The study examines how extra compute during testing affects the performance of large language models.
problem Understanding the conditions under which test-time scaling improves model performance.
method An in-context weight prediction task for linear regression was used to train transformers. The performance was analyzed under varying levels of test-time compute.
result Training transformers on diverse, relevant, and hard tasks leads to the best performance for test-time scaling.
New method optimizes language model performance for test-time strategies.
problem Mismatch between training objectives and test-time deployment of large language models.
method Tail-Extrapolated estimators to approximate best-of-N performance from limited training rollouts.
result Improved performance of best-of-N deployment across various models and datasets.
A probabilistic framework for online test-time adaptation
problem Adapting models to new data under distributional shift
method State-space modelling architecture
result Characterizing parameter learning, time evolution, prior tuning, and prediction
New scaling laws optimize model size, training, and inference for better performance.
problem Trade-off between model size and inference cost in modern LLMs.
method Train-to-Test (T2) scaling laws that jointly optimize model size, training tokens, and inference samples. result Optimal pretraining decisions shift into overtraining regime, leading to stronger performance.
DAD++ improves test-time adversarial defense without access to training data.
problem Improving adversarial robustness of pre-trained models without access to training data.
method Test-time Data-free Adversarial Defense (DAD++) with detection and correction schemes.
result DAD++ significantly improves adversarial robustness with minimal clean accuracy drop.
Simple policy search outperforms advanced learnable test-time augmentation techniques.
problem Improving predictive performance through test-time data augmentation.
method Greedy policy search (GPS) for learning test-time augmentation policies.
result Augmentation policies learned with GPS achieve superior predictive performance and robustness.
New DP training ensures models behave similarly at training and test time.
problem Standard SGD training leads to inconsistent model behavior at training and test time.
method Differentially-Private (DP) training ensures WYSIWYG property through distributional generalization.
result DP training guarantees high-level WYSIWYG property, improving model robustness and privacy.
QAlign improves language model alignment with less compute, outperforming existing methods.
problem Improving language model performance with limited test-time computation.
method QAlign: sampling from optimal aligned distribution using Markov chain Monte Carlo.
result Consistent improvements over existing methods on various benchmarks.
Adapts CNN for robust medical image segmentation across different scanners and protocols.
problem Performance degradation of CNNs in medical image segmentation due to mismatch between training and test images.
method Designs a segmentation CNN as a concatenation of a shallow normalization CNN and a deep CNN. At test time, adapts the normalization sub-network for each test image using a denoising autoencoder.
result Consistently improves performance on multi-center MRI datasets of brain, heart, and prostate.
A new approach for test-time adaptation detects and reacts to distribution shifts.
problem Improving test-time accuracy under distribution shifts.
method Online self-training with a detection tool based on entropy values and betting martingales.
result The classifier's entropy values match those of the source domain, building invariance to distribution shifts.
In-Place TTT enhances LLMs with dynamic parameter updates at inference time.
problem Static training limits LLMs from adapting to new information.
method In-Place TTT updates a subset of model parameters (fast weights) at inference time.
result In-Place TTT enables 4B-parameter models to outperform on tasks with up to 128k contexts.
Method adapts frozen models for few-shot tasks without training.
problem Deployment constraints limit model updates, necessitating new adaptation methods.
method Exponential tilting of latent distribution for inference.
result Method outperforms parameter-update methods across benchmarks.
In this paper, we propose Test-Time Training, a general approach for improving the performance of predictive models when training and test data come from different distributions. We turn a single unlabeled test sample into a self-supervised learning problem, on which we update the model parameters before making a predi…
PriorGuide adapts diffusion models to new priors at test time.
problem Limited applicability of prior distributions in diffusion-based inference.
method PriorGuide uses a guidance approximation to adapt diffusion models to new priors at test time.
result Enhances the versatility of pre-trained inference models by allowing flexible adaptation to new priors.
This paper improves test-time adaptation for distribution shifts using confidence maximization and input transformation.
problem Improving deep networks' performance on data shifted from the training distribution.
method Proposes a novel loss function combining confidence maximization and batch-wise entropy maximization with an input transformation module.
result Significantly improves robustness of pretrained networks to corruptions on benchmarks like ImageNet-C.
Test-time training adapts a pretrained model to each prompt via parameter updates, improving accuracy under pretraining-to-test distribution shifts.
problem Improving accuracy of pretrained models under distribution shifts.
method Explaining TTT behavior through a decision-theoretic lens.
result TTT reduces prediction error when updates are spectrally matched to the prompt's signal-to-noise ratio and aligned with query-relevant eigen-directions.
