A long standing problem in galaxy evolution is whether high-redshift starbursts (selected by their SFR(IR) excess compared to the star-forming Main Sequence) are gas rich galaxies driven by stochastic accretion events, or rather merger systems, in which the infalling gas is consumed at higher star-formation efficiency compared to normal galaxies, as seen in local ULIRGs. In our first published work, we found from a statistical sample of 25 Herschel-detected starbursts at z ~ 0.7 in COSMOS field that their dust attenuation pattern (derived by comparing Paβ, Hα and bolometric IR luminosities) is consistent with an optically thick mixed model with average total optical depth of 9 magnitudes in V band. The presence of highly embedded cores and their disturbed morphology strongly favours a merger-induced starburst scenario at high redshift. Our starbursts also define a one-parameter sequence: we will show, indeed, that they follow relatively tight correlations between total obscuration, emission line ionisation, starburst core size and effective age (as traced by line EWs), which can be interpreted as a time-evolutionary sequence of merger stages. Intriguingly, X-ray emission is detected only for a subset of later-stage mergers, suggesting emergent AGNs leading to a final bright QSO.