Bandit Model Experiment Shows Users Lock Onto Familiar Options, Not the Best Ones

svk_roy · reddit · 2026-09-07

Reddit user svkroy built a simple bandit model to test whether usage converges on quality over time — and found it often doesn't.

• The model doesn't estimate which option is better; it just reinforces whatever gets used more, a miniature of habit formation
• Three update rules produce wildly different outcomes: one locks onto early winners regardless of quality; one mostly self-corrects but can still get stuck under strong reinforcement; a control always finds the true best
• Takeaway: users don't pick the best feature — they pick the one they know, so early exposure bias can permanently lock in a quality disadvantage

A neat demonstration of why product UX design and initial defaults matter so much.

Original post →

More from Research

• COLT 2027, the 40th Annual Conference on Learning Theory, Heads to Tokyo — FrnkNlsn · 2026-09-07
• HoloWorld unifies indoor-outdoor 3D urban generation with cross-scale context — Xiaobin Huang · 2026-09-07
• Dev Open-Sources Design of an LLM Memory Benchmark With Stale-Fact Scoring and Noise Scaling — True_Mongoose_7073 · 2026-09-07
• Engineer's 3-layer MLP classifies automotive radar targets, Macro F1 hits 0.764 — bruno_pinto90 · 2026-09-07
• Developer uses Claude Code to build a tiny 10k tok/s CPU model, shares 3 findings — GregoryDiamos · 2026-09-07
• LoopArena: Open Benchmark Tests Which LLMs Make the Best Runtime Controllers for Coding Agents — PepsiBetter · 2026-09-07

Research channel →