TCLMA: A Two-dimension Contrastive Learning based Multiagent Framework for Scientific Novelty Evaluation

Novelty assessment is a cornerstone of technological innovation and intellectual property protection, profoundly influencing national technology strategies and economic development. Traditional evaluation models, which rely on domain experts, face significant challenges, including subjectivity, inconsistency, and low efficiency. These issues are exacerbated by the surging volume of global patent applications. While generative AI technologies, particularly Large Language Models (LLMs), offer promising avenues for automation, their prevalent use of implicit reasoning often leads to "hallucinations" and weak discrimination capabilities, rendering them unsuitable for the logically rigorous task of patent novelty assessment. To address this gap, this paper introduces TCL-MFSNE (Temporal-Contrastive Learning Multi-Agent Framework with Byzantine Consensus for Scientific Novelty Evaluation), an advanced multi-agent framework designed to systematically enhance the accuracy, robustness, and interpretability of evaluations. The core of TCL-MFSNE rests on three pillars: (1) Explicit Chain-of-Thought (CoT) Reasoning, which guides the model to generate detailed, transparent reasoning steps, overcoming the black-box nature of implicit reasoning. (2) Two-Dimensional Contrastive Learning, which includes "Positive-Negative Contrastive Learning" to sharpen the model's ability to discern key novel features by learning from paired authorized and unauthorized patents, and "Temporal Contrastive Learning" to ensure evaluations are accurately contextualized within the historical technological landscape. (3) Byzantine Consensus Mechanism, which leverages multiple heterogeneous LLM agents for independent evaluations and employs a consensus algorithm to filter out individual agent errors or biases, ensuring a stable and reliable final decision. We conducted comprehensive experiments on a real-world dataset of patents filed between 2008 and 2024. The results demonstrate that the TCL-MFSNE framework achieves an accuracy and F1-score of 85%, significantly outperforming baseline models such as implicit LLMs, single-agent systems, and other ablation variants. This research not only provides a viable solution for automated patent novelty assessment but also offers a new paradigm for applying generative AI to other professional domains that demand high precision and strong logical coherence.