The abstract explores connections between reinforcement learning, scaling, and diffusion.
problem Aligning reinforcement learning with human feedback and scaling techniques.
method Clarifying connections between reinforcement learning, scaling, and diffusion.
result Introducing a resampling approach for alignment and reward-directed diffusion models.
Machine learning models are often used at test-time subject to constraints and trade-offs not present at training-time. For example, a computer vision model operating on an embedded device may need to perform real-time inference, or a translation model operating on a cell phone may wish to bound its average compute tim…
Unified framework for certifying LLM reliability without extra supervision.
problem Improving reliability of large language models without additional supervision.
method Unified framework using majority voting and Martingale Majority Certificate (MMC).
result Certifiable inference in LLMs with statistical guarantees and adaptive stopping rules.
Implicit models can match or exceed explicit models with more test-time compute.
problem Understanding the expressive power and scaling of implicit models.
method Nonparametric analysis of expressive power, mathematical characterization of implicit operators, and test-time scaling experiments.
result Implicit models can progressively express more complex mappings through iteration, matching a richer function class with test-time compute.
ORCA calibrates LLMs for efficient, generalizable reasoning.
problem Miscalibration of large language models leading to inefficiencies.
method Online Reasoning Calibration (ORCA) using conformal prediction and test-time training.
result ORCA provides higher efficiency and generalization across different reasoning tasks.
PG-EVIKAL refines molecular property predictions using neighbor fusion and evidential neural networks.
problem Improving molecular property predictions using test-time neighbor fusion.
method Adapting evidential neural networks to refine predictions by re-ranking structurally similar neighbors.
result PG-EVIKAL reduces RMSE on 14 out of 16 molecular datasets, improving calibration and sequential refinement.
Paper proposes an efficient method for calibrating spatio-temporal forecasts.
problem Real-world spatio-temporal forecasting challenges like signal anomalies and distributional shifts.
method Learning with Calibration (ST-TTC) for real-time bias correction.
result ST-TTC improves spatio-temporal forecasting accuracy with reduced computational cost.
TTT improves transformer models for in-context learning.
problem Improving transformer models for efficient in-context learning.
method Gradient-based TTT method for linear transformers, with theoretical and empirical analysis.
result TTT significantly reduces the sample size required for in-context learning.
M-L2O adapts fast to new tasks by self-adapting during test-time.
problem L2O optimizers struggle with out-of-distribution tasks.
method Meta-training an L2O optimizer to adapt quickly to new tasks.
result M-L2O converges significantly faster than vanilla L2O with only 5 steps of adaptation.
Study compares data-driven vs model-based MRS quantification strategies, focusing on resilience to out-of-distribution effects.
problem Resilience to out-of-distribution effects in data-driven MRS quantification.
method Compared three data-driven strategies (supervised regression, self-supervised learning, test-time adaptation) against model-based fitting tools.
result Test-time adaptation proved most resilient to out-of-distribution effects, while self-supervised learning achieved intermediate performance.
Tent adapts models during testing by minimizing entropy of predictions.
problem Adapting models to new data during testing with limited information.
method Test entropy minimization (tent) and online channel-wise affine transformations.
result Reduces generalization error on various datasets and benchmarks.
We consider the problem of function estimation in the case where the data distribution may shift between training and test time, and additional information about it may be available at test time. This relates to popular scenarios such as covariate shift, concept drift, transfer learning and semi-supervised learning. Th…
PETAL adapts models to changing target domains over time.
problem Lifelong test-time adaptation in changing target domains.
method Probabilistic framework with student-teacher model and data-driven parameter restoration.
result PETAL achieves better results than state-of-the-art for online lifelong test-time adaptation.
Plug-in robust NPE method adapts summaries independently of pretrained NPE.
problem Misspecification of neural posterior estimators under test data distribution.
method Minimum-distance summaries using maximum mean discrepancy (MMD).
result Substantial robustness gains with minimal additional overhead.
Enhances image quality to improve test-time adaptation accuracy.
problem Reducing accuracy loss due to distribution shift in deep networks.
method Integrates image enhancement with TTA methods to reduce prediction uncertainty.
result TECA method increases accuracy of TTA methods without hyperparameters.
Study uses neural fields to improve geophysical inversions by reducing artifacts.
problem Improving geophysical inversions by reducing artifacts and improving model recovery.
method Employing neural fields for test-time learning in geophysical inversions.
result Test-time learning with neural fields eliminates unwanted artifacts in recovered models.
