Finalist
Play-gent
Our agent uses OpenEnv to build a curriculum of video game environments that train a TinyLlama 1.1B agent via GRPO reinforcement learning.
Core trick
TinyLlama trained with GRPO across Diplomacy and poker environments, then transfers zero-shot to negotiating price against an adversarial seller.
Description
Our agent uses OpenEnv to build a curriculum of video game environments that train a TinyLlama 1.1B agent via GRPO reinforcement learning. We implement three OpenEnv-compliant environments — Diplomacy (coalition tactics), webDiplomacy human gameplay (211k real states), and IRC poker (bluff detection) — each with dense reward signals that teach the core primitives of strategic negotiation. The agent learns through RL across this curriculum: Phase 1 trains coalition pressure in Diplomacy, Phase 2 grounds it in human gameplay patterns, Phase 3 unifies all three reward signals in a live arbitrage environment. The result is an agent that transfers video game negotiation skills to real economic interactions — starting with $20, it compounds capital to $80 (4x return) against Groq Llama 3.1 8B adversarial sellers, detecting bluffs at 97% confidence and reaching actual price floors 100% of the time.
README excerpt
# Playgent Curriculum-trained RL negotiation agent. Work in progress. ## What it is TinyLlama 1.1B trained on game data to negotiate in real-world marketplace scenarios. No marketplace data used in training — the model learns strategic behavior from Diplomacy and poker, then transfers zero-shot to price negotiation. ## Training pipeline 1. SFT on 40 Cicero Diplomacy games (5,924 examples) 2. GRPO self-play on Diplomacy (win/loss/draw reward) 3. GRPO self-play on poker (net profit/loss reward) ## Results Evaluated zero-shot against an adversarial Llama-3.1-8B seller with hard floor enforcement. - 98% deal closure rate - 1.302x average capital return - 1.664x best return - Base TinyLlama baseline: 0% closure ## Stack PyTorch, HuggingFace Transformers, TRL ## Status Active development. Poker GRPO currently causes over-folding behavior in marketplace eval — reward signal redesign in progress.