Comment by ivape
3 days ago
What formats? I thought the very schema of json is what allows these LLMs to enforce structured outputs at the decoder level? I guess you can do it with any format, but why stray from json?
3 days ago
What formats? I thought the very schema of json is what allows these LLMs to enforce structured outputs at the decoder level? I guess you can do it with any format, but why stray from json?
Sometimes it will randomly generate something like this in the body of the text: ``` <tool_call>executeshell <arg_key>command</arg_key> <arg_value>echo "" >> novels/AI_Voodoo_Romance/chapter-1-a-new-dawn.txt</arg_value> </tool_call> ```
or this: ``` <｜toolcallsbegin｜><｜toolcallbegin｜>executeshell<｜toolsep｜>{"command": "pwd && ls -la"}<｜toolcallend｜><｜toolcallsend｜> ```
Prompting it to use the right format doesn't seem to work. Claude, Gemini, GPT5, and GLM 4.5, don't do that. To accomodate DeepSeek, the tiny agent that I'm building will have to support all the weird formats.
Can't you use logit bias to help with this? Might depend how they are tokenized.
Maybe you have your temperature turned up too high.
In the modes in APIs, the sampling code essentially "rejects and reinference" any token sampled that wouldn't create valid JSON under a grammar created from the schema. Generally, the training is doing 99% of the work, of course, it's just "strict" means "we'll check it's work to the point a GBNF grammar created from the schema will validate."
One of the funnier info scandals of 2025 has been that only Claude was even close to properly trained on JSON file edits until o3 was released, and even then it needed a bespoke format. Geminis have required using a non-formalized diff format by Aider. Wasn't until June Gemini could do diff-string-in-JSON better than 30% of the time and until GPT-5 that an OpenAI model could. (Though v4a, as OpenAI's bespoke edit format is called, is fine because it at least worked well in tool calls. Geminis was a clown show, you had to post process regular text completions to parse out any diffs)
> In the modes in APIs, the sampling code essentially "rejects and reinference" any token sampled that wouldn't create valid JSON under a grammar created from the schema.
I thought the APIs in use generally interface with backend systems supporting logit manipulation, so there is no need to reject and reinference anything; its guaranteed right the first time because any token that would be invalid has a 0% chance of being produced.
I guess for the closed commercial systems that's speculative, but all the discussion of the internals of the open source systems I’ve seen has indicated that and I don't know why the closed systems would be less sophisticated.
I maintain a cross-platform llama.cpp client - you're right to point out that generally we expect nuking logits can take care of it.
There is a substantial performance cost to nuking, the open source internals discussion may have glossed over that for clarity (see github.com/llama.cpp/... below). The cost is very high, default in API* is not artificially lower other logits, and only do that if the first inference attempt yields a token invalid in the compiled grammar.
Similarly, I was hoping to be on target w/r/t to what strict mode is in an API, and am sort of describing the "outer loop" of sampling
* blissfully, you do not have to implement it manually anymore - it is a parameter in the sampling params member of the inference params
* "the grammar constraints applied on the full vocabulary can be very taxing. To improve performance, the grammar can be applied only to the sampled token..and nd only if the token doesn't fit the grammar, the grammar constraints are applied to the full vocabulary and the token is resampled." https://github.com/ggml-org/llama.cpp/blob/54a241f505d515d62...
4 replies →