Research map / Large language models / Alignment and preferences
Antislop: A Comprehensive Framework for Identifying and Eliminating Repetitive Patterns in Language Models
Samuel J Paech, Allen G Roush, Judah Goldfeder, Ravid Shwartz-Ziv
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 |
|---|---|---|---|
| use FTPO instead of DPO (2k banlist) | gemma-3-12b · Creative writing, 2k banlist | Writing quality rubric: 62.98 -> 68.93 | Improved |
| use FTPO instead of DPO (8k banlist) | gemma-3-12b · Longform writing, 8k banlist | longform score: 26.9 -> 52.3 | Improved |
| use FTPO instead of DPO (8k banlist) | gemma-3-12b · Lexical diversity, 8k banlist | diversity: 73.92 -> 95.09 | Improved |
| use FTPO instead of DPO (2k banlist) | gemma-3-12b · Banlist suppression 2k | ban %: 82.00 -> 92.39 | Improved |
| use FTPO instead of DPO (2k banlist) | gemma-3-12b · GSM8K | accuracy: 0.847 -> 0.876 | Improved |
| use Antislop Sampler instead of token banning | gemma-3-12b · Creative writing, 8k banlist | Writing quality rubric: 28 -> above baseline | Improved |
| add margin-based clipping of preference loss to FTPO (ablation) | FTPO · FTPO hyperparameter ablation | Writing quality rubric: 19.57 -> 67.89 | Improved |
| add target MSE tether to FTPO (ablation) | FTPO · FTPO hyperparameter ablation | Writing quality rubric: 39.65 -> 67.89 | Improved |
| increase lambda_mse_target from 0.05 to 0.4 (ablation) | FTPO · FTPO hyperparameter ablation | Banlist suppression %: 84.51 -> 55.86 | Hurt |
| increase DPO beta from 0.1 to 1.0 (ablation) | DPO · Banlist 1k | Banlist suppression %: 74.7 suppression, <1% quality loss | Mixed |
| train to higher preference accuracy | DPO · gemma-3-12b 1k banlist | Writing quality rubric: degrades after 40% accuracy | Hurt |
| restrict LoRA to lm_head for Llama-3.3-70B | FTPO · Llama-3.3-70B 1k banlist | Banlist suppression %: 66.41 | Mixed |
Closest papers by meaning
- Turning Up the Heat: Min-p Sampling for Creative and Coherent LLM Outputs (ICLR 2025 · similarity 0.86)
- DALD: Improving Logits-based Detector without Logits from Black-box LLMs (NeurIPS 2024 · similarity 0.85)
- LingoLoop Attack: Trapping MLLMs via Linguistic Context and State Entrapment into Endless Loops (ICLR 2026 · similarity 0.85)
- Language Models Can Predict Their Own Behavior (NeurIPS 2025 · similarity 0.85)
- Closing the Curious Case of Neural Text Degeneration (ICLR 2024 · similarity 0.84)
- TROPT: An Open Framework for Unifying and Advancing Discrete Text Optimization (NeurIPS 2026 · similarity 0.84)
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.
