AI Paper Atlas

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

ICML 2026 · Large language models › Language, safety and interpretability · approach: Pruning and circuits in LLMs · 0 citations (OpenAlex)

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 changedWhereMeasuredEffect
add SMMD loss to cross-entropyQwen2.5-1.5B · GSM8Kaccuracy: 54.99 -> 57.77Improved
add SMMD loss to cross-entropyLlama3-8B · SVAMP (transfer)accuracy: 62.20 -> 65.30Improved
add SMMD loss to cross-entropyMinistral-3-3B · Clock-Timeaccuracy: 82.36 -> 97.56Improved
add SMMD instead of GCELlama-3.2-Vision-11B · Clock-Timeaccuracy: 82.15 (GCE) vs 80.76 (SMMD)Hurt
add SMMD to cross-entropy for chart QAQwen2.5-VL / Ministral / Llama VLMs · ChartQAaccuracy: 77.02 -> 78.38 (Qwen2.5-VL-7B)No clear effect
add MMD term to CE (ablation)SMMD · GSM8Kaccuracy: 54.97 -> 57.16Improved
add smoothness regularizer to CE (ablation)SMMD · Clock-Timeaccuracy: 46.39 -> 75.14Improved
add MMD on top of smoothness (ablation)SMMD · Clock-Timeaccuracy: 75.14 -> 74.30Hurt
add smoothness on top of MMD (ablation)SMMD · GSM8Kaccuracy: 57.16 -> 57.77Improved
use distance-induced kernel instead of random PSD kernel (ablation)MMD-only · Clock-Timeaccuracy: 62.01 -> 68.61Improved
use distance-induced kernel instead of shuffled value mapping (ablation)MMD-only · GSM8Kaccuracy: 56.17 -> 57.16Improved
use multi-kernel mixtures of bandwidths instead of single sigma=2.0 (ablation)SMMD · multiple tasksaccuracyNo clear effect

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy