NARS 推理神經符號基準

🔑 增強重點摘要

• •The NARS-Reasoning-v0.1 benchmark addresses the 'symbol grounding problem' by bridging the gap between high-level natural language semantics and the formal, non-axiomatic logic required by Non-Axiomatic Reasoning Systems (NARS).
• •The integration of OpenNARS as a validation engine allows for the automated verification of logical consistency in generated Narsese, effectively turning the benchmark into a self-correcting dataset for neuro-symbolic training.
• •By utilizing a Phi-2 LoRA adapter, the research demonstrates that small language models (SLMs) can achieve specialized symbolic reasoning capabilities with significantly lower computational overhead compared to prompting larger, general-purpose LLMs.

🛠️ 技術深入

• •Architecture: Employs a supervised fine-tuning (SFT) approach on a Phi-2 base model using Low-Rank Adaptation (LoRA) to minimize parameter updates while preserving pre-trained linguistic knowledge.
• •Pipeline: The FOL-to-Narsese compiler utilizes a deterministic mapping function that translates First-Order Logic predicates into Narsese term-logic structures, specifically handling inheritance, similarity, and implication relations.
• •Inference: The system uses a three-label classification head (True, False, Uncertain) to map Narsese truth-value functions (frequency and confidence) into discrete logical states for downstream evaluation.
• •Data Format: The benchmark dataset is structured as JSONL, containing triplets of {NL_Prompt, FOL_Representation, Narsese_Executable} to facilitate multi-modal training.

🔮 前景展望基於引用來源的 AI 分析

⏳ 時間線