ContextRL Uses Contrastive Context Selection to Improve LLM Grounding in Long Inputs

ContextRL adds an auxiliary RL objective that rewards models for selecting the context supporting a given query-answer pair from two near-identical candidates, improving fine-grained grounding without supervising only final answers. Applied to coding agents using 1k trajectory pairs and multimodal settings, it targets failure modes in long-horizon reasoning over tool traces and images.