The First Workshop on Test-time Scaling and Reasoning Models
(ScalR @ COLM 2025)

Abstract: Large Language Models are powerful, yet they often struggle with tasks requiring deep reasoning. This talk traces how we've taught models to "think"—from prompting and decoding to advanced post-training and RL techniques that amplify reasoning depth—showing that reasoning is a crucial scaling dimension alongside pre-training.

Abstract: We cover what it takes to add reasoning to a new 7B open model (Olmo-Thinking) to rival Qwen 3, highlighting fresh results, trade-offs, and methods across midtraining, distillation with high-quality thinking SFT data, and reinforcement learning with verifiable rewards.

Abstract: Since the release of DeepSeek-R1, the open-source AI ecosystem has undergone a remarkable transformation in how it approaches reasoning and inference-time compute. This progress has been driven by three key developments: the emergence of open reasoning datasets that enable reproducible research, a deeper understanding of reinforcement learning for LLMs, and advances in tooling for training and inference that make large-scale experimentation accessible to the community. In this talk, I will cover how these components are converging into a coherent open reasoning stack. I will highlight the leading projects, the lessons learned from recent experimentation, and the challenges that remain for open models to achieve competitive, general-purpose reasoning capabilities.

Abstract: I’ll talk about three recent research directions to make OpenAI models more trustworthy and secure. First, I will do a deep dive into our chain-of-thought reasoning models and how we can align them with human preferences using deliberative alignment. Next, I will discuss how to mitigate prompt injections and jailbreaks by teaching LLMs to follow instructions in a hierarchical manner. Finally, I will discuss the tensions that exist between open model access and system security, whereby providing access to LM output probabilities can allow adversaries to reveal the hidden size of black-box models.