Idanwo ti o sopọ mọ itọsọna · Alabọde Ipele

Idanwo-Aago Iṣiro Ìdánwò

Review inference-time compute, candidate search, adaptive allocation and the cost-latency tradeoffs measured in test-time studies.

Jẹmọ itọsọna awọn ọnaIgbeyewo Time Iṣiro Iwon
Ibeere 1 ti 8

Kini 'iṣiro-akoko idanwo' tọka si?

Awotẹlẹ gbogbo ibeere8 ibeere

Awọn idahun ati awọn alaye han bi o ṣe n mu ibeere naa.

  1. Kini 'iṣiro-akoko idanwo' tọka si?
    • Iṣiro lo ikẹkọ awoṣe lori data
    • Awọn akoko ti o gba lati gba lati ayelujara a awoṣe
    • Iṣiro ti a lo nigbati awoṣe ba dahun ibeere kan ni itọkasi
    • Ina ti a lo lati tutu awọn ile-iṣẹ data
  2. In test-time scaling studies, what can extra inference compute be used for?
    • Changing the model’s training data during the request
    • Increasing the number of model parameters automatically
    • Guaranteeing correctness without evaluation
    • Searching or revising candidate solutions and checking their quality
  3. Kini iṣowo-pipa ti iṣawọn iṣiro akoko idanwo?
    • Iduroṣinṣin ti o ga julọ ni paṣipaarọ fun lairi diẹ sii ati idiyele
    • Kere iranti sugbon losokepupo ikẹkọ
    • Diẹ data sugbon kere si dede
    • Dara eya fun kekere o ga
  4. Kini iṣapẹẹrẹ 'dara julọ-ti-N'?
    • Ikẹkọ lori awọn N tobi datasets
    • Ṣiṣẹda awọn idahun oludije N ati yiyan eyi ti o dara julọ
    • Lilo awọn ọrọ N akọkọ nikan ti itọsi kan
    • Ṣiṣe awọn awoṣe lori N yatọ si awọn kọmputa fun iyara
  5. Kini idi ti iṣiro akoko idanwo ni a ka si 'apa irẹjẹ' tuntun?
    • O rọpo iwulo fun ikẹkọ eyikeyi
    • O ṣiṣẹ nikan lori awọn awoṣe kekere
    • O mu iṣẹ ṣiṣe pọ si nipa fifi iṣiro iṣiro kun kuku ju iwọn awoṣe dagba nikan
    • O din awọn nọmba ti awọn olumulo a awoṣe le sin
  6. Abajade iyalẹnu wo ni iṣiro akoko idanwo ṣiṣẹ?
    • Awọn awoṣe le ṣiṣẹ laisi eyikeyi GPUs
    • Data ikẹkọ di kobojumu
    • Awọn awoṣe nigbagbogbo dahun yiyara
    • Awoṣe kekere ti o ronu gun le dije awoṣe ti o tobi pupọ ti o dahun lẹsẹkẹsẹ
  7. What is one reason a fixed best-of-N strategy may waste inference compute?
    • It spends the same candidate budget on easy and difficult prompts
    • It modifies model weights on every query
    • It cannot generate multiple candidates
    • It eliminates the need for scoring
  8. What did the test-time compute paper compare with best-of-N sampling?
    • A strategy that disables inference
    • A compute-optimal strategy that allocates effort based on expected solution quality
    • A larger tokenizer vocabulary
    • A method that changes image resolution