Skill-Guided Continuation Distillation for GUI Agents

Abstract: Improving GUI agents typically relies on behavior cloning on expert trajectories. However, as the current policy deviates from the expert policy, it inevitably encounters policy-induced off-trajectory states during closed-loop execution, i.e., states that fall outside the expert trajectories. Since expert trajectories provide no demonstrations for these unseen states, such states receive no effective supervision, leaving the policy unable to select the correct action. To close this supervision gap, we propose Skill-Guided Continuation Distillation (SGCD), an iterative self-improvement framework. SGCD first runs the plain policy without skill guidance for a few steps to reach realistic off-trajectory states. From these states, a skill-guided policy then completes the task and produces successful continuations, which are mixed with expert trajectories to supply supervision over policy-induced off-trajectory states. The skills are extracted from both successful and failed rollouts, consisting of Continuation Plans, Critical Targets, Failure Traps, and Success Criteria. On OSWorld-Verified, SGCD improves the success rate of three base models from the low-30\% range to over 50\%, demonstrating its effectiveness and generality.

• The paper introduces Skill-Guided Continuation Distillation (SGCD) to provide explicit supervision over off-trajectory states in GUI agents.
• The method constructs structured continuation skills from task trajectories and uses a skill-conditioned policy for task recovery, significantly improving success rates.
• Empirical results demonstrate over 20–25% gains in task success rates and robust recovery from early policy-induced failures across multiple model backbones.

Skill-Guided Continuation Distillation for GUI Agents: An Expert Review

Motivation and Problem Formulation

The paper addresses the persistent challenge in GUI agent training wherein traditional behavior cloning from expert trajectories fails to provide supervision for policy-induced off-trajectory states, which emerge during closed-loop execution in real deployments. This covariate shift leads to systematic failure modes such as early termination, repeated ineffective actions, hallucination of UI elements, and scope misjudgment. Reinforcement learning approaches have been explored to provide supervision in these states, but reward sparsity makes them inefficient for long-horizon GUI tasks.

Methodology: Skill-Guided Continuation Distillation (SGCD)

SGCD introduces an iterative self-improvement paradigm that supplies explicit supervision over policy-induced off-trajectory states. The framework is composed of four stages:

1. Task Trajectory Sampling: The plain policy is rolled out on the training task suite, recording both successful and failed trajectories. This exposes the agent's own characteristic error distribution and the states responsible for recurring failure.
2. Skill Construction: Gemini-3-Pro abstracts trajectory evidence into structured off-trajectory continuation skills defined by four fields—Continuation Plans, Critical Targets, Failure Traps, and Success Criteria. These abstractions target specific failure modes and avoid path overfitting.
3. Off-Trajectory Continuation Construction: For tasks failed by the plain policy, the first $k$ steps are executed by the policy to induce an off-trajectory state, after which a skill-conditioned policy attempts task recovery. Successful continuations are filtered by an automatic verifier and an LLM judge to ensure both terminal correctness and trajectory-level quality.
4. Mixed Trajectory Training: The final deployment policy is trained on a mixture of expert trajectories and verified successful continuations, discarding supervision on pre-handoff actions to avoid propagating erroneous behavior.

Iterative application of SGCD progressively expands the set of recoverable tasks, realigning the supervision distribution toward the actual policy deployment regime and continually improving agent robustness.

Empirical Analysis

Failure Distribution

The paper empirically validates that more than 90% of failures occur within the first 20 steps of execution, suggesting that early-stage policy errors propagate and dominate failure modes. Consistent recurrence of specific error families substantiates the need for structured skill abstractions.

Numerical Results

SGCD achieves substantial gains across three backbone models (Qwen3-VL-8B, Qwen3-VL-30B-A3B, STEP3-VL-10B), elevating OSWorld-Verified success rates from the low-30% range to over 50% (55.1% for SGCD-8B, 58.4% for SGCD-30B-A3B, 53.2% for SGCD-10B). In backbone-matched comparisons, SGCD surpasses prior GUI-specialized and open-weight baselines by over 20–25% margin in task success rate.

Continuation Success Rate (Cont. SR) analysis on tasks failed by the original policy indicates that the distilled plain policy can recover from policy-induced off-trajectory states: SGCD-8B reaches 39.2% (vs. 32.7% for EvoCUA-8B), SGCD-30B-A3B reaches 50.3% (vs. 27.2% for Qwen3-VL-8B-Thinking).

Ablation studies further demonstrate that enumerating handoff depths (early execution windows) improves performance, and the inclusion of all four skill fields is necessary to remediate dominant failure modes. Excluding pre-handoff supervision leads to improved accuracy.

Iterative Distillation

Multiple rounds of SGCD yield stepwise improvements, as newly recoverable tasks are incorporated into the training pool once the policy’s capabilities expand. The distillation is robust across varying handoff depths and consistently outperforms baselines under strict per-handoff evaluation protocols.

Practical and Theoretical Implications

SGCD solves a critical bottleneck in GUI agent training by directly bridging the supervision gap for policy-induced off-trajectory states. The structured skill abstractions enable the synthesis of compact procedural knowledge from both successful and failed rollouts, facilitating generalization beyond fixed-path imitation. By leveraging skill guidance for continuation generation and distilling this knowledge into the deployment policy, SGCD enables robust recovery from non-trivial off-trajectory states encountered in real environments.

Practically, this framework broadens the horizon for GUI agents, making them more applicable for open-ended, workflow-oriented desktop tasks encountered by practical users. The approach is model-agnostic and generalizes across multiple vision-language backbones, making it compatible with the latest LLM architectures.

Theoretically, SGCD formalizes an iterative, skill-conditioned continuation distillation pipeline tailored for covariate shift in imitation learning with long-horizon sequences—a setting where error accumulation and policy-induced state divergence are fundamental issues.

Limitations and Future Directions

SGCD currently relies on live re-execution for off-trajectory state instantiation, which incurs significant environmental interaction overhead. Coverage and efficiency in acquiring successful continuations for difficult or unrecoverable tasks remain limited. Future developments could incorporate state caching and replay mechanisms, improved skill extraction pipelines, and more scalable continuation acquisition for challenging long-horizon tasks.

SGCD’s skill extraction depends on high-quality successful and failed trajectory pairs. Automating the construction of richer skill schemas or integrating dynamic on-policy expert correction [Lauffer et al., 2025] may further enhance off-trajectory robustness.

Conclusion

Skill-Guided Continuation Distillation (SGCD) advances the training of GUI agents by explicitly supervising policy-induced off-trajectory states through structured skill-guided continuations. It delivers strong empirical improvements across diverse model backbones and aligns supervision distribution with real deployment regimes. SGCD represents a significant methodological step toward robust, generalizable GUI agents capable of reliable open-ended desktop automation.