Research map / Large language models / Language, safety and interpretability
Beyond MLE: Convex Learning for Text Generation
Chenze Shao, Zhengrui Ma, Min Zhang, Yang 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 |
|---|---|---|---|
| add convex-composition loss (exp, k=1) to Transformer | Transformer · WMT14 EN-DE greedy | BLEU: 26.48 -> 26.92 | Improved |
| add convex-composition loss to Transformer, beam5 | Transformer · WMT14 DE-EN | BLEU: 31.21 -> 31.33 | Improved |
| add convex loss: greedy-beam gap | Transformer · WMT14 EN-DE | BLEU gap greedy vs beam5: 1.09 -> 0.86 | Improved |
| add convex loss to CMLM | CMLM · WMT14 EN-DE | BLEU: 11.22 -> 20.45 | Improved |
| add convex loss to Vanilla-NAT | Vanilla-NAT · WMT14 EN-DE | BLEU: 10.41 -> 16.74 | Improved |
| add convex loss to CTC | CTC · WMT14 DE-EN | BLEU: 20.53 -> 26.68 | Improved |
| increase exponent k in CMLM from 1 to 8 (ablation) | CMLM · WMT14 EN-DE | Output NLL: 20.57 -> 4.88 | Improved |
| increase k in CMLM from 1 to 5 (fluency) (ablation) | CMLM · WMT14 EN-DE | External PPL: 939.34 -> 213.84 | Improved |
| increase k in CMLM from 5 to 8 (fluency) (ablation) | CMLM · WMT14 EN-DE | External PPL: 213.84 -> 218.68 | No clear effect |
| raise k for AR model from 1 to 3 (ablation) | Transformer + Convex · WMT14 EN-DE beam5 | BLEU: 27.78 -> 26.76 | Hurt |
| use k=0.5 for AR model (ablation) | Transformer + Convex · WMT14 EN-DE greedy | BLEU: 26.48 -> 26.89 | Improved |
| use power-function convex composition instead of exponential (ablation) | Vanilla-NAT · WMT14 EN-DE | BLEU | Hurt |
Closest papers by meaning
- Language Generation with Strictly Proper Scoring Rules (ICML 2024 · similarity 0.88)
- Error Norm Truncation: Robust Training in the Presence of Data Noise for Text Generation Models (ICLR 2024 · similarity 0.86)
- Model-Based Minimum Bayes Risk Decoding for Text Generation (ICML 2024 · similarity 0.86)
- Guaranteed Generation from Large Language Models (ICLR 2025 · similarity 0.85)
- Ensembling Language Models with Sequential Monte Carlo (NeurIPS 2026 · similarity 0.85)
- Decoding Game: On Minimax Optimality of Heuristic Text Generation Strategies (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.
