| name | ontology-constrained-llm-hypothesis-scoring |
| description | 本体约束多 LLM 假设评分方法论。使用专家本体(36 个概念)约束本地多 LLM 理事会,对跨学科文献(如预测编码神经科学)进行假设支持评分,生成可审计的分歧测量和定量假设空间映射。 |
| version | 1.0.0 |
| category | neuroscience |
| authors | ["Hamed Nejat","Alexander Maier","Jesse Spencer-Smith","André M. Bastos"] |
| arxiv_id | 2606.05206 |
| created | 2026-06-09T00:00:00.000Z |
| activation_keywords | ["multi-LLM","ontology constraint","hypothesis scoring","predictive coding","literature synthesis","disagreement measurement"] |
| related_skills | ["kg-research-workflow","llm-decision-centric-design"] |
Ontology-Constrained Multi-LLM Hypothesis Scoring
Overview
本体约束多 LLM 假设评分是一个本地多 LLM 管道,用于跨学科文献综合。通过专家本体约束提示词生成,验证输出,使用多 LLM 理事会评分假设支持,生成可审计的分歧测量和定量假设空间映射。
Core Methodology
1. Predictive Coding Glossary Construction
专家定义的预测编码本体:
- 36 个概念分组为 3 个假设
- Hypothesis 1:Predictive suppression(预测性抑制)
- Hypothesis 2:Feedforward error propagation(前馈误差传播)
- Hypothesis 3:Ubiquity(普适性)
2. Multi-LLM Pipeline Architecture
四阶段管道:
Stage 1: Read papers → Extract evidence
Stage 2: Incorporate figure descriptions → Assemble prompts
Stage 3: Ontology validation → Glossary compliance check
Stage 4: Multi-LLM council scoring → Agreement/disagreement quantification
3. Local Multi-LLM Council
使用 10 个本地语言模型:
- 独立评分每个研究对本体概念的同意/不同意
- 分歧来源可追溯(模型偏见、证据解读差异)
- 可审计的评分过程
4. Hypothesis-Space Mapping
几何假设空间映射:
- Pairwise study-agreement analysis(成对研究一致性分析)
- Cross-model comparison(跨模型比较)
- Three-dimensional hypothesis space(三维假设空间)
Key Innovation: Hypothesis-Space Temperature
定义:几何离散度度量,衡量研究在假设空间中的紧凑程度
def compute_hypothesis_temperature(study_vectors):
"""计算假设空间温度"""
pairwise_distances = compute_all_distances(study_vectors)
temperature = np.mean(pairwise_distances)
return temperature
发现:
- Local oddball contexts:低温(紧凑聚集)
- Global oddball contexts:高温(分散分布)
- 意义:Global oddball 研究间分歧更大
Implementation Workflow
Step 1: Glossary Definition
glossary:
predictive_suppression:
- error suppression
- prediction error reduction
- sensory suppression