Forgetting Gets Worse With Scale

Here's an uncomfortable question the continual learning literature mostly dodged until scale forced it: does catastrophic forgetting get better or worse as models get bigger? The folk intuition said bigger models should be more robust — more capacity, more flat directions in the loss landscape, more room to fit new tasks without disturbing old ones. Luo et al. tested this directly on large language models, and the answer came back wrong. In the 1B–7B range they studied, forgetting gets worse with scale.

The study

The authors (Luo, Liu, Liu, Wang, and colleagues) ran a clean, systematic experiment: take decoder-only and encoder-decoder LLMs from 1B to 7B parameters, continually fine-tune them on a sequence of five tasks spanning domain knowledge, reasoning, and reading comprehension, and evaluate on four categories — domain knowledge, reasoning, reading comprehension, plus bias/fairness. The design is notable for what it measures besides accuracy: they track the model's general capabilities as it specializes, which is the actual deployment question. Nobody cares if your model learns the new task; they care what it lost to do so.

Three findings stand out:

• Forgetting is real and scale makes it worse. Across 1B–7B models, continual fine-tuning produced clear catastrophic forgetting of general capabilities, and the severity intensified with model size. The folk intuition — capacity as a buffer — fails in this regime. A plausible mechanism: larger models fit the new task's distribution more aggressively, and sharper specialization means larger departures from the pretrained solution.
• Architecture matters: decoder-only forgets less. BLOOMZ (decoder-only) showed less forgetting than mT0 (encoder-decoder) at comparable scales. This is a genuinely useful engineering signal — the dominant architecture of the LLM era happens to be the more continual-learning-friendly one — though the paper doesn't fully isolate why (pretraining objective, instruction tuning history, and architecture are confounded).
• General instruction tuning alleviates forgetting. Models that had undergone broad instruction tuning before continual fine-tuning forgot less. Breadth of the initial solution seems to act as a regularizer: a model trained to do many things has flatter, more robust representations, so task-specific gradients do less damage. This echoes the EWC-era insight about flat minima, rediscovered at scale.

And then the surprise: continual fine-tuning mitigated gender bias. As the models learned the new tasks, stereotypical bias scores went down. Forgetting isn't always a loss — when what's forgotten is a bias rather than a capability, the same mechanism looks like debiasing. It's a small result with a large implication: "forgetting" is a value-neutral description of a mechanism, and whether it's catastrophic depends entirely on what was forgotten.

Why this paper matters more than its citations suggest

Published on arXiv in 2023 and updated through v5 in January 2025, this is an empirical paper in a subfield that loves methods. Its contribution is a set of facts the methods have to reckon with. The scale finding in particular reframes the continual-learning-for-LLMs agenda: if bigger models forget more, then the naive hope that scale solves continual learning is dead, and parameter-efficient methods (LoRA and friends) aren't just about compute — they're about limiting the blast radius of updates. That reframe directly motivates the 2025 papers in this series: sparse memory finetuning (update 11%-worth of damage instead of 89%), CLoRA-style subspace constraints, and merging-based approaches that avoid sequential updates entirely.

The instruction-tuning finding also deserves more weight than it gets. It suggests that the cheapest anti-forgetting intervention isn't a continual learning algorithm at all — it's the starting point. A broadly-trained, flat-minimum initialization is worth more than a clever regularizer on a narrow one. This connects to a thread running from EWC's flat-directions argument through to modern "model soup" and merging work: robustness to subsequent updates is a property you can bake in before continual learning even starts.

Limitations and open questions

The study's scope is honest but narrow: 1B–7B models (small by today's standards — does the trend reverse at 70B+?), five tasks, and a single sequential order. Task order effects in continual learning are notoriously large, and a single ordering can't separate method effects from sequence effects. The BLOOMZ-vs-mT0 comparison confounds architecture with pretraining recipe, so the "decoder-only forgets less" claim needs a controlled ablation. The bias result, while intriguing, is measured on standard bias benchmarks whose relationship to real-world harm is debated — "fine-tuning debiases" is a claim that needs much stronger evidence before anyone acts on it.

The biggest open question is the mechanism behind the scale trend. Why do larger models forget more? Is it sharper minima, more aggressive fitting, or something about how capabilities are entangled in larger models? The paper documents the phenomenon; the explanation — and the intervention — are still open. My bet: it's about the geometry of specialization. Larger models can fit the new task with lower loss, which means the optimizer travels further from the pretrained solution, and distance traveled is what kills old capabilities. If that's right, the fix isn't "smaller models" but "constrain the distance" — which is exactly what subspace methods and sparse updates do.

Scale doesn't buy you memory. In the 1B–7B regime, it buys you sharper specialization — and sharper specialization forgets more. Forgetting is not a capacity problem; it's a geometry problem.

The arc

This paper is the bridge between the classical era and the LLM era of continual learning. McCloskey and Cohen showed that sequential training destroys old knowledge in small networks; Luo et al. showed the same pathology survives — intensifies — at billions of parameters, in models trained on trillions of tokens. The problem didn't get solved by scale; it got more expensive. That fact is the entire justification for the current wave: sparse finetuning, representation-aligned merging, width-upscaled continual pretraining. Every 2025 method in this series is, at bottom, an answer to this paper's implicit challenge: if scale won't save us, what will?

Paper: An Empirical Study of Catastrophic Forgetting in Large Language Models During Continual Fine-tuning — Luo et al., 2025.