Skip to main content

INFO 4340/5440: Sprint 2 Planning

Class 21: Sprint 2 Planning​


  • LLM Randomness
  • Standup Meeting
  • Sprint 2 Planning
  • Studio

Copyright 2026, Kyle J. Harms

LLM Randomness​

Why do responses change for the same prompt?​

LLMs generate responses one token at a time.

  1. Calculate a probability for every possible next token.
  2. Randomly sample one token using those probabilities.
  3. Add it to the response and repeat.

Step 2 can produce different responses to the same prompt.

temperature​


Temperature (0.0 to 2.0) changes the odds of which token is selected.


Lower temperature is more predictable. (Closer to 0)

Higher temperature can be more "creative". (Closer to 2)


temperature: 0 always responds with the most likely token.

Example: temperature​

Prompt: "The best pizza topping is..."

Next token00.512
pepperoni100%73%50%35%
mushrooms0%18%25%25%
sausage0%7%15%19%
onions0%1%7%13%
pineapple0%< 1%3%9%

WebLLM: temperature​


await llmEngine.generate(messages.value, {
  temperature: 0.2,
  onResponse: (text) => (response.value = text),
});

Model Defaults:

Llama-3.2-1B-Instruct-q4f16_1-MLC: temperature 0.6

Example: Consistency, not Creativity​

const messages = [
  {
    role: "system",
    content: `Given a review, assign a rating from 1 to 5,
              where 1 is the worst and 5 is the best.
              Output the rating as only JSON. Example: {"rating": 4}`,
  },
  { role: "user", content: "The food was okay, but the service was slow." },
];

const reply = await llmEngine.generate(messages, { temperature: 0, max_tokens: 20 });

Gotcha: Consistent does not guarantee correctness.

Project​

Activity: Standup Meeting​

Respect your team members by giving them your full attention; put away laptops and phones.

Stand Up.

  • What did I work on since the last time my group met?
  • What am I working on today?
  • What issues are blocking me?

Time limit: 5 minutes (Standup meeting should be short!)

Sprint 2​

Core Functionality​

Review: Sprint Planning​

  1. What is our goal for this sprint?
  2. Which backlog items can we complete in this sprint?
  3. How will the work get done, and who will do it?

Gotcha: Once planning is complete, the sprint plan is locked.

Activity: Sprint 2 Planning​

Working with your team, answer the two sprint planning questions on the handout provided.

Suggestion: Focus on your team's greatest source of uncertainty. (This is probably your "backend" type stuff – the core functionality. Getting the LLM working well is more important right now than the fun "frontend" user interface stuff.)

Studio: Project​

Use this time to work on sprint 1 or 2.


Today: Sprint 1 due

Next Thursday: Sprint 2 due

Last updated on