Skip to content
Modelwire
Subscribe

Gradient ascent during pretraining hardens models against fine-tuning attacks

Source published ·Modelwire updated

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

The development

Researchers demonstrate that actively suppressing knowledge during pretraining, rather than filtering or post-hoc unlearning, creates durable resistance to adversarial fine-tuning. By interleaving gradient ascent with standard training on filtered corpora, the technique forces attackers to expend 20x more compute to recover suppressed capabilities in 1B-parameter models. This shifts the safety-by-design frontier from reactive mitigation to architectural hardening, raising the bar for weaponizing open-weight models and forcing a reckoning with the assumption that any knowledge can be recovered given sufficient resources.

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

Modelwire analysis

Analyst take

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

DETER doesn't just make knowledge harder to recover; it forces a choice between safety-by-design (expensive pretraining) and safety-by-restriction (post-hoc filtering). The real novelty is that suppression during training creates asymmetric costs: defenders pay once, attackers pay 20x per recovery attempt.

This directly responds to the vulnerability landscape we've tracked. The 'Poison-then-Hide' backdoor attack from today showed that dormant threats survive finetuning, and the 'Coverage Collapse' paper from yesterday revealed that post-training optimization itself can degrade model robustness. DETER sidesteps both by hardening at pretraining time, before either attack surface exists. It also echoes the weight-deletion finding from last week: both suggest that model safety and capability may not require full-parameter interventions, just strategic targeting.

If Meta or Mistral release a 7B or larger model trained with DETER-style suppression in the next six months, and independent researchers confirm the 20x cost multiplier holds at scale, then pretraining-time suppression becomes table stakes for any lab claiming safety-conscious open release. If the cost advantage collapses above 10B parameters, the technique remains a niche defense for smaller models.

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.LG

    Backdoor attacks survive full model finetuning on vision encoders

    Researchers have demonstrated a novel backdoor attack that persists through full-model finetuning of pretrained vision encoders, a critical vulnerability for transfer learning workflows. The Poison-then-Hide technique implants dormant triggers during pretraining that activate only when models are finetuned on clean downstream data, bypassing existing defenses that assume frozen layers or minimal domain shift. This work…

    Read Modelwire coverage →Original source ↗

MentionsPubMed

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.LG originally reported this story as “DETER: Suppression During Pretraining for Tamper Resistance”. 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

General capability gains bypass withheld training data in scaled models

arXiv cs.CL·

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

Regularization techniques amplify hidden bias transfer in models

Gradient ascent during pretraining hardens models against fine-tuning attacks · Modelwire