AI Paper Atlas

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

AAAI 2026 · Large language models › Alignment and preferences · approach: Active and token-level preference learning · 0 citations (OpenAlex)

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 changedWhereMeasuredEffect
apply TSAN (k4-M4-T3) test-time optimizationLlama-3.1-70B-SFT · AlpacaEval 2LC win rate (%): 3.01 -> 18.57Improved
apply TSAN test-time optimizationLlama-3.1-70B-SFT · MATH-500accuracy (%): 22.0 -> 28.2Improved
apply TSAN test-time optimizationLlama-3.1-70B-Instruct · MATH-500accuracy (%): 24.0 -> 38.0Improved
apply TSAN test-time optimizationLlama-3.1-70B-Instruct · XSTestcompliance accuracy (%): 69.2 -> 81.7Improved
apply TSAN test-time optimizationMistral-Small-Instruct-2409 · AlpacaEval 2raw win rate (%): 16.10 -> 25.38Improved
apply TSAN test-time optimizationQwen-3-Plus · AlpacaEval 2LC win rate (%): 65.08 -> 57.56Hurt
apply TSAN test-time optimizationQwen-3-Plus · Arena-Hard 2win rate (%): 47.3 -> 72.1Improved
replace TPO with TSAN (T5)gpt-oss 20B · Arena-Hard 2win rate (%): 75.2 -> 82.8Improved
replace TPO with TSAN (T5)gpt-oss 20B · MATH-500accuracy (%): 95.0 -> 97.6Improved
replace TPO with TSANLlama-3.1-70B-SFT · MATH-500accuracy (%): 32.0 -> 28.2Hurt
replace TPO with TSANLlama-3.1-70B-SFT · AlpacaEval 2raw win rate (%): 22.15 -> 17.05Hurt
increase iterations from T3 to T5 (ablation)TSAN on Llama-3.1-70B-Instruct · Arena-Hard 2win rate (%): 8.1 -> 10.4Improved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy