🧑 Human Understanding

🧪 ICML2026 · 5 paper notes

📌 Same area in other venues: 📷 CVPR2026 (151) · 🔬 ICLR2026 (45) · 🤖 AAAI2026 (20) · 🧠 NeurIPS2025 (21) · 📹 ICCV2025 (41) · 🧪 ICML2025 (3)

DiscoForcing: A Unified Framework for Real-Time Audio-Driven Character Control with Diffusion Forcing

DiscoForcing reformulates the "music \(\to\) full-body dance" offline generation problem into a strictly causal, bounded-latency streaming task. It utilizes a VQ-PAE causal music encoder, latent-space Diffusion Forcing, hybrid temporal noise scheduling, and Temporal Guidance sampling to translate music streams into 30 FPS full-body motions that directly drive Unity avatars and Unitree G1 humanoid robots in real-time.

Efficient, Validation-Free Intrinsic Quality Estimation for Large-Scale Face Recognition Datasets

Ours proposes Intrinsic Quality (IQ): after extracting embeddings using a proxy model, it weightedly fuses "Neighborhood Label Consistency (Consis)" and "Normalized Spectral Entropy Effective Rank \(\tilde{r}_{\mathrm{ent}}\)". It provides a "trainability" score for million-scale face recognition datasets without full training or clean validation sets. On WebFace4/12/42M and noise-injected settings, the ranking consistency with downstream MFR-ALL validation accuracy reaches Spearman = 1.0.

Learning Instance-Adaptive Low-Rank Orthogonal Subspaces for Clothes-Changing Person Re-Identification

The "clothing" semantic concept is explicitly modeled as an instance-adaptive low-rank subspace (initialized using the SVD principal components of CLIP text descriptions and refined via cross-attention with image patches). Identity features are then forced to be strictly orthogonal to this subspace through geometric constraints, achieving SOTA results in clothes-changing re-identification (PRCC +5.9% Rank-1) without the need for adversarial training.

MotionGRPO: Overcoming Low Intra-Group Diversity in GRPO-Based Egocentric Motion Recovery

MotionGRPO reformulates first-person full-body motion recovery from head-mounted devices as a Markov Decision Process (MDP) over diffusion sampling. It utilizes Group Relative Policy Optimization (GRPO) post-training with a hybrid reward system consisting of a "trajectory condition-aware perception model + 4 joint-level sub-rewards." Crucially, it identifies that strong input conditions lead to nearly identical intra-group samples, causing advantage variance to vanish—a fatal bottleneck. To resolve this, it injects Perlin noise into the conditioning signal to restore intra-group diversity, reducing MPJPE from EgoAllo's 124.985 mm to 114.207 mm on AMASS/RICH.

WaveVerse: Scalable RF Simulation in Generative 4D Worlds

WaveVerse integrates LLM-driven "4D indoor scene + human motion" generation with a physical ray tracer that preserves spatiotemporal phase coherence into a prompt-to-RF signal pipeline. It significantly enhances downstream RF imaging and activity recognition tasks using synthetic data, with performance scaling continuously as simulation volume increases, unlike existing methods that saturate.