ICLR 2026 Workshop on Scaling Post-training for LLMs

About the Workshop

Post-training—spanning Supervised Fine-Tuning (SFT), Reinforcement Learning (RL), and beyond is no longer a final adaptation step. It has become a compute-intensive, first-class phase that increasingly determines the capabilities, safety, and efficiency of large language models. Yet, despite its growing importance, post-training at scale remains poorly understood.

Recent frontier models allocate a substantial and rapidly growing fraction of total compute to post-training, while academic efforts are only beginning to develop principled, scalable methodologies. Unlike pre-training, where scaling laws and design trade-offs are well studied, post-training lacks a comparable scientific framework.

The Workshop on Scaling Post-Training for LLMs (SPOT) aims to address this gap by focusing on the foundational principles that make post-training scale, across algorithms, data, architectures, systems, and objectives. Rather than targeting specific applications, SPOT centers on understanding the design choices, bottlenecks, and trade-offs that govern efficiency, stability, and asymptotic performance in large-scale post-training.

Call for Papers

We are excited to invite submissions to the ICLR 2026 Workshop on Scaling Post-Training for LLMs (SPOT). SPOT aims to bring together academic and industrial researchers to share experiences, identify open challenges, and move toward a principled science of post-training at scale.

Submission Guidelines

Topics of interest include, but are not limited to:

• Algorithms: SFT and RL formulations, reward modeling, and objectives that enable stable and compute-efficient scaling.
• Systems: Infrastructure challenges unique to post-training, including rollout latency, training–inference mismatches, distributed execution, and KV-cache management.
• Architectures: The impact of design choices such as MoE, sparse attention, and model specialization on sample efficiency and generalization.
• Data: The role of data curation and synthetic data in scaling post-training and shaping generalization.
• Environments & Safety: Post-training in multimodal, embodied, and real-world settings, with emphasis on evaluation fidelity and safety.
• Feedback & Verification: Efficient, dense feedback mechanisms as evaluation costs increasingly rival training costs.

Submission Format

• Submissions must be original, unpublished work, and not under review elsewhere.
• Long Papers should be at most 8 pages long (excluding references and appendices).
• Short/Tiny Papers should be at most 4 pages long (excluding references and appendices).
• Submissions must be in LaTeX format using the ICLR 2026 conference style.
• All submissions will be reviewed in a double-blind manner, and authors must anonymize their manuscripts
• We encourage at least one author per submission to serve as a reviewer for our workshop.

Accepted Submissions

Accepted papers will be presented as talks or posters during the workshop. The workshop will select the best paper(s) to recognize outstanding contributions in the field.

Non-Proceedings Policy

This workshop does not produce formal proceedings. Accepted submissions will appear on OpenReview, but authors remain free to submit and publish their work elsewhere in the future.

Submission Portal

Submit your paper through the OpenReview Submission Portal .

For inquiries, please contact us at spoticlr@gmail.com.