Research Papers 论文研究 2d ago Updated 1d ago 更新于 1天前 48

ClosureBench: A Constructive Benchmark for Compositional Graph Reasoning ClosureBench:组合图推理的建构性基准

ClosureBench is a constructive benchmark for compositional graph-relational reasoning with programmatically verified ground truth via the Ein tensor-logic language, generating instances on demand to eliminate data contamination concerns The benchmark spans 26 task categories across three compositional levels (L1-L3), with difficulty controlled along three independent axes: graph size, edge density, and query depth Fine-tuning on a fixed test set creates a 19.3 percentage-point memorization gap b ClosureBench是首个按需生成实例的组合图推理基准测试,通过执行Ein张量逻辑程序验证答案,可有效检测数据污染和记忆效应 前沿模型在组合推理任务上存在根本性瓶颈:从自然语言描述中错误读取图结构后在错误图上正确推理,准确率随图规模和查询深度增加而下降 4B模型微调为生成可执行程序(Ein或Python+NetworkX)而非直接输出答案,在保留实例上达到94.3%准确率,接近前沿模型水平且成本更低

62
Hot 热度
75
Quality 质量
68
Impact 影响力

Analysis 深度分析

TL;DR

  • ClosureBench is a constructive benchmark for compositional graph-relational reasoning with programmatically verified ground truth via the Ein tensor-logic language, generating instances on demand to eliminate data contamination concerns
  • The benchmark spans 26 task categories across three compositional levels (L1-L3), with difficulty controlled along three independent axes: graph size, edge density, and query depth
  • Fine-tuning on a fixed test set creates a 19.3 percentage-point memorization gap between seen and fresh instances, a flaw invisible to static benchmarks
  • Frontier models degrade from atomic to compositional queries because they misread graphs from natural-language descriptions and then reason correctly over the wrong graph—a bottleneck inherent to reasoning, not input format
  • A 4B model fine-tuned to emit executable programs (Ein or Python+NetworkX) instead of direct answers stays nearly flat across compositional levels and reaches 94.3% accuracy on held-out instances at a fraction of the token cost

Why It Matters

This benchmark directly addresses the growing crisis of data contamination in AI evaluation by replacing static test sets with on-demand programmatic generation, enabling honest measurement of generalization versus memorization. It reveals a fundamental reasoning bottleneck in even frontier models when handling compositional graph tasks, suggesting that current architectures struggle with multi-step relational reasoning regardless of input modality. The program-synthesis approach demonstrates that smaller models can match frontier performance on complex reasoning tasks, offering a cost-effective path forward for production deployments.

Technical Details

  • Benchmark design: ClosureBench generates instances on demand using the Ein tensor-logic language for ground-truth computation, ensuring machine-verified correctness. It covers 26 task categories at three compositional levels (L1 atomic, L2 two-step, L3 multi-step), with difficulty parameterized along graph size, edge density, and query depth.
  • Evaluation scope: Models tested range from 1.5B open-weight models to frontier systems including o3, GPT-4.1, Gemini 2.5, and Claude Sonnet 4, providing a comprehensive comparison across the capability spectrum.
  • Memorization measurement: By evaluating fine-tuned models on both seen and fresh instances, the benchmark quantifies a 19.3 percentage-point accuracy gap, isolating supervised fine-tuning on answer pairs from pretraining contamination.
  • Input format robustness: The reasoning bottleneck persists across natural-language descriptions, JSON edge lists, and adjacency matrices, confirming the limitation is architectural rather than format-dependent.
  • Program synthesis approach: A 4B model fine-tuned to emit executable programs in either Ein or Python+NetworkX achieves 94.3% accuracy on held-out instances with significantly reduced token costs, demonstrating that verified program synthesis generalizes across target languages.

Industry Insight

  • Static benchmarks are becoming increasingly unreliable as model training data overlaps with popular evaluation sets; the industry should adopt constructive benchmarks with on-demand instance generation to maintain evaluation integrity.
  • The finding that reasoning bottlenecks persist regardless of input format suggests that architectural improvements—not just prompt engineering or format optimization—are needed to advance compositional graph reasoning capabilities.
  • Program synthesis fine-tuning of smaller models offers a compelling alternative to scaling frontier models for reasoning-intensive tasks, potentially reducing inference costs by orders of magnitude while maintaining competitive accuracy, making it a viable strategy for production-grade AI systems.

TL;DR

  • ClosureBench是首个按需生成实例的组合图推理基准测试,通过执行Ein张量逻辑程序验证答案,可有效检测数据污染和记忆效应
  • 前沿模型在组合推理任务上存在根本性瓶颈:从自然语言描述中错误读取图结构后在错误图上正确推理,准确率随图规模和查询深度增加而下降
  • 4B模型微调为生成可执行程序(Ein或Python+NetworkX)而非直接输出答案,在保留实例上达到94.3%准确率,接近前沿模型水平且成本更低

为什么值得看

这篇论文揭示了当前大模型在组合图推理任务上的系统性缺陷,证明问题源于推理机制而非输入格式。同时展示了程序合成方法的有效性,为开发更高效可靠的推理系统提供了新方向。

技术解析

  • 基准测试设计:包含26个任务类别和3个组合级别(L1-L3),难度通过图大小、边密度、查询深度三个独立轴控制,答案通过执行Ein程序验证正确性
  • 评估范围:从1.5B开源模型到o3、GPT-4.1、Gemini 2.5、Claude Sonnet 4等前沿系统
  • 记忆效应测量:微调模型在已见实例和新鲜实例间存在19.3个百分点的准确率差距,直接量化监督微调带来的数据污染影响
  • 程序合成方法:4B模型微调为生成可执行程序而非直接答案,在两种目标语言(Ein和Python+NetworkX)上均保持跨组合级别的稳定性能

行业启示

  • 静态基准测试已无法有效评估模型真实能力,构造性基准测试将成为未来评估标准,建议研究机构转向动态生成式评估方法
  • 程序合成是提升复杂推理任务可靠性的有效路径,小模型通过生成可执行代码可达到接近前沿模型的准确率,为资源受限场景提供可行方案
  • 当前模型在图结构解析环节存在系统性缺陷,即使输入格式从自然语言改为JSON或邻接矩阵问题依然存在,提示需要改进模型的底层表示学习能力

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

Benchmark 基准测试 Dataset 数据集 Research 科学研究 Evaluation 评测