Speakers and Panelists

Abstract: Why are LLMs fundamentally so easily jailbroken? We will first analyze how to subvert and "jailbreak" simple, one-layer transformers, formalizing rule-breaking as a vulnerability in the model's attention mechanism. This theoretical work taught us a critical lesson: if attention is the key to breaking rules, it may also be the key to enforcing them. We will then introduce InstABoost, an incredibly simple yet highly effective steering method that directly applies this insight to boost the model's attention on user-provided instructions during generation. This technique, inspired by our theoretical understanding of how a small model is jailbroken, provides state-of-the-art control over large-scale LLMs with just five lines of code. 

Abstract: 

Real-world benchmarks drive a lot of progress, but they cannot capture all key aspects of real-world deployment. How does one study notions like creativity and adaptability to new domains? These are messy, subjective, and cannot be captured easily in static benchmarks. I will argue that carefully constructed minimal examples and stylized settings---"spherical cows"---offer a powerful answer, helping surface important blind spots and limits in current paradigms. 

For adaptation, I will walk through the story of how we discovered a surprising phenomenon of catastrophic overtraining, where pre-training on more tokens can hurt downstream fine-tuning. This challenges the core belief of machine learning that "more data is better". While initially conceptualized via stylized models, we put this to the test at scale on real-world datasets and observe that OLMo-1B model pre-trained on 3T tokens performs worse after fine-tuning than its less trained 2.3T token counterpart.

For creativity, we construct minimal examples that capture notions of creativity inspired by cognitive science. These examples show that next-token prediction is fundamentally myopic, underperforming multi-token approaches like teacherless training and diffusion models in generating diverse, original outputs. These settings also highlight that standard techniques such as temperature sampling may be suboptimal compared to injecting randomness through input prefixes.

Abstract: 

Inference now drives the progress of AI thanks to test-time compute, necessitating efficient redesign of core architectural components such as the attention layer. We examine recent progress in addressing these efficiency challenges through two complementary directions. Starting from first principles of hardware efficiency with arithmetic intensity, we motivate the design of DeepSeek's multi-head latent attention (MLA) and recent variants such as Grouped-Tied Attention and Grouped Latent Attention. These variants reduce memory bandwidth requirements by performing more computation per byte loaded from memory, achieving up to 2× speedup in decoding scenarios. Second, we explore a complementary direction that reduces the FLOPS of attention from quadratic to linear or quasi-linear, bridging the gap between linear attention's efficiency and softmax attention's expressiveness. These new linear and quasi-linear attention methods enable sub-quadratic scaling while maintaining modeling capacity. Through systematic investigation at small scales, we demonstrate how inference-driven design principles can unlock new insights into attention mechanisms and provide practical pathways toward more efficient large-scale deployment.