15 papers
Perception Before Supervision: Self-Contained Visual Distillation from Counterfactual Blind Spots
Shravan Venkatraman, Omkar Thawakar, Ritesh Thawkar +2
Self-improvement for multimodal large language models (MLLMs) is typically driven by reward-based methods that provide only coarse scalar feedback. Distillation offers a richer alt…
Ask, Solve, Generate: Self-Evolving Unified Multimodal Understanding and Generation via Self-Consistency Rewards
Ritesh Thawkar, Shravan Venkatraman, Omkar Thawakar +5
Most unified large multimodal models (LMMs) that support both visual understanding and image generation still rely on curated post-training supervision, such as human annotations,…
EvoLMM: Self-Evolving Large Multimodal Models with Continuous Rewards
Omkar Thawakar, Shravan Venkatraman, Ritesh Thawkar +5
Recent advances in large multimodal models (LMMs) have enabled impressive reasoning and perception abilities, yet most existing training pipelines still depend on human-curated dat…
Not All Modalities Are Equal: Instruction-Aware Gating for Multimodal Videos
Bonan Ding, Umair Nawaz, Ufaq Khan +5
Pre-trained video large language models excel at visual reasoning. However, they struggle when videos arrive with auxiliary streams, such as audio, depth map, or dense temporal evi…
CEPO: RLVR Self-Distillation using Contrastive Evidence Policy Optimization
Ahmed Heakl, Abdelrahman M. Shaker, Youssef Mohamed +4
When a model produces a correct solution under reinforcement learning with verifiable rewards (RLVR), every token receives the same reward signal regardless of whether it was a dec…
WorldCache: Content-Aware Caching for Accelerated Video World Models
Umair Nawaz, Ahmed Heakl, Ufaq Khan +3
Diffusion Transformers (DiTs) power high-fidelity video world models but remain computationally expensive due to sequential denoising and costly spatio-temporal attention. Training…