Meta's World-T³ introduces test-time training as a fixed-size, camera-pose-indexed memory that writes historical observations into fast weights, enabling long-horizon consistent video generation at 20.6 FPS on a single GPU while outperforming baselines on revisit consistency metrics.
TTTIR introduces a test-time training framework that models image restoration as progressive instance-specific state evolution, using a Progressive State Sequence Generator for target trajectories and a State Transition Evolver for adaptive operator updates, achieving state-of-the-art results on low-light enhancement, deraining, and dehazing benchmarks with minimal parameters.
Ilya Sutskever’s new SSI engine uses Test‑Time Training to achieve unprecedented data efficiency, ten‑fold scaling, and persistent digital agents, marking a shift from static, large‑parameter models toward continuously learning systems while raising new safety and interpretability challenges.
The ViT³ paper, a CVPR 2026 best‑paper finalist, introduces test‑time training to compress visual context, achieving 4.6× faster inference and 90 % lower GPU memory on 1248×1248 images, while outlining six design principles and demonstrating its adaptability to classification, detection, segmentation, and generation tasks.
PRISM demonstrates that linear‑attention models can adopt a “write‑before‑think” paradigm by reconstructing the multi‑step step‑size × residual × direction iteration of Test‑Time Training, achieving Transformer‑level quality while delivering up to 174× higher throughput through parallel scan and fused kernels.
The paper systematically studies Test‑Time Training (TTT) for vision, derives six design principles, and introduces ViT³—a pure TTT architecture that uses full‑batch internal training, a learning rate of 1.0, and lightweight SwiGLU‑Depthwise convolution modules, achieving state‑of‑the‑art linear‑complexity performance across classification, detection, segmentation and generation tasks.
Scal3R introduces test‑time training with a global‑context memory and synchronization mechanism that lets models train on and infer over ultra‑long video sequences, achieving accurate camera poses and dense point clouds for kilometer‑scale scenes while outperforming prior SLAM, SfM and streaming baselines on multiple benchmarks.
The paper introduces TEMPO, a test‑time training framework inspired by the Expectation‑Maximization algorithm, which alternates policy optimization (M‑step) with Critic calibration (E‑step) to prevent reward‑signal drift, and demonstrates on Qwen3 and OLMO3 models that it continuously improves reasoning performance and maintains output diversity beyond the saturation point of existing TTT methods.
The article examines the Test‑Time Training to Discover (TTT‑Discover) approach, which applies reinforcement learning during inference to let large language models continuously improve on single test problems, and reports strong results across mathematics, GPU kernel optimization, algorithm design, and biology.
Google’s newly unveiled Titans architecture tackles AI’s “forgetfulness” by embedding a Neural Long‑Term Memory (LMM) module that updates model weights during inference using a test‑time training approach and a MIRAS surprise metric, enabling over 2 million‑token context with linear O(N) computation and superior benchmark results versus GPT‑4 RAG.
MIT researchers argue that the ARC benchmark is essentially a visual problem and present the Vision ARC (VARC) framework, which reformulates ARC as an image‑to‑image translation task using a Vision Transformer, achieving human‑level accuracy through a novel canvas representation and test‑time training.
