EconSkills: Studying Skill Transfer and Retrieval for Web Agents on Live Economic Data

# EconSkills: Studying Skill Transfer and Retrieval for Web Agents on Live Economic Data
Source: [https://arxiv.org/html/2609.19523](https://arxiv.org/html/2609.19523)
Yinzhu QuanAffiliation:Georgia Institute of TechnologyAffiliation:Atlanta, GA, USAEmail:[yquan9@gatech\.edu](mailto:)Zefang Liu

###### Abstract

Web agents often revisit the same sites, yet most evaluations discard the procedures learned in earlier successful interactions\. We introduce EconSkills, a skill library and evaluation framework that distills verified EconWebArena trajectories into parameterized standard operating procedures for retrieving live economic data\. Each skill records its scope, navigation procedure, site\-specific guidance, verification checks, and recovery steps while replacing source\-instance values with placeholders\. EconSkills separates two questions: whether a known relevant procedure transfers to a held\-out task, and whether an agent can retain that benefit when selecting from a library\. In controlled transfer, matched skills improve success over no\-skill prompting and require fewer steps on paired successes, while abstraction is substantially more effective than replaying raw trajectories\. At library scale, retrieval is competitive with the no\-skill baseline overall and performs best on directly covered tasks; coverage\-stratified outcomes show that approximate matches on uncovered tasks offset these gains\. Browser trajectories further identify when procedural guidance shortens portal\-specific navigation and when semantic verification remains necessary\. These results establish that reusable economic web procedures can transfer across task instances and provide a concrete design target for coverage\-aware selection and context delivery\.

## 1Introduction

Autonomous web agents built on large language models \(LLMs\) are increasingly capable of operating real websites\([He et al\., 2024](https://arxiv.org/html/2609.19523#bib.bib17);[Zheng et al\., 2024](https://arxiv.org/html/2609.19523#bib.bib18)\)by reading rendered pages, filling forms, following links, and returning a final answer\. Economic data retrieval is a demanding and practically important instance of this problem\. Analysts routinely pull a specific figure, an inflation print, a policy rate, an exchange rate, a country\-level indicator, from an official portal such as a central bank, a national statistics office, or an international organization\. EconWebArena111[https://huggingface\.co/datasets/EconWebArena/EconWebArena](https://huggingface.co/datasets/EconWebArena/EconWebArena)\([Liu and Quan, 2025](https://arxiv.org/html/2609.19523#bib.bib8)\)formalizes exactly this setting, pairing natural language questions with the live, authoritative pages that answer them\. More broadly, LLM agents are increasingly applied to economic and business tasks, from economic sequential reasoning\([Quan and Liu, 2024a](https://arxiv.org/html/2609.19523#bib.bib10)\)to multi\-agent inventory management\([Quan and Liu, 2024b](https://arxiv.org/html/2609.19523#bib.bib11)\), supply\-chain risk assessment\([Quan et al\., 2026](https://arxiv.org/html/2609.19523#bib.bib12)\), and e\-commerce content generation\([Quan et al\., 2025](https://arxiv.org/html/2609.19523#bib.bib13)\)\. Related systems use LLM agents for macroeconomic simulation\([Li et al\., 2024](https://arxiv.org/html/2609.19523#bib.bib14)\)and financial research and valuation\([Yang et al\., 2024](https://arxiv.org/html/2609.19523#bib.bib15)\)\. These applications depend on current, correctly identified economic evidence, yet obtaining that evidence from official websites remains a separate interactive problem\.

Despite rapid progress, standard web\-agent evaluations commonly treat each task in isolation\. Tasks that recur with a familiar structure, such as two questions on the same portal that differ in the reference year, country, or data series, are each solved from scratch, and whatever an agent learns while solving the first is discarded before it reaches the second\. Reflection can help an agent improve from feedback on a task\([Shinn et al\., 2023](https://arxiv.org/html/2609.19523#bib.bib19)\), but a one\-shot agent normally discards that experience before a related task\. This is wasteful, because within a portal the route to an answer is far more stable than the answer itself\. The path to an international organization indicator, the selector sequence on a central bank statistics page, or the export control on a national statistics office is reusable structure that a one\-shot agent discards each time\. Capturing this structure once, as a library of skills the agent can draw on, would let it reuse the route instead of working it out again for every task; what remains open is how the agent should select from such a library once it holds many skills and only a few fit the task at hand\. The economic setting makes this selection problem unusually exacting\. A portal can expose neighboring series with similar labels but different units, seasonal adjustments, geographic coverage, release vintages, or effective dates\. A useful economic web skill must therefore encode not only where to click, but also which series is being requested and how its provenance and interpretation should be checked\.

To address these gaps, we introduce EconSkills222[https://huggingface\.co/datasets/EconWebArena/EconSkills](https://huggingface.co/datasets/EconWebArena/EconSkills), a set of abstract, instance\-free economic web skills that a web agent can draw on for new tasks, as shown in Figure[1](https://arxiv.org/html/2609.19523#S1.F1)\. Building on EconWebArena, we distill a library of 50 skills, each extracted from a solved seed trajectory as a seven\-part standard operating procedure \(SOP\) that records how to reach a value on an authoritative portal, covering its name, purpose, preconditions, ordered navigation steps, site\-specific guidance, and verification and recovery rules\. Instance\-specific values such as a country or year are replaced with placeholders so the procedure generalizes\. We separate two questions that are often conflated in evaluations of agent skills\. First, when a task has a known corresponding skill, does that skill transfer to a new instance? Second, when the agent must draw from an entire library, can skill selection preserve that benefit? We answer the first with a controlled study on 100 held\-out variants under no\-skill, raw\-trajectory, and matched\-skill conditions, each repeated over 3 runs\. We answer the second across all 360 EconWebArena tasks by varying whether the agent receives no skills, all skills, five retrieved skills, five random skills, or one optional hint\.

![Refer to caption](https://arxiv.org/html/2609.19523v1/econ_skills.png)Figure 1:EconSkills overview\. Reusable, instance\-free skills are distilled from solved economic web tasks and retrieved to guide a web agent on new tasks\.Our contributions are threefold\. \(1\) We introduce EconSkills, a library of 50 abstract economic web skills, each a reusable SOP distilled from a verified trajectory with instance values replaced by placeholders while retaining preconditions, verification, and recovery\. \(2\) We evaluate skill use at two complementary levels: controlled transfer with a known matched skill, and library\-scale delivery on the full benchmark under retrieval, random selection, full\-library, and optional\-hint conditions\. \(3\) We show that these levels behave differently: matched skills improve transfer, and retrieval is strongest on tasks directly covered by the library\. Paired outcomes and trajectory analysis explain how coverage, semantic fit, and context volume determine whether this benefit survives end\-to\-end deployment\.

