ComfyAI·Oct 17TabSyn: Bringing Diffusion Models to Mixed-Type Tabular Data Through a Latent SpaceTabSyn is about tabular data synthesis: training a model to generate rows realistic enough to pass for the real thing, whether to augment a…
ComfyAI·Oct 15CTSyn: Toward a Foundation Model for Generating Tabular DataImages and text both have powerful generative foundation models. Tabular data, however, has always been the awkward one. Tables differ…
ComfyAI·Sep 20Pretrain Once, Skip the Retraining: CausalFM Brings Foundation Models to Causal InferenceIn causal inference — say you’re estimating how a treatment affects different patients — every new dataset with a new causal structure…
ComfyAI·Sep 10Prior-Data Fitted Networks: The Paper That Became TabPFNThis ICLR paper introduced Prior-Data Fitted Networks (PFNs), the work that later grew into TabPFN. Here’s my reading of it.A response icon1A response icon1
ComfyAI·Sep 4Can In-Context Learning Actually Learn Tasks It Never Saw in Training?In-context learning is the ability that made large language models feel magical: you show the model a few examples, and it seems to pick up…
ComfyAI·Jun 7Knowing Which Skills Fine-Tuning Will Break — Before You Fine-TuneThere’s a frustrating law that governs almost all post-training: you push one capability up, and something else quietly slides down. You…
ComfyAI·May 29Is DPO Mathematically Flawed? Inside the ICLR 2026 Oral That Redefines LLM AlignmentDPO (Direct Preference Optimization) has been one of the most popular alignment methods for LLMs over the past couple of years. Its biggest…
InCriteo Tech BlogbyCriteo Tech·May 28What Stood Out at ICLR 2026: Criteo Papers and Research HighlightsAuthors: Ahmed Ben Yahmed, Antoine Schnepf, Karim Kassab, and Mélissa Tamine.
Elias Hossain·Apr 7How I Started Learning to Build Theory for Major AI VenuesWhat rejection taught me about assumptions, proofs, and writing theory that actually supports the method
ComfyAI·Mar 22Beyond Pairwise: How Sapienza’s GRAM is Redefining Multimodal Alignment at ICLR 2025In the rapidly evolving world of multimodal AI, the “CLIP-style” approach has long been the gold standard. We use cosine similarity to pull…
