Research map / Probabilistic and scientific modeling / Molecules and proteins
Non-Monotonic Autoregressive Sequence Model
Tianyi Ma, Yiyue Qian, Yiyang Li, Zehong Wang, Yifan Ding, Zheyuan Zhang, Yan Liang, Chuxu Zhang, Yanfang Ye
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 |
|---|---|---|---|
| mask error tokens in SFT loss (mSFT) instead of standard SFT (ablation) | N-MARS · GSM8K, Llama-3.2-1B | accuracy: 7.8 -> 19.3 | Improved |
| standard SFT on error-augmented data vs few-shot base (ablation) | Llama-3.2-1B · GSM8K | accuracy: 8.8 -> 7.8 | Hurt |
| mSFT instead of SFT (ablation) | N-MARS · MATH-500, Llama-3.1-8B | accuracy: 21.1 -> 30.9 | Improved |
| add GRPO with multi-objective reward after mSFT (ablation) | N-MARS · GSM8K, Llama-3.2-1B | accuracy: 19.3 -> 31.3 | Improved |
| add GRPO after SFT that showed negative learning (ablation) | SFT model · GSM8K, Llama-3.2-1B | accuracy: 7.8 -> 23.9 | Improved |
| add outcome reward Rinc (ablation) | N-MARS · GSM8K, Llama-3.2-1B | accuracy: 28.3 -> 31.3 | Improved |
| add UNDO usage reward/penalty Rpen (ablation) | N-MARS · MATH-500, Llama-3.2-1B | accuracy: 13.5 -> 15.2 | Improved |
| use semantic anchor initialization for UNDO token instead of centroid (ablation) | N-MARS · MATH-500, Llama-3.2-1B | accuracy: 14.1 -> 15.2 | Improved |
| use augmentation from real model errors instead of random error injection (ablation) | N-MARS · MATH-500, Llama-3.2-1B | accuracy: 13.2 -> 15.2 | Improved |
| alter positional encodings so corrections inherit error positions (ablation) | N-MARS · GSM8K, Qwen3-4B-Base | accuracy: 86.1 -> 81.4 | Hurt |
| use explicit UNDO token instead of verbal Self-Reflect | N-MARS · MATH-500, Llama-3.2-1B | accuracy: 11.6 -> 15.2 | Improved |
| increase UNDO budget kappa beyond 0.2 on GSM8K (ablation) | N-MARS · GSM8K | accuracy | Hurt |
Closest papers by meaning
- MARS: Enabling Autoregressive Models Multi-Token Generation (NeurIPS 2026 · similarity 0.88)
- SequenceMatch: Imitation Learning for Autoregressive Sequence Modelling with Backtracking (ICLR 2024 · similarity 0.88)
- Autoregressive Learning in Joint KL: Sharp Oracle Bounds and Lower Bounds (NeurIPS 2026 · similarity 0.87)
- Architectural and Inferential Inductive Biases for Exchangeable Sequence Modeling (NeurIPS 2025 · similarity 0.87)
- Beyond Autoregression: Fast LLMs via Self-Distillation Through Time (ICLR 2025 · similarity 0.87)
- Controllable Generation via Locally Constrained Resampling (ICLR 2025 · similarity 0.86)
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.
