Dictionary

Concept dictionary

The vocabulary you need to read this archive. Each entry is the term as the field uses it, paired with a paper in the corpus where Urban works with it.

The definitions are ours, written to get you into the papers — they are not quotations of Urban. The citation under each one is the part that is his.

Automated theorem prover (ATP)
A program that searches for a proof of a conjecture from a set of axioms with no human in the loop — it either finds one, runs out of time, or reports failure. Powerful in principle and combinatorially explosive in practice, which is the gap most of this archive works in.
In the corpus Learning Guided Automated Reasoning: A Brief Survey (2024) ↗
Proof assistant (ITP)
The interactive counterpart: a system in which a human writes a proof and the machine checks every step. Mizar, HOL Light, HOL4, Isabelle and Coq all appear in this corpus — a hammer is the bridge from one of these to an ATP.
In the corpus Mizar: State-of-the-art and Beyond (2015) ↗
Mizar Mathematical Library (MML)
One of the largest bodies of formalized, machine-checked mathematics, written in the Mizar language over decades. It is the substrate for much of this archive: first as something to translate, later as a training set.
In the corpus Translating Mizar for First Order Theorem Provers (2003) ↗
MPTP
Mizar Problems for Theorem Proving — Urban’s translation of the Mizar library into first-order problems that ordinary automated provers can attempt. The move that turns a library of human proofs into a benchmark and a training corpus.
In the corpus MPTP 0.2: Design, Implementation, and Initial Experiments (2006) ↗
Premise selection
Choosing which of thousands of available facts to hand a prover for a given conjecture. Called out in this corpus as “one of the main bottlenecks in the formalization of mathematics” — get it wrong and a solvable goal becomes unreachable.
In the corpus DeepMath - Deep Sequence Models for Premise Selection (2016) ↗
Large-theory reasoning
Automated reasoning in settings where thousands of theorems, definitions and lemmas are available at once and consistently reused, rather than the handful of axioms a classical prover expects. The setting MaLARea was built for.
In the corpus MaLARea: a Metasystem for Automated Reasoning in Large Theories. (2007) ↗
Hammer
A tool inside a proof assistant that takes the current goal, selects relevant facts, ships the problem to external automated provers, and reconstructs any proof that comes back as a checked proof in the assistant. The user experience of everything above.
In the corpus HOL(y)Hammer: Online ATP Service for HOL Light (2014) ↗
Clause selection
Inside a saturation-style prover, the repeated choice of which clause to process next — the single most consequential decision in the search loop, and the one ENIGMA learns to make from previous proofs.
In the corpus ENIGMA: Efficient Learning-Based Inference Guiding Machine (2017) ↗
Tactic
A proof-assistant command that transforms a goal into subgoals — the unit a human actually writes. TacticToe learns, from human proofs, which tactic suits a situation, and searches at that level rather than over raw inference steps.
In the corpus TacticToe: Learning to Prove with Tactics (2020) ↗
Property-invariant embedding
A way of representing logical formulas for a neural network so that the representation does not depend on the particular names of symbols — so a learner generalizes across problems that use different vocabulary for the same structure.
In the corpus Property Invariant Embedding for Automated Reasoning (2020) ↗
Conjecturing
Producing the statement to be proved rather than the proof of a given statement — here, learning to invent the induction predicates that hard arithmetic problems need, starting from no examples of them.
In the corpus Learning Conjecturing from Scratch (2025) ↗
Autoformalization
Translating ordinary mathematical prose into a formal language a proof assistant can check — the reverse direction from MPTP. In this corpus it moves from a neural-machine-translation experiment on Mizar to a textbook of topology formalized at scale.
In the corpus Exploration of neural machine translation in autoformalization of mathematics in Mizar (2020) ↗
Browse the full corpus →
Are you Josef Urban? Or know Josef Urban? We'd like to hand this room over. →
Join the constellation