acceptodds
Under review as a conference paper at ICLR 2027

PFP-World: Pre-Policy Planning for Web Agents with a Page Functional Primitive-based World Model

Abstract

Large language model-based web agents struggle with long-horizon cross-page web navigation, as they cannot proactively anticipate environmental dynamics. World models could resolve this by simulating page state transitions in advance, yet existing web world models mainly serve post-hoc purposes: they either generate full trajectories for policy training or evaluate atomic actions derived from flawed subgoal reasoning, lacking a mechanism to screen candidate page-level goals prior to execution. To fill this gap, we propose PFP-World, a pre-policy world-model planner built on Page Functional Primitives (PFPs). PFPs denote page-level functional intentions, serving as unified semantic abstractions that group atomic actions by shared functional intent (e.g., search, checkout). We encode each page state as an inventory of available PFPs. PFP-World redefines three core components of conventional web world models: it shifts the world model’s role from a post-hoc trajectory generator or policy evaluator to a pre-policy planner, replaces raw HTML DOM state encoding with abstract PFP inventories, and uses semantic PFPs as high-level control signals for state transition instead of atomic action spaces. With our PFP-World, the planned primitives guide downstream action grounding by mapping each high-level functional intent to executable web operations. We further introduce a page-centric bottom-up paradigm to learn site dynamics from PFP transitions without relying on complete task trajectories. Experiments on the standard WebArena benchmark and a self-constructed cross-page benchmark (namely CrossPageBench) show that PFP-World-SFT outperforms existing world-model-based baselines, bringing a 20.7 percentage-point gain in overall success rate over the reproduced WMA baseline on WebArena and gains of 19.8 and 12.8 percentage points in average target completion over the reproduced RLVR-World baseline on the Shopping and GitLab sites of CrossPageBench, respectively.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
web agent, LLM, world model

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_aRnCOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces PFP-WORLD, a pre-policy world-model planner built on Page Functional Primitives (PFPs). It represents each page state as a primitive inventory, simulates candidate PFP trajectories, and passes the selected trajectory’s parameterized first PFP to action grounding for execution. The authors propose a page-centered, bottom-up data collection procedure and implement Retrieval, SFT, and Prompt variants. On four WebArena sites, PFP-WORLD-SFT achieves 36.9% success compared with 16.2% for the reproduced WMA baseline, under a reporting protocol that takes each task’s highest score across three executions.

Critical strengths

This paper presents a clear separation between page-level functional planning and atomic-action execution. PFPs provide a practical abstraction for reusing functional page types and local transitions across page instances. The evaluation includes WebArena, CrossPageBench, rollout-depth ablations, and cross-site adaptation. In particular, the comparison between direct PFP selection and predictive rollout provides useful evidence that multi-step prediction improves performance within the proposed framework.

Critical weaknesses
  1. DYNAMIC SIMULATOR receives only the primitive inventory and a general PFP, without bound slot values or execution history. The reported inventory-matching accuracy does not establish whether it captures parameter-dependent outcomes needed for planning.
  2. The contribution of predictive planning is not sufficiently isolated. Comparisons change multiple components, including state representation, control granularity, and feedback procedures. Although rollout-depth ablations are provided, inference-budget-matched controls are needed to distinguish predictive planning from additional computation.
Questions for authors

Please see weaknesses above.

Reviewer_j8SNOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces a system that (1) turns webpages into Page Functional Primitives (PFP) inventories, which are lists of small tasks / actions that can be done on a page, (2) chooses, from this list, the PFPs that are most helpful to the overall goal, and (3) uses a world model to predict what future PFP inventories would be generated from the action plans, in order to select the best one. To do this, WebArena web page data are crawled, validated by humans, and used to group web pages by their overall functions. Each group then shares the same PFP inventory. Training and evaluation of this system are mostly done on WebArena, with some small-scale transfer learning results on real-world shopping websites.

Critical strengths
  1. The motivation makes sense: a world model that does not simply rank options offered by a planning policy at test time can be an interesting approach.

  2. The raw-website-to-PFP abstraction seems to be an interesting and intuitive idea worth introducing to the research community.

