Research map / Probabilistic and scientific modeling / Causal discovery
AgentCBO: Causal Belief Routing for Bayesian Optimization under Unknown Graphs
Qiyu Wei, Xin Zhang, Richard Allmendinger, Mauricio A Álvarez
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes AgentCBO, a causal Bayesian optimization method for unknown causal graphs that routes sampling toward structural uncertainty relevant to the optimizer rather than full graph recovery.
Problem
Causal Bayesian optimization usually assumes a known causal graph; existing relaxations couple optimization with structure learning but waste sample budget on structural uncertainty irrelevant to the optimizer.
Approach
Introduces routing-equivalence, shifting the objective from full causal-graph recovery to identifying the parts of the causal structure that matter for routing optimization decisions, with a belief-based routing mechanism for BO under unknown graphs.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add LLM prior (Language-Prior Agent) instead of PC-only discovery (ablation) | AgentCBO · Rosenbrock | best objective Y*: -89.80 -> -90.98 | Improved |
| add LLM prior instead of PC-only discovery (ablation) | AgentCBO · CompleteGraph | best objective Y*: -2.13 -> -2.50 | Improved |
| add LLM prior instead of PC-only discovery (ablation) | AgentCBO · PISHCATGraph | best objective Y*: +0.28 -> +0.30 | No clear effect |
| add calibration and repair agents on top of the prior (ablation) | AgentCBO · CompleteGraph | best objective Y*: -2.50 -> -3.57 | Improved |
| add calibration and repair agents on top of the prior (ablation) | AgentCBO · ten-benchmark suite | best objective Y*: p=0.84 (Calib), p=0.83 (Full) | No clear effect |
| anonymize or shuffle variable names given to the LLM (ablation) | Language-Prior Agent · ten-benchmark suite | best objective Y* | Mixed |
| bypass the deterministic commit gate under a corrupted LLM update (ablation) | AgentCBO · PSAGraph stress test | endpoint degradation: 0 -> +0.1387 | Hurt |
| restrict exploration sets to fixed {B},{D},{B,D} family vs sets containing E (ablation) | AgentCBO ES family · CompleteGraph | low-objective hits: 14/20 vs 0/20 | Improved |
| use random pair-induced ES family instead of fixed BD family (ablation) | AgentCBO ES family · CompleteGraph | low-objective hits: 14/20 -> 4/20 | Hurt |
| delete one edge on a manipulative-to-Y path (ablation) | AgentCBO with ground-truth graph · Rosenbrock | best objective Y*: -90.97 -> -44.78 | Hurt |
| use LLM prior instead of PC-only when coverage rescue is disabled (ablation) | AgentCBO · ALARM 37 nodes | normalized endpoint: 1.001 -> 0.23 | Improved |
| enable default coverage rescue (ablation) | AgentCBO PC-only vs prior · ALARM 37 nodes | normalized endpoint | No clear effect |
Experimental setup
Builds on: causal Bayesian optimization
Closest papers by meaning
- Adversarial Causal Bayesian Optimization (ICLR 2024 · similarity 0.88)
- DCILP: A Distributed Approach for Large-Scale Causal Structure Learning (AAAI 2025 · similarity 0.86)
- Causal LLM Routing: End-to-End Regret Minimization from Observational Data (NeurIPS 2025 · similarity 0.86)
- Near-Optimal Experiment Design in Linear non-Gaussian Cyclic Models (NeurIPS 2025 · similarity 0.86)
- Exploiting Causal Graph Priors with Posterior Sampling for Reinforcement Learning (ICLR 2024 · similarity 0.86)
- Sample Efficient Bayesian Learning of Causal Graphs from Interventions (NeurIPS 2024 · 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.
