xAI Grok 4.20 Beta 2 released with improved instructions following, hallucination reduction and more

xAI has released Grok 4.20 Beta 2 with new improvements for conversations, task execution, and visual generation. This new beta comes after the initial release in late February, featuring a multi-agent system where four specialized parts of Grok work together on each query. They split up the work, check facts against one another, and combine intelligence for better results.

That multi-agent system formed the core of the beta, giving users better reasoning on research, coding, and creative tasks while keeping responses quick. On the other hand, Grok Heavy users get three times the agents’ access compared to SuperGrok users. Additionally, xAI also opened early access to the full 4.20 series for those who want to test the latest model in their projects.

Below is the changelog and explanation of Grok 4.20 beta 2:

• Instruction Following Improvements
• Capability Hallucination Reduction
• Scientific Text Quality (LaTeX)
• Image Search Trigger Precision
• Multiple Image Render Reliability

Instructions following improvements sit front and center in the beta 2. Grok now sticks closer to exactly what you ask for instead of wandering off or adding unrequested extras. You get replies that match your preference, whether you want a summary or a detailed plan.

Capability hallucination reduction reduces cases where the model claims abilities or facts it cannot support. Internal multi-agent checks now detect these issues early and drop them by a significant margin, so answers stay grounded and true.

Scientific text quality receives a boost, especially around LaTeX formatting. Grok produces cleaner equations, symbols, and structured math output that researchers and students can copy straight into documents or notebooks without heavy editing.

Image search helps Grok to decide with precision when a prompt needs pictures. This improvement aims to avoid unnecessary searches while jumping in at the right moments to pull relevant images or start a search.

Multiple image render reliability makes showing several pictures at once much more consistent. Grok handles batches of generated or fetched images without dropping any or mixing up the order.

You can check these beta 2 improvements by starting a conversation with Grok 4.20 beta on the web and mobile apps.