Is Nori-Causal state of the art?Pending benchmarks.Announced Oct 30, 2026; no independent results yet.
A zero-shot causal discovery model for tabular data from Synthefy. It reads one table and, in a single forward pass, returns the probability of a direct causal edge for every ordered pair of columns, so each feature gets a probability of being a direct cause or a direct effect of the target. It is trained only on synthetic tables generated from known causal graphs and needs no per-dataset fitting. About 30M parameters, Apache 2.0 weights on Hugging Face and code in the synthefy-nori library; no API is offered for it. Not an LLM.
The details so far
- Lab
- Synthefy
- Announced
- Oct 30, 2026
- API
- Not yet · None for Nori-Causal; it runs locally (Synthefy's site lists a managed API for its Nori prediction model)
- Open weights
- Published · Synthefy/Nori-Causal on Hugging Face: one PyTorch checkpoint (noricausal_31m.pt, about 121 MB, F32)
- Parameters
- About 30M (30,342,598)
- Task
- Causal discovery on a single observational table: edge probabilities for every ordered pair of columns, plus P(feature causes target) and P(target causes feature); no cutoff is applied, so it does not output a decided graph
- Architecture
- Nori input layer (RBF numeric embedding), six blocks alternating attention across rows and columns, row pooling (mean and standard deviation per column), fifteen column-attention blocks and a bilinear edge head; one optional target slot
- Input limits
- Tables over 2,048 rows are subsampled; up to 25 features in one pass, wider tables as random 30-column subsets with scores averaged
- Targets
- Continuous or 0/1 targets (pass 0/1 with task="regression"); multi-class targets are not supported yet
- Training
- Synthetic tables from simulated structural causal models, from the same data engine that trains Nori; Synthefy says millions of tables and no benchmark data
- Code
- NoriCausal and NoriInterpreter(causal=True) in the synthefy-nori library on GitHub, added Oct 30, 2026; the newest PyPI release (0.22.0, Oct 20) predates it
- License
- Apache 2.0 (model card and the synthefy-nori LICENSE and NOTICE files)
- Lab-stated limits
- Reads only the numbers, not column meaning or timing; hidden confounders, near-duplicate columns and deterministic relationships can mislead it; Synthefy says to treat scores as evidence to check, not causal conclusions
Tags marked unsure are not stated on the lab's own pages.
What would let us call it
Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.
- Publicly announced with a source
- Usable by the public (API or weights)
- Listed on an independent board we track
- Enough results for a rank in a category
- Holds across two snapshots (crowns only)
What the lab claims
Self-reported. Shown for context; not used in any verdict.
Lab claim: on the UCI Credit Default table (30,000 card holders), Nori-Causal scores the latest repayment status as a cause of default at 1.00 and the five earlier months at 0.74 to 0.88, while sex, age and marriage stay below 0.5 in both directions.
Lab claim: on the scikit-learn diabetes table (442 patients), it names triglycerides (0.78) and BMI (0.60) as likely causes of disease progression; Synthefy notes that age, sex and blood pressure lean weakly toward being effects, which cannot be right because they were measured a year earlier.
Sources
LABIntroducing Nori-Causal: Which Features Drive the Target, and Which Follow It (Synthefy blog)synthefy.com/blog/nori-causal · Oct 30, 2026LABSynthefy announcement on Xx.com/synthefyinc/status/2117318189289996392 · Nov 2, 2026REPOSynthefy/Nori-Causal on Hugging Facehuggingface.co/Synthefy/Nori-CausalREPOsynthefy-nori on GitHubgithub.com/Synthefy/synthefy-noriLABNori documentation (Synthefy)docs.synthefy.com/nori/Tracking continues until an independent board lists it.