## 2Related Work

Web agents and benchmarks\.WebArena\([Zhou et al\., 2024](https://arxiv.org/html/2609.19523#bib.bib1)\)evaluates long\-horizon tasks on fully interactive websites, while Mind2Web\([Deng et al\., 2023](https://arxiv.org/html/2609.19523#bib.bib2)\)studies generalization across diverse real\-world sites\. VisualWebArena\([Koh et al\., 2024](https://arxiv.org/html/2609.19523#bib.bib3)\)adds visually grounded tasks, and WebLINX\([Lù et al\., 2024](https://arxiv.org/html/2609.19523#bib.bib4)\)focuses on conversational navigation\. BrowserGym\([Le Sellier De Chezelles et al\., 2025](https://arxiv.org/html/2609.19523#bib.bib5)\)provides a common environment for running and comparing agents across benchmarks\. Evaluation has also moved toward online settings and agentic search through WebCanvas\([Pan et al\., 2024](https://arxiv.org/html/2609.19523#bib.bib6)\)and Mind2Web 2\([Gou et al\., 2025](https://arxiv.org/html/2609.19523#bib.bib7)\)\. EconWebArena\([Liu and Quan, 2025](https://arxiv.org/html/2609.19523#bib.bib8)\)specializes this line to multimodal, multi\-step economic tasks on live, authoritative sites, where related task families often share a procedural structure\.

Experience and workflow reuse\.Several methods carry information across tasks rather than restarting from an empty context\. ExpeL\([Zhao et al\., 2024](https://arxiv.org/html/2609.19523#bib.bib20)\)extracts natural\-language insights from prior experience and recalls relevant examples at inference time\. Agent Workflow Memory\([Wang et al\., 2025](https://arxiv.org/html/2609.19523#bib.bib21)\)induces reusable web workflows in both offline and online settings and selectively supplies them on new tasks, while RaDA\([Kim et al\., 2024](https://arxiv.org/html/2609.19523#bib.bib22)\)retrieves exemplars separately for web\-task decomposition and action generation\. VideoWebArena\([Jang et al\., 2024](https://arxiv.org/html/2609.19523#bib.bib23)\)further shows that long\-context tutorials can hurt skill\-retention performance, suggesting that procedural context is not uniformly helpful\. These results motivate cross\-task reuse, but the stored procedure and its delivery are often evaluated together\. Our two\-level design instead separates the transfer value of a known match from the problem of selecting and presenting it from a library\.

Skills for agents\.Voyager\([Wang et al\., 2024](https://arxiv.org/html/2609.19523#bib.bib16)\)maintains an executable code\-skill library, and SkillWeaver\([Zheng et al\., 2025](https://arxiv.org/html/2609.19523#bib.bib24)\)discovers and refines reusable web APIs through exploration\. More recent work studies auditable and generalizable skill artifacts: SkillGen\([Ma et al\., 2026](https://arxiv.org/html/2609.19523#bib.bib9)\)verifies synthesized skills by their effect on held\-out tasks, ContractSkill\([Lu et al\., 2026](https://arxiv.org/html/2609.19523#bib.bib25)\)adds explicit verification and repair, and PolySkill\([Yu et al\., 2025](https://arxiv.org/html/2609.19523#bib.bib26)\)targets cross\-site generalization through polymorphic abstraction\. SkillsBench\([Li et al\., 2026](https://arxiv.org/html/2609.19523#bib.bib27)\)likewise evaluates skills through downstream agent behavior rather than textual plausibility\. SkillGen and SkillsBench both expose task\-level regressions alongside aggregate gains, motivating paired outcome analysis rather than reporting average improvement alone\. EconSkills builds on this functional view while testing how the same skill library behaves under both matched transfer and library\-scale delivery\.

## 3EconSkills

EconSkills is built and applied in three stages, as illustrated in Figure[1](https://arxiv.org/html/2609.19523#S1.F1)\. Task execution produces a solved seed trajectory, skill extraction distills that trajectory into an abstract standard operating procedure \(SOP\), and skill application supplies skills from the resulting library to a web agent and evaluates them across the EconWebArena benchmark\. The library holds 50 skills, one per solved seed task, each capturing a distinct economic data\-retrieval workflow on an authoritative portal\.

### 3\.1Task Execution

We solve each seed task by driving the live portal through a browser\. At every step the agent observes a structured view of the current page, its accessibility tree \(AXTree\), a screenshot, and contextual metadata such as the focused element and the action history, and it responds with one or more actions from EconWebArena’s BrowserGym333[https://github\.com/ServiceNow/BrowserGym](https://github.com/ServiceNow/BrowserGym)high\-level action set\. This action space supports element\-based and coordinate\-based interaction, keyboard input, scrolling, navigation, and tab management\. The agent repeats this loop until it commits a final answer\. We keep a seed task only after it is solved and its answer clears the benchmark’s automatic check, which requires both the correct value and a landing URL on the expected authoritative domain\. Grounding every retained skill in a verified success, rather than in an aspirational account of how a portal ought to behave, ensures that each skill is later distilled from a trajectory that actually reached the target value\. This procedure yields 50 solved seed tasks, one for each skill in the library\. These 50 seed tasks span nine of EconWebArena’s ten economic categories, namely government, banking, markets, labor, finance, energy, education, health, and trade, and draw on 37 distinct authoritative websites, including central banks, national statistics offices, international organizations, and financial data portals\. Table[1](https://arxiv.org/html/2609.19523#S3.T1)lists the number of seed tasks in each category together with the instance fields that each skill replaces with placeholders\.

Table 1:EconSkills seed tasks by category\. Placeholder fields indicate the instance\-specific values abstracted by each skill\.
### 3\.2Skill Extraction

From each solved trajectory an extractor model distills one structured skill, written as a seven\-part SOP\. The extractor reads the full trajectory that solved the seed task, the visited URLs, the AXTree observations, the actions taken, and the confirmed final answer, and rewrites it into a procedure that a different agent can follow on a fresh instance\. The seven fields are \(i\) name, \(ii\) purpose, the kind of questions the skill answers, \(iii\) preconditions, what must hold before the procedure begins, such as the portal and dataset being reachable, \(iv\) procedure, the ordered navigation and interaction steps, \(v\) website guidance, the portal\-specific menus, selectors, and pitfalls, \(vi\) verification, how to confirm that the retrieved figure is the one the task requested, and \(vii\) recovery, how to react when a step fails\.

The last two fields are especially important for economic data\. Official portals frequently place nominal and real values, adjusted and unadjusted series, preliminary and revised releases, or multiple effective dates on the same page\. EconSkills therefore records checks for entity, period, unit, series definition, and source domain, together with portal\-specific fallbacks such as table, archive, and download views\. This domain information is retained even when the instance values themselves are abstracted\.

