Zekai Wang

While test-time scaling has enabled large language models to solve highly difficult tasks, state-of-the-art results come at exorbitant compute costs. These inefficiencies can be attributed to the miscalibration of post-trained language models, and the lack of calibration in popular sampling techniques. Here, we present Online Reasoning Calibration (ORCA), a framework for calibrating the sampling process that draws upon conformal prediction and test-time training. Specifically, we introduce a meta-learning procedure that updates the calibration module for each input. This allows us to provide valid confidence estimates under distributional shift, e.g. in thought patterns that occur across different stages of reasoning, or in prompt distributions between model development and deployment. ORCA not only provides theoretical guarantees on conformal risks, but also empirically shows higher efficiency and generalization across different reasoning tasks. At risk level delta=0.1, ORCA improves Qwen2.5-32B efficiency on in-distribution tasks with savings up to 47.5% with supervised labels and 40.7% with self-consistency labels. Under zero-shot out-of-domain settings, it improves MATH-500 savings from 24.8% of the static calibration baseline to 67.0% while maintaining a low empirical error rate, and the same trend holds across model families and downstream benchmarks.

Contrastive adversarial training has successfully improved the robustness of contrastive learning (CL). However, the robustness metric in these methods depends on attack algorithms, image labels, and downstream tasks, introducing reliability concerns. To address these issues, this paper proposes a novel Robustness Verification framework for Contrastive Learning (RVCL). Specifically, we define the verification problem of CL from deterministic and probabilistic perspectives, then provide several effective metrics to evaluate the robustness of CL encoder. Furthermore, we use extreme value theory to reveal the relationship between the robust radius of the CL encoder and that of the supervised downstream task. Extensive experiments on various benchmark models and datasets validate theoretical findings, and further demonstrate RVCL’s capability to evaluate the robustness of both CL encoders and images.

It has been recognized that the data generated by the denoising diffusion probabilistic model (DDPM) improves adversarial training. After two years of rapid development in diffusion models, a question naturally arises: can better diffusion models further improve adversarial training? This paper gives an affirmative answer by employing the most recent diffusion model which has higher efficiency (∼20 sampling steps) and image quality (lower FID score) compared with DDPM. Our adversarially trained models achieve state-of-the-art performance on RobustBench using only generated data (no external datasets). Under the L∞-norm threat model with ε=8/255, our models achieve 70.69% and 42.67% robust accuracy on CIFAR-10 and CIFAR-100, respectively, i.e. improving upon previous state-of-the-art models by +4.58% and +8.03%. Under the L2-norm threat model with ε=128/255, our models achieve 84.86% on CIFAR-10 (+4.44%). These results also beat previous works that use external data. We also provide compelling results on the SVHN and TinyImageNet datasets.

Contrastive adversarial training has successfully improved the robustness of contrastive learning (CL). However, the robustness metric used in these methods is linked to attack algorithms, image labels and downstream tasks, all of which may affect the consistency and reliability of robustness metric for CL. To address these problems, this paper proposes a novel Robustness Verification framework for Contrastive Learning (RVCL). Furthermore, we use extreme value theory to reveal the relationship between the robust radius of the CL encoder and that of the supervised downstream task. Extensive experimental results on various benchmark models and datasets verify our theoretical findings, and further demonstrate that our proposed RVCL is able to evaluate the robustness of both models and images.

Randomized smoothing (RS) has provided state-of-the-art (SOTA) certified robustness against adversarial perturbations for large neural networks. Among studies in this field, methods based on adversarial training (AT) achieve remarkably robust performance by applying adversarial examples to construct the smoothed classifier. These AT-based RS methods typically seek a pointwise adversary that generates the worst-case adversarial examples by perturbing each input independently. However, there are unexplored benefits to considering such adversarial robustness across the entire data distribution. To this end, we provide a novel framework called DRF, which connects AT-based RS methods with distributional robustness (DR), and show that these methods are special cases of their counterparts in our framework. Due to the advantages conferred by DR, our framework can control the trade-off between the clean accuracy and certified robustness of smoothed classifiers to a significant extent. Our experiments demonstrate that DRF can substantially improve the certified robustness of AT-based RS.