Ran He

Smaller, cheaper models generate superior negative examples for preference distillation, significantly boosting student performance without requiring larger teacher models.

Most test-time adaptations in vision-language models are either redundant or actively destructive—and skipping up to 85% of them via simple consistency checks actually boosts final accuracy.

TREK transforms the way models tackle challenging prompts by expanding their exploration support, leading to substantial performance gains even in the hardest task scenarios.

Parameter-level defenses against model merging are fundamentally flawed, allowing attackers to exploit their weaknesses with a new Anchor-Guided Attack.

Fine-tuning on the new OmniVideo-100K dataset boosts model performance by over 20% in audio-visual reasoning tasks, revealing the power of structured scripts in enhancing multimodal understanding.

Hallucinated details in super-resolution are not just random noise; they reveal a fundamental spectral mismatch that can be corrected by shaping the generative process itself.

LLMs can reason better and generate more diverse outputs by projecting negative samples onto a positive subspace during reinforcement learning.

Test-time RL's vulnerability to noisy pseudo-labels is amplified by group-relative advantage estimation, but can be mitigated with a surprisingly simple debiasing and denoising approach.

Current audio-language models are surprisingly bad at controlling and interpreting subtle vocal cues, failing in nearly half of situational dialogue scenarios.

EVT achieves 86.6% top-1 accuracy on ImageNet-1k without extra training data, redefining the potential of Vision Transformers in computer vision.

Overconfident tokens, often missed by entropy-based methods, carry surprisingly dense corrective signals in on-policy distillation, allowing for near-baseline performance with <10% of tokens.

Robots can now learn contact-rich manipulation skills like humans by feeling the forces involved, thanks to a new multimodal interface that captures synchronized visual, tactile, and force data.

A principled framework for General World Models reveals the limitations of current systems and the architectural requirements for future progress.

Overconfident errors in RLVR monopolize probability mass and suppress exploration, but a confidence-aware penalty fixes this and boosts mathematical reasoning performance.