Skolemizing Verifiers as Hidden Attractors in Self-Trained Logical Reasoning

When language models are trained on their own generated reasoning, a checker usually decides which samples are kept. For logic and theorem-style tasks that checker is often a symbolic prover. This proposal asks whether such a prover can quietly cap how good the model gets. The concern is a specific, decades-old weakness in how provers handle "there exists" statements. They replace "some bird does not fly" with a made-up name, a Skolem constant (a placeholder symbol standing for the unnamed individual). Doing so throws away the condition that defined that individual. Once the condition is gone, no syntactic trick reliably recovers it, as Variables in Hypotheses shows. We hypothesize that this lost context makes a symbolic checker systematically wrong on a predictable slice of problems. Under repeated generate-check-retrain cycles, the model then converges to the checker's mistakes instead of to correct reasoning.

The claim connects two literatures. Work on synthetic data saturation has proven that retraining on checked data drives a model toward the checker's own "knowledge center," a bias-determined plateau, but it models the checker as an abstract ball in parameter space (Escaping Model Collapse via Synthetic Data Verification). Work on Skolemization has characterized exactly where symbolic checks go wrong on existential facts, but it has never asked what those errors do inside a training loop. Putting the two together makes the knowledge center concrete: it is the set of problems on which a Skolemizing checker disagrees with the true logical answer. That set can be computed in advance, and its effect on the training plateau can be predicted and then tested.

Core Claim

In self-training loops for first-order abductive reasoning, the plateau is set by the checker's disagreement with ground truth on the problems that involve existential facts. Abductive reasoning here means proposing which general rules, assumed by default, explain an observation. The dataset scale at which the plateau appears matters less than that disagreement. Replacing Skolem placeholders with Hilbert epsilon terms removes most of the disagreement and lifts the plateau. An epsilon term, written $\epsilon x.P(x)$ and read "an $x$ such that $P(x)$," is a name that carries its defining property inside it. Neither data diversity nor per-step preference learning achieves the same lift. Three things are new: treating checker logic as the mechanism behind saturation, measuring the checker's error slice before training, and testing whether that slice predicts where training stops.

Stated loosely, let $a_s$ be the checker's agreement with an exact oracle on problem slice $s$ and $w_s$ the slice's share of the test distribution. Yi et al.'s convergence result, carried over from a parameter ball to discrete accept/reject labels, suggests the late-round accuracy

$$\text{Acc}_\infty \approx \sum_s w_s \, a_s .$$

If self-training on correct-checked data can drive accuracy near ceiling on slices where the checker is right, the loop should stall at roughly this value. That holds whatever the volume of synthetic data or the checker's strictness. This is a hypothesis to falsify, not a theorem.

Research Program

Testbed. We first build a generator of abductive theory-formation problems in the style of Poole's setting. Each instance has facts $F$, which mix named individuals with existential facts about unnamed ones. It also has default universal hypotheses $\Delta$, such as "birds fly," and a goal $g$. The student must output a set of hypothesis instances $D$ such that $F \cup D$ entails $g$ and $F \cup D$ is consistent. Everything is rendered in templated natural language, as in ProofWriter. ProofWriter's abduction, however, only recovers one missing ground fact and never meets unnamed individuals or variable-bearing hypotheses. Our generator is built to exercise exactly those.

We restrict the main split to monadic first-order logic, where every predicate takes one argument. This fragment is decidable and has the finite-model property, so an exact oracle exists: it searches for small finite models. A second split adds binary relations with "for all, there exists" patterns. There the oracle is only a semi-decision procedure with timeouts, which tests whether the findings survive outside the clean fragment. Every instance is tagged by whether its correct answer depends on how hypotheses interact with an existential fact. These are the slices $s$.

Checkers. We implement four checkers that differ only in how they handle existentials. The Skolem-ground checker follows the common pipeline: Skolemize, ground the used instances, and check consistency over the resulting terms. Poole's Theorem 1 shows this Herbrand-universe check is inadequate when hypotheses contain variables. The pattern-matching reverse-Skolem checker tries to map Skolem constants back to quantifiers syntactically. Poole's Theorem 2 shows no such procedure is correct, and it can produce spurious contradictions. The whole-universal checker tests consistency of the unrestricted default. It is sound but over-conservative, because a single existential exception rejects the default for every individual. The epsilon checker implements the explanation procedure from Poole and 70) Variables in Hypotheses. It replaces Skolem functions with variables and identifies inconsistent instances through epsilon terms, so the exception is tracked as "the $x$ that is a non-flying bird" rather than as an opaque constant.

We do not assume in advance which way each checker errs. Before any training, we measure each checker's full confusion matrix against the oracle on every slice. That yields the $a_s$ that make the plateau prediction checkable. The over-conservative checker matters for a second reason: its errors are false rejections, which prune valid explanations. That is the diversity-shrinking form of saturation, as opposed to the false-acceptance form.

Loop. A small open language model, initialized with a limited seed of oracle-labeled examples, runs the generate-check-retrain loop for 20 to 30 rounds under each checker. Both rejection fine-tuning and per-step preference learning are run on the checked samples. We track accuracy against the oracle per slice and per round, the share of checker-accepted samples that are wrong, and the diversity of accepted explanations.

