Yisheng Liu | Searcharxiv

arXiv2026

Reflection in the Dark: Exposing and Escaping the Black Box in Reflective Prompt Optimization

Automatic prompt optimization (APO) has emerged as a powerful paradigm for improving LLM performance without manual prompt engineering. Reflective APO methods such as GEPA iteratively refine prompts by diagnosing failure cases, but the optimization process remains black-box and label-free, leading to uninterpretable trajectories and systematic failure. We identify and empirically demonstrate four limitations: on GSM8K with a defective seed, GEPA degrades accuracy from 23.81% to 13.50%. We propose VISTA, a multi-agent APO framework that decouples hypothesis generation from prompt rewriting, enabling semantically labeled hypotheses, parallel minibatch verification, and interpretable optimization trace. A two-layer explore-exploit mechanism combining random restart and epsilon-greedy sampling further escapes local optima. VISTA recovers accuracy to 87.57% on the same defective seed and consistently outperforms baselines across all conditions on GSM8K and AIME2025.

cs.AI↗

arXiv2024

Hybrid Fusion for Battery Degradation Diagnostics Using Minimal Real-World Data: Bridging Laboratory and Practical Applications

Unpredictability of battery lifetime has been a key stumbling block to technology advancement of safety-critical systems such as electric vehicles and stationary energy storage systems. In this work, we present a novel hybrid fusion strategy that combines physics-based and data-driven approaches to accurately predict battery capacity. This strategy achieves an average estimation error of only 0.63% over the entire battery lifespan, utilizing merely 45 real-world data segments along with over 1.7 million simulated data segments derived from random partial charging cycles. By leveraging a thoroughly validated physics-based battery model, we extract typical aging patterns from laboratory aging data and extend them into a more comprehensive parameter space, encompassing diverse battery aging states in potential real-world applications while accounting for practical cell-to-cell variations. By bridging the gap between controlled laboratory experiments and real-world usage scenarios, this method highlights the significant potential of transferring underlying knowledge from high-fidelity physics-based models to data-driven models for predicting the behavior of complex dynamical systems.

cs.CE↗