Zixuan Wang^1,2, Yuchen Yan¹, Hongxing Li¹, Teng Pan^1,2, Dingming Li¹, Ruiqing Zhang²,
Weiming Lu¹, Jun Xiao¹, Yueting Zhuang¹, Yongliang Shen^1,†
1Zhejiang University  ·  ²Baidu Inc.
_{^†Corresponding author}

This repository contains the official implementation of BEACON, a milestone-guided policy learning framework that addresses two pathologies of trajectory-level RL on long-horizon language-agent tasks: credit misattribution (correct early actions penalized by terminal failure) and sample inefficiency (partial successes wasted under sparse rewards). The implementation is built on top of verl-agent; only the components contributed by this work are documented here.

• Consistent gains on three long-horizon benchmarks with a single set of hyperparameters ($\gamma=0.95$, $\lambda=1.0$).
• Horizon-dependent gains. On ALFWorld Long tasks, BEACON reaches 92.9% vs. 53.5% for GRPO. Relative gains over GRPO scale from +26.2% (Short) to +73.6% (Long).
• Recovers learning signal from partial successes. Effective sample utilization improves from 23.7% to 82.0% on ALFWorld.
• Outperforms behavior cloning. 91.4% vs. 43% for SFT on oracle trajectories — gains stem from policy optimization, not milestone imitation.

_{Main results. BEACON outperforms GRPO and GiGPO across ALFWorld, ScienceWorld, and WebShop at both 1.5B and 7B scales.}

BEACON operates in three stages:

1. Trajectory partitioning. A milestone indicator $\Phi$ identifies verifiable subgoal-completion transitions, splitting each trajectory into segments at milestone boundaries. $\Phi$ is environment-defined and requires no learned model: ALFWorld uses object/state predicates, WebShop uses page-transition phases, ScienceWorld exposes subgoal_completed directly.
2. Temporal reward shaping. Within each completed segment, actions receive shaped reward $r_t = R_{\text{ms}} \cdot \gamma^{t_k - t}$, giving graduated positive credit to actions leading up to a milestone and converting partial successes into learning signal.
3. Dual-scale advantage estimation. Trajectory-level advantage (GRPO-style) captures global task performance; segment-level advantage compares only among trajectories that reached the same milestone, isolating local action quality from variance in later segments. The two are combined as $\hat{A}_{i,t} = A^{\text{traj}}i + \lambda \cdot A^{\text{seg}}{i,t}$.

At update time, BEACON automatically routes each batch based on which milestone field is present (trial_id for ALFWorld, milestone_achieved for WebShop, subgoal_completed for ScienceWorld), so a single training pipeline supports all three environments without environment-specific code paths in the trainer.

migpo/                   # BEACON core: advantage / step-reward computation and milestone detector
agent_system/            # ALFWorld, WebShop, and ScienceWorld environment integrations
examples/migpo_trainer/  # Paper-locked training scripts (one per environment)

Everything else is inherited from the upstream verl-agent framework.

conda create -n verl-agent python==3.12 -y
conda activate verl-agent

pip3 install torch==2.6.0 --index-url https://download.pytorch.org/whl/cu124
pip3 install flash-attn==2.7.4.post1 --no-build-isolation
pip3 install -e .
pip3 install vllm==0.8.5

Each environment below is best installed in its own dedicated conda environment to avoid dependency conflicts.

pip3 install gymnasium==0.29.1
pip3 install stable-baselines3==2.6.0
pip install alfworld
pip install vllm==0.8.5

# Download PDDL & Game files and the pre-trained MaskRCNN detector
alfworld-download -f

WebShop requires Python ≤ 3.10:

conda create -n verl-agent-webshop python==3.10 -y
conda activate verl-agent-webshop

cd ./agent_system/environments/env_package/webshop/webshop
./setup.sh -d all

cd repo_root/
pip3 install torch==2.6.0 --index-url https://download.pytorch.org/whl/cu124
pip3 install flash-attn==2.7.4.post1 --no-build-isolation
pip3 install -e .
pip3 install vllm==0.8.2

ScienceWorld requires Java 1.8+ and Python ≤ 3.10:

conda create -n verl-agent-sciworld python==3.10 -y
conda activate verl-agent-sciworld

cd repo_root/
pip3 install torch==2.6.0+cu124 --index-url https://download.pytorch.org/whl/cu124
pip3 install flash-attn==2.7.4.post1 --no-build-isolation
pip3 install -e .
pip3 install vllm==0.8.2

# Java via conda (not system-wide)
conda install -c conda-forge openjdk=11 -y

# ScienceWorld ships its own bundled JAR and the py4j bridge
pip install scienceworld

Variation indices used by our experiments are included at agent_system/environments/env_package/sciworld/variations_idx/.

Sanity check:

python -c "from scienceworld import ScienceWorldEnv; print('ScienceWorld import successful')"

Paper-locked training scripts (Qwen2.5-1.5B-Instruct, single 8-GPU node) live in examples/migpo_trainer/:

bash examples/migpo_trainer/run_alfworld.sh    # ALFWorld
bash examples/migpo_trainer/run_webshop.sh     # WebShop
bash examples/migpo_trainer/run_sciworld.sh    # ScienceWorld

This codebase builds on verl-agent, which itself extends veRL. We thank the authors of those projects, and the maintainers of the supported environments — ALFWorld, WebShop, and ScienceWorld.

@misc{wang2026milestoneguidedpolicylearninglonghorizon,
  title         = {Milestone-Guided Policy Learning for Long-Horizon Language Agents},
  author        = {Zixuan Wang and Yuchen Yan and Hongxing Li and Teng Pan and Dingming Li and Ruiqing Zhang and Weiming Lu and Jun Xiao and Yueting Zhuang and Yongliang Shen},
  year          = {2026},
  eprint        = {2605.06078},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  url           = {https://arxiv.org/abs/2605.06078},
}