acceptodds
Under review as a conference paper at ICLR 2027

Self-Improvement Needs Attribution: Dependency-Aware Causal Memory for Long-Horizon Mobile GUI Agents

Abstract

**Abstract of paper.** Mobile GUI agents are moving toward long-horizon tasks that span dozens or even hundreds of steps across multiple applications, where the working context explodes as execution progresses. Information the agent reads on earlier screens is gradually drowned in the expanding context and becomes hard to identify when needed by subsequent steps. Existing systems fold raw histories into task state or structured memory and recall items by recency or similarity. However, merely summarizing or retrieving past context fails to catch silent step errors, leaving active context polluted with unverified steps and further impairing downstream experience distillation. We introduce ****, a self-improving framework that empowers mobile GUI agents with causal context and attribution-driven evolution, alongside MobileWorld-XL, an evaluation benchmark derived from the MobileWorld ecosystem to assess the generalization of extracted experience across unseen tasks. At its core, **** maintains a causal ledger to log the causal dependencies across actions and observed facts, enabling online step auditing, causal-dependency context assembly, and cause-attributed experience curation. Under identical conditions on three public benchmarks with a Qwen3.5-397B-A17B backbone, **** consistently outperforms competitive frameworks, achieving on MobileWorld ( over backbone), on MemGUI-Bench (), and on AndroidWorld hard slice (). Furthermore, task success rises across five rounds of self-evolution from to , with experience transferring to the unseen tasks of MobileWorld-XL. The code is available at: .

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
GUI Agents, self-improving agent, agent memory

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_hsceOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper conducts research on three key issues faced by long-range Mobile GUI Agent in cross-application and multi-step tasks: the continuous growth of historical context makes it difficult to effectively retrieve early key evidence; the latent errors generated during execution may only be exposed in subsequent steps, making it challenging to accurately pinpoint the root cause; and erroneous or unverified execution trajectories may further contaminate the experience database, affecting the Agent's continuous self-improvement. To address these issues, the authors propose a unified CausGUI framework, which explicitly records sub-goals, execution steps, observed evidence, and their dependencies using Causal Ledger, and design three collaborating modules. The authors have also demonstrated the effectiveness of this framework through experiments.

Critical strengths

The research motivation is clear, and the framework design has good systematicity. The paper conducts research on the problems in long-term GUI Agent, such as the difficulty in effectively utilizing historical evidence, the inability to trace implicit execution errors, and the impact of erroneous experiences on subsequent self-evolution. The author designs a unified Causal Ledger, enabling Dependency Resolver, Attribution Auditor, and Self-Evolution Curator to share structured historical evidence and their dependencies, integrating context management, error attribution, and self-evolution experience into a complete execution and learning loop.

The experiments are quite comprehensive, and the core mechanism has been verified specifically. The author conducted extensive comparisons on multiple benchmarks and different scales of Backbones, achieving stable performance improvements. In addition to incremental ablation to verify the effectiveness of each module, special experiments such as root cause localization, dependency-aware context management, experience admission and elimination, etc., were also conducted to further analyze the contribution of the core mechanism. Moreover, multi-round self-evolution and unseen task migration experiments also demonstrated the potential of the framework in continuous experience accumulation and combination generalization within the same application ecosystem. The overall experimental argumentation is quite solid.

Critical weaknesses

The paper integrates context-aware dependency perception, error attribution, and experiential self-evolution through a unified Causal Ledger, demonstrating certain system innovation value. However, its basic design is still mainly based on existing task planning, structured memory, execution verification, and experience management mechanisms, and causal dependencies are mainly constructed through explicitly declared evidence support relationships in the model. Therefore, compared to a complete system design, the independent innovation of a single algorithm mechanism is relatively limited.

Although the author verified the effectiveness of experience in unseen tasks through MobileWorld-XL, this benchmark shares the application ecosystem with the original MobileWorld and mainly demonstrates the combined generalization ability. Further verification of the effect of experience transfer in different application environments will help to more comprehensively evaluate the generalization ability of the proposed self-evolution mechanism.

Questions for authors

Table 2 demonstrates the improvement in the final task success rate achieved by the Attribution Auditor. Appendix B.6 also confirms the advantage of dependency graph backtracking in root cause identification. However, the Auditor simultaneously includes step-level verification, sub-goal review, and dependency graph backtracking. It is not clear to what extent the final success rate improvement is attributed to the dependency graph backtracking itself. Could the authors further explain or conduct more detailed ablation studies to distinguish the individual contributions of the verification execution and causal backtracking?

