François Chollet: Test-Time Training Shines in ARC-AGI

Reviewing the ARC Prize 2024 competition, Keras founder François Chollet pointed out that Test-Time Training (TTT) technology shined in the event, helping the SOTA score jump from 33% to 55.5%. He believes that using gradient signals for test-time training is a highly promising path that shouldn't be wasted, marking the next direction to unlock the potential of large models.

已确认

During the ARC Prize 2024 competition, Test-Time Training (TTT) was introduced and achieved remarkable results, boosting the SOTA score from 33% to 55.5%.
ARC 1-2 remains one of the few test sets where TTT significantly outperforms traditional methods.
The MindsAI team delivered an outstanding performance in related technical practices.
The current mainstream approach to test-time computation relies on natural language reasoning, which is computationally similar to test-time search (sometimes combined with verifiers or scorers).
TTT is a pure deep learning paradigm that trains by utilizing gradient signals.

为什么重要

Chollet emphasized that the paradigm of merely scaling up pre-trained models has shown its limitations.
TTT represents a technical route different from mainstream language reasoning, offering a new direction to break through the capability bottlenecks of existing large models.
Industry observers (like @sarahcat21) noted that test-time training is quietly booming; predictions from half a year ago that it would become the next major tech trend have now been validated by the industry.

2026-08-13 ~ 2026-08-13 · 5 related posts

Primary sources

François Chollet: Test-Time Training Shows Great Potential to Go Mainstream — fchollet · 2026-08-13 16:41 (UTC+08:00)
François Chollet: Test-Time Training is the Only Pure Deep Learning Adaptation — fchollet · 2026-08-13 16:46 (UTC+08:00)
Keras Creator: Test-Time Training Is the Next Frontier for LLMs — fchollet · 2026-08-13 16:44 (UTC+08:00)

Test-Time Training arrives: the ignored AI prediction becomes reality — sarahcat21 · 2026-08-13
[source] François Chollet: Test-Time Training Shows Great Potential to Go Mainstream — fchollet · 2026-08-13
[source] Keras Creator: Test-Time Training Is the Next Frontier for LLMs — fchollet · 2026-08-13
[source] François Chollet: Test-Time Training is the Only Pure Deep Learning Adaptation — fchollet · 2026-08-13
Test-Time Training Excels on ARC-AGI, Potentially Ushering a New AI Paradigm — fchollet · 2026-08-13