| 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
feedforward_error_propagation:
- error signal transmission
- hierarchical error propagation
- forward error flow
ubiquity:
- predictive coding everywhere
- universal prediction
- across all brain regions
Step 2: Paper Reading & Evidence Extraction
def extract_evidence_from_paper(pdf_path, glossary):
"""从论文提取证据"""
evidence = local_llm.extract(
prompt=f"Extract evidence for concepts: {glossary}",
document=pdf_path
)
return evidence
Step 3: Ontology-Constrained Prompt Assembly
def assemble_constrained_prompt(evidence, figure_desc, glossary):
"""组装约束提示词"""
prompt = f"""
Evidence: {evidence}
Figure Description: {figure_desc}
Score agreement/disagreement with each glossary concept:
- Use ONLY concepts from approved glossary
- Score: -1 (disagree), 0 (neutral), +1 (agree)
- Provide reasoning for each score
Glossary: {glossary}
"""
return prompt
Step 4: Multi-LLM Council Scoring
def multi_llm_council_score(prompt, n_models=10):
"""多 LLM 理事会评分"""
scores = []
for model_id in range(n_models):
model_output = local_models[model_id].generate(prompt)
validated_output = validate_against_glossary(
model_output, glossary
)
scores.append(validated_output)
disagreement_matrix = compute_disagreement(scores)
return scores, disagreement_matrix
Step 5: Hypothesis-Space Mapping
def map_hypothesis_space(all_study_scores):
"""映射假设空间"""
study_vectors = []
for study in all_study_scores:
vector = [
study.score_hypothesis_1,
study.score_hypothesis_2,
study.score_hypothesis_3
]
study_vectors.append(vector)
plot_3d_hypothesis_space(study_vectors)
return study_vectors
Key Findings
1. Structured Disagreement
预测编码文献中的结构化分歧:
- Local oddball paradigms:高一致性
- Global oddball paradigms:显著分歧
- 原因:实验范式差异导致证据解读分歧
2. Hypothesis-Space Temperature
温度差异揭示:
- Local oddball:假设空间紧凑(共识高)
- Global oddball:假设空间分散(分歧大)
- 应用:识别需要进一步研究的分歧点
3. Cross-Model Comparison
10 个 LLM 模型间比较:
- 模型偏见可识别
- 部分模型倾向严格评分
- 部分模型倾向宽松解读
Applications
- 跨学科文献综合:解决碎片化问题
- 假设验证:定量评估假设支持度
- 研究缺口识别:高分歧区域指向研究缺口
- 文献综述自动化:替代传统元分析
Advantages Over Traditional Meta-Analysis
| Approach | Common Comparison Space | Disagreement Measurement | Interdisciplinary |
|---|
| Traditional meta-analysis | Requires | Limited | Difficult |
| Ontology-constrained multi-LLM | Creates | Quantitative | Facilitated |
Limitations
- 本体依赖:专家本体质量影响结果
- LLM 偏见:模型训练偏见影响评分
- 计算成本:多 LLM 理事会计算昂贵
- 证据遗漏:部分证据可能未提取
Future Directions
- 本体自动化:使用 NLP 自动构建本体
- 模型校准:开发偏见校准方法
- 扩展应用:应用于其他跨学科领域(量子物理、AI 安全)
- 实时更新:动态更新假设空间
Example Use Case: Predictive Coding Neuroscience
预测编码神经科学的碎片化问题:
- 方法:计算理论、电生理、成像、行为、建模
- 传统元分析:无法统一比较
- Ontology-constrained multi-LLM:创建统一假设空间
结果:
- 31 篇研究的假设空间映射
- Local vs Global oddball 的分歧结构
- 识别预测编码假设的关键分歧点
References
- arXiv:2606.05206
- Friston K. The free-energy principle: a unified brain theory? Nat Rev Neurosci (2010)
- Bastos AM, et al. Canonical microcircuits for predictive coding. Neuron (2012)
Related Work
- kg-research-workflow:知识图谱研究工作流
- llm-decision-centric-design:决策中心 LLM 设计
- autoresearch:自动化研究循环