Research map / Large language models / Language, safety and interpretability
Words & Weights: Streamlining Multi-Turn Interactions via Co-Adaptation
Chenxing Wei, Hong Wang, Ying He, Zhongxiang Dai, Bo Jiang, Fei Yu, Yao Shu
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 |
|---|---|---|---|
| add semantic (Words) stream to weights-only ROSA (ablation) | ROSA · MATH | accuracy (%): 62.2 -> 80.8 | Improved |
| add weight updates to words-only TextGrad (ablation) | TextGrad · MATH | accuracy (%): 63.4 -> 80.8 | Improved |
| co-adapt words and weights instead of ROSA alone | ROSA · MATH | average turns: 6.3 -> 4.4 | Improved |
| co-adapt words and weights | Qwen3-8B · MATH | avg time per problem (s): 334.5 -> 297.6 | Improved |
| co-adapt words and weights | Qwen3-8B · MATH | peak memory (GB): 90.6 -> 93.7 | Hurt |
| add words and weights co-adaptation to UI-TARS-7B-SFT versus ROSA | UI-TARS-7B-SFT · OSWorld | success rate: 17.8 -> 23.6 | Improved |
| TextGrad prompt optimization alone | UI-TARS-7B-SFT · OSWorld | success rate: 13.2 -> 13.7 | No clear effect |
| TextGrad prompt optimization alone | Qwen3-0.6B-Instruct · SuperGPQA | accuracy (%): 3.8 -> 3.8 | No clear effect |
| TextGrad prompt optimization alone | Qwen2.5-0.5B-Instruct · MT-AIME24 | accuracy (%): 2.6 -> 2.2 | Hurt |
| replace Best-of-16 voting with co-adaptation | Qwen3-8B · MATH | accuracy (%): 67.8 -> 80.8 | Improved |
| add switch from prompt to parameter optimization at turn 5 (ablation) | TextGrad / ROSA conditional optimization · MATH 10-turn | accuracy | Improved |
Closest papers by meaning
- Regressing the Relative Future: Efficient Policy Optimization for Multi-turn RLHF (ICLR 2025 · similarity 0.88)
- RIZZ: Routing Interactions to Near Zero-Interference Zones for Continual Adaptation of Black-Box Agents (NeurIPS 2026 · similarity 0.87)
- ArCHer: Training Language Model Agents via Hierarchical Multi-Turn RL (ICML 2024 · similarity 0.87)
- MINT: Evaluating LLMs in Multi-turn Interaction with Tools and Language Feedback (ICLR 2024 · similarity 0.87)
- Boosting Multi-Domain Fine-Tuning of Large Language Models through Evolving Interactions between Samples (ICML 2025 · similarity 0.87)
- On Predictability of Reinforcement Learning Dynamics for Large Language Models (ICLR 2026 · similarity 0.87)
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.
