Gemma 4, small models fail basic grammar checks

Despite high reasoning benchmarks, small models remain surprisingly bad at surgical text editing because they are trained to rewrite rather than correct.

• –Gemma 4 E4B "corrects" valid words to themselves, demonstrating an inability to execute a simple no-op when text is already correct
• –Instruct-tuning conditions these models to always be "helpful," leading them to force unnecessary edits and add conversational filler
• –Smaller architectures struggle to maintain strict negative constraints (like "do not change the meaning") across longer sentences
• –The over-correction issue is widespread across new releases, affecting the recently launched GPT-OSS 20B and Qwen 3 Next 80B models