๐
๐
Old Age
Beyond Shallow Alignment: How Post-Training Methods Determine Refusal Circuits And Steering Robustness
September 03, 2026 ยท Grace Period ยท ๐ EMNLP 2026 Main Conference
Authors
Hoang Cuong Nguyen, Mark Dras, Usman Naseem
arXiv ID
2609.03887
Category
cs.CL: Computation & Language
Citations
0
Venue
EMNLP 2026 Main Conference
Abstract
How do the methods used to train language models to refuse harmful requests shape how that refusal actually works inside the model? We compare three post-training methods - supervised fine-tuning, reasoning-augmented fine-tuning (training on reasoning chains that justify a safety decision), and preference optimization (ORPO) - across three architecturally distinct models (Llama-3.1-8B, Gemma-2-9B, Qwen3-8B). We find that training method, not just data, reshapes how refusal is computed internally: reasoning-augmented training consistently produces a distinct kind of refusal computation, visible across all three models, while architecture independently shapes internal structure and how reliably refusal can be steered. Most importantly, no method we study achieves all three properties we would want from safe alignment at once: refusal that isn't concentrated in a few fragile components, safety gains that don't cost general capability, and safety behavior correctable through small, targeted edits. We caution against treating current post-training methods as a solved, reliable defense, especially for security-critical use. Code and models are available in https://github.com/hoangcuongnguyen2001/Beyond-Shallow-Alignment.
Community Contributions
Found the code? Know the venue? Think something is wrong? Let us know!
๐ Similar Papers
In the same crypt โ Computation & Language
๐
๐
Old Age
BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
๐
๐
Old Age
XLNet: Generalized Autoregressive Pretraining for Language Understanding
๐ฎ
๐ฎ
The Ethereal
Effective Approaches to Attention-based Neural Machine Translation
๐
๐
Old Age
A large annotated corpus for learning natural language inference
๐
๐
Old Age