Skip to content
Modelwire
Subscribe

Reasoning history deletion costs vary wildly across open models

Source published ·Modelwire updated

Original coverage: arXiv cs.CL ↗·How Modelwire adds context

The development

Researchers quantified a previously unmeasured failure mode in multi-turn reasoning workflows: when conversation history is stripped during cross-provider forwarding, open-weight reasoning models degrade unpredictably. Testing seven models across 794 tasks revealed performance swings from minus 4.4 to plus 51.3 points depending on deletion scope and model architecture. This finding exposes a hidden cost in federated AI deployments where reasoning traces cannot follow requests across system boundaries, forcing practitioners to choose between architectural simplicity and reasoning continuity.

Modelwire’s AI-generated summary of coverage from arXiv cs.CL.

Modelwire analysis

Explainer

Our AI-generated reading of the wider context and the next developments to watch.

The paper isolates deletion sensitivity as a model-intrinsic property rather than a deployment accident. The 51-point swing suggests some architectures retain reasoning coherence across context breaks while others collapse, implying this isn't fixable at the infrastructure layer alone.

This connects directly to the November 8th finding that models lose capabilities in predictable order under perturbation. Here, deletion of conversation history is the perturbation, and the 0.90 correlation framework from that work suggests deletion sensitivity may follow the same capability hierarchy. The earlier paper on weight deletion (November 7th) showed that models are surprisingly robust to surgical parameter removal, but this work reveals they're fragile to information removal during inference. Together these paint a picture: models can tolerate structural changes but not reasoning continuity breaks.

If the same seven models tested here show deletion sensitivity rankings that correlate with their task-failure hierarchies from the November 8th compression study, that confirms deletion sensitivity is a manifestation of underlying capability fragility rather than a separate failure mode. If rankings diverge, it signals a distinct architectural vulnerability worth investigating separately.

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.

  1. ·arXiv cs.CL

    Model capabilities fail in predictable order under compression

    Researchers have discovered that language models lose capabilities in a consistent, predictable sequence when compressed or perturbed, regardless of the degradation method used. Testing seven compression techniques across eight instruction-tuned models revealed that task failures correlate at 0.90 across different perturbation types, suggesting an underlying fragility hierarchy independent of compression approach. This finding has immediate…

    Read Modelwire coverage →Original source ↗

MentionsShareGPT

MW

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 “Deletion Sensitivity Is a Property of the Model: A Three-Condition Measurement across Seven Open Reasoning 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.

More to explore

Conversation history degrades LLM reasoning accuracy by up to 29 points

arXiv cs.CL·

OpenAI research links frontier model opacity to capability optimization tradeoffs

The Decoder·

Selective weight deletion matches full LLM fine-tuning at 0.05% parameter change

Reasoning history deletion costs vary wildly across open models · Modelwire