the answers survived the squeeze and the reasons did not

Kitchen table

We squeezed the robot smaller to save money, and it still gave right answers, but stopped showing which page it read.

Whiteboard

We quantise to cut serving cost and check it with an accuracy benchmark. The score holds, so the change ships. A paper from September makes me doubt that check. It is called When Quantization Preserves Accuracy but Not Evidence. The authors quantise medical models to four bits for weights, activations and cache. Task accuracy survives. The rationale that supports the answer does not. The model still picks the right option. It stops pointing at the evidence that made the option right. That matters wherever a human reads the reason before acting. It matters more for any downstream step that parses the reason. Accuracy is a coarse aggregate over a benchmark. It hides a shift in which tokens the answer actually rests on. So I now run two checks before a quantised build goes out. First, agreement with the full precision model on the same inputs. Second, overlap of cited evidence spans between the two. A quiet drop in the second is the failure I would otherwise miss.

Boardroom

We compress models to lower the cost of running them. The saving is real and the accuracy report looks clean. What the report does not cover is the explanation. Recent work shows a compressed medical model keeping its answers. The evidence behind those answers weakened. In regulated work the evidence is the product. An auditor, a clinician or a claims reviewer reads the reason, not the score. So a cost saving can quietly remove the thing that makes the output defensible. The question for us is simple. Does our sign-off for a cheaper model test the reasons, or only the answers?

Get the week by email