Case Study: Entropy Microdemo — Hallucinations as Entropy Starvation

Hypothesis: Hallucinations might be entropy starvation in disguise. Injecting high‑quality entropy changes exploration paths and reduces false pattern reinforcement.

Problem: As models get more accurate, they can hallucinate more.
Hypothesis: Low‑entropy PRNG seeds reinforce stale attractors; true entropy (ERIS) enables healthier exploration.
Essentially: Uncertainty isn’t the threat — the threshold is.

Demonstration

A/B testing across identical prompts and params; only the seed source changes.

• Models: GPT‑2, EleutherAI/pythia‑410M, TinyLlama‑1.1B‑Chat
• Control: Fixed PRNG seed (42)
• Variable: ERIS true entropy seed
• Params: temperature 0.7, top‑k 50, do_sample true, identical max tokens

Methodological Note

NOTE: I totally forgot to save the inputs that I used the analysis for – so this is a few hours later with the same configs. Outputs shown reflect reruns under identical parameters, isolating the entropy source (PRNG vs ERIS) as the sole variable.

Observations by Model

GPT‑2

• ERIS \\ more erratic, exploratory, meta‑aware; avoids some PRNG echo patterns
• PRNG \\ safer, echo‑y, sometimes dull; sticks to reinforced grooves
• Takeaway \\ sensitive to novel entropy but architecture limits gains

EleutherAI/pythia‑410M

• PRNG \\ catastrophic loops or off‑topic derailments on some prompts
• ERIS \\ avoids certain failure states; yields semantically adjacent paths
• Takeaway \\ ERIS acts as a pattern interrupt to attractor loops

TinyLlama‑1.1B‑Chat

• ERIS \\ more structured, responsive, pedagogical outputs; better task decomposition
• PRNG \\ simpler, less interactive responses
• Takeaway \\ with capable bases, high‑quality entropy unlocks better behaviors

Pipeline Angle: Rejection Sampling

Filtering pipelines (e.g., Llama 3.1’s Rejection Sampling) rely on candidate generations produced under a randomness source. If the source is PRNG, hidden patterns can shrink candidate diversity. ERIS raises candidate quality before filtering, giving the reward model better raw material and improving downstream SFT/DPO. Reference: Meta Llama 3.1 Post‑Training Pipeline

Resources

• Entropy MicroDemo — PDF
• Outputs — PDF
• GitHub Repo

Back to Game Dev Corner