Zoe Wu

Diffusion models achieve strong generative performance across domains. However, achieving high sample quality under hard physical and dynamical constraints, as in protein design and robotic manipulation, remains difficult. Existing inference-time methods either require a costly or unavailable verifier, or avoid a verifier at the cost of reduced diversity. To relieve this tension, we introduce Score Tempering with Exchange between Exploratory Replicas (STEER), a verifier-free inference-time method that replaces single-chain sharpening with global search using multiple diffusion chains. During sampling, neighboring chains, each at different sharpening strengths, exchange particles according to an energy-based acceptance rate, propagating exploration to outputted samples. Across a diverse range of domains, from synthetic multimodal distributions to image generation, STEER achieves a better fidelity–diversity trade-off and improves success rates on continuous-control tasks in both verifier-based and verifier-free settings. Our findings suggest that effective diffusion tempering requires going beyond local updates to population-based global search that enhance fidelity and diversity.

Recent advances in image generative models have enabled rich collaborations between humans and AI systems. Among these, Energy-Based Models (EBMs) learn an energy landscape that guides noised samples toward high-probability regions. Unlike diffusion models that use fixed time schedules, EBMs possess equilibrium properties that enable user feedback during generation without destabilizing the distribution. However, current EBM research primarily optimizes for high-fidelity images, offering little control over the trade-off between semantic realism and fine-grained diversity—an essential feature for interactive creative applications. Artists and creatives thus lack a modality to balance semantic coherence (e.g., “a red apple”) with creative variation (e.g., apples of different shapes or colors). To address this, we introduce a geometry-aware annealing framework for EBMs. We propose a directionally-aware annealing variable that leverages local geometric information to directionally adjust the effective noise level during sampling.

Recorded from prefrontal (PFC) and motor (M1) cortex during delayed visually and memory guided reaches to study how each region encodes sensory input and motor plans. Population decoding shows PFC shifting from processing sensory input to maintaining a motor plan, while M1 maintains the motor plan in both visual and memory trials.