AI

How Fragile Is On-Device Language Model Safety? Localizing Safety-Critical Parameters for Sparse Fault Analysis

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 08/10/2026, 04:03:25 UTC.

Briefings

This study investigates whether safety-critical behaviors in LLaMA-2-7B-Chat are concentrated in a sparse subset of parameters, enabling targeted fault analysis. Using low-rank subspace analysis and parameter-level importance filtering, researchers identified the MLP down_proj as the primary safety-sensitive component, while o_proj contributed less. Parameter-level modification of just 0.19% of down_proj weights improved Basic ASR by 53% and GCG ASR by 56%, with tinyBenchmarks accuracy remaining at 51.6% versus a 52.2% unmodified baseline. These findings suggest that selective integrity protection for on-device models can be highly effective by focusing on specific, non-uniform safety-sensitive regions.

Key points

• Low-rank safety-associated subspace analysis and parameter-level safety--utility importance filtering both revealed highly non-uniform safety sensitivity across the LLaMA-2-7B-Chat network.
• The MLP down_proj consistently emerged as the most prominent safety-sensitive component, while o_proj provided a smaller contribution to safety-critical behavior.

Context and limitations

Results are based on a single preprint (arXiv:2610.09000v1) from October 2026, with snapshot limitations regarding the specific deployment context of agentic systems and resource-constrained platforms.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 08/10/2026, 04:51:12 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeicl7ap3djd5xkh5k72kmy2opqby3iap4go7fy3mcchhzt7fcnsrmq