Appau Gideon Kofi Amo

Appau Gideon Kofi Amo

M.S. student at USTC working on optimizer-level methods for long-tailed visual recognition and on world models for decision-making.

Language-Guided Navigation Agent

An end-to-end language-agent project (github.com/Amo-Gideon/rl-portfolio/language-guided-navigation): a Qwen2.5-0.5B-Instruct model learns to act in a custom text-based grid world, following instructions like “Go to the red house and report its color” and emitting structured JSON actions (move_forward, turn_left, turn_right, look, answer).

  • Environment: a deterministic grid world with landmarks and verifiable, decomposed rewards (answer accuracy, format validity, action validity)
  • SFT stage: LoRA fine-tuning on shortest-path expert trajectories generated with BFS
  • RL stage: REINFORCE with mean baseline and a KL penalty against a frozen reference model - the same SFT-then-RL recipe as large-scale RLHF, reproduced from scratch
  • Results: success rate improves from 9/20 (45%) after SFT to 11/20 (55%) after RL fine-tuning, with seeded evaluation for reproducibility

Honest current limits, documented in the repo README: the 15-step episode cap and sparse reward leave headroom (the agent sometimes wanders before locating the target).