Forgetting concentrates in rare-token embeddings during LLM continual learning
Source published ·Modelwire updated
Original coverage: arXiv cs.CL ↗·How Modelwire adds context
The development
Researchers have pinpointed where continual learning breaks down in large language models: the output embeddings of infrequent tokens. By freezing parameters across multiple model scales up to 1.4B parameters, they discovered that catastrophic forgetting concentrates in a narrow, predictable region tied directly to vocabulary gaps in new training data. This finding enables practitioners to forecast forgetting risk from token frequency alone, without access to original training data. The mechanism involves asymmetric gradient flow in the softmax layer for unseen tokens. This localization insight reshapes how teams approach data-free continual pre-training and fine-tuning, offering a path toward targeted mitigation strategies that don't require replay buffers.
Modelwire’s AI-generated summary of coverage from arXiv cs.CL.
Modelwire analysis
ExplainerOur AI-generated reading of the wider context and the next developments to watch.
The paper doesn't just identify that catastrophic forgetting happens; it shows forgetting is NOT distributed across the model but concentrates in a narrow, predictable region tied to vocabulary gaps. This localization is what enables practitioners to forecast risk without access to original training data, a practical constraint that prior work on forgetting mechanics didn't directly address.
This work sits in direct conversation with last week's finding that forgetting is often reversible (the October 6 Spurious Forgetting paper). That work showed representations shift rather than erase; this paper goes deeper by pinpointing exactly where the irreversible damage concentrates. The two together suggest a hierarchy: some forgetting is recoverable through normalization, but output embedding forgetting for rare tokens is the hard case. The connection also echoes the October 3 quantization study, which found semantic collapse on rare vocabulary; here we see the mechanism in full-precision models that explains why rare tokens are structurally vulnerable during adaptation.
If teams implementing the output embedding freeze strategy report measurable retention gains on held-out rare-token benchmarks within the next two quarters, that confirms the localization thesis is actionable. If instead forgetting still spreads to other layers despite the freeze, the mechanism is more distributed than claimed.
This interpretation is generated from the summary above and the archive coverage cited below. Our methodology · Report an error
Coverage behind this analysis
These archive entries ground the connection in our analysis. They are ordered by source publication date, with links to our coverage and the original sources.
·arXiv cs.LG
Language models hide rather than erase old knowledge during finetuning
Researchers have identified why language models appear to forget old knowledge during finetuning, yet often retain it in recoverable form. The work reveals that finetuning shifts representations along shared directions rather than erasing them, and that normalization mechanisms can restore access to buried facts once new training stabilizes. This mechanistic insight matters for practitioners: it…
MentionsLarge Language Models · catastrophic forgetting · output embeddings · softmax gradients
How this coverage is produced
Modelwire uses AI to generate summaries and context from source headlines, snippets, and selected archive coverage. Automated checks do not verify every claim, and items are not routinely reviewed by a person before publication. Zacaria Solis operates the site. Read the linked source for the full evidence and report errors through our corrections process.
Modelwire summarizes, we don’t republish. arXiv cs.CL originally reported this story as “A Deafening Silence: Catastrophic Forgetting Lives in the Output Embeddings of Tokens the Data Never Speaks”. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.