficiency on general-purpose vision-language benchmarks, reaching or even exceeding human-level performance. However, these evaluations typically rely on standard in-distribution data, leaving the robustness of MLLMs largely unexamined when faced with scenarios that defy common-sense priors. To address this gap, we introduce VIA-Bench, a challenging benchmark designed to probe model performance on visual illusions and anomalies. It includes six core categories: color illusions, motion illusions, gestalt illusions, geometric and spatial illusions, general visual illusions, and visual anomalies. Through careful human-in-the-loop review, we construct over 1K high-quality question-answer pairs that require nuanced visual reasoning. Extensive evaluation of over 20 state-of-the-art MLLMs, including proprietary, open-source, and reasoning-enhanced models, uncovers significant vulnerabilities. Notably, we find that Chain-of-Thought (CoT) reasoning offers negligible robustness, often yielding ``brittle mirages'' where the model's logic collapses under illusory stimuli. Our findings reveal a fundamental divergence between machine and human perception, suggesting that resolving such perceptual bottlenecks is critical for the advancement of artificial general intelligence. The benchmark data and code will be released.

Hunt Instead of Wait: Evaluating Deep Data Research on Large Language Models

The agency expected of Agentic Large Language Models goes beyond answering correctly, requiring autonomy to set goals and decide what to explore. We term this investigatory intelligence, distinguishing it from executional intelligence, which merely completes assigned tasks. Data Science provides a natural testbed, as real-world analysis starts from raw data rather than explicit queries, yet few benchmarks focus on it. To address this, we introduce Deep Data Research (DDR), an open-ended task where LLMs autonomously extract key insights from databases, and DDR-Bench, a large-scale, checklist-based benchmark that enables verifiable evaluation. Results show that while frontier models display emerging agency, long-horizon exploration remains challenging. Our analysis highlights that effective investigatory intelligence depends not only on agent scaffolding or merely scaling, but also on intrinsic strategies of agentic models.

FutureMind: Equipping Small Language Models with Strategic Thinking-Pattern Priors via Adaptive Knowledge Distillation

Small Language Models (SLMs) are attractive for cost-sensitive and resource-limited settings due to their efficient, low-latency inference. However, they often struggle with complex, knowledge-intensive tasks that require structured reasoning and effective retrieval. To address these limitations, we propose FutureMind, a modular reasoning framework that equips SLMs with strategic thinking-pattern priors via adaptive knowledge distillation from large language models (LLMs). FutureMind introduces a dynamic reasoning pipeline composed of four key modules: Problem Analysis, Logical Reasoning, Strategy Planning, and Retrieval Guidance. This pipeline is augmented by three distinct retrieval paradigms that decompose complex queries into tractable subproblems, ensuring efficient and accurate retrieval execution. Extensive experiments on multi-hop QA benchmarks, including 2WikiMultihopQA, MuSiQue, Bamboogle, and Frames, demonstrate the superiority of FutureMind. It consistently outperforms strong baselines such as Search-o1, achieving state-of-the-art results under free training conditions across diverse SLM architectures and scales. Beyond empirical gains, our analysis reveals that the process of thinking-pattern distillation is restricted by the cognitive bias bottleneck between the teacher (LLMs) and student (SLMs) models. This provides new perspectives on the transferability of reasoning skills, paving the way for the development of SLMs that combine efficiency with genuine cognitive capability.

SP^2DPO: An LLM-assisted Semantic Per-Pair DPO Generalization

Direct Preference Optimization (DPO) controls the trade-off between fitting preference labels and staying close to a reference model using a single global temperature beta, implicitly treating all preference pairs as equally informative. Real-world preference corpora are heterogeneous: they mix high-signal, objective failures (for example, safety, factuality, instruction violations) with low-signal or subjective distinctions (for example, style), and also include label noise. We introduce our method, SP2DPO (Semantic Per-Pair DPO), a generalization that replaces the global temperature with an instance-specific schedule beta_i pre-decided offline from structured semantic-gap annotations (category, magnitude, confidence) produced by teacher language models. We instantiate this procedure on the UltraFeedback preference corpus (59,960 pairs), enabling large-scale construction of an auditable beta_i artifact, and incur zero training-time overhead: the inner-loop optimizer remains standard DPO with beta set per pair. We focus our empirical study on AlpacaEval 2.0, reporting both raw win rate and length-controlled win rate. Across four open-weight, instruction-tuned student backbones (4B-8B), SP2DPO is competitive with a tuned global-beta DPO baseline and improves AlpacaEval 2.0 length-controlled win rate on two of four backbones, while avoiding per-model beta sweeps. All code, annotations, and artifacts will be released.

