Full technical workshop papers
Submissions up to 6 pages (excluding references) using the official ICML'26 template, with optional appendices not reviewed.
Machine Learning/CORE Unranked
AdaptFM 2026
AdaptFM: Resource-Adaptive Foundation Model Inference
AdaptFM 2026 is a workshop co-located with ICML'26 in Seoul, South Korea, focused on advancing resource-adaptive inference for foundation models. It brings together researchers from machine learning, systems, and hardware communities to develop flexible, efficient techniques that adjust model computation based on resource constraints like memory, latency, energy, or cost, while preserving output quality.
Paper fit
Contribution paths
A strong submission should clearly identify its contribution and evaluate it appropriately.
Short position/experience papers
Submissions up to 6 pages (excluding references) using the official ICML'26 template, with optional appendices not reviewed.
Research areas in scope
01
Flexible & Elastic Architectures
Once-for-all and supernet approachesSlimmable & any-width/any-depth networksMatryoshka representation learningMatFormer and elastic transformersLayer skipping and dynamic widthModular and reconfigurable architectures
02
Adaptive Test-Time Compute
Input-adaptive inferenceAdaptive reasoningUniversal transformer & recursive inference
03
Dynamic Networks & Routing
Early-exit & adaptive-depth networksMoEs & conditional computationCascade & routing systems
04
Efficient Decoding & Token-Level Adaptation
Speculative & parallel decodingToken pruning and mergingAdaptive token computationAdaptive attention mechanismsDynamic KV cache compression
05
Model Compression & Optimization
Quantization, pruning, and sparsityKnowledge distillationLow-rank decomposition and factorizationHardware-aware model optimization
06
Systems & Hardware
H/W–S/W co-design for adaptive inferenceRuntime systems for flexible computationBenchmarking and profiling across resource budgets
07
Analysis & Trade-offs
Quality-resource tradeoff analysisEnergy-efficient and sustainable inferenceEmerging applications of adaptive inference
Policies worth checking twice
- Submissions are non-archival and may be under active submission to other venues, but not previously published.
- Simultaneous submission to other ICML '26 workshops is actively discouraged.
- All submissions must be double-blind; author names and identifying features must be removed.
- Generative AI tools (e.g. LLMs) may be used to assist with writing, but cannot be listed as authors.
- Authors are encouraged to disclose any significant use of generative AI in their methodology.
- Submissions must adhere to standards of accuracy, originality, and scientific integrity.
Official sources
Compiled from the official call for papers. The organizers’ pages remain authoritative.
Last verified September 9, 2026