New Implementation Framework for Saturation-Based Reasoning

The saturation-based reasoning methods are among the most theoretically developed ones and are used by most of the state-of-the-art first-order logic reasoners. In the last decade there was a sharp increase in performance of such systems, which I attribute to the use of advanced calculi and the intensified research in implementation techniques. However, nowadays we are witnessing a slowdown in performance progress, which may be considered as a sign that the saturation-based technology is reaching its inherent limits. The position I am trying to put forward in this paper is that such scepticism is premature and a sharp improvement in performance may potentially be reached by adopting new architectural principles for saturation. The top-level algorithms and corresponding designs used in the state-of-the-art saturation-based theorem provers have (at least) two inherent drawbacks: the insufficient flexibility of the used inference selection mechanisms and the lack of means for intelligent prioritising of search directions. In this position paper I analyse these drawbacks and present two ideas on how they could be overcome. In particular, I propose a flexible low-cost high-precision mechanism for inference selection, intended to overcome problems associated with the currently used instances of clause selection-based procedures. I also outline a method for intelligent prioritising of search directions, based on probing the search space by exploring generalised search directions. I discuss some technical issues related to implementation of the proposed architectural principles and outline possible solutions.

💡 Research Summary

The paper opens by recalling the impressive performance gains achieved over the last decade by saturation‑based first‑order theorem provers. These gains are attributed to advances in calculi (complete variants of resolution and paramodulation with ordering restrictions), sophisticated redundancy elimination, term indexing, and heuristic search control. However, recent benchmark results (e.g., CASC‑20) show a clear slowdown, prompting some researchers to claim that saturation‑based reasoning has reached its inherent limits.

The author argues that this pessimistic view is premature. He identifies two fundamental design shortcomings shared by virtually all modern saturation‑based provers: (1) insufficient flexibility in inference selection, and (2) lack of intelligent prioritisation of search directions. Both problems stem from the dominance of the given‑clause (clause‑selection) paradigm.

In the given‑clause approach, the prover maintains two sets of clauses: passive (waiting to be selected) and active (already selected). At each iteration a passive clause is chosen (according to a clause‑quality heuristic) and immediately combined with all active clauses, generating every possible inference. This coarse‑grained selection has several adverse effects. First, the heuristic that evaluates a clause’s quality is only loosely correlated with the quality of the inferences it will produce. A clause deemed “good” may interact with many previously selected “bad” clauses, yielding a flood of low‑quality inferences. Conversely, two “good” clauses can still generate many useless inferences if the quality metric does not penalise the presence of “bad” sub‑terms. Second, a single selected clause can trigger an explosion of inference attempts with a large active set, consuming most of the available time while a handful of other inferences might lead to a proof. The paper distinguishes two classic variants of the given‑clause algorithm: the OTTER style, which aggressively simplifies both passive and active clauses, and the DISCOUNT style, which restricts simplification to the active set. Both suffer from the same coarse‑grained control, either by over‑simplifying (potentially discarding useful clauses) or by allowing the passive set to grow unchecked.

To overcome these limitations the author proposes two architectural innovations:

1. Low‑cost, high‑precision inference selection – Instead of selecting an entire clause, the prover would evaluate the individual inferences that could be generated from a clause‑active pair. A lightweight cost model would estimate, for each potential inference, metrics such as expected clause size, literal count, symbol complexity, and redundancy likelihood. Only inferences whose estimated cost falls below a configurable threshold would be performed. This fine‑grained selection preserves the flexibility of the OTTER approach while avoiding the prohibitive overhead of the DISCOUNT approach. Implementing this requires efficient term indexing, fast unification cost estimation, and a dynamic cost‑adjustment mechanism that can be updated on‑the‑fly.

2. Intelligent prioritisation via probing of generalized search directions – The active clause set would be partitioned into a small number of abstract “search directions”. Each direction groups clauses that share structural or semantic characteristics (e.g., similar predicate symbols, similar term patterns). A meta‑heuristic layer would maintain statistics for each direction: success rate, average inference cost, depth of derivations, etc. Before committing resources, the prover would perform a shallow probing phase—limited‑depth simulations or sampling—to assess the potential of each direction. Directions with higher estimated payoff would receive higher priority, while less promising directions would be delayed or pruned. This mechanism replaces the static clause‑size or literal‑count based ordering currently used in most provers.

The paper also discusses practical implementation challenges. Building an accurate yet cheap cost model for individual inferences demands extensive profiling and possibly machine‑learning techniques. Managing generalized directions requires data structures such as direction trees or histograms, and efficient updates as new clauses are generated. Integration with existing simplification and redundancy elimination pipelines must be handled carefully to avoid contradictions (e.g., a clause removed by simplification should also be withdrawn from its direction). Finally, the author stresses the need for systematic experimental evaluation: new benchmark suites, ablation studies, and comparison against state‑of‑the‑art provers (E, Vampire, etc.) to validate the proposed ideas.

In conclusion, the paper posits that saturation‑based reasoning has not hit a theoretical ceiling; rather, its current performance plateau is a symptom of outdated architectural choices. By moving from clause‑level to inference‑level selection and by introducing a dynamic, probing‑based prioritisation of search directions, future provers could regain the rapid progress seen a decade ago. The work is presented as a position paper, inviting the community to explore these concepts experimentally and to develop the necessary tooling for their realization.