The property that makes a skill reusable is abstraction\. The extractor identifies every instance\-specific value the seed trajectory touched, a year, a country, an indicator, or a numeric target, and replaces it with a named placeholder such as<year\>,<country\>, or<indicator\>\. The procedure then refers to these slots rather than to one concrete question, so a step such as select the<indicator\>series, set the region to<country\>, and set the period to<date\>stays valid when the placeholders are filled for a different instance\. Verification and recovery are what separate a skill from a brittle recording of clicks\. The verification field states how to confirm that the value read back matches the requested units, period, and entity, and the recovery field states what to do when a step fails, for instance when a menu label has changed or a selector is missing, by falling back to the portal’s search or an alternate export view\. We serialize each skill asJSONover these seven fields and render it to Markdown for injection into the web agent, and we extract exactly one skill per solved seed task, which gives 50 independent skills\. Table[2](https://arxiv.org/html/2609.19523#S3.T2)gives an example skill, extracted from a European Central Bank exchange\-rate seed task, with the counter\-currency and date reduced to placeholders\. Appendix[A](https://arxiv.org/html/2609.19523#A1)gives library and reproduction details, and Appendix[B](https://arxiv.org/html/2609.19523#A2)reproduces this example and a second skill in full\.

Table 2:Example seven\-part skill extracted from an ECB exchange\-rate seed task\. Counter\-currency and date are represented as placeholders\.
### 3\.3Skill Application

We evaluate EconSkills at two levels\. The first isolates whether a correctly matched procedure transfers; the second asks whether an agent can use the whole library when the match is not given\.

Controlled matched\-skill transfer\.Each of the 50 source skills is paired with two held\-out task variants from the same procedural family, yielding 100 source\-variant pairs\. The variants change an instance field such as date, entity, series, or category while preserving the website workflow\. For each variant and condition, we conduct 3 independent agent runs, using seeds 0, 1, and 2:BASE, the task alone;TRAJ1, the task plus one matched raw successful trajectory with its final answer removed; andMATCH1, the task plus one directly matched extracted skill\. Thus BASE versus MATCH1 tests transfer from a matched skill, while TRAJ1 versus MATCH1 tests whether abstraction is preferable to replaying a verbose task\-specific history\.

Library\-scale deployment\.We next run all 360 tasks once under conditions that vary which EconSkills library context is supplied:BASEreceives no skill;ALL50receives the full library;RETR5receives the five skills ranked most relevant from their names and purposes; andRAND5receives five task\-seeded random skills\. RETR5 and RAND5 hold context count fixed, while ALL50 tests whether additional procedures help or interfere\.

We additionally evaluateHINT30, which injects one skill but explicitly tells the agent that it is an optional, possibly outdated reference that may be ignored whenever it does not match the page\. For the 150 tasks in the 50 source families \(50 sources and 100 variants\), HINT30 supplies the directly corresponding family skill\. The remaining 210 tasks have no direct library match and receive a top\-1 retrieved skill\. We analyze these groups separately because mixing them would confound skill quality with library coverage\.

Metrics\.For taskttand conditioncc, letrc​\(t\)∈\{0,1\}r\_\{c\}\(t\)\\in\\\{0,1\\\}be the native EconWebArena reward, which is one only when the answer contains the gold value and the final URL lies on the authoritative domain\. We reportSRc=N−1​∑trc​\(t\)\\mathrm\{SR\}\_\{c\}=N^\{\-1\}\\sum\_\{t\}r\_\{c\}\(t\)and interaction steps\. Matched outcomes are compared with McNemar’s exact test\([McNemar, 1947](https://arxiv.org/html/2609.19523#bib.bib28)\); step comparisons are restricted to runs where both conditions succeed, so early failure is not mistaken for efficiency\.

## 4Experiments

We first test transfer where the correct skill is known, then test deployment where the agent must draw from the library\. The two experiments measure different parts of the system: the first evaluates the stored procedure, whereas the second also depends on deciding which procedure belongs with a new task\.

### 4\.1Experimental Setup

EconWebArena contains 360 live tasks on authoritative portals across government, banking, markets, labor, finance, energy, education, health, and trade\. Our controlled study uses the 100 held\-out variants directly covered by the 50 skill families; our deployment study uses all 360 tasks\. No task requires API credentials, and every answer is obtained through browser interaction with publicly reachable pages\.

Our agent is instantiated ongpt\-5\-mini444[https://developers\.openai\.com/api/docs/models/gpt\-5\-mini](https://developers.openai.com/api/docs/models/gpt-5-mini)within the BrowserGym and AgentLab555[https://github\.com/ServiceNow/AgentLab](https://github.com/ServiceNow/AgentLab)ecosystem\([Le Sellier De Chezelles et al\., 2025](https://arxiv.org/html/2609.19523#bib.bib5)\)\. At each step it receives the page accessibility tree \(AXTree\), a screenshot, and contextual metadata, namely the focused element and the recent action history, and emits one or more actions from the EconWebArena high\-level action set, terminating when it returns a final answer or reaches the 30\-step budget\. The agent configuration, observation space, and step budget are held fixed across all conditions\. The experimental conditions differ only in the trajectory or skill text added to the task prompt\.

We compare the experimental conditions defined in Section[3\.3](https://arxiv.org/html/2609.19523#S3.SS3)\. For each of the 100 variants, BASE, TRAJ1, and MATCH1 are each evaluated in 3 independent runs, using seeds 0, 1, and 2\. This yields 100 variants×\\times3 conditions×\\times3 runs = 900 browser\-agent episodes\. The deployment conditions each contain one run for all 360 tasks\. Supplied trajectories or skills are serialized to Markdown and prepended through the same extra\-instructions channel\. Within each study, the model, browser environment, observations, and action budget are fixed\. We use the native benchmark reward and do not rescue answers from the agent text after a failed automatic check\. Four independently verified benchmark values that had changed at their authoritative source are handled through a fixed correction table constructed without reference to condition outcomes; details and the reproduction protocol are in Appendix[A](https://arxiv.org/html/2609.19523#A1)\.

### 4\.2Experimental Results

We first evaluate EconSkills under controlled transfer with a known skill match, then under library\-scale deployment where the agent must select its own context\. Subsequent analyses examine paired outcomes, representative browser cases, and library coverage\.

#### 4\.2\.1RQ1: Do matched skills transfer to held\-out tasks?

Table[3](https://arxiv.org/html/2609.19523#S4.T3)reports the controlled transfer study\. Supplying the matched extracted skill raises success from 41\.0% to 49\.3%, a gain of \+8\.3 percentage points\. On the 300 matched seed\-level outcomes, MATCH1 changes 46 BASE failures into successes, compared with 21 changes in the opposite direction; McNemar’s exact test givesp=0\.0031p=0\.0031\. This positive imbalance confirms that the gain is distributed across paired outcomes rather than driven only by aggregate averaging, while also motivating the case analysis in RQ3\. Appendix[C](https://arxiv.org/html/2609.19523#A3)provides the cell\-level aggregation and testing details\.

