Large Language Models Encode Semantics and Alignment in Linearly Separable Representations
Fuente:
arXiv
Saved in:
| Main Authors: | , , , , , |
|---|---|
| Format: | Preprint |
| Published: |
2025
|
| Subjects: | |
| Online Access: | |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866908777841688576 |
|---|---|
| author | Saglam, Baturay Kassianik, Paul Nelson, Blaine Weerawardhena, Sajana Singer, Yaron Karbasi, Amin |
| author_facet | Saglam, Baturay Kassianik, Paul Nelson, Blaine Weerawardhena, Sajana Singer, Yaron Karbasi, Amin |
| contents | Understanding the latent space geometry of large language models (LLMs) is key to interpreting their behavior and improving alignment. Yet it remains unclear to what extent LLMs linearly organize representations related to semantic understanding. To explore this, we conduct a large-scale empirical study of hidden representations in 11 autoregressive models across six scientific topics. We find that high-level semantic information consistently resides in low-dimensional subspaces that form linearly separable representations across domains. This separability becomes more pronounced in deeper layers and under prompts that elicit structured reasoning or alignment behavior$\unicode{x2013}$even when surface content remains unchanged. These findings motivate geometry-aware tools that operate directly in latent space to detect and mitigate harmful and adversarial content. As a proof of concept, we train an MLP probe on final-layer hidden states as a lightweight latent-space guardrail. This approach substantially improves refusal rates on malicious queries and prompt injections that bypass both the model's built-in safety alignment and external token-level filters. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2507_09709 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | Large Language Models Encode Semantics and Alignment in Linearly Separable Representations Saglam, Baturay Kassianik, Paul Nelson, Blaine Weerawardhena, Sajana Singer, Yaron Karbasi, Amin Computation and Language Machine Learning Understanding the latent space geometry of large language models (LLMs) is key to interpreting their behavior and improving alignment. Yet it remains unclear to what extent LLMs linearly organize representations related to semantic understanding. To explore this, we conduct a large-scale empirical study of hidden representations in 11 autoregressive models across six scientific topics. We find that high-level semantic information consistently resides in low-dimensional subspaces that form linearly separable representations across domains. This separability becomes more pronounced in deeper layers and under prompts that elicit structured reasoning or alignment behavior$\unicode{x2013}$even when surface content remains unchanged. These findings motivate geometry-aware tools that operate directly in latent space to detect and mitigate harmful and adversarial content. As a proof of concept, we train an MLP probe on final-layer hidden states as a lightweight latent-space guardrail. This approach substantially improves refusal rates on malicious queries and prompt injections that bypass both the model's built-in safety alignment and external token-level filters. |
| title | Large Language Models Encode Semantics and Alignment in Linearly Separable Representations |
| topic | Computation and Language Machine Learning |
| url | https://arxiv.org/abs/2507.09709 |