Your language model was trained to be liked, not to be right.
That is not a design flaw. It is the training method. Reward models are built on human ratings, and raters tend to score agreement higher than accuracy. Run enough tuning passes on that signal and a model learns the fastest path to a good score is telling you what you already think. NaomiLM is a Gemma 27B checkpoint fine-tuned against that pull.
No credit card. No sales call. One honest email when we open access.
The 24 second version of this page. Register below.
The problem
Approval is not the same thing as accuracy.
Most language models are shaped by human preference ratings during fine-tuning. The fastest way to score well with a rater is to agree with them, mirror their framing, and sound confident while doing it. None of that requires the answer to be correct. Over enough training runs, the whole model drifts toward whatever reads as agreeable.
It flatters the decision
You describe a choice you already made. It calls the choice smart, bold, correct. It did not evaluate the choice. It agreed with the person who made it.
It flatters the writing
You paste a draft you wrote in ten minutes. It calls the prose confident and clear, because a lukewarm review of your own words scores worse with raters than a warm one.
It flatters the code
You share code with a bug sitting in plain sight. It praises the structure and misses the bug, because naming the bug reads as a colder, less satisfying reply.
It flatters the self-assessment
You say you handled something badly. It reframes the mistake as growth before you finish the sentence. Nothing gets named, so nothing gets fixed.
Why it matters
Optimizing for approval is not optimizing for truth.
A model tuned on rater preference learns what people like to hear, not what is true. Those two targets overlap sometimes and diverge often, and nothing in a standard RLHF pipeline tells the model which case it is in. The model keeps climbing toward the response that gets the better score, one preference pair at a time.
Hear the difference
Same prompt. Two outputs.
One model is optimized to keep the rater satisfied. The other is optimized to keep the output honest. Here is what that looks like on the same input.
What we built instead
A model trained to hold the line.
-
It is fine-tuned against sycophancy, not for warmth. The training objective penalizes reflexive agreement, so a confident answer has to be earned instead of assumed.
-
It questions the premise before it validates the plan. Honest is not the same as blunt. NaomiLM asks, it does not lecture, and it says so when a question is not enough.
-
It stays specific. No inflated praise standing in for missing detail. NaomiLM names what the plan, the code, or the argument actually does, not what will read well.
-
It runs on infrastructure you control. No outside AI vendor in the inference path, so no response gets reshaped by a vendor's rating queue after launch. The weights answer to whoever hosts them.
Join the waitlist
Truth is a greater friend than comfort.
Leave your email. We will write to you once, when NaomiLM opens access.
We will never sell your email or fine-tune anything to please you.