Логит-линза и тест на декодирование промежуточного слоя

Ответы и пояснения появляются по мере прохождения теста.

1. What readout does a plain logit lens reuse to obtain vocabulary scores from an intermediate state?
   • A required new image classifier for every layer
   • The model's appropriate final normalization and unembedding output projection
   • A random dictionary sort with no model parameters
   • A separate retrained tokenizer for each prompt
2. Why can the same output projection be applied mathematically at multiple residual-stream depths?
   • Every layer must contain an identical copy of the final state
   • The entire transformer is a single linear mapping
   • All intermediate token scores are guaranteed identical
   • The states have compatible hidden dimensions, although their representations may not be equally calibrated for that readout
3. Какую проблему решает «настроенный объектив» по сравнению с обычным логит-объективом?
   • Он исправляет несоответствие каждого слоя с помощью изученного аффинного преобразования.
   • Это устраняет необходимость в каком-либо извлечении
   • Это делает модель быстрее при выводе
   • Он заменяет softmax на argmax.
4. If an intermediate lens readout favors a correct factual token, what has been directly observed?
   • The unique causal location of the model's factual knowledge
   • A guarantee that all later output will be correct
   • That the particular probe maps that state to a high score for the token
   • The exact reasoning process of a conscious system
5. Для чего больше всего полезен объектив Logit?
   • Сжатие веса модели
   • Перевод между языками
   • Активации модели шифрования
   • Наблюдение за тем, как и где формируется прогноз по глубине
6. Why can a plain final-head readout be misleading on early transformer states?
   • Early states always use a different vocabulary file
   • Their representation distributions can differ from the final state for which the head was trained
   • Every early-layer vector is necessarily unstructured noise
   • No matrix operation can be applied before the final layer
7. A lens uses a model with a fifty-thousand-entry vocabulary and a 768-dimensional hidden state. How many vocabulary logits does its unembedding produce?
   • 768, one score per hidden coordinate only
   • One, because the token must be chosen before projection
   • Fifty thousand, one score per vocabulary entry
   • Exactly the number of transformer layers
8. What does softmax add to a logit-lens vocabulary-score display?
   • A normalized distribution over the projected vocabulary scores
   • Proof that every displayed probability is a calibrated belief
   • A causal intervention on the model's hidden states
   • A replacement for the output projection's vocabulary dimension