Aditya Grover

Contrastive Prompting Enables Token-Level Credit Assignment in Discrete Policy Optimization

GCPO (Guidance Contrastive Policy Optimization) addresses a core weakness in group-advantage RL methods like GRPO and DAPO: uniform token credit assignment. By contrasting model predictions under positive and negative prompts, GCPO derives per-token advantage estimates proportional to the prediction difference, yielding finer-grained learning signals. This approach consistently outperforms GRPO and DAPO on both text-to-image generation and chain-of-thought reasoning benchmarks, and generalizes across modalities without architectural changes.

Contrastive Prompting Enables Token-Level Credit Assignment in Discrete Policy Optimization

Group-advantage RL methods like GRPO and DAPO suffer from coarse credit assignment — they broadcast a single sample-level reward uniformly across all tokens, obscuring which tokens actually drove performance. GCPO (Guidance Contrastive Policy Optimization) resolves this by computing token-level advantages as the delta between model predictions conditioned on positive vs. negative prompts, analogous to classifier-free guidance but applied to the RL objective. This yields sharper, semantically grounded learning signals — attending to visual regions in image generation and critical keywords in chain-of-thought traces — and consistently outperforms GRPO and DAPO baselines across both domains.

Sparse-LaViDa: Sparse Multimodal Discrete Diffusion Language Models

MobileWorldBench: Towards Semantic World Modeling For Mobile Agents

Self-Distilled Reasoner: On-Policy Self-Distillation for Large Language Models

LaViDa-R1: Advancing Reasoning for Unified Multimodal Diffusion Language Models

SNCE: Geometry-Aware Supervision for Scalable Discrete Image Generation