Research map / Label-efficient and robust learning / Recognition and long-tail
Annotations Mitigate Post-Training Mode Collapse
Jacob Mitchell Springer, Madhu Advani, Lukas Aichberger, Arwen Bradley, Eran Malach, Omid Saremi, Sinead Williamson, Preetum Nakkiran, Etai Littwin, Aditi Raghunathan
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 |
|---|---|---|---|
| annotation-anchored training instead of standard SFT | 2.5B model pretrained and post-trained on Tulu3 · STORIES | semantic entropy: 6.88 -> 8.30 | Improved |
| add annotated pretraining before annotation-anchored post-training (ablation) | annotation-anchored 2.5B model · STORIES | semantic entropy: 7.07 -> 8.30 | Improved |
| sample annotations at inference time (ablation) | annotation-anchored 2.5B model · STORIES | semantic entropy: 7.01 -> 8.30 | Improved |
| increase model size under standard SFT | Qwen2.5 and Llama 3 instruction-tuned models · STORIES | semantic entropy | Hurt |
| increase number of SFT post-training examples (18K to 464K) | standard PT + SFT · STORIES | semantic entropy | Hurt |
| prompt for diversity (brainstorm, multiple responses) instead of direct prompting | instruction-tuned Qwen2.5 / Llama 3 · STORIES | semantic entropy | Mixed |
| use annotation-anchored training instead of standard SFT | 2.5B model · GSM8K | accuracy: 35.4% -> 36.4% | No clear effect |
| use annotation-anchored training instead of standard SFT | 0.6B, 1B, 2.5B models · Zero-shot suite average | average accuracy: 62.3 vs 62.4 at 2.5B | No clear effect |
| use annotation-anchored training instead of SFT across temperatures | standard PT + SFT · STORIES diversity-quality | diversity at given quality | Improved |
| narrow post-training data entropy, with annotation anchoring vs standard SFT | 2.5B models on SimpleStories · SimpleStories topic/persona entropy | output entropy | Improved |
Closest papers by meaning
- Steering Information Utility in Key-Value Memory for Language Model Post-Training (NeurIPS 2025 · similarity 0.87)
- IDEAL: Influence-Driven Selective Annotations Empower In-Context Learners in Large Language Models (ICLR 2024 · similarity 0.86)
- A Closer Look at Model Collapse: From a Generalization-to-Memorization Perspective (NeurIPS 2025 · similarity 0.86)
- Boosting Multi-Domain Fine-Tuning of Large Language Models through Evolving Interactions between Samples (ICML 2025 · similarity 0.86)
- Reinforced Fast Weights via Next-Sequence Prediction (NeurIPS 2026 · similarity 0.86)
- An Empirical Study on Noisy Data and LLM Pretraining Loss Divergence (NeurIPS 2026 · 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.