MobileBench-OL: A Comprehensive Chinese Benchmark for Evaluating Mobile GUI Agents in Real-World Environment

Recent advances in mobile Graphical User Interface (GUI) agents highlight the growing need for comprehensive evaluation benchmarks. While new online benchmarks offer more realistic testing than offline ones, they tend to focus on the agents' task instruction-following ability while neglecting their reasoning and exploration ability. Moreover, these benchmarks do not consider the random noise in real-world mobile environments. This leads to a gap between benchmarks and real-world environments. To addressing these limitations, we propose MobileBench-OL, an online benchmark with 1080 tasks from 80 Chinese apps. It measures task execution, complex reasoning, and noise robustness of agents by including 5 subsets, which set multiple evaluation dimensions. We also provide an auto-eval framework with a reset mechanism, enabling stable and repeatable real-world benchmarking. Evaluating 12 leading GUI agents on MobileBench-OL shows significant room for improvement to meet real-world requirements. Human evaluation further confirms that MobileBench-OL can reliably measure the performance of leading GUI agents in real environments. Our data and code will be released upon acceptance.

Youtu-VL: Unleashing Visual Potential via Unified Vision-Language Supervision

Despite the significant advancements represented by Vision-Language Models (VLMs), current architectures often exhibit limitations in retaining fine-grained visual information, leading to coarse-grained multimodal comprehension. We attribute this deficiency to a suboptimal training paradigm inherent in prevailing VLMs, which exhibits a text-dominant optimization bias by conceptualizing visual signals merely as passive conditional inputs rather than supervisory targets. To mitigate this, we introduce Youtu-VL, a framework leveraging the Vision-Language Unified Autoregressive Supervision (VLUAS) paradigm, which fundamentally shifts the optimization objective from ``vision-as-input'' to ``vision-as-target.'' By integrating visual tokens directly into the prediction stream, Youtu-VL applies unified autoregressive supervision to both visual details and linguistic content. Furthermore, we extend this paradigm to encompass vision-centric tasks, enabling a standard VLM to perform vision-centric tasks without task-specific additions. Extensive empirical evaluations demonstrate that Youtu-VL achieves competitive performance on both general multimodal tasks and vision-centric tasks, establishing a robust foundation for the development of comprehensive generalist visual agents.

R^3: Replay, Reflection, and Ranking Rewards for LLM Reinforcement Learning

Large reasoning models (LRMs) aim to solve diverse and complex problems through structured reasoning. Recent advances in group-based policy optimization methods have shown promise in enabling stable advantage estimation without reliance on process-level annotations. However, these methods rely on advantage gaps induced by high-quality samples within the same batch, which makes the training process fragile and inefficient when intra-group advantages collapse under challenging tasks. To address these problems, we propose a reinforcement learning mechanism named \emph{\textbf{R^3}} that along three directions: (1) a \emph{cross-context \underline{\textbf{R}}eplay} strategy that maintains the intra-group advantage by recalling valuable examples from historical trajectories of the same query, (2) an \emph{in-context self-\underline{\textbf{R}}eflection} mechanism enabling models to refine outputs by leveraging past failures, and (3) a \emph{structural entropy \underline{\textbf{R}}anking reward}, which assigns relative rewards to truncated or failed samples by ranking responses based on token-level entropy patterns, capturing both local exploration and global stability. We implement our method on Deepseek-R1-Distill-Qwen-1.5B and train it on the DeepscaleR-40k in the math domain. Experiments demonstrate our method achieves SoTA performance on several math benchmarks, representing significant improvements and fewer reasoning tokens over the base models. Code and model will be released.