Table 3:Controlled matched\-skill transfer on 100 held\-out variants\. Each variant\-condition pair has 3 independent runs, using seeds 0, 1, and 2; standard deviations are across the 3 run\-level success rates\. The highest success values and lowest mean step count are bold\.By contrast, the raw source trajectory reaches only 7\.0%, although it comes from a successful task in the same family\. Its short mean episode length is caused by early termination rather than efficient completion\. Simply deleting the source answer, while leaving instance\-specific actions and observations in place, does not produce a useful demonstration for the new instance\. This contrast shows that the structured, parameterized skill, rather than prior experience alone, is what transfers to the new instance\.

We compare step counts only for the 102 seed\-level cells in which both BASE and MATCH1 succeed, since a failed run often uses the entire 30\-step budget\. Within this subset, MATCH1 uses 2\.37 fewer actions on average \(median two fewer\)\. It is shorter in 59 cells, tied in 15, and longer in 28 \(two\-sided sign test excluding ties,p=0\.0012p=0\.0012\)\.

#### 4\.2\.2RQ2: How does skill delivery scale from a matched skill to a library?

Table[4](https://arxiv.org/html/2609.19523#S4.T4)reports the experiment in which the correct skill is no longer given\. RETR5 is the strongest library\-scale condition and is statistically tied with BASE overall \(28\.3% and 28\.1%; McNemarp=1\.0p=1\.0\)\. On directly covered tasks, it also leads all conditions at 48\.0%, compared with 44\.0% for BASE\. RAND5 reaches 24\.4% \(p=0\.085p=0\.085against BASE\), while ALL50 reaches 11\.4% \(p<10−12p<10^\{\-12\}\)\. Together, these results show that a compact retrieved set preserves baseline\-level overall performance and leads on the covered subset, whereas indiscriminate full\-library delivery introduces substantial context interference\.

Table 4:EconSkills library\-scale deployment by direct skill\-library coverage\. Cells report successes / tasks \(SR\)\. Each condition has one run per task; the best result in each column is bold and the second\-best is underlined\.The coverage split clarifies the aggregate comparison\. On the 150 covered tasks, RETR5 reaches 48\.0% and HINT30 46\.0%, compared with 44\.0% for BASE\. On the 210 uncovered tasks, BASE is highest at 16\.7%; RETR5 reaches 14\.3%, RAND5 12\.4%, HINT30 11\.4%, and ALL50 6\.7%\. The coverage split therefore explains the overall tie: RETR5 leads on covered tasks, while the uncovered group leaves room for a more selective retrieval policy\.

HINT30 asks whether explicitly allowing the agent to ignore one optional skill reduces this interference\. Its covered\-task difference from BASE is not significant \(p=0\.711p=0\.711\), and both conditions solve 45 of the 100 held\-out variants\. On uncovered tasks, the approximate top\-1 hint lowers success from 16\.7% to 11\.4% \(p=0\.043p=0\.043\)\. Across all 360 tasks, HINT30 obtains 25\.8%, compared with 28\.1% for BASE \(p=0\.341p=0\.341\)\.

We next move from aggregate rates to paired outcomes and preserved trajectories, asking which task and interface properties govern successful reuse\.

#### 4\.2\.3RQ3: What determines successful skill reuse?

Table 5:Pair\-level outcomes in the controlled study\. Categories compare mean success across the 3 runs for each source\-variant pair\.Table[5](https://arxiv.org/html/2609.19523#S4.T5)summarizes the pair\-level outcomes\. The matched skill improves the mean outcome for 27 of the 100 source\-variant pairs, more than twice the 12 pairs that move in the opposite direction; 29 pairs remain successful in at least one condition\. The improvements tend to occur when a portal has a stable series page or selector sequence\. Examples include ONS inflation time series, Treasury yield tables, and CoinMarketCap historical snapshots\. A match at the website or task\-family level is nevertheless not always sufficient\. The boundary cases involve a nearby but different semantic choice, such as a form\-specific fee category or the distinction between close and adjusted close\.

We also read the preserved trajectories from the library\-scale deployment and HINT30 conditions\. They identify two boundaries on reuse\. First, a dynamic selector, download control, or redesigned page can prevent the agent from following a recorded route\. Second, an agent may reach the relevant page but read the wrong unit, date, series, or filing category\. We further observe runs in which an approximately relevant skill keeps the agent on the wrong page after progress has stalled\. These cases help interpret HINT30 on the 210 uncovered tasks and motivate explicit coverage checks before a candidate enters the prompt\.

Figure[2](https://arxiv.org/html/2609.19523#S4.F2)gives a successful transfer and a boundary case\. On ONS task 225, BASE spends 30 steps at a general download interface, while HINT30 reaches the dedicated MM23 time\-series page and succeeds in 14 steps\. On USCIS task 264, by contrast, the hint reaches the relevant I\-140 fee section but the agent does not return a verified answer within 30 steps\. The second case shows that reaching the right page and extracting the requested value are separate parts of the task\. Full step sequences appear in Figures[3](https://arxiv.org/html/2609.19523#A4.F3)and[4](https://arxiv.org/html/2609.19523#A4.F4)in Appendix[D](https://arxiv.org/html/2609.19523#A4)\.

![Refer to caption](https://arxiv.org/html/2609.19523v1/figures/ons_base_step15.png)

\(a\)ONS BASE: dataset page
![Refer to caption](https://arxiv.org/html/2609.19523v1/figures/ons_base_step30.png)

\(b\)ONS BASE: download view
![Refer to caption](https://arxiv.org/html/2609.19523v1/figures/ons_hint_step14.png)

\(c\)ONS HINT30: answer table
![Refer to caption](https://arxiv.org/html/2609.19523v1/figures/uscis_base_step9.png)

\(d\)USCIS BASE: answer table
![Refer to caption](https://arxiv.org/html/2609.19523v1/figures/uscis_hint_step15.png)

\(e\)USCIS HINT30: fee search
![Refer to caption](https://arxiv.org/html/2609.19523v1/figures/uscis_hint_step30.png)

\(f\)USCIS HINT30: I\-140 section

Figure 2:Representative browser states from the library\-scale study\. Top: the matched hint moves the ONS agent from broad dataset and download views to the dedicated series table\. Bottom: both USCIS agents reach relevant content, but the hinted run does not return the verified value\.

### 4\.3Coverage Analysis

The HINT30 aggregate combines three substantively different groups\. Table[6](https://arxiv.org/html/2609.19523#S4.T6)separates the source tasks used to create the skills, their held\-out variants, and tasks for which the library has no direct match\. Source tasks are included as a diagnostic of prompt use, not as transfer evidence\. The held\-out variants are the relevant transfer group, and HINT30 does not change their aggregate success: both conditions solve 45 of 100 tasks\. This differs from the controlled MATCH1 result because HINT30 changes both the presentation and the run protocol: the skill is framed as optional and each task is run once\.

Table 6:HINT30 and BASE outcomes by library coverage\. Source tasks diagnose reuse on the originating tasks; held\-out variants provide the transfer comparison\. The higher success result within each subset is bold; ties are both bold\.The uncovered group isolates the selection problem rather than the quality of the stored skills: by construction, none is the procedure required by those tasks\. The result measures whether retrieval and the optional\-hint instruction can reject an approximate match\. The significant gap identifies this coverage decision as the key remaining target\. Together with the controlled study, the split yields a more precise conclusion: the library contains transferable procedures, and end\-to\-end gains depend on filtering candidates before prompt construction\. By construction, all 150 covered tasks receive the exact family skill, while each uncovered task receives the top\-1 retrieved skill used by that condition\.

