Lesson 15 — Beginner

Reasoning Models & Test-Time Compute

Kuch models jawab dene se pehle andar hi kaafi soch lete hain — aur us sochne ka compute tum par hi aata hai. Is lesson mein hum dekhenge test-time compute kya hai, thinking budget kaise set hota hai, kaunse tasks par extra soch paisa vasool hai aur kaunse par bekaar ka kharcha. Saath mein — reasoning models ko prompt kaise karein.

⏱ 10 min• Beginner• Sampling & Generation Parameters

Reasoning models kya karte hain

WHAT

Reasoning model wo model hai jo jawab dene se pehle apne andar hi kaafi steps soch leta hai — internal reasoning tokens bana kar. Test-time compute ka matlab yahi hai: training ke baad, jawab ke waqt kharch hone wala extra compute.

WHEN

Mushkil tasks par — complex maths, multi-step coding, planning — jahan galat jawab ka nuksaan zyada ho. Aur tab jab latency ki fikr kam ho, quality ki zyada. Aasan rozmarra tasks par iski zarurat nahi.

WHERE

Providers ke alag reasoning ya thinking wale models, ya usi model ke alag reasoning-effort modes. API mein thinking budget, reasoning-effort jaise parameters milte hain — naam alag-alag hote hain, docs padho.

Normal model turant likhna shuru karta hai. Reasoning model pehle andar ek draft (reasoning) bana leta hai, usse check karta hai, phir final jawab deta hai. Yahi test-time compute hai — training-time compute ke baad, inference ke waqt kharch hone wala extra compute. Jitna zyada sochne ka budget, utni zyada der aur mehengi ticket.

Kahan reasoning helpful hai

Difficult mathematics aur word problems — jahan steps todna hi kaam hai.
Complex coding — architecture, edge cases, multi-file debugging jaise kaam.
Planning aur multi-step logic — jahan pehle sochna zaroori hai, warna raasta hi galat ho jaayega.
Constraint-heavy problems — jab ek saath 5-6 conditions lagani hon aur unke beech conflict bhi ho sakta hai.

Common sense rule: Jab galat jawab ka nuksaan zyada ho aur waqt ka pressure kam ho — tab reasoning model ka extra paisa vasool hai. Batch job mein deri chalegi, customer ke chat mein nahi.

Kahan reasoning bekaar kharcha hai

TASK CHECK (quick)
  Multi-step? Constraints zyada? Har galti mehngi?
    → REASONING MODEL  (sochne ka paisa do)

  Sentiment? Simple extraction? Rewrite? Chhota summary?
    → FAST MODEL        (reasoning bekaar kharcha)

Yaad rakho: reasoning ke baad bhi jawab galat ho sakta hai —
extra soch guarantee nahi, sirf behtar chances.

Sentiment classification, simple extraction, rewriting, straightforward summarization — yahan extra soch ka faayda nahi, sirf bill aur deri.
High-volume pipelines mein har request par reasoning = cost explosion. Chhote kaam ke chhote model, mushkil ke bade — yahi routing ki neev hai (L9 mein model routing).

Quality, cost aur latency ka trade-off

Reasoning ka trade-off ekdum seedha hai: quality upar ja sakti hai, par cost aur latency dono badhte hain. Kai providers isse control karne ka raasta dete hain — reasoning effort ya thinking budget set karo (low/medium/high jaisa scale). Kam budget = tez aur sasta; zyada = dheere aur mehnga.

Har task par high reasoning dena ek common galti hai — user-facing chat mein deri dikhegi, bill bhi.
Batch pipelines mein budget thoda badha sakte ho — wahan latency ki fikr nahi, quality ki.
Compare karke dekho: kabhi-kabhi chhota fine-tuned model ya achha few-shot prompt bhi reasoning model jaisa result de sakta hai — saste mein.

Reasoning models ko prompt kaise karein

"Think step by step" likhne ki zarurat nahi — wo trick purane models ke liye thi; reasoning model andar khud sochta hai.
Uski jagah: goal saaf batao, relevant context do, constraints likho, aur batao ki achha jawab kaisa dikhta hai.
Reasoning ka output alag section mein aata hai — final answer hi use karo, sochne ka hissa parse karne ki zarurat nahi.

Interview line: Reasoning models ke saath prompt ka tarika badalta hai — "think step by step" ki zarurat nahi; goal, context, constraints aur good-answer definition do. Extra soch ka kharcha model par aa raha hai, control bhi tumhare paas hai (budget).

Exercise

Question: Inference ke waqt jawab se pehle kharch hone wala extra compute kya kehlaata hai?

Question: Sentiment classification ke liye reasoning model chahiye — haan ya na?

Question: Reasoning models ko prompt mein "think step by step" likhne ki zarurat hoti hai — haan ya na?

Question: Reasoning budget badhane se do cheezein sure-shot badhti hain — kaunsi? (spelling dhyan se)

Common mistakes

Har task pe reasoning model chunna: Sentiment aur extraction par extra soch bekaar kharcha hai — cost aur latency dono phati hai, quality same. Task ki complexity pehle dekho, phir model chuno.
Sochna reasoning se model ke facts better ho jaate hain: Reasoning model lamba sochta hai, par jaanta wahi hai jo training se aaya. Extra soch ke baad bhi galti ho sakti hai — guarantee nahi, sirf behtar chances.
Purani "think step by step" tricks ratna: Reasoning model khud internally sochta hai — uske bajaye goal, context aur constraints saaf likho. Uska asar zyada hai aur token bhi bachte hain.

Desi example: Exam hall ka topper — test-time compute ka desi analogy

Normal student seedha jawab likh deta hai; topper pehle rough work karta hai. Simple fill-in-the-blank ke liye rough work bekaar; word problem ke liye zaroori. Rough work hi test-time compute hai — aur rough work karke bhi answer galat ho sakta hai:

Fill-in-the-blank  : seedha jawab (fast model, no rough work)
Word problem       : rough work + check + jawab (reasoning model)
Rough work = extra compute = extra time = (kabhi kabhi) behtar marks
Aur haan — rough work guarantee nahi hai, sirf chances.

Interview angle — yahan se sawaal aate hain

Q: Test-time compute kya hai? — Inference ke waqt jawab se pehle kharch hone wala extra compute (reasoning tokens) — quality badh sakti hai, cost-latency bhi.
Q: Kab reasoning model chunna chahiye? — Multi-step, constraint-heavy, mushkil tasks par; simple extraction, classification, rewriting par nahi.
Q: Reasoning models ko prompt kaise karein? — "Think step by step" ki zarurat nahi; goal, context, constraints aur good-answer definition do.
Q: Reasoning ka trade-off kya hai? — Quality behtar ho sakti hai par cost aur latency definitely badhte hain — budget set karke control karo.

Lesson complete?

Reasoning ka trade-off clear. Agla lesson — The Context Window — mein dekhenge request ka workspace kya khaata hai, context kyun kat-ta hai, aur bada window kyun kaafi nahi hota.