Research map / Probabilistic and scientific modeling / Molecules and proteins
The Shape of Addition: Geometric Structures of Arithmetic in Large Language Models
Liuyuan Wen, Xun Zhu, Lihao Huang, Wenbin Li, Yang Gao
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 dual-stream consistency check with tolerance delta=0.1 at inference | Qwen3-4B · 3-term 10-digit addition | token accuracy: 86.26% -> 89.56% | Improved |
| Use strict integer carry (delta = 0) in the consistency check (ablation) | Dual-stream consistency · 3-term 10-digit addition | token accuracy / TP correction: accuracy 89.56% -> 87.27%; TP corr 30.46% -> 44.39% | Mixed |
| Re-prompt the model instead of correcting internally | Qwen3-4B · 3-term 10-digit addition | token accuracy: 86.26% -> 79.90% | Hurt |
| Apply linear steering along carry direction | Qwen3-4B · 3-term 10-digit addition | token accuracy: 86.26% -> 88.27% | Improved |
| Use true carry instead of probed carry in dual-stream correction (ablation) | Dual-stream correction (delta = 0) · 3-term 10-digit addition | token accuracy: 86.7% -> 96.0% | Improved |
| Use true raw sum instead of probed raw sum in dual-stream correction (ablation) | Dual-stream correction (delta = 0) · 3-term 10-digit addition | token accuracy: 86.7% -> 90.5% | Improved |
| Probe for ground truth digit instead of model output digit | Logistic regression probes · Qwen3-4B final layer, p=4 | probe accuracy: 98.81% -> 94.85% | Hurt |
| Probe for correctness instead of other targets | Logistic regression probes · Qwen3-4B final layer, p=4 | probe accuracy: 98.81% -> 82.41% | Hurt |
Closest papers by meaning
- Interpreting and Improving Large Language Models in Arithmetic Calculation (ICML 2024 · similarity 0.87)
- Arithmetic Without Algorithms: Language Models Solve Math with a Bag of Heuristics (ICLR 2025 · similarity 0.86)
- Controlling Equational Reasoning in Large Language Models with Prompt Interventions (AAAI 2025 · similarity 0.86)
- Tracing the Representation Geometry of Language Models from Pretraining to Post-training (NeurIPS 2025 · similarity 0.85)
- Failure by Interference: Language Models Make Balanced Parentheses Errors When Faulty Mechanisms Overshadow Sound Ones (NeurIPS 2025 · similarity 0.85)
- Emergence of a High-Dimensional Abstraction Phase in Language Transformers (ICLR 2025 · 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.