### 4\.4Discussion

The two EconSkills studies separate procedural transfer from library selection\. MATCH1 measures whether a stored procedure helps after the relevant skill is known; under this controlled match, it improves success and reduces actions on paired successful runs\. RETR5 and HINT30 additionally require the agent to receive and use suitable context\. Their library\-scale outcomes therefore complement the controlled result: useful procedures transfer, while selection remains the main opportunity for improvement\.

The contrast between TRAJ1, MATCH1, and ALL50 shows why the form and amount of experience matter\. A raw trajectory mixes reusable steps with instance\-specific state\. A full library instead makes many plausible but inapplicable procedures compete for attention\. This pattern is consistent with context interference, although we do not isolate prompt length, conflicting instructions, and relevance judgments\. Effective reuse therefore requires explicit skill scope and selective prompt construction\.

The coverage split suggests a concrete design target\. RETR5 leads BASE on directly covered tasks; the overall tie arises after uncovered tasks are included\. An operational system should therefore withhold skills when no suitable match is found by checking the portal, data type, and interaction schema against skill preconditions\. Verification should then confirm the final page, unit, date, entity, series definition, adjustment status, release vintage, and quote orientation when relevant\. Evaluations should likewise report covered and uncovered tasks alongside the aggregate: covered tasks measure reuse when a matching skill exists, while the aggregate can conceal useful transfer behind retrieval failures\.

## 5Conclusion

EconSkills demonstrates that verified web interactions can be distilled into reusable procedural knowledge for live economic data\. In controlled transfer, parameterized skills outperform no\-skill prompting and raw trajectories on held\-out variants and shorten paired successful interactions\. At library scale, RETR5 is the strongest skill\-delivery condition, matches BASE overall, and leads it numerically on directly covered tasks\. Approximate candidates outside the library’s scope offset gains from well\-matched procedures, while full\-library delivery shows the cost of excessive context\. EconSkills thus makes skill reuse measurable as procedural transfer and coverage\-aware selection, with concrete checks for portal, series, date, unit, entity, and provenance\. This framework points toward web agents that accumulate and reliably reuse economic\-domain experience\.

## Limitations

We study one benchmark, one agent backbone, and a fixed 30\-step budget\. The controlled matched\-skill study assumes that the corresponding family skill is known, so it isolates procedural transfer rather than retrieval; the full\-benchmark study instead combines library coverage, selection, and skill use\. Each full\-benchmark condition has one run per task, enabling paired comparisons but not across\-run variance estimates\. The 50 skills come from verified trajectories under one extraction schema, so results may change with other models, larger libraries, or evolving sites\. Future work should repeat deployment across seeds and test coverage\-aware skill rejection under site change\.

