IC-089OLMo 0724 7B base learns formatted multiple-choice question answering between 80k and 100k training steps, transitioning from near-random to near-perfect accuracy on the synthetic colors task

Sarah Wiegreffe, Oyvind Tafjord, Yonatan Belinkov, Hannaneh Hajishirzi, Ashish Sabharwal

SourceAnswer, Assemble, Ace: Understanding How LMs Answer Multiple Choice Questions

Using a synthetic colors task that disentangles symbol binding from dataset-specific knowledge, the paper shows that OLMo 0724 7B base undergoes a sharp transition in formatted MCQA ability between 80k and 100k training steps. Accuracy goes from 25% (random) at 50k steps to 97% at 100k steps. However, vocabulary projection reveals that while accuracy is high at 100k steps, the logit differences between answer symbols remain small; only with sustained training (final checkpoint at ~652k steps) does the model widen the logit difference substantially, assigning high probability to the predicted answer alone. This small logit difference persists even at the final checkpoint on harder datasets like MMLU and HellaSwag.

Evidence
correlational
Key metric
25% accuracy at 50k steps, 97% at 100k steps, 81% at 200k steps, 100% at final checkpoint (~652k steps) on 3-shot colors task
Caveat
The non-monotonic accuracy (97% at 100k, 81% at 200k) is not explained; the finding is on a single model family and a synthetic task
Model
OLMo / OLMo base OLMo 0724 7B
Datasets
MMLU / MMLU-Math [eval], HellaSwag [eval]
Related work
Li & Gao 2024 [context]
Related findings
IC-087, IC-088
Extraction
automatic-extraction