Research map / Large language models / Language, safety and interpretability
An Information-Theoretic Parameter-Free Bayesian Framework for Probing Labeled Dependency Trees from Attention Score
Hongxu Liu, Jing Ma, Xiaojie Wang, Caixia Yuan, Fangxiang Feng
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use posterior-based tree reconstruction instead of raw score | IPBP · UD 2.9 English | UAS: 32.3 -> 49.1 | Improved |
| use posterior-based tree reconstruction instead of raw score | IPBP · UD 2.9 English | LAS: 16.6 -> 30.6 | Improved |
| add positive MI (MIpos) mixed with binary MI (ablation) | IPBP · UD 2.9 English | LAS: 30.6 -> 34.8 | Improved |
| add positive MI (MIpos) mixed with binary MI (ablation) | IPBP · UD 2.9 English | UAS: 49.1 -> 49.9 | Improved |
| probe transposed attention instead of original orientation (ablation) | IPBP · UD 2.9 English | UAS: 49.1 -> 42.6 | Hurt |
| probe undirected trees instead of directed (ablation) | IPBP · UD 2.9 English | UAS: 49.1 -> 45.3 | Hurt |
| probe labeled trees directly instead of arc-first only (arc only baseline) (ablation) | IPBP · UD 2.9 English | UAS: 36.5 -> 49.1 | Improved |
| use IPBP instead of V-Information supervised probe | IPBP · UD 2.9 English | LAS: 20.9 -> 30.6 | Improved |
| KDE bandwidth 0.5x instead of rule-of-thumb (ablation) | IPBP · UD 2.9 English | UAS: 49.1 -> 49.2 | No clear effect |
| use IPBP head selection instead of ElasticNet/IoU/Probeless | IPBP · head selection consistency | Spearman-R: 0.309 -> 0.398 | Improved |
Closest papers by meaning
- A Polar coordinate system represents syntax in large language models (NeurIPS 2024 · similarity 0.86)
- Dependency Parsing is More Parameter-Efficient with Normalization (NeurIPS 2025 · similarity 0.86)
- Unveiling and Harnessing Hidden Attention Sinks: Enhancing Large Language Models without Training through Attention Calibration (ICML 2024 · similarity 0.86)
- Learning the Wrong Lessons: Syntactic-Domain Spurious Correlations in Language Models (NeurIPS 2025 · similarity 0.86)
- Order-Level Attention Similarity Across Language Models: A Latent Commonality (NeurIPS 2025 · similarity 0.85)
- Syntax vs. Semantics: How Transformers Learn Deep Dependencies (ICML 2026 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
