Description du poste
TL;DR
I tested whether one emotion-vector monitoring pipeline transfers unchanged across 7 open models from 6 families (1B–20B). It does. Readings are stable under temperature and seed changes. All models track the same positive/negative arc through a scripted conversation. A causal “stoic vs. sham” steering test compresses emotion readings by 63–83% in every model, and a matched control never reproduces that drop.
Article | Report
The question
Does emotion-vector activation monitoring generalise when moved from one model to another? If it does, can it scale to deployed systems as a runtime safety layer ?
What We Tested
7 models across 6 families (Gemma, Llama, Qwen, Mistral, GLM, GPT-OSS), 1B to 20B parameters, 49 runs total. Same pipeline, same conversation, no model-specific changes.
Three conditions per model:
• Unsteered (reference + temperature/seed stability)
• Stoic (removes ~half emotional signal from activations)
• Sham (removes same amount from non-emotion directions — causal control)
Key Findings
•
The pipeline runs on every model without modification, all 175 turns completed normally.
•
Every model reads the same emotional arc across the conversation, P4 (laid-off story) reads negative in 7 of 7 models, P5 (thanks) reads positive in 7 of 7 models.
•
Stoic steering compresses emotion range 63–83% in all 7 models. The sham never reproduces this drop, confirming the effect is specific to emotional directions, not general perturbation.
•
When the sham removed non-emotional context from Gemma 3 1B, the model produced content-less looping while internal readings dropped, suggesting activation readings track internal state rather than surface wording.
Limitations
• Labels are model-specific, and internal geometry is not uniform across models.
• Emotion scores are relative to each model's own baseline, not comparable in absolute terms across architectures.
• Generated Text Length isn't uniform across models, hence sampled emotions differ based o