Program

12 sessions

Training

• Curriculum Schedules for Trillion-Token Pretraining

  September 15, 2026 09:00 · Aurora Hall

  How the order you feed data in changes what a trillion-token model ends up knowing.

• Gradient Noise as a Signal: Adaptive Batch Sizing

  September 15, 2026 11:00 · Cobalt Theatre

  Using the geometry of gradient noise to grow the batch size on the fly.

• Data-Efficient Distillation from Mixture-of-Experts Teachers

  September 16, 2026 09:00 · Aurora Hall

  Compressing a sparse expert model into a dense student without losing calibration.

Inference

• Speculative Decoding Beyond Draft Models

  September 15, 2026 10:00 · Room 210

  Getting speculative-decoding speedups without training and hosting a separate draft model.

• KV-Cache Compression at Long Context

  September 15, 2026 14:00 · Room 200

  Keeping the KV cache small enough to serve 200k-token contexts affordably.

• Serving 100 Models on One GPU: Elastic Routing

  September 16, 2026 10:00 · Room 210

  Multiplexing a fleet of fine-tuned models on shared hardware with elastic routing.

Agents

• Planning with Verifier-Guided Tree Search

  September 15, 2026 13:00 · Cobalt Theatre

  Combining a learned verifier with explicit search for reliable multi-step planning.

• Tool Synthesis: Agents that Write Their Own APIs

  September 16, 2026 11:00 · Room 200

  Agents that author, test, and reuse their own tools instead of relying on a fixed toolset.

• Long-Horizon Credit Assignment in Multi-Agent Teams

  September 16, 2026 13:00 · Cobalt Theatre

  Figuring out which agent in a team deserves credit for an outcome many steps later.

Safety

• Mechanistic Auditing of Deceptive Circuits

  September 15, 2026 15:00 · Room 200

  Interpretability tools for finding circuits that behave differently when a model thinks it is watched.

• Scalable Oversight via Debate and Recursive Reward Models

  September 16, 2026 14:00 · Aurora Hall

  Supervising models on tasks where humans can no longer check the answer directly.

• Robustness Certificates for Open-Weight Deployments

  September 16, 2026 15:00 · Room 210

  What guarantees you can and cannot make once model weights are public.