Cycle-Consistent Neural Explanations: 90.0% soundness

Cycle-Consistent Neural Explanation of Formal Verification Certificates, submitted to arXiv on 23 Jun 2026 by Andoni Rodriguez, Alberto Pozanco and Daniel Borrajo, presents a trained neural architecture that generates faithful natural-language explanations of verification certificates and verifies them end-to-end. The model was evaluated on 420 test certificates from a financial compliance domain with 207 named states and achieved 90.0% cycle-verified soundness.

What did the authors build and how?

They built a cycle-consistent neural architecture with a forward network NN1 that maps certificates to explanations and an inverse network NN2 that reconstructs certificates from explanations; a symbolic verifier closes the loop and provides a differentiable faithfulness proxy. The system uses a pointer-generator mechanism to copy state names from the certificate for lexical grounding, and the authors apply a hybrid inference-time routing strategy to improve results.

The architecture is explicitly designed to enforce faithfulness: explanations are not only generated but are checked by reconstructing the original certificate and running a symbolic verifier, which yields a cycle-verified soundness metric used during training and evaluation.

How does it compare to multi-LLM baselines and what was the evaluation?

On a test set of 420 certificates spanning six verification methods, the trained architecture achieved 90.0% cycle-verified soundness, outperforming a multi-LLM few-shot baseline whose best result was 76.1% across 16 LLM combinations and four frontier models, a gap of 13.9 percentage points. The evaluation covered YES and NO verdict variants and six verification methods: bounded proof, k-induction, inductive invariant, lasso, reachability and witness pair, drawn from a financial compliance domain containing 207 named states.

The neural model wins on 10 of 12 verdict/kind categories, and in three categories it reached 100% soundness. The trained specialization also produced dramatic latency improvements: inference runs in 185 ms per certificate versus 160 s for the full multi-LLM baseline, an 860x speedup. The authors highlight that the model operates offline, produces deterministic outputs and incurs zero per-inference cost compared with cloud-based LLM prompting.

Why it matters

Non-specialist stakeholders often receive machine-checkable verification certificates they cannot read. This work targets that gap by producing explanations that are both generated and validated against the original certificate, which raises the bar for faithfulness over mere natural-language plausibility. The combination of higher cycle-verified soundness and orders-of-magnitude faster, offline inference makes the approach practical for settings where latency, determinism and per-inference cost matter, such as automated compliance workflows.

The results also argue for trained specialization: a model trained for the structured task, with a verifier in the loop, outperformed a general-purpose multi-LLM few-shot approach on the authors' dataset and metrics.

What to watch

Whether the cycle-consistent architecture generalizes beyond the paper's financial compliance dataset and its 207 named states is the key question. Future signals will include replication on other verification domains, broader coverage of verification methods, and public releases of code or datasets tied to the 420-certificate evaluation.

Submission and technical details: the paper is 15 pages of main text and is available on arXiv as arXiv:2606.24414 (submitted 23 Jun 2026).