Research map / Large language models / Alignment and preferences
TAMPO: Task- and Model-Aware Automatic Prompt Optimization for Auto-Routing in LLM-based Systems
Yuejun Jiao, Jun Xia, Yanxin Yang, Yonghao Yang, Hao Shen, Mingsong Chen
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 |
|---|---|---|---|
| replace online iterative APO (APE/OPRO) with a single-pass conditional policy | TAMPO · reasoning tasks | latency (ms): 2098 (APE) -> 285 (latency-priority TAMPO) | Improved |
| condition prompt generation on model metadata (task- and model-aware) instead of task-only | RLPrompt vs TAMPO · reasoning tasks | worst-case accuracy (%): 37.4 (RLPrompt) -> 39.7 (TAMPO quality) | Improved |
| use preference vector to prioritise quality | TAMPO · reasoning tasks | average accuracy (%): 40.0 (Static) -> 48.4 | Improved |
| prioritise token efficiency via preference vector | TAMPO · reasoning tasks | tokens: 131 (Static) -> 21 | Improved |
| use model-aware reward model instead of model-agnostic reward model (ablation) | uncertainty-aware reward model · heterogeneity-aware dataset | Kendall's tau / NDCG | Improved |
| add Gaussian NLL uncertainty term and epistemic penalty to an MSE-only reward model (ablation) | reward model · MMLU degraded prompt | final reward vs actual accuracy 0.12: 0.85 -> 0.37 | Improved |
| generalise to unseen task and unseen model | TAMPO · AQuA + Gemma-1B | accuracy (%): 26.7 -> 30.6 (vs Static) | Improved |
| quantise models from FP16 to INT8 | TAMPO quality-focused · heterogeneous models | accuracy | No clear effect |
| use few-shot prompting instead of zero-shot | Static · reasoning tasks | average accuracy (%): 40.0 -> 36.8 | Hurt |
Later papers that compared with this method
1 later accepted paper reports TAMPO as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- TGRL: Temperature-Grouped Reinforcement Learning for Efficient Exploration in LLMs (NeurIPS 2026 · better on 1 of 1 shared result)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- MASPO: Joint Prompt Optimization for LLM-based Multi-Agent Systems (ICML 2026 · similarity 0.86)
- Cost-Aware Contrastive Routing for LLMs (NeurIPS 2025 · similarity 0.86)
- Router-R1: Teaching LLMs Multi-Round Routing and Aggregation via Reinforcement Learning (NeurIPS 2025 · similarity 0.85)
- Automated Reformulation of Robust Optimization via Memory-Augmented Large Language Models (NeurIPS 2026 · similarity 0.85)
- Unleashing the Potential of Large Language Models as Prompt Optimizers: Analogical Analysis with Gradient-based Model Optimizers (AAAI 2025 · similarity 0.85)
- MESS+: Dynamically Learned Inference-Time LLM Routing in Model Zoos with Service Level Guarantees (NeurIPS 2025 · 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.