New method learns models to adapt to domain shifts at test time.
problem Learning models robust to distribution shifts in practical applications.
method Adaptive Risk Minimization (ARM) framework.
result Performance gains of 1-4% on image classification problems.
TTT improves model adaptation to test data, especially for nonlinear models.
problem Improving model performance in adapting to test data, especially for nonlinear models.
method Combining Test-time Training (TTT) with In-context Learning (ICL) for nonlinear models.
result TTT enables models to adapt to both feature vector and link function shifts, improving performance.
This paper tackles over-certainty in test-time adaptation models, proposing a solution to improve calibration.
problem Over-certainty in predictions caused by domain shifts, leading to misplaced trust.
method Introduces a certainty regularizer that dynamically adjusts pseudo-label confidence based on backbone entropy and logit norm.
result Achieves state-of-the-art performance in terms of Expected Calibration Error and Negative Log Likelihood, while maintaining accuracy.
We propose a novel adaptive approximation approach for test-time resource-constrained prediction. Given an input instance at test-time, a gating function identifies a prediction model for the input among a collection of models. Our objective is to minimize overall average cost without sacrificing accuracy. We learn gat…
Test-time training improves sampling efficiency in generative AI.
problem Efficient sampling from complex probability distributions.
method Test-time training (TTT) updates model weights based on partial generations and feedback.
result TTT can adapt to specific sampling tasks, improving sampling efficiency.
We study the problem of structured prediction under test-time budget constraints. We propose a novel approach applicable to a wide range of structured prediction problems in computer vision and natural language processing. Our approach seeks to adaptively generate computationally costly features during test-time in ord…
Low bit-width weights and activations are an effective way of combating the increasing need for both memory and compute power of Deep Neural Networks. In this work, we present a probabilistic training method for Neural Network with both binary weights and activations, called BLRNet. By embracing stochasticity during tr…
Prunes neural networks at initialization to save resources, achieving high accuracy.
problem Efficiently reducing resource requirements for neural networks at both training and test time.
method Gradient Signal Preservation (GraSP) to prune networks at initialization.
result Pruning 80% of VGG-16 weights on ImageNet with only a 1.6% drop in accuracy.
A significant advance in accelerating neural network training has been the development of normalization methods, permitting the training of deep models both faster and with better accuracy. These advances come with practical challenges: for instance, batch normalization ties the prediction of individual examples with o…
TTLSA adapts models to label shifts across domains with nuisance factors.
problem Adapting models to changes in label distributions with nuisance factors.
method TTLSA uses EM on unlabeled samples to adapt a trained model to new label distributions.
result TTLSA improves model performance over invariance methods and baseline methods.
Convolutional neural networks (CNNs) are commonly trained using a fixed spatial image size predetermined for a given model. Although trained on images of aspecific size, it is well established that CNNs can be used to evaluate a wide range of image sizes at test time, by adjusting the size of intermediate feature maps.…
While Convolutional Neural Networks (CNNs) trained for image and video super-resolution (SR) regularly achieve new state-of-the-art performance, they also suffer from significant drawbacks. One of their limitations is their lack of robustness to unseen image formation models during training. Other limitations include t…
STAD adapts models to evolving time-based data shifts.
problem Gradual distribution shifts over time challenge existing test-time adaptation methods.
method Bayesian filtering method that learns time-varying dynamics in hidden features.
result STAD excels in handling small batch sizes and label shift on real-world data.
Study robustness of split conformal prediction under adversarial attacks.
problem Ensuring distribution-free coverage guarantees in CP under adversarial conditions.
method Theoretical analysis and extensive experiments on split conformal prediction robustness.
result Prediction coverage varies with calibration-time attack strength, enabling control over coverage under adversarial tests.
A method to automatically learn proposal distributions for energy-based regression models.
problem Manual design and initial estimate of proposal distributions for energy-based regression models.
method Introduces a method to learn an effective proposal distribution automatically, parameterized by a separate network head, and derives a unified training objective to minimize KL divergence and negative log-likelihood.
result Consistently outperforms conventional MDN training on four real-world regression tasks within computer vision.
We propose a method to impose homogeneous linear inequality constraints of the form Ax≤0 on neural network activations. The proposed method allows a data-driven training approach to be combined with modeling prior knowledge about the task. One way to achieve this task is by means of a projection step at test time…