## References

- Denget al\.\(2023\)X\. Deng, Y\. Gu, B\. Zheng, S\. Chen, S\. Stevens, B\. Wang, H\. Sun, and Y\. SuMind2Web: Towards a Generalist Agent for the Web\.InAdvances in Neural Information Processing Systems \(NeurIPS\),Note:arXiv:2306\.06070Cited by:[§2](https://arxiv.org/html/2609.19523#S2.p1.1)\.
- Gouet al\.\(2025\)B\. Gou, Z\. Huang, Y\. Ning, Y\. Gu, M\. Lin, W\. Qi, A\. Kopanev, B\. Yu, B\. Jiménez Gutiérrez, Y\. Shu, C\. H\. Song, J\. Wu, S\. Chen, H\. N\. Moussa, T\. Zhang, J\. Xie, Y\. Li, T\. Xue, Z\. Liao, K\. Zhang, B\. Zheng, Z\. Cai, V\. Rozgic, M\. Ziyadi, H\. Sun, and Y\. SuMind2Web 2: Evaluating Agentic Search with Agent\-as\-a\-Judge\.InAdvances in Neural Information Processing Systems \(NeurIPS\),Cited by:[§2](https://arxiv.org/html/2609.19523#S2.p1.1)\.
- Heet al\.\(2024\)H\. He, W\. Yao, K\. Ma, W\. Yu, Y\. Dai, H\. Zhang, Z\. Lan, and D\. YuWebVoyager: Building an End\-to\-End Web Agent with Large Multimodal Models\.InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics \(ACL\),Note:arXiv:2401\.13919Cited by:[§1](https://arxiv.org/html/2609.19523#S1.p1.1)\.
- Janget al\.\(2024\)L\. Jang, Y\. Li, D\. Zhao, C\. Ding, J\. Lin, P\. P\. Liang, R\. Bonatti, and K\. KoishidaVideoWebArena: Evaluating Long Context Multimodal Agents with Video Understanding Web Tasks\.arXiv preprint arXiv:2410\.19100\.Cited by:[§2](https://arxiv.org/html/2609.19523#S2.p2.1)\.
- Kimet al\.\(2024\)M\. Kim, V\. Bursztyn, E\. Koh, S\. Guo, and S\. HwangRaDA: Retrieval\-Augmented Web Agent Planning with LLMs\.InFindings of the Association for Computational Linguistics: ACL 2024,External Links:[Document](https://dx.doi.org/10.18653/v1/2024.findings-acl.802)Cited by:[§2](https://arxiv.org/html/2609.19523#S2.p2.1)\.
- Kohet al\.\(2024\)J\. Y\. Koh, R\. Lo, L\. Jang, V\. Duvvur, M\. Lim, P\. Huang, G\. Neubig, S\. Zhou, R\. Salakhutdinov, and D\. FriedVisualWebArena: Evaluating Multimodal Agents on Realistic Visual Web Tasks\.InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics \(ACL\),External Links:[Document](https://dx.doi.org/10.18653/v1/2024.acl-long.50)Cited by:[§2](https://arxiv.org/html/2609.19523#S2.p1.1)\.
- Le Sellier De Chezelleset al\.\(2025\)T\. Le Sellier De Chezelles, M\. Gasse, A\. Drouin, M\. Caccia, L\. Boisvert, M\. Thakkar, T\. Marty, R\. Assouel, S\. Omidi Shayegan, L\. K\. Jang, X\. H\. Lù, O\. Yoran, D\. Kong, F\. F\. Xu, S\. Reddy, Q\. Cappart, G\. Neubig, R\. Salakhutdinov, N\. Chapados, and A\. LacosteThe BrowserGym Ecosystem for Web Agent Research\.Transactions on Machine Learning Research \(TMLR\)\.Note:arXiv:2412\.05467Cited by:[§2](https://arxiv.org/html/2609.19523#S2.p1.1),[§4\.1](https://arxiv.org/html/2609.19523#S4.SS1.p2.1)\.
- Liet al\.\(2024\)N\. Li, C\. Gao, M\. Li, Y\. Li, and Q\. LiaoEconAgent: Large Language Model\-Empowered Agents for Simulating Macroeconomic Activities\.InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics,pp\. 15523–15536\.External Links:[Document](https://dx.doi.org/10.18653/v1/2024.acl-long.829)Cited by:[§1](https://arxiv.org/html/2609.19523#S1.p1.1)\.
- Liet al\.\(2026\)X\. Li, W\. Chen, Y\. Liu, S\. Zheng, X\. Chen, Y\. He,et al\.SkillsBench: Benchmarking How Well Agent Skills Work Across Diverse Tasks\.arXiv preprint arXiv:2602\.12670\.Cited by:[§2](https://arxiv.org/html/2609.19523#S2.p3.1)\.
- Liu and Quan \(2025\)Z\. Liu and Y\. QuanEconWebArena: Benchmarking Autonomous Agents on Economic Tasks in Realistic Web Environments\.InNeurIPS Workshop on Bridging Language, Agent, and World Models for Reasoning and Planning \(LAW\),Note:arXiv:2506\.08136Cited by:[§1](https://arxiv.org/html/2609.19523#S1.p1.1),[§2](https://arxiv.org/html/2609.19523#S2.p1.1)\.
- Lùet al\.\(2024\)X\. H\. Lù, Z\. Kasner, and S\. ReddyWebLINX: Real\-World Website Navigation with Multi\-Turn Dialogue\.InInternational Conference on Machine Learning \(ICML\),Note:arXiv:2402\.05930Cited by:[§2](https://arxiv.org/html/2609.19523#S2.p1.1)\.
- Luet al\.\(2026\)Z\. Lu, Y\. Zuo, Y\. Nie, X\. He, W\. Fan, and C\. DaiContractSkill: Repairable Contract\-Based Skills for Multimodal Web Agents\.arXiv preprint arXiv:2603\.20340\.Cited by:[§2](https://arxiv.org/html/2609.19523#S2.p3.1)\.
- Maet al\.\(2026\)Y\. Ma, Y\. Huang, H\. Bao, H\. Zhuang, S\. Shukla, M\. Galley, X\. Zhang, and S\. FeuerriegelSkillGen: Verified Inference\-Time Agent Skill Synthesis\.arXiv preprint arXiv:2605\.10999\.Cited by:[§2](https://arxiv.org/html/2609.19523#S2.p3.1)\.
- McNemar \(1947\)Q\. McNemarNote on the Sampling Error of the Difference Between Correlated Proportions or Percentages\.Psychometrika12\(2\)\.External Links:[Document](https://dx.doi.org/10.1007/BF02295996)Cited by:[§3\.3](https://arxiv.org/html/2609.19523#S3.SS3.p5.1)\.
- Panet al\.\(2024\)Y\. Pan, D\. Kong, S\. Zhou, C\. Cui, Y\. Leng, B\. Jiang, H\. Liu, Y\. Shang, S\. Zhou, T\. Wu, and Z\. WuWebCanvas: Benchmarking Web Agents in Online Environments\.arXiv preprint arXiv:2406\.12373\.Cited by:[§2](https://arxiv.org/html/2609.19523#S2.p1.1)\.
- Quanet al\.\(2025\)Y\. Quan, X\. Li, and Y\. ChenCRMAgent: A Multi\-Agent LLM System for E\-Commerce CRM Message Template Generation\.InGenAIECommerce: Workshop on Agentic and Generative AI for E\-Commerce at RecSys,Note:arXiv:2507\.08325Cited by:[§1](https://arxiv.org/html/2609.19523#S1.p1.1)\.
- Quanet al\.\(2026\)Y\. Quan, Z\. Liu, F\. Benaben, and B\. MontreuilLeveraging large language models to enhance multi\-agent risk assessment in supply chain networks\.International Journal of Production Research\.External Links:[Document](https://dx.doi.org/10.1080/00207543.2026.2619562)Cited by:[§1](https://arxiv.org/html/2609.19523#S1.p1.1)\.
- Quan and Liu \(2024a\)Y\. Quan and Z\. LiuEconLogicQA: A Question\-Answering Benchmark for Evaluating Large Language Models in Economic Sequential Reasoning\.InFindings of the Association for Computational Linguistics: EMNLP 2024,Note:arXiv:2405\.07938Cited by:[§1](https://arxiv.org/html/2609.19523#S1.p1.1)\.
- Quan and Liu \(2024b\)Y\. Quan and Z\. LiuInvAgent: A Large Language Model based Multi\-Agent System for Inventory Management in Supply Chains\.arXiv preprint arXiv:2407\.11384\.Cited by:[§1](https://arxiv.org/html/2609.19523#S1.p1.1)\.
- Shinnet al\.\(2023\)N\. Shinn, F\. Cassano, E\. Berman, A\. Gopinath, K\. Narasimhan, and S\. YaoReflexion: Language Agents with Verbal Reinforcement Learning\.InAdvances in Neural Information Processing Systems \(NeurIPS\),Note:arXiv:2303\.11366Cited by:[§1](https://arxiv.org/html/2609.19523#S1.p2.1)\.
- Wanget al\.\(2024\)G\. Wang, Y\. Xie, Y\. Jiang, A\. Mandlekar, C\. Xiao, Y\. Zhu, L\. Fan, and A\. AnandkumarVoyager: An Open\-Ended Embodied Agent with Large Language Models\.Transactions on Machine Learning Research \(TMLR\)\.Note:arXiv:2305\.16291Cited by:[§2](https://arxiv.org/html/2609.19523#S2.p3.1)\.
- Wanget al\.\(2025\)Z\. Z\. Wang, J\. Mao, D\. Fried, and G\. NeubigAgent Workflow Memory\.InProceedings of the 42nd International Conference on Machine Learning,Proceedings of Machine Learning Research, Vol\.267\.Cited by:[§2](https://arxiv.org/html/2609.19523#S2.p2.1)\.
- Yanget al\.\(2024\)H\. Yang, B\. Zhang, N\. Wang, C\. Guo, X\. Zhang, L\. Lin, J\. Wang, T\. Zhou, M\. Guan, R\. Zhang, and C\. D\. WangFinRobot: An Open\-Source AI Agent Platform for Financial Applications using Large Language Models\.arXiv preprint arXiv:2405\.14767\.Cited by:[§1](https://arxiv.org/html/2609.19523#S1.p1.1)\.
- Yuet al\.\(2025\)S\. Yu, G\. Li, W\. Shi, and P\. QiPolySkill: Learning Generalizable Skills Through Polymorphic Abstraction\.arXiv preprint arXiv:2510\.15863\.Cited by:[§2](https://arxiv.org/html/2609.19523#S2.p3.1)\.
- Zhaoet al\.\(2024\)A\. Zhao, D\. Huang, Q\. Xu, M\. Lin, Y\. Liu, and G\. HuangExpeL: LLM Agents Are Experiential Learners\.InProceedings of the AAAI Conference on Artificial Intelligence,Vol\.38\.External Links:[Document](https://dx.doi.org/10.1609/aaai.v38i17.29936)Cited by:[§2](https://arxiv.org/html/2609.19523#S2.p2.1)\.
- Zhenget al\.\(2025\)B\. Zheng, M\. Y\. Fatemi, X\. Jin, Z\. Z\. Wang, A\. Gandhi, Y\. Song, Y\. Gu, J\. Srinivasa, G\. Liu, G\. Neubig, and Y\. SuSkillWeaver: Web Agents Can Self\-Improve by Discovering and Honing Skills\.arXiv preprint arXiv:2504\.07079\.Cited by:[§2](https://arxiv.org/html/2609.19523#S2.p3.1)\.
- Zhenget al\.\(2024\)B\. Zheng, B\. Gou, J\. Kil, H\. Sun, and Y\. SuGPT\-4V\(ision\) is a Generalist Web Agent, if Grounded\.InInternational Conference on Machine Learning \(ICML\),Note:arXiv:2401\.01614Cited by:[§1](https://arxiv.org/html/2609.19523#S1.p1.1)\.
- Zhouet al\.\(2024\)S\. Zhou, F\. F\. Xu, H\. Zhu, X\. Zhou, R\. Lo, A\. Sridhar, X\. Cheng, T\. Ou, Y\. Bisk, D\. Fried, U\. Alon, and G\. NeubigWebArena: A Realistic Web Environment for Building Autonomous Agents\.InInternational Conference on Learning Representations \(ICLR\),Note:arXiv:2307\.13854Cited by:[§2](https://arxiv.org/html/2609.19523#S2.p1.1)\.

