textworld

This dataset contains sequences of states generated by playing Microsoft TextWorld. Each state contains graphs, descriptions, previous action and valid actions.

Deep reinforcement learning (PPO/DQN) agents that follow natural-language route instructions in indoor environments. Code for IJGIS 2025 paper.

ReAct (Reasoning and Acting) agent built from scratch in Python, adapted to interact with Microsoft’s TextWorld environments.

Reproducible MSc thesis materials on experience retrieval for an LLM agent in TextWorldExpress CookingWorld.

twagents is a framework for building and comparing language model agents with different memory architectures in interactive, text-based environments. It is designed to work with TextWorld and focuses on exploring how working, episodic, and semantic memory influence agent behavior, recall, and decision-making.

LoRA fine-tuning of LFM2.5 (230M–2.6B) on TextWorld cooking games. DAgger takes the 1.2B model from 75% to 97.8% of test games won, and the 230M model with DAgger (87.8%) beats the 2.6B model trained without it.

A frozen decision model (JEV) as an LLM agent's world model, compared with zero-shot and LoRA-fine-tuned generative LLMs in TextWorld. Code, logs, analysis and manuscript for KIIS 2026 Fall (oral). LLM 에이전트의 월드모델: 결정모델과 범용 LLM의 비교

Adaptive memory for LLM agents in sparse environments: FadeMem extended with fact compression, predictive shielding & dormant reactivation. Benchmarked 6 memory designs × 3 LLMs on TextWorld & BabyAI.

A repo for my MSc project