SCAFFOLD: Self-Improving Web Agents via Recursive Parametric Skill Abstraction

Bowei He; Xiaokun Zhang; Meng Ding; Xue Liu

cs.AIAug 31, 2026

SCAFFOLD: Self-Improving Web Agents via Recursive Parametric Skill Abstraction

Authors: Bowei He, Xiaokun Zhang, Meng Ding, Xue Liu

Organizations: MBZUAI · McGill University · CityUHK · UMass Boston

Abstract

Web agents need to navigate visually rich, long-horizon interfaces that change across sites, yet most previous agents still learn each task in isolation and discard the procedural knowledge they accumulate. Recent skill-augmented frameworks take an important first step, but they treat the skill library as a flat or two-tier prompt-side cache and offer no principled mechanism for compressing redundancy or composing skills recursively. We introduce Scaffold, a self-improving framework for visual web agents that (i) induces parametric, executable skills from successful trajectories under a multi-instance abstraction constraint, (ii) maintains a recursively composed hierarchy in which higher-level skills invoke lower-level ones, (iii) compacts the library via a minimum-description-length (MDL) criterion and behavioral equivalence checking, and (iv) periodically distills skill-augmented trajectories back into model weights to internalize the abstractions. Across WebArena, VisualWebArena, and a held-out split of Online-Mind2Web, Scaffold improves success rate by $11.1$ -- $17.2$ absolute points over the strongest skill-augmented baseline and shows monotonic gains across five self-improvement iterations without library collapse. We release the code and documents in the Github repository.

Figures & tables

Figure 1: Overview of Scaffold framework. The five stages are executed sequentially in each iteration: rollout, multi-instance induction, recursive composition, MDL-driven compaction, and distillation. The library growing in both depth and width is periodically compressed.

(A) Zero-shot agents (Qwen2.5-VL-7B)
Method	WebArena	VWA	OM2W-X
ReAct Yao et al. (2023)	8.1 $\pm$ 0.6	6.2 $\pm$ 0.4	11.2 $\pm$ 0.9
SeeAct Zheng et al. (2024a)	12.4 $\pm$ 0.5	9.7 $\pm$ 0.6	14.0 $\pm$ 1.0
WebVoyager He et al. (2024)	15.3 $\pm$ 0.7	11.8 $\pm$ 0.5	17.5 $\pm$ 1.2
AgentOccam Yang et al. (2025)	16.8 $\pm$ 0.4	12.4 $\pm$ 0.6	18.1 $\pm$ 0.8
(B) Memory/Workflow-augmented

Table 1: Main results (Success Rate %) on WebArena, VisualWebArena (VWA), and Online-Mind2Web held-out test split (OM2W-X). Mean

\pm

std over 3 seeds.

\dagger

p<0.05

over the strongest competitor ( SkillRL ) by paired bootstrap. All rows in (A)–(C) and the bold Scaffold row use Qwen2.5-VL-7B as the base; the last row is included as an upper reference with a stronger backbone.

Variant	WA	VWA	OM2W-X
Full Scaffold	42.7	36.3	43.5
$-$ multi-instance ( $n_{\min}{=}1$ )	38.1	31.7	38.4
$-$ recursive composition (depth $\leq$ 1)	36.4	30.5	36.1
$-$ MDL compaction	39.6	33.8	40.0
$-$ distillation (in-context only)	34.7	28.4	34.0
$-$ holdout validation in induction	39.0	32.6	39.7

Table 2: Ablation on the four core innovations. Each row removes a single component while keeping the rest unchanged. Last row is one of strongest baselines.

Figure 2: Success rate vs. self-improvement iteration on WebArena. Scaffold continues to improve through

K{=}5

while baselines saturate around

k{=}2

–

3

. Shaded regions are 1 std over 3 seeds.

Figure 3: Library growth dynamics across iterations on WebArena. Top : total skill count without compaction (dashed) the library balloons; with compaction (solid) it stabilizes near 180 skills. Bottom : mean abstraction depth, which grows monotonically as recursive composition kicks in.

Figure 4: Cross-site transfer on Online-Mind2Web. Cell $(i,j)$ shows SR (%) on test site $j$ after training only on site $i$ . The ziprecr and ticketm are the abbreviations for ziprecruiter and ticketmaster, respectively. The sites in redlines are the held-out test ones in main experiments, also indicated with ∗ .

