Unit testing is a core practice in programming, enabling systematic evaluation of programs produced by human developers or large language models (LLMs). Given the challenges in writing comprehensive unit tests, LLMs have been employed to automate test generation, yet methods for training LLMs to produce high-quality tests remain underexplored. In this work, we propose UTRL, a novel reinforcement learning framework that trains an LLM to generate high-quality unit tests given a programming instruction. Our key idea is to iteratively train two LLMs, the unit test generator and the code generator, in an adversarial manner via reinforcement learning. The unit test generator is trained to maximize a discrimination reward, which reflects its ability to produce tests that expose faults in the code generator's solutions, and the code generator is trained to maximize a code reward, which reflects its ability to produce solutions that pass the unit tests generated by the test generator. In our experiments, we demonstrate that unit tests generated by Qwen3-4B trained via UTRL show higher quality compared to unit tests generated by the same model trained via supervised fine-tuning on human-written ground-truth unit tests, yielding code evaluations that more closely align with those induced by the ground-truth tests. Moreover, Qwen3-4B trained with UTRL outperforms frontier models such as GPT-4.1 in generating high-quality unit tests, highlighting the effectiveness of UTRL in training LLMs for this task.

Reference-free evaluation metrics for grammatical error correction (GEC) have achieved high correlation with human judgments. However, these metrics are not designed to evaluate adversarial systems that aim to obtain unjustifiably high scores. The existence of such systems undermines the reliability of automatic evaluation, as it can mislead users in selecting appropriate GEC systems. In this study, we propose adversarial attack strategies for four reference-free metrics: SOME, Scribendi, IMPARA, and LLM-based metrics, and demonstrate that our adversarial systems outperform the current state-of-the-art. These findings highlight the need for more robust evaluation methods.

Reinforcement learning with verifiable rewards (RLVR) has proven effective in eliciting complex reasoning in large language models (LLMs). However, standard RLVR training often leads to excessively verbose processes (in reasoning tasks) and inefficient exploration trajectories (in agentic settings), as outcome-only rewards provide no incentive for efficiency and the high variance in response length within relatively small rollout groups results in noisy optimization signals. To address this, we propose Rollout Response Recomposition (RoRecomp), a plug-and-play method that guides models toward concise reasoning by strategically recomposing the training data. RoRecomp separates responses into two distinct batch types: 1) priority batches, which combine short-correct and long-incorrect responses selected from online batches to provide a clear gradient signal for brevity, and 2) compensation batches, which utilize remaining responses from a replay buffer to maintain stability and prevent model collapse. To comprehensively evaluate effectiveness, we test RoRecomp across three settings where results demonstrate substantial efficiency gains: reducing reasoning length by 27.7% in zero RL training, reducing unnecessary tool calls by 46.8% while improving accuracy in agentic RL, and achieving up to 52.5% length reduction in thinking compression, all with minimal performance impact.

Test-time scaling methods improve the capabilities of large language models (LLMs) by increasing the amount of compute used during inference to make a prediction. Inference-time compute can be scaled in parallel by choosing among multiple independent solutions or sequentially through self-refinement. We propose Recursive Self-Aggregation (RSA), a test-time scaling method inspired by evolutionary methods that combines the benefits of both parallel and sequential scaling. Each step of RSA refines a population of candidate reasoning chains through aggregation of subsets to yield a population of improved solutions, which are then used as the candidate pool for the next iteration. RSA exploits the rich information embedded in the reasoning chains -- not just the final answers -- and enables bootstrapping from partially correct intermediate steps within different chains of thought. Empirically, RSA delivers substantial performance gains with increasing compute budgets across diverse tasks, model families and sizes. Notably, RSA enables Qwen3-4B-Instruct-2507 to achieve competitive performance with larger reasoning models, including DeepSeek-R1 and o3-mini (high), while outperforming purely parallel and sequential scaling strategies across AIME-25, HMMT-25, Reasoning Gym, LiveCodeBench-v6, and SuperGPQA. We further demonstrate that training the model to combine solutions via a novel aggregation-aware reinforcement learning approach yields significant performance gains. Code available at https://github.com/HyperPotatoNeo/RSA.

