TLA-Prover: Verifiable TLA+ Specification Synthesis via Preference-Optimized Low-Rank Adaptation

Researchers have developed TLA-Prover, a 20-billion-parameter AI model that significantly improves the synthesis of TLA+ formal specifications for distributed systems, achieving 30% correctness on verified benchmarks—roughly 3.5x better than previous baselines. The model combines supervised fine-tuning with repair-based policy optimization and uses TLC model checker feedback directly as a reward signal, eliminating the need for learned reward models.

TLA-Prover represents a meaningful advance in applying large language models to formal verification, a domain where correctness is critical and failure rates historically high. Previous approaches struggled because LLMs generated syntactically correct but semantically flawed specifications that failed verification checks. The new model addresses this through a two-stage training approach: initial supervised fine-tuning on verified examples establishes baseline capability, then group-relative policy optimization teaches the model to repair its own failed attempts using TLC feedback. This self-correction mechanism proves more effective than static training alone.

Formal specification languages like TLA+ guard against catastrophic failures in distributed systems and safety-critical protocols, yet remain difficult for LLMs to generate reliably. The tool bridges a gap between natural language reasoning and formal verification requirements, potentially accelerating development cycles for blockchain systems, payment networks, and other safety-critical applications where distributed consensus matters.

The 30% pass rate at the highest verification tier (Diamond) substantially outperforms the 8.6% baseline achieved by the best untuned models. Crucially, the researchers implemented a detection mechanism for trivial properties—specifications that always pass verification because they make no meaningful claims—preventing a common failure mode. The alignment between Gold (passes TLC) and Diamond tiers across all checkpoints suggests the model genuinely learned correctness rather than gaming the verification process.

For the cryptocurrency and blockchain space, improved TLA+ synthesis could accelerate secure protocol design and reduce audit costs. However, the 30% success rate indicates significant limitations remain before autonomous specification generation can replace expert review.

• →TLA-Prover achieves 30% correctness on formal specifications, a 3.5x improvement over previous LLM baselines
• →The model uses direct TLC model checker feedback as reward signal, eliminating the need for separate learned reward models
• →Training combines supervised fine-tuning with repair-based group-relative policy optimization to teach self-correction
• →A novel Diamond-tier verification mechanism prevents trivial specifications from falsely passing verification
• →The tool has potential applications in blockchain and distributed systems where formal verification is critical