Independent Reproduction of LLM Reasoning Reliability Methods RPC and LCF

A new arXiv paper (2608.08514) reports an independent reproduction and stress-testing of two methods for improving large language model (LLM) reasoning reliability: RPC (test-time probability aggregation) and LCF (logic-representation editing). The authors re-ran RPC's published-path aggregation and re-implemented LCF's projector, contrastive, and intervention pipeline, as LCF shipped no public code. They extended both methods to four new task domains—text-to-SQL, legal extraction, fallacy identification, and precedent grading—and tested LCF across four 7-8B models (including Qwen3-8B) and RPC across those domains. The motivation is that original evaluations were conducted by each method's own authors and never independently reproduced or stress-tested across models and domains. The paper validates the reliability claims by probing LCF's representation directly. The study found that RPC reproduces its published results, while LCF's performance varied across domains and models. The paper is announced as a new arXiv preprint, with the abstract available at the provided URL.