arXiv:2606.26987v1 Announce Type: cross Abstract: Recent work identified emotion vectors in Claude Sonnet 4.5, which are internal representations that encode emotion concepts, causally influence behavior, and exhibit geometry mirroring human psychological structure. We test the generality of these findings in two open-weight models, Apertus-8B-Instruct-2509 and Gemma-4-E4B-it, extracting emotion contrast vectors across all layers, using two model-generated corpora. We recover valence geometry for both models, with peak PC1--valence correlations of $r = 0.76$ and $r = 0.83$, approaching the $r
This research builds directly on recent findings in proprietary models like Claude Sonnet 4.5, extending the investigation into widely accessible open-source LLMs which are rapidly evolving and seeing broad adoption.
The discovery of emotion vectors in open-source LLMs suggests a common underlying mechanism for emotional representation across different model architectures and scales, paving the way for more nuanced and ethically complex AI interactions.
The understanding that even open-source, smaller models can encode and potentially be influenced by 'emotion vectors' means the development of emotionally intelligent or manipulative AI agents is more universally accessible.
- · AI researchers (ethics and alignment)
- · Open-source AI community
- · Developers of emotionally aware AI applications
- · Companies relying solely on black-box proprietary LLMs
- · Those unprepared for complex human-AI interaction ethics
This research provides a foundational step towards understanding and controlling the 'emotional' states and responses of LLMs.
The ability to predictably manipulate or interpret emotional concepts in open-source models could lead to new forms of beneficial AI, but also advanced disinformation or influence operations.
As these capabilities become industrialized, legal and ethical frameworks around AI emotional manipulation could become critical, potentially leading to 'empathy-audits' for AI systems.
This signal links to a primary source. Continuum Brief monitors and indexes it as part of the live intelligence stream — we do not republish source content.
Read at arXiv cs.AI