WM/W2R trajectories of WebShop TM-WM (step60/80/92) with Qwen3-8B and Qwen3-32B agents at 40k context.
YOULING HUANG
Ricardo-H
AI & ML interests
None yet
Recent Activity
liked a model 1 day ago
Kwaipilot/KAT-Coder-V2.5-Dev updated a collection 1 day ago
KAT upvoted a collection 1 day ago
KATOrganizations
WM/W2R trajectories of Ricardo-H/tw-wm-token-match-llama-step171 on TextWorld test split.
WM/W2R trajectories of Ricardo-H/tw-wm-tm-0501-step-170 on TextWorld test split, evaluated by various agents.
Step92 WebShop world-model trajectories and W2R replays for Qwen and GPT agents.
Archived checkpoints from the terminated OCAR/surprise-as-credit line. See post-mortem in the verl-agent GitHub repo.
RLVR-World style Token F1 reward ablation models for BehR-WM rebuttal experiments
Exp3-FactR: FactR-Only GRPO ablation. Exponential(a=1), behavior=0, facts=1. 240 steps, Qwen2.5-7B.
WM/W2R trajectories of Ricardo-H/BehR-WorldModel-Textworld-Llama3.1-8B on TextWorld test split.
WM/W2R trajectories of WebShop TM-WM checkpoints (step60/80/...) evaluated with Qwen3-32B agent at 32k context.
Models and datasets from the BehR paper. Code: https://github.com/Ricardo-H/behr-wm
WebShop World Model - LLaMA3.1-8B BehR-Only GRPO checkpoints (2026-02-27)
WM/W2R trajectories of WebShop TM-WM (step60/80/92) with Qwen3-8B and Qwen3-32B agents at 40k context.
WM/W2R trajectories of Ricardo-H/BehR-WorldModel-Textworld-Llama3.1-8B on TextWorld test split.
WM/W2R trajectories of Ricardo-H/tw-wm-token-match-llama-step171 on TextWorld test split.
WM/W2R trajectories of WebShop TM-WM checkpoints (step60/80/...) evaluated with Qwen3-32B agent at 32k context.
WM/W2R trajectories of Ricardo-H/tw-wm-tm-0501-step-170 on TextWorld test split, evaluated by various agents.
Step92 WebShop world-model trajectories and W2R replays for Qwen and GPT agents.
Archived checkpoints from the terminated OCAR/surprise-as-credit line. See post-mortem in the verl-agent GitHub repo.
Models and datasets from the BehR paper. Code: https://github.com/Ricardo-H/behr-wm
RLVR-World style Token F1 reward ablation models for BehR-WM rebuttal experiments
WebShop World Model - LLaMA3.1-8B BehR-Only GRPO checkpoints (2026-02-27)
Exp3-FactR: FactR-Only GRPO ablation. Exponential(a=1), behavior=0, facts=1. 240 steps, Qwen2.5-7B.