Research map / Large language models / Language, safety and interpretability
Enhancing Numerical Prediction in LLMs via Smooth MMD Alignment
Zhuo Zuo, Li Yue, Wenhao Zheng, Chenpeng Wang, Xianggen Liu
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 SMMD loss to cross-entropy | Qwen2.5-1.5B · GSM8K | accuracy: 54.99 -> 57.77 | Improved |
| add SMMD loss to cross-entropy | Llama3-8B · SVAMP (transfer) | accuracy: 62.20 -> 65.30 | Improved |
| add SMMD loss to cross-entropy | Ministral-3-3B · Clock-Time | accuracy: 82.36 -> 97.56 | Improved |
| add SMMD instead of GCE | Llama-3.2-Vision-11B · Clock-Time | accuracy: 82.15 (GCE) vs 80.76 (SMMD) | Hurt |
| add SMMD to cross-entropy for chart QA | Qwen2.5-VL / Ministral / Llama VLMs · ChartQA | accuracy: 77.02 -> 78.38 (Qwen2.5-VL-7B) | No clear effect |
| add MMD term to CE (ablation) | SMMD · GSM8K | accuracy: 54.97 -> 57.16 | Improved |
| add smoothness regularizer to CE (ablation) | SMMD · Clock-Time | accuracy: 46.39 -> 75.14 | Improved |
| add MMD on top of smoothness (ablation) | SMMD · Clock-Time | accuracy: 75.14 -> 74.30 | Hurt |
| add smoothness on top of MMD (ablation) | SMMD · GSM8K | accuracy: 57.16 -> 57.77 | Improved |
| use distance-induced kernel instead of random PSD kernel (ablation) | MMD-only · Clock-Time | accuracy: 62.01 -> 68.61 | Improved |
| use distance-induced kernel instead of shuffled value mapping (ablation) | MMD-only · GSM8K | accuracy: 56.17 -> 57.16 | Improved |
| use multi-kernel mixtures of bandwidths instead of single sigma=2.0 (ablation) | SMMD · multiple tasks | accuracy | No clear effect |
Closest papers by meaning
- Learning from Students: Applying t-Distributions to Explore Accurate and Efficient Formats for LLMs (ICML 2024 · similarity 0.88)
- Regress, Don't Guess: A Regression-like Loss on Number Tokens for Language Models (ICML 2025 · similarity 0.88)
- Kernel Language Entropy: Fine-grained Uncertainty Quantification for LLMs from Semantic Similarities (NeurIPS 2024 · similarity 0.87)
- Understanding and Mitigating Numerical Sources of Nondeterminism in LLM Inference (NeurIPS 2025 · similarity 0.87)
- Any-Precision LLM: Low-Cost Deployment of Multiple, Different-Sized LLMs (ICML 2024 · similarity 0.87)
- A Probabilistic Perspective on Unlearning and Alignment for Large Language Models (ICLR 2025 · similarity 0.87)
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.
