DPO trained cleanly but the model got worse. What are DPO's real failure modes?
DPO is simple and stable, but it can drive down the probability of chosen responses, overfit the preference margin, and magnify length bias. The signal is naming these and the variants that fix each.
Updated Sep 2026 · Grounded in real GenAI, LLM, and AI/ML engineering interview loops and written to a senior-engineer editorial bar.
DPO is simple and stable, but it can drive down the probability of chosen responses, overfit the preference margin, and magnify length bias. The signal is naming these and the variants that fix each.
Lead with where the obvious approach breaks, because that is the judgment they are screening for — most candidates jump straight to the happy path and lose the room.
Then walk the failure back through the pipeline in order, naming the one metric the customer's exec sponsor actually cares about before you propose the fix.
