Question

When this process is performed post-training, the AdaRound algorithm may be used to minimize loss. For 10 points each:
[10h] Name this process that is classified as either symmetric or asymmetric, depending on whether a zero-point is specified. Applying this process to LLMs has dramatically reduced their VRAM (“vee-RAM”) requirements.
ANSWER: quantization [accept binarization; accept more specific answers such as post-training quantization or quantization-aware training; prompt on PTQ or QAT]
[10m] Quantized neural networks often utilize “straight-through estimators” to estimate “surrogate” examples of this function. In a process named for it, the product of this function with a constant is repeatedly subtracted from the objective function.
ANSWER: gradient [accept surrogate gradient or gradient descent; prompt on SGD; prompt on del]
[10e] Quantization works by reducing the precision of this data type, which is defined in IEEE 754. Doubles are a more precise variant of this type, which is the most common type used to represent decimal numbers.
ANSWER: floats [accept floating-point numbers or FP]
<Other Science>

Back to bonuses

Summary

2025 ACF Nationals04/19/2025Y2115.71100%57%0%

Data

VirginiaChicago B001010
Cornell BIllinois B0101020
RutgersIllinois A0101020
FloridaIndiana001010
OttawaLSE001010
WUSTL BMichigan001010
British ColumbiaMinnesota0101020
NYURIT0101020
HarvardNorth Carolina A001010
NorthwesternUC Berkeley B001010
Johns HopkinsOhio State0101020
Virginia TechStanford001010
TexasArizona State0101020
Toronto AColumbia B0101020
Penn StateToronto C0101020
Georgia TechUC Berkeley A0101020
UCFGeorgia State0101020
Cornell AVanderbilt0101020
Waterloo BNorth Carolina B0101020
Winona StateChicago A001010
Iowa StateYale001010