Research map / Large language models / Language, safety and interpretability
Text Has Curvature
Karish Grover, Hanqing Zeng, Yinglong Xia, Christos Faloutsos, Geoffrey J. Gordon
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 |
|---|---|---|---|
| add Texture curvature re-weighting to BM25 span selection | BM25 pruning, Llama-3.1-8B-Instruct · HotpotQA (LongBench, B=2048) | F1: 28.1 -> 38.4 | Improved |
| add Texture curvature re-weighting to BM25 span selection | BM25 pruning, Llama-3.1-8B-Instruct · 2WikiMultiHopQA (LongBench, B=2048) | F1: 9.4 -> 18.2 | Improved |
| add Texture curvature re-weighting to BM25 span selection | BM25 pruning · Qasper, GovReport, QMSum (LongBench) | F1 / ROUGE-L: 8.4 -> 7.9 (Qasper); 19.8 -> 20.3 (GovReport); 15.1 -> 14.9 (QMSum) | No clear effect |
| use curvature-only span scoring without BM25 (CurvPrune-Pure) (ablation) | CurvPrune · HotpotQA (LongBench, B=2048) | F1: 32.5 -> 14.8 | Hurt |
| combine curvature with BM25 (CurvPrune-Hybrid) instead of curvature alone (ablation) | CurvPrune · 2WikiMultiHopQA (LongBench, B=2048) | F1: 14.6 -> 15.0 | No clear effect |
| replace Self-Route with CurvFlag curvature-triggered routing | retrieval routing · HotpotQA | F1 and tokens: F1 45.4 -> 49.1; tokens 7096 -> 3344 | Improved |
| replace Self-Route with CurvFlag curvature-triggered routing | retrieval routing · 2WikiMultiHopQA | F1: 37.7 -> 43.0 | Improved |
| add Texture trigger to FLARE | FLARE · HotpotQA and 2WikiMQA | F1: 39.1 -> 45.6 (HotpotQA); 38.2 -> 41.5 (2Wiki) | Improved |
| add Texture trigger to Self-Route | Self-Route · HotpotQA and 2WikiMQA | F1 and tokens: 45.4 -> 47.6 (HotpotQA); 37.7 -> 40.8 (2Wiki); tokens 7096 -> 256 | Improved |
| use natural text instead of suffix-swap control (ablation) | holonomy certificate · WikiText-2 | median holonomy magnitude: 0.305 -> 0.596 | Improved |
| use natural text instead of suffix-swap control (ablation) | CEI certificate · WikiText-2 | median CEI (nats): 0.697 -> 1.514 | Improved |
| compare relational (transversal) curvature to longitudinal curvature | Texture · 200 WikiText-2 slots | Spearman correlation: Spearman rho ~0.20 | No clear effect |
Closest papers by meaning
- From Topology to Retrieval: Decoding Embedding Spaces with Unified Signatures (NeurIPS 2026 · similarity 0.86)
- Textural or Textual: How Vision-Language Models Read Text in Images (ICML 2025 · similarity 0.85)
- HELM: Hyperbolic Large Language Models via Mixture-of-Curvature Experts (NeurIPS 2025 · similarity 0.85)
- Correlation Dimension of Autoregressive Large Language Models (NeurIPS 2025 · similarity 0.84)
- Invariant Features in Language Models: Geometric Characterization and Model Attribution (NeurIPS 2026 · similarity 0.84)
- The Double-Ellipsoid Geometry of CLIP (ICML 2025 · similarity 0.84)
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.
