Research map / Large language models / Alignment and preferences
Optimas: Optimizing Compound AI Systems with Globally Aligned Local Rewards
Shirley Wu, Parth Sarthi, Shiyu Zhao, Aaron Lee, Herumb Shandilya, Adrian Mladenic Grobelnik, Nurendra Choudhary, Edward W Huang, Karthik Subbian, Linjun Zhang, Diyi Yang, James Zou and 1 more
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 |
|---|---|---|---|
| optimize with globally aligned local reward functions | compound AI system · HotpotQA (RAG) | F1: 33.80 -> 50.48 | Improved |
| optimize with Optimas | compound AI system · Amazon product rec. | accuracy: 21.21 -> 24.24 | Improved |
| optimize with DSPy | compound AI system · Amazon product rec. | accuracy: 21.21 -> 18.18 | Hurt |
| optimize with Optimas vs best baseline | compound AI system · PubMedQA | accuracy: 67.93 -> 69.13 | Improved |
| optimize with HBC | compound AI system · BigCodeBench (instruct) | pass rate: 36.67 -> 27.78 | Hurt |
| use learned local reward instead of LLM judge (ablation) | Optimas · Local-global alignment (average) | pairwise ranking accuracy: 49.52% -> 77.96% | Improved |
| higher alignment quality of local reward (ablation) | Optimas · component optimization | global reward | Improved |
| reduce LRF training data to 12.5% (ablation) | Optimas · HotpotQA | pairwise ranking accuracy: 65.46 (92.7% of full) | No clear effect |
| sweep retriever top-k (ablation) | local reward · HotpotQA | local vs global reward: local 0.4247 -> 0.6124; global 0.3398 -> 0.3598 (k=1 to 5) | No clear effect |
Closest papers by meaning
- Aligning Compound AI Systems via System-level DPO (NeurIPS 2025 · similarity 0.90)
- Are More LLM Calls All You Need? Towards the Scaling Properties of Compound AI Systems (NeurIPS 2024 · similarity 0.85)
- CO-Bench: Benchmarking Language Model Agents in Algorithm Search for Combinatorial Optimization (AAAI 2026 · similarity 0.84)
- MASPOB: Bandit-Based Prompt Optimization for Multi-Agent Systems with Graph Neural Networks (ICML 2026 · similarity 0.84)
- MetaBox-v2: A Unified Benchmark Platform for Meta-Black-Box Optimization (NeurIPS 2025 · similarity 0.84)
- AgentGrad: Intervention-guided Prompt Optimization for Multi Agent Systems (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.
