🔍Problem
Safety-aligned LLMs can become unsafe again with minimal fine-tuning — even on non-malicious data. The paper asks why alignment is so fragile and whether post-alignment effects are robust or superficial.
Paste any arXiv link. Get the full picture in seconds.
Runtime app needs Groq API + arXiv URL. This page is a filled static card for the ACL 2025 paper (no live API call).
Paper Digest
ACL 2025 Long Papers · pages 23411–23432 · Project: pku-lm-resist-alignment.github.io
Safety-aligned LLMs can become unsafe again with minimal fine-tuning — even on non-malicious data. The paper asks why alignment is so fragile and whether post-alignment effects are robust or superficial.
Models LLMs as lossless compressors; uses compression rate as a proxy for alignment loss. Defines elasticity (resistance + rebound) and inverse alignment, then proves Theorem 4.2 under a Pareto mass assumption and validates with SFT / PPO / DPO / KTO / SimPO.
Inverse loss < forward loss across Llama2/3 × Alpaca/Truthful/Beavertails (Table 1). More positive alignment data → steeper rebound under fewer negatives (Fig. 4). Elasticity grows with model size and pretrain volume (Figs. 5–6). Cross-algorithm rebound in Fig. 7.
Theory relies on a Pareto quality-distribution assumption. Experiments do not cover a full pretrain–align lifecycle. Multimodal generality and quantitative scaling laws for when elasticity appears remain open.
High-signal ACL long paper if you work on alignment robustness, safety fine-tuning, or open-weight governance: it reframes “shallow safety” as data-volume-driven elasticity with theory + multi-model evidence.
