Depth pruning breaks function calling uniformly across small models
Source published ·Modelwire updated
Original coverage: arXiv cs.CL ↗·How Modelwire adds context
The development
Researchers systematically mapped how depth pruning damages small language models, revealing that capability loss is highly non-uniform across tasks. When Transformer blocks are removed from 1.2-1.7B parameter models, function calling and mathematical reasoning collapse almost entirely, while other skills degrade more gracefully. Critically, brief fine-tuning can selectively restore some lost abilities, but recovery patterns differ across model families. This work matters for practitioners deploying pruned models as agents: it shows that naive compression can silently break mission-critical features like tool use, and that healing strategies must be capability-aware rather than one-size-fits-all.
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 critical finding isn't just that pruning breaks things unevenly, but that some capabilities (function calling, math) are structurally fragile to depth reduction in ways that fine-tuning cannot fully repair across model families. This suggests the damage reflects architectural dependency, not just learned weights.
This work directly complicates the optimism from 'Rethinking LLM Pruning via Test-Time Scaling' (late October), which showed pruned models retain reasoning capacity when given multiple attempts. That paper measured pass@k on math and coding; this one reveals those same tasks collapse under depth pruning specifically, suggesting the earlier finding may not generalize to all compression strategies. The tension matters: test-time scaling works for certain pruning methods but not others. This also echoes 'What Survives Distillation' (late October), which found compression preserves performance through different internal mechanisms. Here we see some mechanisms don't survive at all, and selective fine-tuning recovery suggests practitioners need diagnostic tools before deploying pruned agents.
If the authors release a diagnostic toolkit that predicts which capabilities will survive pruning before compression happens (using the recovery patterns they mapped), adoption by edge deployment teams in the next six months would confirm this moves from academic observation to practical mitigation. If instead the work remains descriptive without actionable recovery recipes, it's a warning label rather than a solution.
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
Pruned models retain reasoning coverage despite lower single-attempt accuracy
Researchers challenge the conventional wisdom that pruned language models lose reasoning capacity, finding instead that substantial model compression preserves the ability to solve problems across multiple solution attempts. By measuring pass@k performance on math and coding tasks, the work reveals pruned models can match or exceed originals in solution diversity despite lower single-attempt accuracy. The…
MentionsLlama-3.2-1B · Qwen2.5-1.5B · SmolLM2-1.7B · OLMo-2-1B
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 “What Breaks and What Comes Back: Capability Damage and Recovery in Depth-Pruned Small Language Models”. 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.