Steering Black-Box Large Language Model Agents with Pitfall Anticipation World Model

arXiv 2026 33 method

TLDR

Introduces Pitfall Anticipation World Model (PAWM), a selectively-invoked world model that predicts impending failures to steer black-box LLM agents, improving performance on ScienceWorld and WebArena-Lite.

评分理由

The paper presents a focused contribution: a world model that predicts pitfalls only when needed, with a two-stage training pipeline and empirical gains on established benchmarks. Strengths include the selective intervention design and clear task motivation; weaknesses are limited to abstract-level evidence, with no details on synthetic data generation or broader limitations.

Read-first 评分解释

综合优先阅读分 33，由主题、引用、图谱、方法、可复现性和近期性等信号加权得到。原始总分保留为 34。

近期性 5%

100

使用温和的时间衰减，让近期论文更容易浮现，同时保留较早基础工作的价值。年份：2026

方法质量 15%

检查可见的摘要与分析字段，寻找实验、数据集、基线、指标和局限性等方法证据。命中信号：基线、结果

主题相关性 25%

48.6

使用现有 LLM 关键词相关性评分，并归一化到 0-100。关键词：world model、world simulator、generative world model、interactive world model、video world model、world dynamics prediction、model-based reinforcement learning world model

可复现性 15%

检查链接和可见文本中的论文、代码、数据集、工件与仓库信号。论文：有；代码：无；数据：无；命中信号：无

图谱声望 15%

5.7

基于可用 OpenAlex 引用关系构建局部引用图，并计算 PageRank 风格的图谱声望。局部 PageRank：0.0016

引用影响力 15%

使用 OpenAlex 形态的引用元数据作为文献关注度信号，并与论文本身质量分开处理。引用数：0

引用速度 10%

引用速度按发表年限估算年均引用，降低旧论文天然占优的偏差。年均引用：0.00

研究版图角色

前沿论文方法锚点

排序敏感性

稳定性：volatile；排名波动范围：251。

关键词评分

world model

world dynamics prediction

model-based reinforcement learning world model

interactive world model

generative world model

world simulator

video world model

深度分析

创新点

Pitfall Anticipation World Model (PAWM) that selectively intervenes when a potential pitfall arises and predicts impending failure to steer the agent away
Shifts world model guidance from predicting next observations at each step to anticipating pitfalls within a bounded multi-step future
Two-stage training: Supervised Fine-Tuning on synthetic data labeled with agent pitfalls, followed by GRPO co-training with a black-box LLM
Effective world model guidance for black-box LLM agents without requiring internal access to the agent

方法

PAWM is a world model that selectively intervenes when a potential pitfall arises and predicts impending failure to steer a black-box LLM agent away from it. It is trained in two stages: Supervised Fine-Tuning on synthetic data labeled with agent pitfalls, followed by GRPO co-training with a black-box LLM in the target environment. The method is evaluated with Claude Sonnet 4.5 on ScienceWorld and WebArena-Lite against a range of baselines.

关键结果

PAWM improves Claude Sonnet 4.5 performance by 24.3% on ScienceWorld and 7.7% on WebArena-Lite, outperforming a range of baselines.

局限性

No explicit limitations are stated in the title or abstract.

资源

论文