Short papers
4 pages of content (excluding references), with an optional unlimited appendix
Machine Learning/CORE Unranked
TTU at ICLR 2026
Third Workshop on Test-Time Updates (Main Track)
The 3rd Workshop on Test-Time Updates (TTU) at ICLR 2026 focuses on methods for updating models after training, including test-time adaptation, model editing, and post-training adjustments. It aims to foster cross-pollination across learning settings and domains, with an emphasis on reliability, efficiency, and practical deployment of updates in dynamic environments.
Key deadlines
Verified September 9, 2026Full paper
February 7, 2026
AoE
Workshop timeline
Submission and decisions
Full paperKey deadline
February 7, 2026 · AoE
Paper fit
Contribution paths
A strong submission should clearly identify its contribution and evaluate it appropriately.
Tiny papers
2 pages of content (excluding references), no appendix allowed
Research areas in scope
01
Foundations & Objectives
Unsupervised/self-supervised losses at test timeImplicit/explicit regularizationStability–plasticity trade-offsTheory of adaptation and generalization under shift
02
Parameterizations & Interfaces
Input-space updates (learnable augmentations, prompts)Feature-space adapters (BN/affine, LoRA adapters)Head-level editsRetrieval-augmented updatesBlack-box query strategies for closed foundation models
03
Shift, Attacks, & Tasks
Coping with domain and style shiftDistribution driftAdversarial perturbationsLabel shiftOnline continual learning and task switchesModel availability attacks
04
Adaptation of Foundational Models (FM)
Adapting LLMs/VLMs and domain FMs to specialized/personalized settings via in-context learningAdapters/LoRATTU-RLModel editing and unlearning
05
Safety, Reliability, & Alignment
UncertaintyConformal prediction at test timeFallback/abstentionGuardrails and risk monitorsPrivacy-preserving updatesAuditabilityRoll-back
06
Dynamic Architectures
Recurrent depth modelsLooped transformersDynamically allocating compute (early-exit networks, mixture-of-depth)Iterative test-time optimization (deep equilibrium networks, implicit computation)
07
Metrics, Datasets, & Benchmarks
End-to-end metrics that couple utility (accuracy, calibration) with costs (compute, memory, wall-clock, energy)Realistic streams and recurrencesReproducible TTU pipelines
08
Cost-Aware & Green TTU
Methods and evaluations under compute/energy budgetsLatency/throughput targetsEdge constraintsCarbon accountingCost–quality frontiersAny improvement must justify its operational footprint
Policies worth checking twice
- Submissions must use the ICLR 2026 paper kit
- Short papers are limited to 4 pages of content (excluding references)
- Tiny papers are limited to 2 pages of content (excluding references)
- No appendix is allowed for tiny papers
- Accepted submissions will be selected for poster or oral presentation
- The workshop will not include proceedings
Official sources
Compiled from the official call for papers. The organizers’ pages remain authoritative.
Last verified September 9, 2026