Test-Time Training Done Right

• URL: http://arxiv.org/abs/2505.23884v1
• Date: Thu, 29 May 2025 17:50:34 GMT
• Title: Test-Time Training Done Right
• Authors: Tianyuan Zhang, Sai Bi, Yicong Hong, Kai Zhang, Fujun Luan, Songlin Yang, Kalyan Sunkavalli, William T. Freeman, Hao Tan,
• Abstract summary: Test-Time Training (TTT) models context by adapting part of the model's weights (referred to as fast weights) during inference.<n>Existing TTT methods struggled to show effectiveness in handling long-context data.<n>We develop Large Chunk Test-Time Training (LaCT) which improves hardware utilization by orders of magnitude.
• Score: 61.8429380523577
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Test-Time Training (TTT) models context dependencies by adapting part of the model's weights (referred to as fast weights) during inference. This fast weight, akin to recurrent states in RNNs, stores temporary memories of past tokens in the current sequence. Existing TTT methods struggled to show effectiveness in handling long-context data, due to their inefficiency on modern GPUs. The TTT layers in many of these approaches operate with extremely low FLOPs utilization (often <5%) because they deliberately apply small online minibatch sizes (e.g., updating fast weights every 16 or 64 tokens). Moreover, a small minibatch implies fine-grained block-wise causal dependencies in the data, unsuitable for data beyond 1D ordered sequences, like sets or N-dimensional grids such as images or videos. In contrast, we pursue the opposite direction by using an extremely large chunk update, ranging from 2K to 1M tokens across tasks of varying modalities, which we refer to as Large Chunk Test-Time Training (LaCT). It improves hardware utilization by orders of magnitude, and more importantly, facilitates scaling of nonlinear state size (up to 40% of model parameters), hence substantially improving state capacity, all without requiring cumbersome and error-prone kernel implementations. It also allows easy integration of sophisticated optimizers, e.g. Muon for online updates. We validate our approach across diverse modalities and tasks, including novel view synthesis with image set, language models, and auto-regressive video diffusion. Our approach can scale up to 14B-parameter AR video diffusion model on sequences up to 56K tokens. In our longest sequence experiment, we perform novel view synthesis with 1 million context length. We hope this work will inspire and accelerate new research in the field of long-context modeling and test-time training. Website: https://tianyuanzhang.com/projects/ttt-done-right

Related papers

• Towards Universal Modal Tracking with Online Dense Temporal Token Learning [66.83607018706519]
  We propose a universal video-level modality-awareness tracking model with online dense temporal token learning.<n>We expand the model's inputs to a video sequence level, aiming to see a richer video context from a near-global perspective.
  arXiv  Detail & Related papers  (2025-07-27T08:47:42Z)
• Next-Scale Autoregressive Models are Zero-Shot Single-Image Object View Synthesizers [4.015569252776372]
  ArchonView is a method that significantly exceeds state-of-the-art methods despite being trained from scratch with 3D rendering data only and no 2D pretraining.<n>Our model also exhibits robust performance even for difficult camera poses where previous methods fail, and is several times faster in inference speed compared to diffusion.
  arXiv  Detail & Related papers  (2025-03-17T17:59:59Z)
• TREAD: Token Routing for Efficient Architecture-agnostic Diffusion Training [20.712053538676116]
  Diffusion models typically suffer from sample inefficiency and high training costs.<n>We show that TREAD reduces computational cost and simultaneously boosts model performance.<n>We achieve a competitive FID of 2.09 in a guided and 3.93 in an unguided setting.
  arXiv  Detail & Related papers  (2025-01-08T18:38:25Z)
• Truncated Consistency Models [57.50243901368328]
  Training consistency models requires learning to map all intermediate points along PF ODE trajectories to their corresponding endpoints.<n>We empirically find that this training paradigm limits the one-step generation performance of consistency models.<n>We propose a new parameterization of the consistency function and a two-stage training procedure that prevents the truncated-time training from collapsing to a trivial solution.
  arXiv  Detail & Related papers  (2024-10-18T22:38:08Z)
• Power Scheduler: A Batch Size and Token Number Agnostic Learning Rate Scheduler [34.416299887009195]
  We study the correlation between optimal learning rate, batch size, and number of training tokens for the recently proposed WSD scheduler. We propose a new learning rate scheduler, Power scheduler, that is agnostic about the number of training tokens and batch size. Our 3B dense and MoE models trained with the Power scheduler achieve comparable performance as state-of-the-art small language models.
  arXiv  Detail & Related papers  (2024-08-23T20:22:20Z)
• Training-Free Long-Context Scaling of Large Language Models [114.53296002607993]
  We propose Dual Chunk Attention, which enables Llama2 70B to support context windows of more than 100k tokens without continual training. By decomposing the attention for long sequences into chunk-based modules, DCA manages to effectively capture the relative positional information of tokens.
  arXiv  Detail & Related papers  (2024-02-27T12:39:23Z)
• YaRN: Efficient Context Window Extension of Large Language Models [1.024113475677323]
  Rotary Position Embeddings (RoPE) have been shown to effectively encode positional information in transformer-based language models. We present YaRN, a compute-efficient method to extend the context window of such models. We show that LLaMA models can effectively utilize and extrapolate to context lengths much longer than their original pre-training would allow.
  arXiv  Detail & Related papers  (2023-08-31T18:18:07Z)
• Parameter-efficient Tuning of Large-scale Multimodal Foundation Model [68.24510810095802]
  We propose A graceful prompt framework for cross-modal transfer (Aurora) to overcome these challenges. Considering the redundancy in existing architectures, we first utilize the mode approximation to generate 0.1M trainable parameters to implement the multimodal prompt tuning. A thorough evaluation on six cross-modal benchmarks shows that it not only outperforms the state-of-the-art but even outperforms the full fine-tuning approach.
  arXiv  Detail & Related papers  (2023-05-15T06:40:56Z)
• eP-ALM: Efficient Perceptual Augmentation of Language Models [70.47962271121389]
  We propose to direct effort to efficient adaptations of existing models, and propose to augment Language Models with perception. Existing approaches for adapting pretrained models for vision-language tasks still rely on several key components that hinder their efficiency. We show that by freezing more than 99% of total parameters, training only one linear projection layer, and prepending only one trainable token, our approach (dubbed eP-ALM) significantly outperforms other baselines on VQA and Captioning.
  arXiv  Detail & Related papers  (2023-03-20T19:20:34Z)
• Efficient Large-Scale Language Model Training on GPU Clusters [19.00915720435389]
  Large language models have led to state-of-the-art accuracies across a range of tasks. Memory capacity is limited, making it impossible to fit large models on a single GPU. The number of compute operations required to train these models can result in unrealistically long training times.
  arXiv  Detail & Related papers  (2021-04-09T16:43:11Z)
• PGT: A Progressive Method for Training Models on Long Videos [45.935259079953255]
  Main-stream method is to split a raw video into clips, leading to incomplete temporal information flow. Inspired by natural language processing techniques dealing with long sentences, we propose to treat videos as serial fragments satisfying Markov property. We empirically demonstrate that it yields significant performance improvements on different models and datasets.
  arXiv  Detail & Related papers  (2021-03-21T06:15:20Z)
• Convolutional Tensor-Train LSTM for Spatio-temporal Learning [116.24172387469994]
  We propose a higher-order LSTM model that can efficiently learn long-term correlations in the video sequence. This is accomplished through a novel tensor train module that performs prediction by combining convolutional features across time. Our results achieve state-of-the-art performance-art in a wide range of applications and datasets.
  arXiv  Detail & Related papers  (2020-02-21T05:00:01Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.