Salvato in:
| Autore principale: | |
|---|---|
| Natura: | Preprint |
| Pubblicazione: |
2025
|
| Soggetti: | |
| Accesso online: | https://arxiv.org/abs/2510.01228 |
| Tags: |
Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
|
| _version_ | 1866918253516816384 |
|---|---|
| author | Zeng, Siqi |
| author_facet | Zeng, Siqi |
| contents | Large language models should follow hierarchical instructions where system prompts override user inputs, yet recent work shows they often ignore this rule while strongly obeying social cues such as authority or consensus. We extend these behavioral findings with mechanistic interpretations on a large-scale dataset. Linear probing shows conflict-decision signals are encoded early, with system-user and social conflicts forming distinct subspaces. Direct Logit Attribution reveals stronger internal conflict detection in system-user cases but consistent resolution only for social cues. Steering experiments show that, despite using social cues, the vectors surprisingly amplify instruction following in a role-agnostic way. Together, these results explain fragile system obedience and underscore the need for lightweight hierarchy-sensitive alignment methods. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2510_01228 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | Who is In Charge? Dissecting Role Conflicts in Instruction Following Zeng, Siqi Computation and Language Machine Learning Large language models should follow hierarchical instructions where system prompts override user inputs, yet recent work shows they often ignore this rule while strongly obeying social cues such as authority or consensus. We extend these behavioral findings with mechanistic interpretations on a large-scale dataset. Linear probing shows conflict-decision signals are encoded early, with system-user and social conflicts forming distinct subspaces. Direct Logit Attribution reveals stronger internal conflict detection in system-user cases but consistent resolution only for social cues. Steering experiments show that, despite using social cues, the vectors surprisingly amplify instruction following in a role-agnostic way. Together, these results explain fragile system obedience and underscore the need for lightweight hierarchy-sensitive alignment methods. |
| title | Who is In Charge? Dissecting Role Conflicts in Instruction Following |
| topic | Computation and Language Machine Learning |
| url | https://arxiv.org/abs/2510.01228 |