When does SoftPrompt-IR not help?

[tobs-code]

SoftPrompt-IR is intentionally minimal and probabilistic.

This issue collects cases where:

• explicit weighting had no effect
• results got worse
• the model ignored the structure entirely

Any model, any domain, any short example is welcome.
Negative results are especially valuable.

[tobs-code]

I’ll summarize findings here once we have a few data points.