Hallucination

Hallucination is the tendency of a generative language model to produce output that reads fluently but is not supported by its input or by the world: an invented citation, a made-up product name, a JSON field that was never requested, or a class label that is not in the list you supplied.

The root cause is structural. An autoregressive model samples one token at a time from an open vocabulary, so nothing in the decoding process forces the output to stay inside a closed answer set. Constrained decoding, JSON mode and retry loops reduce the problem, but they are guards added around a generator, not properties of the model.

Why it matters for classification

When an LLM is used as a classifier, hallucination shows up as:

• a label that is not one of the options ("billing_issue" when you offered "billing")
• malformed output that fails to parse
• a confident-sounding explanation that does not match the chosen label
• verbalised confidence ("I am 90% sure") that is not a probability at all

Each of these needs defensive code, and each adds latency when it triggers a retry.

How Laya avoids it

Laya is a non-autoregressive decision model. It never generates text. Every option you pass is scored at its own option marker, and the answer is the argmax of a softmax over exactly those options. It is therefore impossible for Laya to return a label you did not define, or to return unparseable output: the response is always a typed choice, score or noul answer with probabilities.

That removes one failure mode, not all of them. A decision model can still be wrong, and it can be wrong with high confidence (the English checkpoint scores 0.000 accuracy on Khmer at 0.952 confidence). The fix for that is calibration plus language routing, not more generation guards. See hallucination-free decisions for the full argument.