Drop relative to clean ( $\Delta$ in points)
Method	Clean	Latency	DOM-shuf.	+7 d
SkillWeaver	29.1	25.3	24.7	26.5
SkillRL	31.6	28.0	27.6	28.9
Scaffold	42.7	40.9	40.6	41.4
SkillWeaver	n/a	$-3.8$	$-4.4$	$-2.6$
SkillRL	n/a	$-3.6$	$-4.0$	$-2.7$

Table 3: Stability under environment perturbations on WebArena (SR %). Clean : unperturbed baseline. Latency : 0–3s random delay per action. DOM-shuf. : non-functional DOM attributes randomly permuted on each page load. +7 d : same task suite re-evaluated 7 days later. Scaffold degrades by at most 2.1 points across all regimes, roughly half the sensitivity of the strongest baselines. Each cell is mean over 3 seeds (std

\leq 1.0

Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

WebArena
Site	SkillRL	Scaffold	$\Delta$
OneStopShop	30.5	42.6	+12.1
Reddit	33.2	41.9	+8.7
GitLab	28.4	45.1	+16.7
CMS	27.0	42.9	+15.9
VisualWebArena

Appendix

Table 4: Per-site success rate (%) on WebArena Zhou et al. (2024) (four primary sites) and VisualWebArena Koh et al. (2024) . Scaffold ’s largest per-site gains occur on long-horizon, multi-step sites (GitLab, CMS, VWA Shopping), where recursive composition contributes most. Differences are bootstrap-significant (

p<0.05

) for all rows.

Hyperparameter	Range tested	$\Delta$ SR (WA)
Multi-instance threshold $n_{\min}$	$\{1,2,3,4\}$	$-3.1$ / $0.0$ / $-0.8$ / $-2.4$
Compaction interval $M$	$\{1,2,3,5\}$	$-0.6$ / $0.0$ / $-0.4$ / $-1.7$
Equivalence threshold $\rho$	$\{0.7,0.8,0.9,0.95\}$	$-1.2$ / $-0.3$ / $0.0$ / $-0.4$
LoRA rank for distillation	$\{16,32,64,128\}$	$-1.6$ / $-0.5$ / $0.0$ / $-0.2$

Appendix

Table 5: Sensitivity to Scaffold ’s main hyperparameters.

\Delta

SR is relative to the default setting (bold values in the experiments section:

n_{\min}{=}2

M{=}2

\rho{=}0.9

, rank=64).

Training signal (same trajectories, $K{=}5$ )	WA
No SFT, in-context library only (Tab. 2 )	34.7
Standard SFT, trajectories flattened to primitives	38.6
Plan-augmented SFT, no auxiliary skill-name loss	41.6
Full Scaffold distillation	42.7
STaR-style loop: SFT on raw trajectories, no library	30.4

Appendix

Table 6: Supervision-format study on WebArena (SR %). All rows share the base model, the rollout pipeline and the exact same successful trajectories, differing only in how those trajectories supervise the policy.

Configuration	WA
SkillWeaver	29.1
$+$ multi-instance induction with holdout validation	31.9
$+$ recursive composition	34.0
$+$ distillation	39.6
$+$ MDL compaction (= full Scaffold )	42.7

Appendix

Table 7: Incremental build-up on WebArena (SR %), starting from SkillWeaver and adding one component at a time. Every addition is bootstrap-significant (

p<0.05

). The intermediate point

39.6

coincides with the “

-

MDL compaction” row of Table 2 by construction.

Max primitive steps	10	20	30
SkillWeaver	11.5	24.0	29.1
SkillRL	13.2	26.4	31.6
Scaffold	19.8	36.9	42.7

Appendix

Table 8: WebArena SR (%) under a capped primitive-action budget (§ 4.9 ). The margin is largest at the tightest cap, the opposite of what an inflated action budget would produce.

Method	Pol.	Grd.	Jdg.	Tok. (K)	Time (s)
ReAct	18.7	0	0	224	102
SkillWeaver	9.6	0	0	118	74
Scaffold	5.3	7.1	4.6	79	61

Appendix

Table 9: Inference cost per successful WebArena task: policy, grounder and judge calls, total tokens and wall clock. A policy call costs about 12K tokens (screenshot, DOM and history), a grounder call about 1.5K, and a judge call about 0.8K.

Grounding mode	Clean	DOM-shuf.	Grd./task
VLM grounding (main paper)	42.7	40.6	7.1
Hybrid with selector cache	42.5	40.1	2.7
Static selectors only	41.2	33.5	0

Appendix

Table 10: Grounding modes on WebArena (SR %) and grounder calls per successful task. The hybrid cache removes

62\%

of grounder calls at essentially no cost in accuracy or robustness, whereas static selectors reintroduce baseline brittleness under DOM-shuffle.

Aspect	PolySkill	SkillEvo	Scaffold
Skill unit	abstract interface plus per-site implementations	skill paths in a graph	parametric executable program with pre/postconditions
Composition	compositional, prompt-side	composite skills over the skill path graph	recursive, unbounded depth (5 observed)
Library governance	none	graph updates, no compression objective	MDL merge / refactor / prune with validation safety net
Policy weights	frozen	GRPO with a learned reward model	SFT distillation of skill plans
Setting	HTML agents, Mind2Web / WebArena	text agents, WebArena-Lite	visual agents, WA / VWA / OM2W

Appendix

Table 11: Component-level comparison with the two concurrent skill-centric frameworks discussed in § 2 .

Web Agents

Web agents use language and vision models to navigate websites, gather information, and complete tasks through interactive interfaces.

