I am positive I know exactly how this works, I made something similar a few months back. But the problem is without generation you are extremely limited in the use cases. And while the model can't hallucinate, it can still be wrong. It just can't make up data.
Last year I also had a rather similar idea, but dropped it before I went very far in working on it. I wonder if you and I had similar ideas?
1. Start with an LLM, so that your model understands natural language.
2. Replace RoPE with a tree embedding scheme, and causal attention with a sparse attention on the graph structure. (You could use full attention... but it's cheaper to use graph attention.)
3. Chop off the final unembedding layer, replacing it with a projection down to two scalars, one for logits and one for confidence.
4. Each option of a choice is represented by a number of tokens in leaf position; average these tokens' logit outputs to get the option's logit. Average all of the confidences from all of the options to get the choice's confidence.
5. Train the logits by KL divergence from a true distribution (or NLL on samples from a true distribution).
6. Train the confidences on a subset of the data in which you know the entire true distribution.
The hardest part is getting real world data for workflows, but I wildly speculate that you can get by with only ~50,000 documents if you first adapt domains using synthetic data.
Yeah saying it can't hallucinate is crazy. It can still forward a billing query to the dev department incorrectly. It can still get an obvious yes/no question completely wrong
Are you able to share how it works in that case?
I'll tell you this. Output isn't too cheap to meter, there is no decoder.
So an encoder-only model with a classifier trained on the heads or something? DeepSeek recently switched to an encoder-decoder architecture in an attempt to get the best of both worlds (fast prefill while preserving generation capability), I wonder if that might be the future?
That was the first thing that come into my head. OK I can train very simple model, that can generate json's for specific tasks, so what? How we can be sure that this "limited use cases" not just overfitting for particular outputs (or even distillation?)
Except this, this thing looks like revolution.