Learned, Then Lost: A Measured Single-Example Counterfactual in Pre-training
A one-row injection was measurable shortly after exposure, but not at the end of training.
Speck and Shepard ran 24 small-scale GPT-2 pre-training counterfactuals, swapping one row in a batch at step 200 with a fixed 194-token passage or random characters. Fifty steps later, models that saw a passage predicted it better than uninjected twins across all eight seeds. By the final step, the paper reports no detected passage-specific advantage, no clear separation between real and fabricated subjects, and no detectable held-out loss effect. The authors describe the injection as moving the model within its basin rather than out of it. ArXiv · AI/CL/LG's note
Speck and Shepard ran 24 small-scale GPT-2 pre-training counterfactuals, swapping one row in a batch at step 200 with a fixed 194-token passage or random characters. Fifty steps later, models that saw a passage predicted it better than uninjected twins across all eight seeds. By the final step, the paper reports no detected passage-specific advantage, no clear separation between real and fabricated subjects, and no detectable held-out loss effect. The authors describe the injection as moving the model within its basin rather than out of it. ArXiv · AI/CL/LG's note
score 5