Critical weaknesses
  1. It seems that the majority of the training and testing data are collected from WebArena. This raises the question of whether a model trained on WebArena alone can generalize to other websites, especially real-world ones that could be more complex. The paper does offer a transfer study, but it uses a very small dataset and shows mixed results, if I understand correctly. In my opinion, the paper also does not put enough emphasis on this transfer capability, given that most of the results on it are in the Appendix.

  2. Related to the point above, to strengthen the generalization claim, more standard web / CUA benchmarks (Mind2Web, and OSWorld if possible) should be tested.

  3. Continuing with the generalization argument: if page types are collected through crawling, and every page of one type shares the same PFP inventory, then generalization to other websites / domains could be an issue, since it is likely that new page types will be introduced, and page types are reviewed and edited by humans.

  4. The framing of the paper overall seems slightly off / disconnected from the premise. It does not seem that the paper is (just) introducing a CUA / web "world model"; rather, it seems to be proposing a new approach to agent design. For example, the use of PFPs and the webpage-to-PFP-inventory "abstractor" is a central contribution of the paper, and cannot be ablated from the agent / planner design in the paper. If the claim is specifically to offer a "world model" of the web environment, then it seems that the focus should be on the transition prediction capability of the "dynamic simulator", which does not seem to be the case at the moment.

Minor:

  • Writing: the PFP concept really needs an example early on to help readers understand what it is. Right now, it is very confusing for a first-time reader.
  • Related work: the paper should probably include prior work related to PFP-like ideas. This sounds a lot like subtask decomposition during planning.
Questions for authors

WebArena also has a Map-related site. What is the reason for not using it?

Reviewer_sdRZOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies how world models can help LLM web agents with long, multi-page tasks. Existing web world models score atomic actions after the policy proposes them. The authors instead propose PFP-WORLD, which plans before the policy acts. Each page is represented by its inventory of Page Functional Primitives (PFPs), which are page-level functions such as search or checkout. A selector proposes candidate PFPs, and a world model predicts how each one changes the page’s inventory over a few steps. The first PFP of the best plan is then passed to the action policy to execute. On WebArena, the fine-tuned variant reaches 36.9% success, compared with 16.2% for a reproduced WMA baseline.

Critical strengths
  • The idea of planning over page-level functions before generating atomic actions is intuitive and well-motivated. Representing a page by what can be done on it is also appealing for robustness to layout changes.
  • Collecting site dynamics page by page, independent of task goals, is a reasonable way to reuse transitions across tasks. The controlled comparison with Plan–Action–Reflect shows a clear gain from planning over PFPs (+12.8 and +15.7 points), since both share the same backbone, grounding, and task evaluation.
Critical weaknesses

Main concerns are:

  • Weak WebArena baselines: The only WebArena baselines are WMA and RLVR-World reproduced with gpt-5.2, which score 16.2% and 14.1%. No plain gpt-5.2 agent is reported. For reference, AWM reports 35.5% on WebArena with GPT-4 in its own paper. That is comparable to PFP-WORLD’s 34.0–36.9% on four WebArena sites, which are also the best of three runs per task.
  • Unclear what drives the gains: PFP-WORLD changes many things at once: the PFP representation, lookahead, a judge that retries failed steps, human-validated site data, and gpt-5.2 modules. Only rollout depth and judge feedback are ablated. The three variants also score within 3 points of each other on CrossPageBench, even though the Prompt variant uses no collected data. So it is unclear how much the learned site dynamics matter.
  • Reliance on the authors’ own benchmark: The most controlled comparison and all ablations are on CrossPageBench. Its checkpoints are written by a model, and the paper does not report validating them. Its example goals also list the required steps in order, which reduces planning to following the given steps.
  • Costly, site-specific data: Humans validate every page type, inventory, and transition (339 types and 5,766 transitions for four sites), but the paper does not report the effort.
  • Limited novelty: The planner uses the same propose–simulate–select loop as action-level planners such as WebDreamer. PFPs are also close to parameterized actions, since 61% take a single atomic action.
Questions for authors
  • Could the authors report a plain gpt-5.2 agent without a world model, and run the Prompt variant, Plan–Action–Reflect, and AWM on WebArena?
  • Could the authors report mean success over runs, with variance, rather than the best of three?
  • How large is the depth-ablation subset? How does depth-3 planning perform with an uninformative simulator, such as one that returns the current inventory? This would test whether accurate dynamics matter.
  • How often does the Retrieval variant fall back to the simulator for unseen type–PFP pairs?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.