See all 176 papers

Apr 14, 2026

WebXSkill: Skill Learning for Autonomous Web Agents

Autonomous web agents powered by large language models (LLMs) remain brittle on long-horizon browser workflows. A key bottleneck is a grounding gap in existing skill formulations: textual workflow skills provide natural language guidance but cannot be directly executed, while code-based skills execute without giving the agent step-level guidance for adaptation or recovery. We introduce WebXSkill, a framework that bridges this gap with executable skills, each pairing a parameterized action program with step-level natural-language guidance. WebXSkill operates in three stages: skill extraction mines reusable action subsequences from readily available synthetic agent trajectories and abstracts them into parameterized skills, skill organization indexes them into a URL-based graph for context-aware retrieval, and skill deployment exposes two complementary modes, grounded mode for fully automated execution and guided mode where skills serve as step-by-step instructions the agent follows with its native planning. WebXSkill demonstrates consistent improvements on WebArena, WebVoyager, and Online-Mind2Web. We further find that better skill deployment mode depends on a model's plan and execution capability. The code is available at https://github.com/aiming-lab/WebXSkill.

Zhaoyang Wang, Qianhui Wu, Xuchao Zhang +12

University of North Carolina at Chapel Hill · Microsoft

LLM Agent Skill Learning Long-Horizon LLM Agents Web Agents

Ask PDF

Oct 26, 2026

Back to Basics: A Skill-Centric Approach to Generalist Web Agents

Browser agents do not just need more practice. They need to practice the right things. We introduce MiniWeb, a realistic miniature of the web comprising 65 interconnected websites across 15 domains, where agents navigate, search, and transact as people do. Its 369 human-annotated tasks cover everyday workflows and label 37 reusable interaction primitives grouped into 7 skill families, allowing us to diagnose which skills fail rather than only whether a task succeeds. Benchmarking ten open-weight agents, we find that drag & gesture is consistently the weakest family, with no agent solving more than 24% of its instances. To address this gap, we introduce MiniWebGen, a framework that autonomously generates verified, skill-centric training data from MiniWeb. Fine-tuning Qwen3.5-4B on just 3.1K verified trajectories improves single-primitive success by 24.9pp across 13 unseen websites, with gains across every skill family. However, these trajectories teach short-horizon interactions, which do not directly translate to full-task success. We build on this result with MiniWebAgent, which pairs the trained executor with a planner that decomposes full tasks into the skill-level steps it learned to perform. This approach improves task success by 19.4pp points on unseen MiniWeb sites and transfers to external benchmarks, with gains of 9.3pp on WebArena-Lite, 16.1pp on Online-Mind2Web, and 20.0pp on WebVoyager, without training on their trajectories. We provide a simple recipe for compact browser agents: practice reusable skills directly, then compose them to solve unfamiliar tasks.

Minh Pham-Dinh, Md Farhan Ishmam, Hernan Yupanqui +2

University of Utah · Islamic University of Technology

Web Agents LLM Agent Skill Learning Agent Skill Learning LLM Agent Training Web Agent Benchmarks

Ask PDF

Apr 28, 2026

DRIVE: Modeling Skills at the Reasoning and Interaction Levels for Web Agents under Continual Learning

Web agents require both high-level reasoning (for task decomposition) and low-level interactions (for page elements manipulation) to conduct different tasks. However, these knowledge types differ fundamentally: reasoning knowledge (e.g., booking a flight requires first searching for routes) is abstract and transferable across websites, while interaction knowledge (e.g., clicking the Search button at a specific coordinate on Site A) depends heavily on page-specific contexts. Existing methods store experiences uniformly. This creates a dilemma: abstract representations lose executability on concrete pages, while concrete representations fail to generalize across domains. This entanglement limits capability accumulation: on new websites, agents either fail to recognize reusable task logic due to surface-level differences or attempt infeasible actions from outdated page structures. To disentangle them, we propose DRIVE, a dual-level skill modeling framework separating historical experience into natural language reasoning skills, which capture transferable task logic, and programmatic interaction skills, grounding abstract actions to executable operations. A scene-aware coordination mechanism adaptively retrieves and invokes these dual-level skills based on task semantics. DRIVE also uses skill-level reflection to identify hierarchy-specific failure modes, enabling targeted skill library expansion and refinement. Experiments across five WebArena domains show DRIVE attains an average task success rate of 52.8%, exceeding the skill-free baseline by 7.3 percentage points. Further ablations show reasoning and interaction skills provide distinct, complementary benefits, supporting separation of transferable task logic from executable page-level operations.

Xirui Liu, Sihang Zhou, Yanning Hou +6

College of Intelligence Science and Technology, National University of Defense Technology, Changsha, 410073, Hunan, China · College of Computer Science and Technology, National University of Defense Technology, Changsha, 410073, Hunan, China

Continual Learning for LLM Agents LLM Agent Skill Learning Agent Skill Learning Web Agents

Ask PDF

Abstract

Web Agents

Explore similar work

WebXSkill: Skill Learning for Autonomous Web Agents

Back to Basics: A Skill-Centric Approach to Generalist Web Agents

DRIVE: Modeling Skills at the Reasoning and Interaction Levels for Web Agents under Continual Learning