Token choice can change measured hallucination scores
Cette page n’a pas encore de traduction vérifiée. La version anglaise est affichée. Lire la version anglaise
En termes simples
POPE asks a vision-language model whether a named object is in a picture. The paper checks whether the score changes when evaluators use different token IDs to turn the model’s first answer into yes or no.
On LLaVA-1.5-7B’s adversarial questions, the model produced space-prefixed “ Yes” and “ No” tokens. Including those forms gave a higher F1 score than comparing two unprefixed IDs on the same predictions. The audit also found that one fixed set of IDs did not work across all tested models.
La recherche
When a test reports whether a vision-language model recognizes objects, should the score depend on how an evaluator reads the model’s first answer token? The paper examines this question on POPE, where yes/no answers are extracted from token scores.
The authors compare yes/no token readouts and find that deriving them from each model’s tokenizer changes POPE scores and avoids failures from fixed IDs.
The authors compare two ways to extract yes/no answers from LLaVA-1.5-7B’s first answer token. Its prompt context leads the model to generate space-prefixed forms, which the eight-token rule includes; the two-token rule compares unprefixed IDs instead. Neither of those two IDs was generated across the 9,000 questions.
On the same adversarial predictions, the eight-token rule scored 0.8221 F1 versus 0.7608 for the two-token rule. The comparison supports the conclusion that readout choice can alter POPE’s measured score, without changing the model’s predictions.
The main comparison uses one LLaVA-1.5-7B checkpoint at one quantization level. The authors cannot determine how much of the published methods’ gains would remain under matched evaluation, because those papers do not report their readout and may use different precision or other settings.
Analyse
Détails
- Auteurs de l’article original
- Kesav Kumar Jayakumar, Karthigeyan Thilak
- Soumis à arXiv
- Licence
- CC BY 4.0