TTT, TAO, δ-mem: When to Spend a Gradient
summary
Compares three concrete methods and when to pay for a gradient: Test-Time Training is ephemeral per inference, Test-Time Adaptive Optimization is offline critic-scored fine-tuning, and δ-mem is a zero-weight sidecar that gains ~10 points with no latency cost. Proposes a gated human-verified LoRA that “rents” corrections then “owns” them after eight verified corrections; demo drills a small LoRA in ~50 seconds, simulations show a 95% performance gain with experts consulted 26% of the time but required 12 rollbacks for forgetting.
AI-written summary · about
chapters (18)
- 0:00The unseen failure in agent learning
- 1:15Learning at inference time
- 1:31When is a correction worth a gradient?
- 1:42Prompt, memory, or weights?
- 4:18An analogy for three kinds of learning
- 5:23Why fixed weights repeat mistakes
- 6:06Cramming, hunches, and instincts
- 7:42Test-Time Training
- 9:30Test-Time Adaptive Optimization
- 12:02δ-mem
- 13:50Comparing the methods
- 14:15What existing approaches do not solve
- 15:27A gated, human-verified LoRA method
- 16:34Confidence gates, rent-to-own learning, and rollback
- 18:14Demo
- 19:40Results and tradeoffs
- 20:26When to spend a gradient
- 21:24Deployment options and future work
related talks
01h 04mPredicting and optimizing the behavior of large ML models
46mIncreasing Participant Agency by Understanding the Impact of Training Data
01h 02mThe Key Ingredients of Optimizing Test-Time Compute and What's Still Missing
42mPost-Training Small Language Models: the adventures of a practitioner
44mParsimonious Learning-Augmented Algorithms
01h 05mOptimization for Machine Learning I
01h 05mTutorial: Federated Optimization, Part III
45mScale Learning and Reasoning Across Heterogeneous Gradients and Semantics