Research map / Probabilistic and scientific modeling / Bayesian optimization
In-Context Black-Box Optimization with Unreliable Feedback
Read the paperCheck your idea against 108,967 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 FICBO, a feedback-aware transformer pretrained on synthetic tasks with a structured feedback prior, so in-context black-box optimizers can estimate auxiliary-source reliability from history and exploit or discount cheap feedback at test time.
Problem
Black-box optimization with cheap auxiliary feedback (expert heuristics, simulators, proxies) that may be biased, input-dependent, or misleading; existing in-context optimizers use only the optimization history and feedback-aware BO methods handle one task at a time with static trust weights.
Approach
Pretrain an amortized transformer optimizer over a structured feedback prior (feature-split additive and reweighted mixture priors plus bias operators) so it infers source reliability in context and jointly learns posterior prediction and query policy conditioned on candidate-wise feedback.
Key results
- On Airfoil Lift, FICBO (Mixture) reaches final cumulative regret 3.617 vs 24.179 for feedback-only and 7.807 for pi-BO-EI.
- On Airfoil Small, FICBO (Additive) final regret 0.014 vs 0.097 for pi-BO-EI.
- GP-based pi-BO variants improve over in-context pi-BO, but FICBO remains best in both best-so-far and cumulative regret on marginalized benchmarks.
Surprising findings
- Training with the true objective as feedback (y_true variant) made the source over-trusted and gave regret on Airfoil Lift comparable to feedback-only (24.649).
- A naive feedback-as-input-feature baseline still improves over the feedback-agnostic amortized model, showing feedback carries exploitable information even without special architecture.
Limitations
- Feedback prior coverage is central; smooth but globally rank-misaligned feedback (Branin a=0.75) is a difficult out-of-prior regime.
- Assumes feedback is available for every candidate in the pool; low-dimensional experiments only.
- Feedback source is fixed during the BO episode and does not update from new evaluations; horizon distribution shift at test time.
Experimental setup
Datasets: synthetic GP tasks with feedback prior Branin multi-fidelity Airfoil NeuralFoil NAS-Bench-201 style CIFAR-10 to ImageNet16-120 sequential reaction kinetics
Benchmarks: 13 analytic benchmark functions (Sphere, Ackley, Rastrigin, Rosenbrock, Schwefel, Levy, Hartmann 3D) Branin multi-fidelity Airfoil Small Airfoil Lift ML hyper-parameter (architecture) task Reactor
Metrics: normalized best-so-far objective normalized cumulative regret final cumulative regret top-5 policy-feedback agreement
Baselines: GP-prior ICBO GP-prior ICBO + feedback feature pi-BO-UCB pi-BO-EI Feedback greedy Random pi-BO-UCB GP pi-BO-EI GP
Builds on: in-context black-box optimization pi-BO prior-data fitted networks REINFORCE
Closest papers by meaning
- Learning Surrogates for Offline Black-Box Optimization via Gradient Matching (ICML 2024 · similarity 0.86)
- MetaBox-v2: A Unified Benchmark Platform for Meta-Black-Box Optimization (NeurIPS 2025 · similarity 0.86)
- Posterior Inference with Diffusion Models for High-dimensional Black-box Optimization (ICML 2025 · similarity 0.86)
- SYMBOL: Generating Flexible Black-Box Optimizers through Symbolic Equation Learning (ICLR 2024 · similarity 0.86)
- ROOT: Rethinking Offline Optimization as Distributional Translation via Probabilistic Bridge (NeurIPS 2025 · similarity 0.86)
- Posterior Inference in Latent Space for Scalable Constrained Black-box Optimization (NeurIPS 2026 · 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.