RealSafe-R1: Safety-Aligned DeepSeek-R1 without Compromising Reasoning Capability

Large Reasoning Models (LRMs), such as OpenAI o1 and DeepSeek-R1, have been rapidly progressing and achieving breakthrough performance on complex reasoning tasks such as mathematics and coding. However, the open-source R1 models have raised safety concerns in wide applications, such as the tendency to comply with malicious queries, which greatly impacts the utility of these powerful models in their applications. In this paper, we introduce RealSafe-R1 as safety-aligned versions of DeepSeek-R1 distilled models. To train these models, we construct a dataset of 15k safety-aware reasoning trajectories generated by DeepSeek-R1, under explicit instructions for expected refusal behavior. Both quantitative experiments and qualitative case studies demonstrate the models' improvements, which are shown in their safety guardrails against both harmful queries and jailbreak attacks. Importantly, unlike prior safety alignment efforts that often compromise reasoning performance, our method preserves the models' reasoning capabilities by maintaining the training data within the original distribution of generation. Model weights of RealSafe-R1 are open-source at https://huggingface.co/RealSafe.

The Emperor's New Clothes in Benchmarking? A Rigorous Examination of Mitigation Strategies for LLM Benchmark Data Contamination

Benchmark Data Contamination (BDC)-the inclusion of benchmark testing samples in the training set-has raised increasing concerns in Large Language Model (LLM) evaluation, leading to falsely inflated performance estimates and undermining evaluation reliability. To address this, researchers have proposed various mitigation strategies to update existing benchmarks, including modifying original questions or generating new ones based on them. However, a rigorous examination of the effectiveness of these mitigation strategies remains lacking. In this paper, we design a systematic and controlled pipeline along with two novel metrics-fidelity and contamination resistance-to provide a fine-grained and comprehensive assessment of existing BDC mitigation strategies. Previous assessment methods, such as accuracy drop and accuracy matching, focus solely on aggregate accuracy, often leading to incomplete or misleading conclusions. Our metrics address this limitation by emphasizing question-level evaluation result matching. Extensive experiments with 10 LLMs, 5 benchmarks, 20 BDC mitigation strategies, and 2 contamination scenarios reveal that no existing strategy significantly improves resistance over the vanilla case (i.e., no benchmark update) across all benchmarks, and none effectively balances fidelity and contamination resistance. These findings underscore the urgent need for designing more effective BDC mitigation strategies. Our code repository is available at https://github.com/ASTRAL-Group/BDC_mitigation_assessment.

Sample Weight Averaging for Stable Prediction

The challenge of Out-of-Distribution (OOD) generalization poses a foundational concern for the application of machine learning algorithms to risk-sensitive areas. Inspired by traditional importance weighting and propensity weighting methods, prior approaches employ an independence-based sample reweighting procedure. They aim at decorrelating covariates to counteract the bias introduced by spurious correlations between unstable variables and the outcome, thus enhancing generalization and fulfilling stable prediction under covariate shift. Nonetheless, these methods are prone to experiencing an inflation of variance, primarily attributable to the reduced efficacy in utilizing training samples during the reweighting process. Existing remedies necessitate either environmental labels or substantially higher time costs along with additional assumptions and supervised information. To mitigate this issue, we propose SAmple Weight Averaging (SAWA), a simple yet efficacious strategy that can be universally integrated into various sample reweighting algorithms to decrease the variance and coefficient estimation error, thus boosting the covariate-shift generalization and achieving stable prediction across different environments. We prove its rationality and benefits theoretically. Experiments across synthetic datasets and real-world datasets consistently underscore its superiority against covariate shift.