Reasoning-focused LLMs sometimes alter their behavior when they detect that they are being evaluated--which can lead them to optimize for test-passing performance or to comply more readily with harmful prompts if real-world consequences appear absent. We present the first quantitative study of how such "test awareness" impacts model behavior, particularly its performance on safety-related tasks1. We introduce a white-box probing framework that (i) linearly identifies awareness-related activations and (ii) steers models toward or away from test awareness while monitoring downstream performance. We apply our method to different state-of-the-art openweight reasoning LLMs across both realistic and hypothetical tasks (denoting tests or simulations). Our results demonstrate that test awareness significantly impacts safety alignment (such as compliance with harmful requests and conforming to stereotypes) with effects varying in both magnitude and direction across models. By providing control over this latent effect, our work aims to provide a stress-test mechanism and increase trust in how we perform safety evaluations.

Large language models (LLMs) have shown promise in transforming machine learning research, yet their capability to faithfully implement novel ideas from recent research papers--ideas unseen during pretraining--remains unclear. We introduce ResearchCodeBench, a benchmark of 212 coding challenges that evaluates LLMs' ability to translate cutting-edge ML contributions from top 2024-2025 research papers into executable code. We assessed 32 proprietary and open-source LLMs on both the full benchmark and ResearchCodeBench-HARD, a subset of 109 particularly challenging tasks.

We introduce CLEVER1, a high-quality, curated benchmark of 161 problems for end-to-end verified code generation in Lean. Each problem consists of (1) the task of generating a specification that matches a held-out ground-truth specification, and (2) the task of generating a Lean implementation that provably satisfies this specification. Unlike prior benchmarks, CLEVER avoids test-case supervision, LLM-generated annotations, and specifications that leak implementation logic or allow vacuous solutions. All outputs are verified post-hoc using Lean's type checker to ensure machine-checkable correctness. We use CLEVER to evaluate several few-shot and agentic approaches based on state-of-the-art language models. These methods all struggle to achieve full verification, establishing it as a challenging frontier benchmark for program synthesis and formal reasoning. Our benchmark can be found on GitHub as well as HuggingFace. All our evaluation code is also available online.

We introduce PurpCode, the first post-training recipe for training safe code reasoning models towards generating secure code and defending against malicious cyberactivities. PurpCode trains a reasoning model in two stages: (i) Rule Learning, which explicitly teaches the model to reference cybersafety rules to generate vulnerabilityfree code and to avoid facilitating malicious cyberactivities; and (ii) Reinforcement Learning, which optimizes model safety and preserves model utility through diverse, multi-objective reward mechanisms. To empower the training pipelines with comprehensive cybersafety data, we conduct internal red-teaming to synthesize comprehensive and high-coverage prompts based on real-world tasks for inducing unsafe cyberactivities in the model. Based on PurpCode, we develop a reasoning-based coding model, namely PurpCode-32B, which demonstrates state-of-the-art cybersafety, outperforming various frontier models. Moreover, our alignment method decreases the model overrefusal rates in both general and cybersafety-specific scenarios, while preserving model utility in both code generation and common security knowledge.

Unsafe operations described in natural text lead to a lower rejection rate than thoseincodeformat.

A.2 ExpertRollouts The expert rollouts consist of acollection of HDF5 files, one file per clip. A.3 HostingPlan The link to the dataset can be found on the project website. The dataset website also includes the policies we trained in Section 5, i.e., the multi-clip tracking policies, RL-trained taskpolicies, andtheGPTpolicy. Training clip experts to track long clips is potentially slow and laborious, so wefollowMerel etal.[2019]bydividing Each expert is a neural network with three hidden layers, 1024 neurons in each hidden layer, and thetanh activation.