Full Papers
Original research contributions presenting substantial theoretical, empirical, or methodological results; up to 9 pages at submission, extendable to 10 pages for camera-ready.
NLP/CORE Unranked
Sci-FM 2026
Workshop on Scientific Understanding of Foundation Models
Sci-FM 2026 is a workshop at COLM 2026 focused on building a systematic scientific understanding of foundation models, moving beyond empirical observations to develop predictive, testable theories about their training dynamics, alignment, and evaluation. It brings together researchers to uncover laws, invariants, and causal structures in foundation models through rigorous theoretical and empirical work.
Paper fit
Contribution paths
A strong submission should clearly identify its contribution and evaluate it appropriately.
Short Papers
Preliminary findings, negative results, position papers, and focused contributions that advance the workshop's scientific goals; up to 4 pages.
Research areas in scope
01
Training Dynamics, Data, and Optimization
Data curation, high-quality data mixtures, and the role of open models in driving capabilitiesOptimization at scale: learning rate schedules, gradient flow, and hyperparameter transfer across model and data sizesHow optimization choices affect quantization, post-training, and downstream model behaviorTheoretical and empirical limits of scaling laws, including domain-specific scaling and breakdown regimes
02
Post-Training, Reward Modeling, and Alignment
RL, self-improvement, and how pre-training enables effective post-trainingReward systems, reward model overoptimization, and utility engineering for value systemsScaling and designing RL environments for evaluating agentic behaviorHigh-quality post-training datasets, preference pairs and reasoning traces
03
Evaluation Science and Reliability
Measurement methodology and fluid benchmarking for rapidly changing language modelsCharacterizing model capabilities: discontinuous capability gains, compositional generalization, and skill acquisition dynamicsReproducibility, determinism in inference, and reliable conclusions from imperfect dataScalable and automated analysis of model behavior and population-level phenomena
Policies worth checking twice
- All submissions undergo double-blind peer review.
- Submissions must use the default COLM template.
- The workshop is non-archival — accepted papers may be submitted elsewhere.
- Full papers may extend from 9 to 10 pages for camera-ready version; short papers remain at 4 pages.
- Authors must include \lhead{Published as a workshop paper at Sci-FM@COLM 2026} in the .sty file for camera-ready submissions.
- Negative results, careful reproductions, and position papers are valued.
Official sources
Compiled from the official call for papers. The organizers’ pages remain authoritative.
Last verified September 9, 2026