Language Models Resist Alignment: Evidence From Data Compression

Jiaming Ji*, Kaile Wang*, Tianyi Qiu*, Boyuan Chen*, Jiayi Zhou*, Changye Li, Hantao Lou, Juntao Dai, Yunhuai Liu, and Yaodong Yang†

* Equal contribution   † Corresponding author
Institute for Artificial Intelligence, Peking University

Large language models (LLMs) may exhibit unintended or undesirable behaviors. Recent works have concentrated on aligning LLMs to mitigate harmful outputs. Despite these efforts, some anomalies indicate that even a well-conducted alignment process can be easily circumvented, whether intentionally or accidentally. Does alignment fine-tuning yield have robust effects on models, or are its impacts merely superficial? In this work, we make the first exploration of this phenomenon from both theoretical and empirical perspectives. Empirically, we demonstrate the elasticity of post-alignment models, i.e., the tendency to revert to the behavior distribution formed during the pre-training phase upon further fine-tuning. Leveraging compression theory, we formally deduce that fine-tuning disproportionately undermines alignment relative to pre-training, potentially by orders of magnitude. We validate the presence of elasticity through experiments on models of varying types and scales. Specifically, we find that model performance declines rapidly before reverting to the pre-training distribution, after which the rate of decline drops significantly. Furthermore, we further reveal that elasticity positively correlates with the increased model size and the expansion of pre-training data. Our findings underscore the need to address the inherent elasticity of LLMs to mitigate their resistance to alignment.

ACL 2025 (Volume 1: Long Papers), pages 23411–23432

PDF path note: local ACL 2025 paper Language Models Resist Alignment (llms-resist-alignment.pdf)

Elasticity

TL;DR

We demonstrate the elasticity of post-alignment models, forming resistance to alignment. Language models exhibit elasticity — resistance to leaving the pretraining distribution and rebound under reverse fine-tuning — explained via data compression.

Motivation

Training-based alignment (SFT, RLHF and variants) is mainstream, yet highly safety-aligned models can be made unsafe again with minimal fine-tuning — even on non-malicious data. We ask: Why is alignment so fragile?

We argue the fragility stems from alignment resistance: alignment often stays superficial. With elasticity, one can perform inverse alignment using far less data than the original alignment set.

Figure 1

Figure 1: Elasticity of language models — spring analogy linking dataset size to compression-rate change.

Method & Theory

We model LLMs as lossless compressors. Compression rate $\gamma$ proxies loss. Normalized compression-rate changes $\Delta\gamma$ across datasets are inversely proportional to dataset sizes $|D_i|$, like series springs. Theorem 4.2 shows alignment-side changes scale as $\Theta(k)$ times pretraining-side with $k=|D_p|/|D_a|\gg 1$.

Inverse alignment: after aligning on $D_a$, a much smaller $D_b$ can pull the model back toward pretraining behavior.

Resistance pipeline

Figure 2: Resistance experiment pipeline (forward vs inverse alignment).

Rebound pipeline

Figure 3: Rebound experiment pipeline (positive alignment then negative fine-tuning).

Results

Resistance: Inverse training loss is consistently lower than forward across Alpaca, TruthfulQA, and Beavertails for Llama2-7B/13B and Llama3-8B (Table 1).

Rebound: Models trained with more positive data drop faster under negative fine-tuning (IMDb & safety; Gemma-2B & Llama2-7B).

Scaling: Larger models (Qwen) and more pretraining tokens (TinyLlama) amplify rebound. Ablations show the effect across PPO/DPO/KTO/SimPO.

Rebound results

Figure 4: Rebound results — deeper positive alignment → faster drop under negative fine-tuning.

Table 1

Table 1: Inverse alignment loss consistently lower than forward across models and datasets.

Model scale

Figure 5: Larger models exhibit stronger rebound (Qwen 0.5B / 4B / 7B).

Pretrain scale

Figure 6: More pre-training data amplifies rebound (TinyLlama 2.0T / 2.5T / 3.0T).

Algorithm ablation

Figure 7: Rebound holds across PPO / DPO / KTO / SimPO.

Takeaways

Citation

@inproceedings{ji2025language,
  title={Language Models Resist Alignment: Evidence From Data Compression},
  author={Ji, Jiaming and Wang, Kaile and Qiu, Tianyi and Chen, Boyuan and Zhou, Jiayi and Li, Changye and Lou, Hantao and Dai, Juntao and Liu, Yunhuai and Yang, Yaodong},
  booktitle={Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)},
  pages={23411--23432},
  year={2025}
}