TTT discover

Here is a technical LinkedIn post summarizing the TTT-Discover framework:

🚀 Beyond Frozen LLMs: Learning to Discover at Test Time

Why settle for a "frozen" model when solving complex scientific problems? Standard LLM search (like Best-of-N) is limited by the model’s pre-existing knowledge. Enter TTT-Discover (Test-Time Training to Discover), a new framework that performs Reinforcement Learning at test time to push the boundaries of what's possible.

Unlike traditional RL, which optimizes for average performance, TTT-Discover is laser-focused on finding a single, maximum-reward solution—the ultimate "discovery".

Key Technical Pillars: 🔹 Entropic Utility Objective: A specialized training objective that prioritizes the "max" reward rather than the mean. By setting an adaptive factor ($\beta$), the model shifts focus toward the most promising solutions. 🔹 PUCT-based State Reuse: Instead of starting every attempt from scratch, the system reuses promising historical states. It balances exploitation (selecting high-reward candidates) with exploration to ensure the model doesn't get stuck in local optima. 🔹 Test-Time Weight Updates: The model continues to learn using LoRA (Low-Rank Adaptation) as it generates and evaluates solutions, effectively internalizing new insights for the specific problem at hand.

Real-World Impact:

This method has already established new SOTA benchmarks:

• Mathematics: Improved Erdős' minimum overlap problem bound to 0.380876.

• GPU Engineering: Discovered kernels up to 2x faster than human-expert baselines.