In-depth Analysis of Generative Behavior in Large Language Models: How Temperature Parameters and Sampling Strategies Shape Output Diversity

This article uses controlled experiments to conduct an in-depth analysis of the generative behavior of the locally deployed llama3:8b model, exploring how temperature parameters and nucleus sampling (top_p) affect output diversity and consistency, and providing empirical insights into understanding the randomness and controllability of LLMs. The experiment focuses on creative writing tasks, comparing output differences under different sampling configurations, and revealing how parameter interactions balance creativity and coherence.

The generative process of large language models is essentially a probabilistic sampling process, but users lack an intuitive understanding of the actual impact of parameters like temperature and top_p. This study takes variation as its object, systematically exploring how different sampling configurations shape output diversity through local experiments, helping developers precisely control model behavior and providing observable counterparts to theoretical concepts.

Model and Environment Configuration

• Model: llama3:8b via Ollama local service
• Environment: Python3.10+, no external API dependencies
• Randomness: No fixed seed, fresh sampling each time

Comparative Experiment Setup

Configuration	Temperature	top_p	Number of Runs
Low Variation	0.2	0.9	5 times
High Variation	0.9	0.95	5 times

Test Prompt

Write a 120-180 word product description for the fictional snack "Midnight Maple Pretzel Bites", including 3 sensory details, with a single-sentence slogan at the end.

Consistency Elements

• Task Structure: Strictly follows prompt format (description + slogan)
• Core Concepts: "Midnight" = late-night imagery, "Maple" = sweet tone, "Pretzel" = baked form
• Sensory Details: All meet the 3 requirements

Dimensions of Variation

• Surface Wording: Differences in sentence structure and adjectives
• Flavor Extension: High variation configuration adds black pepper, bourbon maple syrup, etc., beyond smoked sea salt
• Packaging Description: High variation shows variants like deep navy blue, gold foil crescent, etc.
• Slogan Creativity: Low variation converges to repetition; high variation is different each time
• Tone Style: Low variation is marketing copy; high variation is more casual and poetic

Temperature Parameter Mechanism

• Low Temperature (0.2): Sharp distribution, selects high-probability tokens, outputs are similar
• High Temperature (0.9): Flat distribution, low-probability tokens are selected, diversity increases

top_p Synergistic Effect

• 0.9 restricts to a tight nucleus, while 0.95 opens up rare tokens; when combined with high temperature, it amplifies variation

Configuration Comparison

Dimension	Low Variation	High Variation
Diversity	Low (approximate rewriting)	High (unique)
Creativity	Safe and predictable	Unexpected
Stability	Stable across runs	Independent outputs
Repetition Risk	High	Low
Drift Risk	Low	Relatively high

Value of Variation

• Supports open-ended tasks (creative writing, brainstorming)
• Reflects honest uncertainty, avoiding pretending there is a single correct answer

Disadvantages of Forcing Identical Outputs

• Discards model knowledge and impairs creativity
• Hides uncertainty and makes it hard to recover from poor completions

Application Scenarios

• Low Temperature: Factual Q&A, code generation, structured extraction
• High Temperature: Creative writing, marketing variations, art projects
• Balanced Strategy: Medium temperature (0.5-0.7) + top_p (0.9-0.95)

This study translates abstract probabilistic sampling theory into observable behavior, revealing that LLM outputs can be understood and controlled through parameters. For developers, it allows adjusting the balance between creativity and stability; for researchers, it provides a replicable framework; for users, it is a valuable lesson in controllable randomness—finding a balance between structural constraints and free creation.