Vision-language-action models (VLAs) have shown generalization capabilities in robotic manipulation tasks by inheriting from vision-language models (VLMs) and learning action generation. Most VLA models focus on interpreting vision and language to generate actions, whereas robots must perceive and interact within the spatial-physical world. This gap highlights the need for a comprehensive understanding of robotic-specific multisensory information, which is crucial for achieving complex and contact-rich control. To this end, we introduce a multisensory language-action (MLA) model that collaboratively perceives heterogeneous sensory modalities and predicts future multisensory objectives to facilitate physical world modeling. Specifically, to enhance perceptual representations, we propose an encoder-free multimodal alignment scheme that innovatively repurposes the large language model itself as a perception module, directly interpreting multimodal cues by aligning 2D images, 3D point clouds, and tactile tokens through positional correspondence. To further enhance MLA's understanding of physical dynamics, we design a future multisensory generation post-training strategy that enables MLA to reason about semantic, geometric, and interaction information, providing more robust conditions for action generation. For evaluation, the MLA model outperforms the previous state-of-the-art 2D and 3D VLA methods by 12% and 24% in complex, contact-rich real-world tasks, respectively, while also demonstrating improved generalization to unseen configurations. Project website: https://sites.google.com/view/open-mla

This paper introduces \textit{measurement trees}, a novel class of metrics designed to combine various constructs into an interpretable multi-level representation of a measurand. Unlike conventional metrics that yield single values, vectors, surfaces, or categories, measurement trees produce a hierarchical directed graph in which each node summarizes its children through user-defined aggregation methods. In response to recent calls to expand the scope of AI system evaluation, measurement trees enhance metric transparency and facilitate the integration of heterogeneous evidence, including, e.g., agentic, business, energy-efficiency, sociotechnical, or security signals. We present definitions and examples, demonstrate practical utility through a large-scale measurement exercise, and provide accompanying open-source Python code. By operationalizing a transparent approach to measurement of complex constructs, this work offers a principled foundation for broader and more interpretable AI evaluation.

Reinforcement Learning (RL) has shown remarkable success in enhancing the reasoning capabilities of Large Language Models (LLMs). Process-Supervised RL (PSRL) has emerged as a more effective paradigm compared to outcome-based RL. However, existing PSRL approaches suffer from limited exploration efficiency, both in terms of branching positions and sampling. In this paper, we introduce a novel PSRL framework (AttnRL), which enables efficient exploration for reasoning models. Motivated by preliminary observations that steps exhibiting high attention scores correlate with reasoning behaviors, we propose to branch from positions with high values. Furthermore, we develop an adaptive sampling strategy that accounts for problem difficulty and historical batch size, ensuring that the whole training batch maintains non-zero advantage values. To further improve sampling efficiency, we design a one-step off-policy training pipeline for PSRL. Extensive experiments on multiple challenging mathematical reasoning benchmarks demonstrate that our method consistently outperforms prior approaches in terms of performance and sampling and training efficiency.

