Tiny and Short Paper
3-5 pages in ICLR format + references and appendix; welcomes late-breaking results, replication studies, and conceptual explorations with emphasis on clarity and potential impact rather than extensive experiments.
Machine Learning/CORE Unranked
LIT Workshop @ ICLR 2026
Workshop on Latent & Implicit Thinking – Going Beyond CoT Reasoning
The LIT Workshop @ ICLR 2026 focuses on latent and implicit reasoning in vision and language models, exploring alternatives to explicit Chain-of-Thought (CoT) reasoning. It aims to unify research on non-autoregressive, continuous, and stateful reasoning mechanisms to advance more efficient and capable AI reasoning systems.
Key deadlines
Verified September 9, 2026Full paper
February 9, 2026
AoE
Workshop timeline
Submission and decisions
Full paperKey deadline
February 9, 2026 · AoE
Paper fit
Contribution paths
A strong submission should clearly identify its contribution and evaluate it appropriately.
Regular Paper
Up to 10 pages in ICLR format + references and appendix.
Research areas in scope
01
Training Strategy for Implicit Reasoning
curriculum learningdistillationreinforcement learningpretraining from scratch
02
Looped Architectures
recurrence mechanisms (loop unrolling, dynamic halting)training curricula and stability for deep iterative models
03
Mechanical Interpretability of Implicit Reasoning
layer-wise specialization and attribution of reasoning functionscausal interventions on intermediate representations
04
Special Thinking Tokens
text CoT compressed to special tokens (e.g., continuous thought tokens, VQ-VAE codes, gist tokens)CoT augmentation via filler or planning tokens
05
KV-Cache and Hybrid Stateful Reasoning
leveraging key/value caches for multi-step latent inferencecomparisons between pure activation vs. cache-augmented loops
06
Inference Paradigms Beyond Autoregression
text diffusion for bidirectional, iterative denoising-based reasoningfractal generative frameworks and next-block prediction
07
Theoretical Results on Depth, Scaling, and Efficiency
theoretical bounds on reasoning depth vs. layer countparameter- and compute-efficient designs for implicit reasoning
08
Evaluation and Benchmarks
metrics and probes for implicit vs. explicit CoT capabilitiesdatasets and tasks that stress ultra-deep or multi-hop latent reasoning
09
Limitations and Safety
understanding pros and cons of implicit and explicit CoTinterpretability & faithfulness of reasoning from a safety and alignment perspective
Policies worth checking twice
- Submissions must be anonymized for double-blind review.
- Submissions must not have been previously published at ICLR or other machine learning or related venues.
- Dual submissions are allowed if not violating the guidelines of the other venue.
- AI-generated papers are prohibited in the tiny and short paper tracks.
- Limited AI assistance (e.g., grammar correction) is permitted only if human-authored intellectual contribution remains primary.
- All submissions must disclose substantial use of AI tools in metadata or acknowledgments.
- LLMs are not permitted for automated reviewing or decision-making; all reviews are conducted by humans.
- Organizers may use LLMs only for logistical purposes (e.g., website drafting) with full human oversight.
Official sources
Compiled from the official call for papers. The organizers’ pages remain authoritative.
Last verified September 9, 2026