🟡 Evaluating 'Confidence First' vs 'Reasoning Fir...
# questionsi
iakashvekariya
12/05/2025, 6:22 AMNON-BAML QUESTION !!!
I see a lot of prompts asking LLMs to output a confidence score (0-10) or a label (High/Low) before generating the answer. Given that these models are stateless and autoregressive, isn't this counter-productive?
If the model predicts a token like Confidence: Low at the start, it biases the attention mechanism for the subsequent generation, potentially degrading the quality of the answer to match that low score. Conversely, if it guesses "High," it might hallucinate to maintain that persona.
I’ve been experimenting with forcing a Chain of Thought first and putting the Verdict/Confidence as the very last step. This allows the model to "read" its own reasoning before assigning a score.
Does anyone rely on "Confidence First" prompts in production, or is "Reasoning First -> Verdict Last" the only viable path without accessing raw logprobs?
---
Or better yet, is asking for confidence even a real question? Help me understand.
n
nikitoshina
12/05/2025, 6:58 AMi put cot, answer, confidence
nikitoshina
12/05/2025, 6:58 AMit is a decent indicator of problematic responses
nikitoshina
12/05/2025, 6:59 AMi had tried a confidence fist a couple years ago, it did biad itself
v
Vaibhav (BAML 🐑)
12/05/2025, 7:18 AM^for best quality that is the way
n
nikitoshina
12/05/2025, 7:24 AMAn experiment you can do is enable logprobs and look at probability for confidence response 😛 it will change on the position (at least that’s what I remember)
i
iakashvekariya
12/05/2025, 1:29 PMGreat. Thanks both of you.
But even at the last you said looking at probability of logprob. So which exact token? Is that the one where it says HIGH/medium/low? the exact token or average of past outputs? And even tho we take last token, would that be correct tho?
n
nikitoshina
12/05/2025, 5:39 PMhere is a good article:
https://cookbook.openai.com/examples/using_logprobs
nikitoshina
12/05/2025, 5:43 PMSo, you want to use a non-reasoning model, which is very important here. Get the logprobs for your test case, look at the structure, and ask the agent to build you a demo viewer. I’d suggest a prompt that separates your confidence section, answer, and CoT with new lines. Then you can split the logprobs by new lines tokens, inspect perplexity for the CoT, and look at the probabilities for the confidence and answer tokens.
The placement of your answer also matters—I once accidentally put the answer before the reasoning, and it biased the model. Also, in the API call, specify that you want probabilities for all potential tokens so you can compare how likely each response was.
You can just ask Cursor to build it out for you, by the way.
nikitoshina
12/05/2025, 5:45 PMand maybe check in pairs 😛 instead of 3 at the same time. Depending on your provider and task complexity you could even replace confidence field with probability of answer token
i
iakashvekariya
12/06/2025, 1:53 AMThanks for sharing! Will check it out 🫡🙌🏼