Research map / Large language models / Alignment and preferences
Textual Self-Attention Network: Test-Time Preference Optimization Through Textual Gradient-Based Attention
Shibing Mo, Haoyang Ruan, Kai Wu, Jing Liu
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 |
|---|---|---|---|
| apply TSAN (k4-M4-T3) test-time optimization | Llama-3.1-70B-SFT · AlpacaEval 2 | LC win rate (%): 3.01 -> 18.57 | Improved |
| apply TSAN test-time optimization | Llama-3.1-70B-SFT · MATH-500 | accuracy (%): 22.0 -> 28.2 | Improved |
| apply TSAN test-time optimization | Llama-3.1-70B-Instruct · MATH-500 | accuracy (%): 24.0 -> 38.0 | Improved |
| apply TSAN test-time optimization | Llama-3.1-70B-Instruct · XSTest | compliance accuracy (%): 69.2 -> 81.7 | Improved |
| apply TSAN test-time optimization | Mistral-Small-Instruct-2409 · AlpacaEval 2 | raw win rate (%): 16.10 -> 25.38 | Improved |
| apply TSAN test-time optimization | Qwen-3-Plus · AlpacaEval 2 | LC win rate (%): 65.08 -> 57.56 | Hurt |
| apply TSAN test-time optimization | Qwen-3-Plus · Arena-Hard 2 | win rate (%): 47.3 -> 72.1 | Improved |
| replace TPO with TSAN (T5) | gpt-oss 20B · Arena-Hard 2 | win rate (%): 75.2 -> 82.8 | Improved |
| replace TPO with TSAN (T5) | gpt-oss 20B · MATH-500 | accuracy (%): 95.0 -> 97.6 | Improved |
| replace TPO with TSAN | Llama-3.1-70B-SFT · MATH-500 | accuracy (%): 32.0 -> 28.2 | Hurt |
| replace TPO with TSAN | Llama-3.1-70B-SFT · AlpacaEval 2 | raw win rate (%): 22.15 -> 17.05 | Hurt |
| increase iterations from T3 to T5 (ablation) | TSAN on Llama-3.1-70B-Instruct · Arena-Hard 2 | win rate (%): 8.1 -> 10.4 | Improved |
Closest papers by meaning
- Test-Time Preference Optimization: On-the-Fly Alignment via Iterative Textual Feedback (ICML 2025 · similarity 0.90)
- Self-alignment of Large Video Language Models with Refined Regularized Preference Optimization (NeurIPS 2025 · similarity 0.87)
- Selective Attention: Enhancing Transformer through Principled Context Control (NeurIPS 2024 · similarity 0.87)
- Reward-Augmented Data Enhances Direct Preference Alignment of LLMs (ICML 2025 · similarity 0.87)
- Token-level Direct Preference Optimization (ICML 2024 · similarity 0.87)
- Latent Preference Coding: Aligning Large Language Models via Discrete Latent Codes (ICML 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.