Baselines

The naive reference is unfiltered self-training on the model's own outputs. Yi et al. find this collapses (an MNIST VAE degenerates to noise within 40 rounds) or fluctuates without gain (XSUM summarization with SmolLM2-135M). If our checked loops do not clearly beat this, the testbed is broken.

The upper reference is the oracle checker in the same loop. It shows what self-training can reach when checking is exact. The epsilon checker's job is to come close to it.

The strongest existing method is the Skolem-ground prover-as-checker loop. It mirrors how LINC uses Prover9 as the trusted judge and how Mixture-of-Thought retrains on traces that pass verification filters. We pair it with the best training objective in the file, per-step preference optimization at the "first pit" from RL on Incorrect Synthetic Data. That paper reports an 8x gain in scaling efficiency over positive-only training on GSM8K and MATH. On our own testbed there is no reported result found on Lacuna for any method. The closest reported numbers are ProofWriter's abduction accuracy of over 85% on single-missing-fact problems and, on FOLIO, Logic-LM's 78.1%, which Mixture-of-Thought matches with Gemma-9B (FOLIO).

For fairness, all checker conditions share the same seed data, generator, student model, number of rounds, samples per round, and hyperparameter search budget. The search is tuned once on the oracle loop and then frozen. Wall-clock checking cost is reported separately.

The key ablation isolates the epsilon ingredient. We give the Skolem-ground checker a context table that records, for each Skolem constant, the formula it came from, but the checker does not use that formula in consistency checks. If this table alone closes the gap, the benefit comes from bookkeeping rather than from epsilon semantics.

Two results would count as beating the strongest baseline. First, the epsilon loop must exceed the Skolem-ground loop on the existential-interaction slice by more than seed-to-seed variance. Second, the Skolem-ground loop's plateau must fall within the band predicted from its pre-training $a_s$ values.

Why not just…?

Why not just propositionalize over a finite domain, as Mixture-of-Thought does? Grounding over the named individuals assumes a closed world. Poole's Theorem 1 says checking over the Herbrand universe is inadequate exactly when variable-bearing hypotheses meet existential facts. Grounding is in effect our Skolem-ground checker, so the experiment measures its cost directly. A finite-model check that also adds fresh unnamed elements is correct on the monadic fragment, and we use it as the oracle. The real question is therefore whether the epsilon checker matches it more cheaply and still works on the binary split, where finite search has no completeness guarantee. We report checking cost against domain size for both. If finite search stays cheap and accurate on the binary split, the practical case for epsilon terms narrows to explanation quality, and we will say so.

Why not just call a complete first-order prover on the full, ungrounded theory? This is the most serious threat. A full refutation prover is sound, and it will detect that "all birds fly" contradicts "some bird does not fly." What it cannot do is say which instances of the default remain safe to assume, so it rejects every explanation that relies on the default. That is our whole-universal checker. We predict its loop shows the false-rejection form of saturation: high precision, a shrinking set of accepted explanations, and a plateau below the oracle on the existential slice. If this checker instead matches the oracle, the Skolemization story is wrong for this task, and the experiment will show it.

Why not just add diversity or harder problems, as Simula or CHIMERA do? Simula widens coverage with taxonomies, and CHIMERA concentrates on hard problems with long reasoning chains, reaching 70.1% on GPQA-Diamond with a 4B model. Neither changes who decides correctness. Yi et al. show that changing checker strictness changes how fast the loop converges but not where it ends up. We test this directly by doubling generator diversity under the Skolem-ground checker. The theory predicts the plateau stays put while the error mix shifts.

Why not just use first-pit preference learning? Setlur et al. rely on a correct final-answer checker to assign credit to individual steps. With a biased checker, the first pit is located wherongly: valid steps get penalized and invalid ones rewarded on the error slice. We cross every training objective with every checker. If per-step learning under the Skolem-ground checker still reaches the oracle plateau, our claim fails.

Decisive Experiments

Four checks settle the claim, and a fifth tests transfer. First, the pre-training agreement values $a_s$ must predict each loop's late-round accuracy. Second, the gap between the Skolem-ground loop and the oracle loop must sit mostly in the existential-interaction slice, not spread evenly across slices. Third, a checker swap from Skolem-ground to epsilon at round 10 should move the plateau toward the oracle within a few rounds, which is the signature of an attractor rather than of permanent damage. Fourth, the benefit should hold on the binary split despite oracle timeouts. Finally, models self-trained under each checker are evaluated on FOLIO and on ProofWriter's out-of-distribution Birds-Electricity theories. FOLIO's 78.1% Logic-LM figure is a reference point, not a target we expect to match, since our training data is templated. The outcome to watch is whether the gap between checkers survives the move to expert-written text.

Two outcomes would refute the proposal. One is that all checkers plateau at the same level. The other is that each plateau is uncorrelated with its checker's measured disagreement. Either would show that saturation in this setting has nothing to do with checker logic.