Transformers and LSTMs fail differently when neurons are removed
Source published ·Modelwire updated
Original coverage: arXiv cs.CL ↗·How Modelwire adds context
The development
Researchers systematically removed computational units from LSTMs and Transformers to model how neural degradation affects language model performance, mimicking neurodegenerative disease patterns. Results reveal architectural differences in failure modes: LSTMs degrade smoothly across syntactic and semantic tasks, while Transformers show threshold-like collapse in feedforward layers but gradual decline in attention heads. The work bridges neuroscience and deep learning interpretability, offering insights into model robustness and suggesting that architectural choices fundamentally shape how systems fail under stress, with implications for understanding both AI resilience and biological neural loss.
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 study doesn't just measure degradation; it reveals that failure isn't uniform across architectures. Transformers don't degrade gracefully like LSTMs do, suggesting that attention-based designs may have hidden brittleness thresholds that recurrent systems lack.
This work extends the mechanistic interpretability thread from our October 20 coverage on linguistic and arithmetic circuit dissociation. That research showed LLMs organize capabilities into modular subsystems; this ablation study now reveals those modules fail differently depending on the underlying architecture. The finding also connects to our October 19 piece on competing learning in hybrid architectures, which identified how recurrent and attention layers suppress each other during training. Here we see the downstream consequence: when units degrade, attention layers collapse suddenly while recurrence degrades smoothly, suggesting the architectural choice fundamentally shapes not just learning dynamics but robustness properties.
If researchers apply the same ablation protocol to modern hybrid architectures (combining recurrence and attention), watch whether the failure mode splits between layer types or whether one dominates the collapse pattern. If attention-based collapse persists in hybrids, that confirms architectural brittleness is intrinsic to the mechanism rather than a scaling artifact.
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.CL
LLMs partition linguistic and arithmetic processing into separate circuits
Mechanistic interpretability research has uncovered that linguistic and arithmetic processing in large language models operate through largely separate neural circuits, mirroring dissociations observed in human neuroscience. Testing five LLMs with controlled materials and circuit discovery methods reveals that task domain, not vocabulary overlap or syntax, explains circuit organization. This finding reshapes how researchers think about…
MentionsLSTM · Transformer · arXiv
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 “Can Language Models Get "Aphasia" From Neuronal Ablation?”. 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.