GAVEL 用图世界模型让 Qwen3-8B 长程机器人任务成功率升至 91.8%

DAIR.AI · @dair_ai · X·2026-09-20 10:10·1小时前
AI 导读

GAVEL 在不改动模型的前提下,将 Qwen3-8B 在长程机器人任务上的成功率从 41.2% 提升至 91.8%,增益全部来自外部 harness。它维护一个显式图世界模型,记录物体关系、动作前置条件与效果,以及未观测物体位置的概率信念,在动作执行前预测后果,可直接修复的违规无需再次调用模型。

DAIR.AI@dair_ai
46AI 编辑部评分,满分 100

GAVEL 用图世界模型让 Qwen3-8B 长程机器人任务成功率升至 91.8%

2026-09-20 10:10· 1小时前
AI 导读

GAVEL 在不改动模型的前提下,将 Qwen3-8B 在长程机器人任务上的成功率从 41.2% 提升至 91.8%,增益全部来自外部 harness。它维护一个显式图世界模型,记录物体关系、动作前置条件与效果,以及未观测物体位置的概率信念,在动作执行前预测后果,可直接修复的违规无需再次调用模型。

Impressive paper showing the impact of a good harness.

Improves Qwen3-8B from 41.2% to 91.8% on long-horizon robot tasks without any change to the model.

The gain comes from the harness.

GAVEL keeps an explicit graph world model holding object relations, action preconditions and effects, and probabilistic beliefs about where unobserved objects are.

Before the robot executes an LLM-generated action, the graph predicts what that action would do.

Violations get caught, and the ones whose fix follows directly from the world model get repaired without calling the model again.

Only errors that need semantic reasoning go back to the LLM.

On BEHAVIOR-1K across 500 multi-task instructions, success rises from 19.9% to 92.6%. Reasoning over the distribution of possible object locations also reorders the remaining subtasks and cuts travel distance about 5.4%.

Why does it matter?

Many long-horizon agent failures are state-tracking failures rather than reasoning failures, and a symbolic model sitting outside the LLM catches them cheaply.

Paper: https://academy.dair.ai/papers/gavel-graph-world-models-for-verified-and-efficient-long-horizon-llm-task-planni-2609.19315

来源:DAIR.AI· x.com