EmotionScope probes emotion vectors in Gemma 2 2B

Hot take: this is more valuable as a methodology harness than as a finished scientific claim, and that is exactly where the interesting work is.

• –The project’s main contribution is operational: it makes probing, layer sweeps, validation gates, and visualization easier to run on open-weight models, which is the right way to lower the barrier for follow-up interpretability work.
• –The repo claims strong early signals on Gemma 2 2B IT, including a high-performing emotion layer around 84.6% depth, but the sample size is still small enough that the results should be treated as promising, not settled.
• –The “response-preparation” probing detail matters more than the visuals: if that methodological choice holds up, it suggests emotion structure is easier to read at the point where the model is about to answer than while it is still processing user content.
• –The visual layer is aspirational but useful for demos and communication; it will help people inspect trajectories, but the real scientific value is in whether the extraction/probing pipeline is reproducible.
• –The biggest open question is generalization across model sizes and families, especially whether richer social representations like dual-speaker structure emerge reliably outside this 2B-scale setup.