ScienceWorld
A long-horizon embodied science task benchmark requiring agents to complete complex scientific experiment tasks within a simulated environment. It is used to evaluate agents' long-horizon planning and embodied reasoning capabilities.
Timeline (last 90 days)
This reproduction study covers four agent task environments—ALFWorld, ScienceWorld, Search-QA, and WebShop—spans three model scales, and compares 10 representative self-distillation methods against a GRPO baseline.
CoPE was validated on five benchmarks: ALFWorld, ScienceWorld, WebShop, τ²-bench, and AppWorld.
On ScienceWorld, CoPE's sample efficiency improves by up to 1.9x compared to ECHO and up to 2.0x compared to GRPO.
SkillScaffold was evaluated on four benchmarks—ALFWorld, WebShop, ScienceWorld, and τ²-Bench—and surpassed strong agentic RL baselines and skill-driven RL baselines on both short-horizon and long-horizon tasks.
ScienceWorld is built on the TextWorld engine and requires agents to complete scientific reasoning tasks in physics, chemistry, and other disciplines within a text-simulated laboratory.
The research team validated SAGA on two interactive environment benchmarks, ScienceWorld and ALFWorld, achieving improved task completion performance.
The authors evaluated RLDS on four agent benchmarks: FrozenLake, HotpotQA, ScienceWorld, and DeepResearch.
RLDS reduces wall-clock time per step on ScienceWorld by 10.9% compared to scalar GRPO, because long trajectories amortize the fixed overhead of reflection and scoring.
RLDS achieved an improvement of +11.5 points on ScienceWorld (95% confidence interval [+9.8, +13.3]) and +9.8 points on FrozenLake ([+7.0, +12.8]).
All Facts (9)
This reproduction study covers four agent task environments—ALFWorld, ScienceWorld, Search-QA, and WebShop—spans three model scales, and compares 10 representative self-distillation methods against a GRPO baseline.
50%UnverifiedCoPE was validated on five benchmarks: ALFWorld, ScienceWorld, WebShop, τ²-bench, and AppWorld.
50%UnverifiedOn ScienceWorld, CoPE's sample efficiency improves by up to 1.9x compared to ECHO and up to 2.0x compared to GRPO.
50%UnverifiedSkillScaffold was evaluated on four benchmarks—ALFWorld, WebShop, ScienceWorld, and τ²-Bench—and surpassed strong agentic RL baselines and skill-driven RL baselines on both short-horizon and long-horizon tasks.
50%UnverifiedScienceWorld is built on the TextWorld engine and requires agents to complete scientific reasoning tasks in physics, chemistry, and other disciplines within a text-simulated laboratory.
50%UnverifiedThe research team validated SAGA on two interactive environment benchmarks, ScienceWorld and ALFWorld, achieving improved task completion performance.
50%UnverifiedThe authors evaluated RLDS on four agent benchmarks: FrozenLake, HotpotQA, ScienceWorld, and DeepResearch.
50%UnverifiedRLDS reduces wall-clock time per step on ScienceWorld by 10.9% compared to scalar GRPO, because long trajectories amortize the fixed overhead of reflection and scoring.
50%UnverifiedRLDS achieved an improvement of +11.5 points on ScienceWorld (95% confidence interval [+9.8, +13.3]) and +9.8 points on FrozenLake ([+7.0, +12.8]).
50%