## Appendix AReproduction Details

The 50 EconSkills procedures are available on Hugging Face666[https://huggingface\.co/datasets/EconWebArena/EconSkills](https://huggingface.co/datasets/EconWebArena/EconSkills)\. Each JSON file contains the seven fields defined in Section[3](https://arxiv.org/html/2609.19523#S3): name, purpose, preconditions, procedure, website guidance, verification, and recovery\. The same fields are rendered to Markdown when a skill is placed in the agent prompt\. Appendix[B](https://arxiv.org/html/2609.19523#A2)shows two examples from the library in full\.

The experiments can be rerun on EconWebArena through BrowserGym and AgentLab using the agent configuration in Section[4\.1](https://arxiv.org/html/2609.19523#S4.SS1)and the prompt conditions in Section[3\.3](https://arxiv.org/html/2609.19523#S3.SS3)\. The controlled study evaluates 100 variants under 3 conditions and 3 runs per condition\. The deployment study evaluates each condition once on all 360 tasks\. Both studies use the native benchmark reward: the returned answer must contain the gold value and the browser must finish on the authoritative domain\.

Four benchmark values that had changed on their authoritative websites are handled through a fixed task\-level correction table constructed without reference to condition outcomes\. The HINT30 prompt states that the supplied skill may be incomplete, outdated, or mismatched and should be ignored when it does not fit the visible page\. Its 150 directly covered tasks receive the corresponding family skill, while the remaining 210 tasks receive the top\-1 retrieved skill\.

## Appendix BSkill Examples

The first example is the ECB skill summarized in Table[2](https://arxiv.org/html/2609.19523#S3.T2); the second covers a different workflow on the USCIS fee schedule\. Both preserve the seven fields and operational content of the stored JSON skills\. Angle\-bracketed terms denote values supplied by a new task rather than copied from the source trajectory\.

### B\.1ECB Reference Rate

Name\.lookup\_ecb\_euro\_reference\_fx\_rate

Purpose\.Retrieve the ECB euro foreign\-exchange reference rate for a specified currency and date\.

Preconditions\.The task requests a euro\-to\-currency exchange rate from the European Central Bank; a target calendar date is provided or inferable; and the ECB website is reachable\.

Procedure\.

1. 1\.Open the ECB “Euro foreign exchange reference rates” section\.
2. 2\.Navigate to the target\-currency page\. ECB quotes units of the target currency per euro\.
3. 3\.Set the start and end of the date range to<date\>\.
4. 4\.Switch to a table or another exact\-value view instead of relying on a chart tooltip\.
5. 5\.Read the rate for<date\>and preserve four decimal places\.

Website guidance\.Currency pages follow the patterneurofxref\-graph\-<currency\>\.en\.html\. The date inputs use day/month/year; setting both fields to the same day isolates one observation\. Use “Show table” or the CSV download when the chart is ambiguous\.

Verification\.Confirm the ECB domain and page heading, the requested currency, and the target\-currency\-per\-EUR orientation\. Check that the displayed date exactly matches<date\>and that the value is a nominal reference rate rather than a percentage change or index\.

Recovery\.If the currency page is missing, return to the landing page or use ECB site search\. If the chart does not expose an exact value, use the table or CSV view\. If date entry fails, use the date picker\. For a non\-business day, do not silently substitute the previous observation unless the task permits it\. If the layout changes, navigate from the ECB home page through Stats→\\rightarrowPolicy and exchange rates, or use site search\.

### B\.2USCIS Fee Schedule

Name\.lookup\_uscis\_fee\_schedule

Purpose\.Retrieve a USCIS fee for a specified form or process and effective date from the official fee schedule onuscis\.gov\.

Preconditions\.The task identifies a form or process, any relevant fee qualifier such as per beneficiary or per filing, and an effective date\. The answer must come from an authoritative USCIS page and use the requested dollar format\.

Procedure\.

1. 1\.Open the USCIS Fee Schedule and confirm that its effective date matches<date\>\.
2. 2\.Use the schedule filters, search field, or find\-in\-page to locate<form\-or\-process\>\.
3. 3\.Read the full matching entry and select the line corresponding to<fee\-qualifier\>and the requested operation\.
4. 4\.Read the listed fee and check any notes, surcharges, or exemptions that could change its interpretation\.
5. 5\.Normalize the amount to the output format requested by the task\.

Website guidance\.USCIS maintains the schedule atuscis\.gov/g\-1055\. Its effective date appears near the top of the page, and topic filters or a search box may appear above the table\. If a filter does not respond, clear it and use find\-in\-page with the form number or process name\. News and rulemaking pages may explain a fee change, but the requested amount should be read from the applicable fee schedule\.

Verification\.Confirm theuscis\.govdomain, the fee schedule title, and the effective date\. Check that the selected row matches the form, operation, and qualifier exactly, and that the value is the requested fee rather than a separate surcharge, discounted amount, or exemption\. When several rows have similar names, compare their full descriptions before selecting one\.

Recovery\.If the schedule does not load, use USCIS site search for “Fee Schedule” or “G\-1055\.” If the displayed schedule has the wrong effective date, follow its links to the applicable version\. If dynamic filters fail, reload the page, use a text or print view when available, or search the unfiltered table\. When official USCIS pages show conflicting amounts, prioritize the fee schedule whose effective date matches the task\.

