Somatic Marker Analogues in Language Models via Emotion Vector Re-Injection
A new research paper published on arXiv introduces a novel method to bridge the gap between semantic and episodic memory in large language models. Addressing the limitation that current models store factual data without emotional context, the study utilizes Gemma 3 1B-IT and sparse autoencoders to identify emotion-exclusive features. The researchers construct distinctive-feature emotion vectors during experience and partially re-inject them during recall, triggered by context similarity. This approach replicates Antonio Damasio’s somatic marker hypothesis, demonstrating that while emotional echoes alone do not drive decision-making, they significantly amplify knowledge into action when combined with semantic labels. Experimental results show that the combination of semantic memory and emotion echoes yields an 80% rate of good choices, compared to 52% for semantic labels alone. The findings suggest that integrating emotional markers steepens threat-safety gradients and enhances decision-making capabilities in AI, effectively mimicking human autonoetic consciousness without replacing factual knowledge.
Wire timeline
Somatic Marker Analogues in Language Models via Emotion Vector Re-Injection
A new research paper published on arXiv introduces a novel method to bridge the gap between semantic and episodic memory in large language models. Addressing the limitation that current models store factual data without emotional context, the study utilizes Gemma 3 1B-IT and sparse autoencoders to identify emotion-exclusive features. The researchers construct distinctive-feature emotion vectors during experience and partially re-inject them during recall, triggered by context similarity. This approach replicates Antonio Damasio’s somatic marker hypothesis, demonstrating that while emotional echoes alone do not drive decision-making, they significantly amplify knowledge into action when combined with semantic labels. Experimental results show that the combination of semantic memory and emotion echoes yields an 80% rate of good choices, compared to 52% for semantic labels alone. The findings suggest that integrating emotional markers steepens threat-safety gradients and enhances decision-making capabilities in AI, effectively mimicking human autonoetic consciousness without replacing factual knowledge.
cs.AI updates on arXiv.org