Reviewer_iu1JOpenReview ↗
Rating
4
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents CausGUI, a self-improvement framework for long-horizon mobile interaction tasks. At its core is a causal ledger that logs, for each action step, the sub-goal, supporting evidence, expected outcome, and actual result. Building on this design, the system includes three key modules. The Attribution Auditor inspects the execution process and, when a task fails, traces the failure back to the earliest incorrect step. The Dependency Resolver selects only the historical information that is relevant to the current task, helping reduce interference from irrelevant or outdated context. The Self-Evolution Manager distills reusable memories and action skills from verified successful trajectories and corrected failures, while pruning experiences that repeatedly lead to mistakes.

The paper also introduces MobileWorld-XL, a benchmark designed to test whether acquired experience can transfer to unseen tasks. Experiments show that CausGUI outperforms baseline methods on MobileWorld, MemGUI-Bench, and AndroidWorld. After five rounds of self-improvement, it boosts the success rate on MobileWorld from 52.1% to 59.0%.

Critical strengths
  1. The paper is clearly written and easy to follow. The motivation is especially well articulated: existing long-horizon GUI agents often have trouble keeping track of important information, pinpointing errors that only surface many steps later, and learning reliably from past trajectories.

  2. The proposed causal ledger is a compelling and intuitive idea. The framework is also evaluated thoroughly on multiple benchmarks, which provides strong empirical evidence for its effectiveness. In addition, the paper introduces MobileWorld-XL, a new benchmark for testing whether learned experience can transfer to previously unseen tasks, which further strengthens the experimental contribution.

Critical weaknesses

The framework is fairly complex from a system design perspective. CausGUI brings together several closely connected components, including the causal ledger, attribution auditor, dependency resolver, manager, and self-evolution curator. While the paper shows that the extra token cost is limited, it would still be useful to discuss the practical overhead of implementing, integrating, and maintaining the full framework in real-world deployments.

The generalization setting is also somewhat narrow. MobileWorld-XL is a valuable addition, but it is built on the same apps and backends as MobileWorld. As a result, the current results mainly show generalization to new task compositions, rather than robustness to truly new apps, interfaces, or environments. Evaluating the method under stronger distribution shifts would make the claims about generality more convincing.

In fact, the two weaknesses I mentioned are intended more as suggestions to further strengthen the paper; the current version is already excellent.

Questions for authors
  1. How sensitive is CausGUI to incorrect causal records? Since the causal ledger depends on the Actor’s declared evidence and the Manager’s judgments, could an incorrect attribution propagate into the memory store and negatively affect future tasks?
  2. How does the causal ledger scale with longer trajectories? As the number of steps, applications, and dependencies increases, does ledger construction or dependency resolution introduce significant computational or context-management overhead?
Reviewer_eYb7OpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes CausGUI, a self-improving framework for long-horizon mobile GUI agents built around a shared "causal ledger." Every executed step records its targeted subgoal, the evidence it cites, and the observed effect, forming a directed acyclic support graph. Three components read this ledger: an attribution auditor that backtracks from a surfaced failure to the earliest step whose execution was wrong despite valid grounds; a dependency resolver that mounts context into three zones (resident, recency, mount) governed by declared subgoal dependencies and entry scopes rather than recency or surface similarity; and a self-evolution curator that distills verified causes, effects, and contrasts into reusable memories and skills, admitting entries only when grounded in a localized root cause and retiring them via reuse counters across tasks.

Critical strengths

The two-level verifier (deterministic post-action check plus subgoal-boundary ruling) and the admission rule requiring a localized cause, observed effect, and stated contrast are well-motivated responses to the credit-assignment problem, and the frozen-generation store protocol makes each self-evolution round attributable to one store version.

Critical weaknesses

First, the reported gains are not credible under a fixed benchmark with simple attribute dependencies. The paper claims +23.1 points on MobileWorld for the 397B backbone (35.9 → 59.0) and attributes +11.1 to a ledger-plus-resolver that merely reorganizes context; my own experience is that dependency-scoped mounting over simple attribute lookups yields limited improvements. The context audit (Table A7) reports stale-value use falling from 11.4% to 2.3%, but this metric is defined by the authors' own supersession annotation, and the flat-dump condition is a strawman (it scores below the bare backbone). Without an independent replication or a controlled test isolating dependency scope from prompt-length effects, the headline numbers read as harness artifacts.

Second, novelty is thin against the existing GUI-memory literature. Structured task states, experience stores with admission and retirement, skill/macro-action induction, and planner-executor verification are all established (Liu et al. 2026b; Mi et al. 2026; Wang et al. 2024b; Jiang et al. 2025; Xiao et al. 2026). The contribution reduces to relabeling these as "causal" and adding a DAG bookkeeping layer; the causal claim is never formally validated—no intervention or counterfactual test shows the graph identifies true causes rather than plausible citations.

The "causal" framing is unvalidated and collapses into attribute matching. No intervention or counterfactual test shows the support graph identifies true causes rather than plausible citations; "causality" here is simple attribute dependency encoded as a DAG.

Questions for authors

see the above

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.