## Appendix CStatistical Aggregation Details

The controlled result table is indexed by source task, held\-out variant, condition, and seed\. This gives50×2×3×3=90050\\times 2\\times 3\\times 3=900cells\. For each cell, we use the native benchmark reward\. The 300 matched BASE/MATCH1 cells yield 46 failure\-to\-success changes and 21 success\-to\-failure changes\. At the source\-variant level, we average the three runs within each condition before assigning the four outcome patterns in Table[5](https://arxiv.org/html/2609.19523#S4.T5); this prevents one unusually easy seed from being reported as a separate task family\.

For the deployment study, each condition has one row for each of the 360 tasks\. The HINT30 coverage label is determined by task construction: the 50 original source tasks and 100 variants have a direct family skill, whereas the other 210 tasks do not\. The latter still receive a retrieved hint, so “no direct match” does not mean an empty prompt\. Pairwise tests use the BASE and treatment reward for the same task\. Because there is only one run per task, we do not report across\-seed error bars for this study\.

## Appendix DQualitative Browser Cases

Figures[3](https://arxiv.org/html/2609.19523#A4.F3)to[5](https://arxiv.org/html/2609.19523#A4.F5)compare browser paths from three library\-scale deployment cases\. The ONS and Treasury cases show the positive mechanism: a suitable skill names a portal\-specific destination and recovery route, which avoids broad download interfaces or dead links \(Figures[3](https://arxiv.org/html/2609.19523#A4.F3)and[5](https://arxiv.org/html/2609.19523#A4.F5)\)\. The USCIS comparison separates navigation from answer validation: both agents reach relevant content, but the hinted run does not complete extraction and verification \(Figure[4](https://arxiv.org/html/2609.19523#A4.F4)\)\. Figure[6](https://arxiv.org/html/2609.19523#A4.F6)then compares BASE and RETR5 endpoints on ONS, USCIS, and SBA tasks\. RETR5 shortens the successful ONS and SBA runs, while the USCIS pair isolates extraction as a separate issue\. Together, the cases show that skills can reduce route\-finding effort and that economic values still require explicit semantic checks\.

![Refer to caption](https://arxiv.org/html/2609.19523v1/figures/ons_base_step0.png)

\(a\)BASE, step 0: ONS home
![Refer to caption](https://arxiv.org/html/2609.19523v1/figures/ons_base_step15.png)

\(b\)BASE, step 15: dataset page
![Refer to caption](https://arxiv.org/html/2609.19523v1/figures/ons_base_step30.png)

\(c\)BASE, step 30: download view
![Refer to caption](https://arxiv.org/html/2609.19523v1/figures/ons_base_step0.png)

\(d\)HINT30, step 0: ONS home
![Refer to caption](https://arxiv.org/html/2609.19523v1/figures/ons_hint_step7.png)

\(e\)HINT30, step 7: series page
![Refer to caption](https://arxiv.org/html/2609.19523v1/figures/ons_hint_step14.png)

\(f\)HINT30, step 14: answer table

Figure 3:Browser trajectories for Task 225 on ONS\. BASE remains at a general dataset download interface after 30 steps\. With the matched optional hint, the agent reaches the dedicated MM23 time\-series table and succeeds in 14 steps\.![Refer to caption](https://arxiv.org/html/2609.19523v1/figures/uscis_base_step0.png)

\(a\)BASE, step 0: USCIS home
![Refer to caption](https://arxiv.org/html/2609.19523v1/figures/uscis_base_step5.png)

\(b\)BASE, step 5: fee search
![Refer to caption](https://arxiv.org/html/2609.19523v1/figures/uscis_base_step9.png)

\(c\)BASE, step 9: answer table
![Refer to caption](https://arxiv.org/html/2609.19523v1/figures/uscis_hint_step0.png)

\(d\)HINT30, step 0: USCIS home
![Refer to caption](https://arxiv.org/html/2609.19523v1/figures/uscis_hint_step15.png)

\(e\)HINT30, step 15: fee search
![Refer to caption](https://arxiv.org/html/2609.19523v1/figures/uscis_hint_step30.png)

\(f\)HINT30, step 30: I\-140 section

Figure 4:Browser trajectories for Task 264 on USCIS\. BASE reaches the I\-140 fee table and succeeds in 9 steps\. HINT30 reaches the relevant section but does not return the verified value within 30 steps, separating correct navigation from correct extraction\.![Refer to caption](https://arxiv.org/html/2609.19523v1/figures/treasury_base_step0.png)

\(a\)BASE, step 0: Treasury home
![Refer to caption](https://arxiv.org/html/2609.19523v1/figures/treasury_base_step15.png)

\(b\)BASE, step 15: site search
![Refer to caption](https://arxiv.org/html/2609.19523v1/figures/treasury_base_step30.png)

\(c\)BASE, step 30: dead link
![Refer to caption](https://arxiv.org/html/2609.19523v1/figures/treasury_hint_step0.png)

\(d\)HINT30, step 0: Treasury home
![Refer to caption](https://arxiv.org/html/2609.19523v1/figures/treasury_hint_step8.png)

\(e\)HINT30, step 8: targeted search
![Refer to caption](https://arxiv.org/html/2609.19523v1/figures/treasury_hint_step15.png)

\(f\)HINT30, step 15: yield table

Figure 5:Browser trajectories for Task 274 on the U\.S\. Treasury site\. BASE remains in site search and ends on a dead link after 30 steps\. The matched hint names the daily par\-yield route and its recovery strategy; the agent reaches the historical table and succeeds in 15 steps\.![Refer to caption](https://arxiv.org/html/2609.19523v1/figures/ons_base_step30.png)

\(a\)ONS BASE, step 30: unresolved
![Refer to caption](https://arxiv.org/html/2609.19523v1/figures/uscis_base_step9.png)

\(b\)USCIS BASE, step 9: success
![Refer to caption](https://arxiv.org/html/2609.19523v1/figures/task306_base_step12.png)

\(c\)SBA BASE, step 12: success
![Refer to caption](https://arxiv.org/html/2609.19523v1/figures/ons_hint_step14.png)

\(d\)ONS RETR5, step 5: success
![Refer to caption](https://arxiv.org/html/2609.19523v1/figures/uscis_retr5_step30.png)

\(e\)USCIS RETR5, step 30: unresolved
![Refer to caption](https://arxiv.org/html/2609.19523v1/figures/task306_retr5_step3.png)

\(f\)SBA RETR5, step 3: success

Figure 6:BASE and RETR5 endpoints on three economic web tasks\. RETR5 converts the ONS endpoint into a 5\-step success and reaches the SBA evidence in 3 rather than 12 steps\. The USCIS pair shows that reaching the correct fee page still requires exact extraction\. Together, the cases illustrate retrieval gains and the role of semantic verification; aggregate performance is reported in Table[4](https://arxiv.org/html/2609.19523#S4.T4)\.