Persona Policies (PPol): evolutionary framework for diverse LLM user simulators
natashajaques · x · 2026-10-06
Persona Policies (PPol), presented at the COLM 2026 Social Simulation Workshop, is now pip-installable (pip install ppol). Problem: LLM-based user simulators inherit their underlying models' cooperative, clear, homogeneous behavior — unlike messy real humans who falter, forget, and push back; hand-written personas are brittle and hard to scale. PPol is an evolutionary framework grounded in real dialogue traces that automatically discovers behaviors and instructions to generate diverse human-like user personas for any task, with examples on GitHub.
More from coding & agent
- Verifier Agents and Placebo Tests: Catching an AI Causal Analysis That Reasons Itself Wrong — hugobowne · 2026-10-06
- Claude Code 2.1.290 Ships 190 CLI Changes, Adds Sign-in Deny Button and Partial Session Names — ClaudeCodeLog · 2026-10-06
- OpenAI Median Researcher Burns $601/Day on Coding-Agent Inference, Up From Under $1 — rohanpaul_ai · 2026-10-06
- Custom agent harnesses are losing their edge as models train on official ones — pvncher · 2026-10-06
- Google Docs beats local markdown files for syncing agent context, dev finds — rseroter · 2026-10-06
- Sonbal: a local MCP execution substrate with bounded tools, written in Ada — hodong-kim · 2026-10-06