While previous AI Scientist systems can generate novel findings, they often lack the focus to produce scientifically valuable contributions that address pressing human-defined challenges. We introduce DeepScientist, a system designed to overcome this by conducting goal-oriented, fully autonomous scientific discovery over month-long timelines. It formalizes discovery as a Bayesian Optimization problem, operationalized through a hierarchical evaluation process consisting of "hypothesize, verify, and analyze". Leveraging a cumulative Findings Memory, this loop intelligently balances the exploration of novel hypotheses with exploitation, selectively promoting the most promising findings to higher-fidelity levels of validation. Consuming over 20,000 GPU hours, the system generated about 5,000 unique scientific ideas and experimentally validated approximately 1100 of them, ultimately surpassing human-designed state-of-the-art (SOT A) methods on three frontier AI tasks by 183.7%, 1.9%, and 7.9%. This work provides the first large-scale evidence of an AI achieving discoveries that progressively surpass human SOT A on scientific tasks, producing valuable findings that genuinely push the frontier of scientific discovery.Figure 1: Comparison of research progress timelines for AI text detection on the RAID (Dugan et al., 2024). The right panel shows that DeepScientist achieves progress in two weeks that is comparable to three years of human research (Su et al.; Bao et al., a;b; Hu et al., 2023) (left panel). All zero-shot methods, including the system-generated T -Detect, TDT, and P A-Detect, uniformly adopt Falcon-7B (Almazrouei et al., 2023) as the base model. Additionally, all methods produced by DeepScientist demonstrate higher throughput than the previous SOT A method, Binoculars (Hans et al., 2024). 1 Scientific discovery is inherently a process of continuous exploration and trial-and-error, where vast amounts of time and effort are invested to push the boundaries of human knowledge forward by a small step. This principle of persistent, incremental advancement is visible across the history of technology. For example, the decades-long optimization of semiconductor manufacturing has seen the feature size of transistors systematically reduced from micrometers to single-digit nanometers (Moore, 1965). Similarly, the efficiency of photovoltaic cells has been continuously advanced over half a century, with myriad material and architectural innovations pushing conversion rates from nascent single-digit percentages ever closer to their theoretical limits (Green, 1993). These historical trajectories underscore a process where human scientists engage in decades of goal-directed, iterative work to advance the SoT A artifacts continuously. Recently, the emergence of Large Language Models (LLMs) has propelled automated scientific discovery, where LLM-based AI Scientist systems take the lead in exploration (Xie et al., 2025b).

As large language models (LLMs) begin to saturate existing benchmarks, automated benchmark creation using LLMs (LLM-as-a-benchmark) has emerged as a scalable alternative to slow and costly human curation. While these generated test sets have to potential to cheaply rank models, we demonstrate a critical flaw. LLM-generated benchmarks systematically favor the model that created the benchmark: they exhibit self-bias on low resource languages to English translation tasks. We show three key findings on automatic benchmarking of LLMs for translation: First, this bias originates from two sources: the generated test data (LLM-as-a-testset) and the evaluation method (LLM-as-an-evaluator), with their combination amplifying the effect. Second, self-bias in LLM-as-a-benchmark is heavily influenced by the model's generation capabilities in the source language. For instance, we observe more pronounced bias in into-English translation, where the model's generation system is developed, than in out-of-English translation tasks. Third, we observe that low diversity in source text is one attribution to self-bias. Our results suggest that improving the diversity of these generated source texts can mitigate some of the observed self-bias. The rapid advancements in Large Language Models (LLMs) have led to an unprecedented saturation of existing, meticulously human-curated benchmarks. This phenomenon exposes two critical, intertwined challenges: traditional benchmark creation is too laborious and expensive to keep pace with rapid model development, and this challenge is compounded by the inherent difficulty of constructing high-quality benchmarks for low-resource languages, even with human labor, which further strains existing benchmark resources.

Model fingerprinting has emerged as a promising paradigm for claiming model ownership. However, robustness evaluations of these schemes have mostly focused on benign perturbations such as incremental fine-tuning, model merging, and prompting. Lack of systematic investigations into {\em adversarial robustness} against a malicious model host leaves current systems vulnerable. To bridge this gap, we first define a concrete, practical threat model against model fingerprinting. We then take a critical look at existing model fingerprinting schemes to identify their fundamental vulnerabilities. Based on these, we develop adaptive adversarial attacks tailored for each vulnerability, and demonstrate that these can bypass model authentication completely for ten recently proposed fingerprinting schemes while maintaining high utility of the model for the end users. Our work encourages fingerprint designers to adopt adversarial robustness by design. We end with recommendations for future fingerprinting methods.