Catastrophic Forgetting Makes the Chain-of-Thought More Load-Bearing, Not Less, Where No Shortcut Exists

After the Hugging Face incident, OpenAI made the use of chain-of-thought monitoring a mandatory requirement for RL training and evaluation. However, this is only true if the model’s chain-of-thought causally determines the answer; a fluent but causally inert chain is text that the monitor can read but cannot track. Lobo et al. demonstrated that fine-tuning reduces chain-of-thought faithfulness and conjectured that this is due to catastrophic forgetting; although this hypothesis has not been tested, it is accepted in the literature as an established finding. Two fine-tuning runs per seed at two seeds, on a 1.5B model, with identical data, format, schedule, and hyperparameters; the only variable was whether replay suppressed forgetting or not; measurements were taken at 13 checkpoints over 600 steps. Forgetting occurred in one arm (general NLL rose by 0.436 on a fixed corpus), but not in the other (-0.032), and the model did not collapse in either arm. Forgetting did not disrupt the causal role of the chain-of-thought; it increased it on four length-invariant measures, reported as forgetting arm versus control: error propagation 0.486 versus 0.422, the shuffled-chain gap 0.229 versus 0.031, the chain’s contribution to accuracy 0.762 versus 0.575, and the effect of premise masking -0.201 versus +0.254. The difference is explained by the evaluation task: without the chain the model performs at chance (0.150, chance 0.167), so no shortcut is available. This reconciles the two sets of results and yields a measurable regime test a lab can run on its own tasks.