When LLMs don’t change stuff you don’t want changed

Posted by silenceimpaired@reddit | LocalLLaMA | View on Reddit | 8 comments

Perhaps other models have improved in this area and I need to look more carefully, but the recent GPT-OSS 120b model is pretty good editing a specific part of a writing without changing the rest of the text. Maybe it was a sampling choice on my part… what is your take on LLM models and their ability to return you entire original text with specific changes only?

Reply to Post

8 Comments

[-]

phree_radical@reddit

Place a marker around the area you want edited, "pre-fill" the parts you don't want edited, generate only the parts you want edited and stop generation at the marker

[-]

zyxwvu54321@reddit

Better prompt engineering.

[-]

silenceimpaired@reddit (OP)

Could be. As I recall, often times the LLM (Llama and Qwen) would change very minor stuff in areas I didn’t ask to be changed in the past. I hate to think OpenAI has some secret sauce.

[-]

zyxwvu54321@reddit

It’s mostly about prompt engineering. Which OpenAI model are you talking about? Free versions of GPT-4o and even GPT-5 do basically the same thing. I haven’t used the o-series. The models that give minimal changes to original code without needing much prompting: Gemini Pro 2.5 and Qwen3-Coder have been solid. Sonnet was like that too, but for some reason, I don’t know if it’s just me, I can’t get good answers from Sonnet anymore. Feels like they nerfed the free version or something. And I’m not sure which frontend stack you’re using, but most LLMs seem weak at React compared to backend languages.

[-]

silenceimpaired@reddit (OP)

Hmm. I am pasting 1800 words with a small one sentence set of instructions. Not sure how to think of this prompt in light of your suggestions.

[-]

Sativatoshi@reddit

They all do. Ive been really experimenting lately, and with proper prompt engineering, youd be blown away at what even a smaller model can do

[-]

colin_colout@reddit

Have you tried playing with temperature? I also find lower quants seem to increase likeliness of conflating similar tokens

[-]

silenceimpaired@reddit (OP)

You might be right. I did try a much lower temperature at one point… but I’m thinking my temperature was at 1 when it worked well.