Daily Paper Digest
- LLMs show alignment elasticity (resistance + rebound); Δγ ∝ 1/|Di|.
- Inverse alignment with |Db| ≪ |Da| can undo alignment; Theorem 4.2 links this to compression.
- Elasticity rises with model size and pretraining volume; holds across SFT/PPO/DPO/KTO/SimPO.
Focus Brief
alignmentcompressionACL2025
Language Models Resist Alignment: Evidence From Data Compression
Published: ACL 2025 Long Papers · pp. 23411–23432
Authors: Jiaming Ji*, Kaile Wang*, Tianyi Qiu*, Boyuan Chen*, Jiayi Zhou*, Changye Li, Hantao Lou, Juntao Dai, Yunhuai Liu, Yaodong Yang†
Source: ACL Anthology / project page (local PDF pack)
Categories: cs.CL, cs.LG, alignment
Relevance: high — theoretical + empirical account of why safety alignment is shallow
Conclusion: Language models are elastic and thus naturally resist alignment; robust alignment must mitigate this intrinsic elasticity.
Contributions: (1) Phenomenon: elasticity = resistance + rebound with Δγ ∝ 1/|Di|; (2) Mechanism: compression protocol + Theorem 4.2; (3) Validation across model scales, pretrain volumes, and preference algorithms.
Best For: alignment researchers, safety evaluators, open-weight governance
Limitations: Pareto assumption in theory; incomplete lifecycle experiments; multimodal / scaling-law quantification open
Summary: Empirically shows post-alignment models rebound toward pretraining under further fine-tuning; theoretically argues fine-tuning undermines alignment relative to pretraining potentially by orders of magnitude via compression rates.
PDF / Project: pku-lm-resist-alignment.github.io
Key numbers (from paper)
- Table 1: every inverse vs forward SFT pair — inverse loss lower (e.g. Llama2-7B Alpaca θ₃→θ₁ 0.1666 ↓ vs θ₁→θ₃ 0.2346 ↑).
- Table 3(b): Llama2-7B unsafe samples to recover KL ε=10⁻² after 1000/2000/5000/10000 safe: 961 / 844 / 801 / 729.
- Figs 5–6: larger Qwen / more TinyLlama pretrain tokens → stronger rebound.
Figures
