Select the changes that genuinely reduce object hallucination in a vision-language model

Object hallucination happens when a model names things that are not in the image — a fork on an empty table — because its language side expects them. So sort each option by what it does to two competing signals: the visual evidence and the text-side guess. **Correct.** Hard-negative grounding data trains the model on images where an expected object is absent, so it stops assuming. Higher resolution or tiling keeps fine detail that would otherwise be lost, strengthening the evidence. Contrastive or grounding-aware decoding directly down-weights the language prior. A verification pass re-checks each claimed object against the image and catches confident misses. **Wrong.** Raising the temperature only makes outputs more random — it does not make them more faithful, and it tends to increase invention. Dropping the encoder and captioning from the filename throws away the visual evidence entirely, which is the exact opposite of grounding. The rule: strengthen the eyes, weaken the habit, or check the answer. Anything that adds randomness or removes evidence is a distractor.

The root cause sets the filter: object hallucination is the **language prior** overpowering weak visual grounding. So a real fix has to strengthen the visual signal, weaken the prior, or verify after the fact. Score each option against that. Hard-negative grounding data is correct — training on scenes where an expected object is absent teaches the model not to complete from co-occurrence. Higher resolution or tiling is correct — it preserves the fine detail that, when lost, forces the model to guess. Contrastive or grounding-aware decoding is correct — it directly down-weights the text prior at generation time. A verification pass is correct — re-checking each claimed object against the image catches the confident misses that slip through. The two distractors fail in opposite ways. Raising the sampling temperature changes diversity, not faithfulness; on a model already too willing to invent, more randomness makes hallucination worse, not better. Dropping the encoder and captioning from the filename removes the visual evidence entirely — it is the literal opposite of grounding. So the four-part taxonomy is the answer: more grounding data, more visual detail, prior-suppressing decoding, post-hoc verification. Any option that adds noise or deletes evidence is a trap.