Prueba de decodificación paralela del esqueleto del pensamiento

Las respuestas y explicaciones aparecen a medida que realiza el cuestionario.

1. ¿Qué produce la primera etapa del Esqueleto del Pensamiento?
   • Una respuesta final completa
   • Un punto de control del modelo afinado
   • Un breve esqueleto de encabezados de puntos.
   • Un conjunto de etiquetas de entrenamiento.
2. ¿Por qué la etapa de expansión de puntos puede realizarse en paralelo?
   • Porque cada punto depende del anterior
   • Porque los puntos son en gran medida independientes entre sí.
   • Porque el modelo se vuelve a entrenar para el paralelismo.
   • Porque los tokens se generan al revés
3. ¿Cuál es el principal cuello de botella de los objetivos SoT en la decodificación LLM normal?
   • Latencia secuencial token por token
   • Velocidad de lectura del disco
   • Tamaño del vocabulario del tokenizador
   • Ancho de banda de red para el usuario.
4. Which answer structure is a good fit for Skeleton-of-Thought's parallel expansion?
   • A proof whose next step must use the previous step's newly derived result
   • A single fixed one-word output with no expansion
   • A procedure that requires each section to wait for all earlier outputs
   • Several sections that can be developed largely independently
5. What extra work should a Skeleton-of-Thought latency comparison include?
   • Only the longest point while ignoring every other stage
   • Generating the outline, coordinating point expansion, and assembling the final response
   • Mandatory full model retraining for every answer
   • A guarantee that overhead is always exactly the same fixed duration
6. How can local batched Skeleton-of-Thought decoding expand multiple points?
   • Generate every token of every paragraph in one complete forward pass
   • Retrain a new model for each heading
   • Advance the point sequences together across repeated decoding steps
   • Remove autoregressive dependencies within every point automatically
7. If Skeleton-of-Thought reduces answer latency, what other assessment remains necessary before adopting the workflow?
   • Treat the speedup as proof that all factual claims are correct
   • Measure only the spelling of the model's name
   • Assume outline ordering guarantees that every section agrees
   • Check whether the assembled response remains coherent and meets answer-quality requirements
8. After concurrent point expansion, how should Skeleton-of-Thought outputs be assembled into the answer?
   • Associate expansions with their skeleton points and combine them in the intended outline order
   • Shuffle the sections randomly regardless of their headings
   • Treat the first finished point as the entire answer
   • Discard the outline and keep only token counts