A balanced classification test is meant to prevent a model from looking good merely by choosing the most common class. But balancing the labels does not prove the model uses the question. On a binary dataset, a system that always says the same thing can still get half the answers right. The headline score conceals whether any meaningful discrimination is happening.
This matters when teams use a few labelled examples to turn a general language model into a classifier. Prompt order and label wording become part of the decision system. The label-collapse study examines cases where those surface choices dominate the query itself, creating a failure that ordinary accuracy reporting can overlook.
Balanced examples do not remove label dependence
An October 12 study documents label collapse across tested autoregressive and diffusion language models at the 8–9B scale. It finds that demonstration arrangements and label words can change the behaviour, even when class counts are balanced. Read the research paper.
The authors test five models spanning autoregressive and discrete-diffusion families at the 8–9B scale. They observe query-independent labels even under a class-balanced alternating arrangement. Some label-word pairs collapse across almost every tested condition; others change their preferred label when the arrangement changes.
That makes the behaviour more specific than a simple majority-class bias. The prompt can contain equal counts and still induce a strong preference. An engineer who responds only by balancing the examples may therefore leave the actual mechanism untouched. The tested architectures share the failure, but the paper does not claim every scale and prompting setup must behave identically.
Diffusion changes generation, not the need for diagnosis
The inclusion of discrete-diffusion language models broadens the test beyond ordinary left-to-right decoding. Their iterative denoising process raises additional questions about how context and label preferences interact, but it does not remove the basic need to establish dependence on the query. Both families can produce surface answers whose average looks ordinary while the underlying decision has collapsed.
The paper deliberately limits its architectural observation to the tested models and scale. That qualification is useful: sharing a failure mode does not establish an identical internal cause, and a clean behavioural diagnostic should not be mistaken for a complete mechanistic explanation. For a practitioner comparing architectures, the first task is to measure the same query-sensitivity property under a careful extraction protocol. Only then is it meaningful to investigate whether iterative denoising, prompt arrangement or another factor explains why a particular model ignores the evidence.
A probability can move without the displayed answer moving
The proposed complementary diagnostic is separation: the log-odds between the two label words at the query position. It provides a view of the model’s relative preference beneath the final extracted label. The measurement protocol also guards against tokenisation and formatting artefacts that could make an extraction error look like a reasoning failure.
This is a reason to inspect both the decision surface and the decoding interface. A system may show weak but query-dependent separation, or almost constant separation regardless of input. Those cases can share an accuracy score while requiring different remedies. The diagnostic helps determine which problem is actually present.
The first check is whether evidence changes the decision
A practical audit can pair inputs that differ in the one feature determining the label, keep the demonstrations fixed and inspect the probability change as well as the answer. It can then vary label names and example order independently. This proposed intervention isolates query sensitivity from prompt sensitivity.
The larger implication is uncomfortable for benchmark reporting: chance accuracy is not always a weak version of competence. It can be the visible average of a system that has stopped performing the intended task. Before spending on more demonstrations or calibration, establish that there is an input-dependent decision process worth improving.
Questions, news tips or corrections? Contact [email protected].
