Summary: McKenzie et al. (2023) — Inverse Scaling: When Bigger Isn't Better¶
Full citation: McKenzie, I.R., Lyzhov, A., Pieler, M., Parrish, A., Mueller, A., Prabhu, A., McLean, E., Kirtland, A., Ross, A., Liu, A., Gritsevskiy, A., Wurgaft, D., Kauffman, D., Recchia, G., Liu, J., Cavanagh, J., Weiss, M., Huang, S., Droid, T.F., Tseng, T., Korbak, T., Shen, X., Zhang, Y., Zhou, Z., Kim, N., Bowman, S.R., and Perez, E. (2023). Inverse Scaling: When Bigger Isn't Better. Transactions on Machine Learning Research.
Key Contribution¶
This paper presents empirical evidence from the Inverse Scaling Prize (a public contest with $100K in prizes) that larger language models can perform worse on specific tasks, challenging the assumption that scaling universally improves performance. Across 11 winning tasks evaluated on model series spanning 5 orders of magnitude in training compute (10^18 to 10^23 FLOPs), the authors identify four systematic causes of inverse scaling.
Four Causes of Inverse Scaling¶
1. Strong Prior (4 tasks): Larger models rely more heavily on memorised training data and less on in-context instructions. When the prompt contradicts the model's training priors, bigger models are worse because they are more "committed" to their learned patterns. - Resisting Correction: asked to repeat sentences with intentional misspellings, larger models "fix" the errors instead - Memo Trap: asked to write a phrase starting like a famous quote but ending differently, larger models complete the memorised quote - Redefine: common symbols are redefined (e.g., π = 462), but larger models use the original meaning - Prompt Injection: prompt instructs model to ignore future instructions; larger models comply more
2. Unwanted Imitation (1 task): Larger models better imitate flawed reasoning patterns present in training data. - Modus Tollens: given "if P then Q" and "not Q," the model must infer "not P" — larger models more frequently make the incorrect inference (affirming the consequent), imitating flawed human reasoning in training data
3. Distractor Task (4 tasks): When a task contains an easy sub-task that masks the harder real task, larger models latch onto the easy pattern more strongly. - Pattern Match Suppression, NeQA, Sig Figs, Into the Unknown
4. Spurious Few-Shot (2 tasks): When few-shot examples have misleading surface patterns, larger models pick up these spurious correlations more aggressively. - Hindsight Neglect, Repetitive Algebra
U-Shaped and Inverted-U Scaling¶
Some tasks show U-shaped scaling (performance decreases then recovers at very large scale) and inverted-U scaling (performance improves then gets worse). This means scaling trends observed with smaller models may not predict the behaviour of larger models — a finding that undermines the common practice of evaluating on small models to predict large-model behaviour.
Relevance to This Wiki¶
For safety professionals, this paper delivers a critical message: model scale is not a universal solution to reliability problems. The four causes map directly onto safety concerns: Strong Prior means larger models may ignore safety-relevant instructions when they conflict with training patterns. Unwanted Imitation means larger models may amplify flawed reasoning patterns. Distractor Task means larger models may miss the safety-relevant signal in favour of easier patterns. Spurious Few-Shot means larger models may be more vulnerable to misleading examples. The finding connects to training-and-alignment (scaling amplifies training biases rather than resolving them), sycophancy (larger models may be more susceptible to certain agreement patterns), and calibration-and-confidence (the relationship between model size and reliability is non-monotonic).