Solving Physics Olympiad via Reinforcement Learning on Physics Simulators

We have witnessed remarkable advances in LLM reasoning capabilities with the advent of DeepSeek-R1. However, much of this progress has been fueled by the abundance of internet question-answer (QA) pairs, a major bottleneck going forward, since such data is limited in scale and concentrated mainly in domains like mathematics. In contrast, other sciences such as physics lack large-scale QA datasets to effectively train reasoning-capable models. In this work, we show that physics simulators can serve as a powerful alternative source of supervision for training LLMs for physical reasoning. We generate random scenes in physics engines, create synthetic question-answer pairs from simulated interactions, and train LLMs using reinforcement learning on this synthetic data. Our models exhibit zero-shot sim-to-real transfer to real-world physics benchmarks: for example, training solely on synthetic simulated data improves performance on IPhO (International Physics Olympiad) problems by 5-10 percentage points across model sizes. These results demonstrate that physics simulators can act as scalable data generators, enabling LLMs to acquire deep physical reasoning skills beyond the limitations of internet-scale QA data. Code available at: https://sim2reason.github.io/.

An Embarrassingly Simple Defense Against Backdoor Attacks On SSL

Self Supervised Learning (SSL) has emerged as a powerful paradigm to tackle data landscapes with absence of human supervision. The ability to learn meaningful tasks without the use of labeled data makes SSL a popular method to manage large chunks of data in the absence of labels. However, recent work indicates SSL to be vulnerable to backdoor attacks, wherein models can be controlled, possibly maliciously, to suit an adversary's motives. Li et. al (2022) introduce a novel frequency-based backdoor attack: CTRL. They show that CTRL can be used to efficiently and stealthily gain control over a victim's model trained using SSL. In this work, we devise two defense strategies against frequency-based attacks in SSL: One applicable before model training and the second to be applied during model inference. Our first contribution utilizes the invariance property of the downstream task to defend against backdoor attacks in a generalizable fashion. We observe the ASR (Attack Success Rate) to reduce by over 60% across experiments. Our Inference-time defense relies on evasiveness of the attack and uses the luminance channel to defend against attacks. Using object classification as the downstream task for SSL, we demonstrate successful defense strategies that do not require re-training of the model. Code is available at https://github.com/Aryan-Satpathy/Backdoor.