The main finding is about measurement itself: an exact oracle gives not one correctness number but a range, which is called the correctness envelope, and the findings inside it are stable.
Abstract
Large language models increasingly write TLA$^{+}$ formal specifications from natural-language descriptions, but progress is hard to measure: existing resources grade by resemblance to a reference or by whether the output parses, neither of which shows correctness. We present TLA$^{+}$-Bench, a dataset and benchmark that grades by execution. Every gold specification ships a configuration the TLA$^{+}$ model checker runs over the full reachable state space, deciding exactly whether the specification holds the properties that configuration names. The dataset holds 403 model-checked gold and 897 parse-only silver specifications from 13 public repositories, subsumes prior TLA$^{+}$ generation data, and carries four model-written descriptions in two styles from two providers, with difficulty and category labels. Our main finding is about measurement itself: an exact oracle gives not one correctness number but a range. Varying only the grading choices earlier benchmarks leave unstated, on one fixed set of model outputs, the correct rate moves sixfold, from 10.0\% to 1.7\%; adding the interface-supply choice, where the model is told the configuration's names, widens the range to elevenfold, from 18.7\% to 1.7\%. We call this range the correctness envelope and measure each of its bounds. The findings inside it are stable. Every model writes valid TLA$^{+}$ far more often than correct TLA$^{+}$: the strongest is correct 16\% of the time by default and 26\% when given the interface names, open models at most 1\%, and correctness falls sharply with difficulty.
The ORCA-bench benchmark is introduced, a benchmark that puts general-purpose coding agents in a production-fidelity oncall setting and is a lower bound on the engineering investment required before frontier coding agents can be safely entrusted with production reliability.
Albert Gong, Kyuseong Choi, Abhineet Agarwal et al.· 0 citations
Can large language models write faithful formal specifications of real systems, and does it matter whether they write in a formal language they have seen rarely or in a mainstream language abundant in their training data? We study this on SysMoBench, which grades a generated specification in four phases, the decisive one replaying execution traces captured from the running system. We add JS-SAM, its first non-formal backend, in which a specification is executable JavaScript written in the SAM pattern, a pattern whose semantics mirror TLA+, and run a controlled comparison that separates three variables an ordinary head-to-head entangles: the language, the specification contract (the shape the model must fill), and the prompt. The study spans four frontier models and three systems (an operating-system spinlock, a distributed lock service, and the Etcd Raft consensus implementation), with counterexample-driven repair. Three findings emerge. First, conformance against the real system is the only phase that discriminates among models; internal consistency is inexpensive to satisfy, and a specification that looks right is not thereby right. Second, once the comparison is drawn like for like, the specification contract, not the language, governs fidelity: JavaScript in the shape of the TLA+ transition relation is as faithful as TLA+. Third, a minimal contract carries transcription but not semantic derivation: at consensus scale the difficulty becomes understanding the protocol, which no contract shape and no language supplies. We frame executable JavaScript as a checkable specification substrate that complements, rather than replaces, the verification TLA+ provides, and present the study as a case study.
ContainmentBench, a sandboxed benchmark comprising a 504-scenario specification dataset, a shared rollout-trace schema, and stage-scoped metrics for endpoint violations, logged propagation, and explicitly authorized taint-exposed proposals that commit, is introduced.
The first systematic study of imperfect specifications is presented and an automated framework to repair them to enhance the quality of resulting Verilog design is proposed, demonstrating the capabilities of specification repair by {VClare} as well as further potential of LLMs in front-end hardware design.
An automated TCS research framework that generates, formalizes, filters, and proves new claims, and further develops an automated TCS research framework that generates, formalizes, filters, and proves new claims.
Dingzirui Wang, Xuanliang Zhang, Keyan Xu et al.· 0 citations
Large Language Models (LLMs) agents deployed in vehicles must satisfy a written operating policy on every turn: a single hallucinated identifier, omitted mandatory side-effect, or premature completion claim fails the task. We present AgentGuardUtil, our entry to CAR-bench Track~1, which treats the AI planer (LLM) as a fallible proposer inside a grounded verify-and-revise loop. Its core novelty is a runtime policy compiler: the natural-language policy shipped with each conversation is compiled, once per policy, into typed machine-checkable rules, a subset of which receive an executable form. A deterministic obligation engine interprets these rules against live tool results and the simulated post-write state of the draft itself, emitting the exact remedial calls with computed arguments rather than natural-language reminders. Around this engine, 25 deterministic gates (identifier provenance, schema and enum validity, gather-before-act, confirmation and future-time protocols) and an LLM critic produce tiered findings that drive a bounded revision loop tuned for the pass k metric.
R. Bouchekir, Damir Safin, Tomas Bueno Momcilovic· 0 citations