PaperlayerView exact revision on arXiv

Source-linked AI summary

What Ethereum Developers Ask For, and Whether They Get It: Ten Years of Ethereum Stack Exchange

Tamer Abdelaziz

arXiv:2610.30829v1cs.SE

TL;DR

Earlier ESE studies covered individual topics or platform facets, but none tracked resolution across its full history, leaving open what developers ask and whether the venue still resolves it. The study labels every question by topic and arithmetically separates changes in resolution due to task mix from changes in whether the same tasks are resolved, using the complete record. Acceptance spans 47 points across asker-reputation buckets versus 17 across topics, while the shift toward newer askers accompanies up to 59% of the acceptance fall.

  • Problem

    Earlier ESE studies covered individual topics or platform facets, but none tracked resolution across its full history, leaving open what developers ask and whether the venue still resolves it.

    Source: §I · p. 1

  • Method

    The study labels every question by topic and arithmetically separates changes in resolution due to task mix from changes in whether the same tasks are resolved, using the complete record.

    Source: §I · p. 1

  • Results

    Acceptance spans 47 points across asker-reputation buckets versus 17 across topics, while the shift toward newer askers accompanies up to 59% of the acceptance fall.

    Source: §I · p. 2

  • Takeaways & Limitations

    The study locates the support gap in the toolchain and among newcomers, with answers still arriving but closure falling further.

    Source: §VII · p. 10

    ESE users are English-speaking developers who post publicly, and the archive underrepresents problems developers solve alone.

    Source: B. External Validity · p. 9

Abstract

from arXiv · show

The Ethereum community learns where developers struggle through surveys and interviews, which record what developers say when asked about a chosen theme. Ethereum Stack Exchange (ESE) records what they ask for unprompted, and whether anyone resolves it. We study every question posted to ESE from Jan 2016 to Dec 2025, 54,637 in all, and label each with a prompt-based topic pipeline seeded with eight Ethereum-specific topics and open to new ones (22 topics in all). A three-provider committee of language models screens the labels, and manual adjudication of every disputed case puts their accuracy at 87.9 to 93.5%. Because the corpus is the complete record, no rate carries sampling error. We ask what developers ask for help with, whether the venue still resolves it, and, where it does not, whether the questions changed or their resolution did. Demand is concentrated on writing, deploying, and debugging contracts; the toolchain is the second-largest topic and growing, while security, consensus, and scaling are thin. Supply has weakened, and unevenly: answers still arrive, closure falls further. Across five estimators of the decade, the share of questions drawing any answer fell by 9.5 to 18.6 percentage points, but among answered questions the share reaching an accepted answer fell by 16.2 to 24.5. Rebuilding the later acceptance rate at the earlier topic mix moves the fall only from 9.40 to 9.32 points, so the topic mix does not explain it. Who asks separates outcomes far more sharply than what is asked: acceptance runs from 19.8% for askers below 10 reputation to 66.8% at 10k and above, a 47- point spread with a gradient in every year, against 17 across topics, and a shift toward newer askers accounts for up to 59% of the acceptance fall. The largest of seven properties an asker controls accompanies 5 points. The record places the support gap in the toolchain and among newcomers.

I. INTRODUCTION

This study uses Ethereum Stack Exchange’s unprompted questions and their resolution outcomes to examine what Ethereum developers need help with and whether the venue still provides it. It finds that resolution has weakened, especially at the acceptance stage, and that asker reputation separates outcomes more sharply than topic does.

I. INTRODUCTION

  • ESE records problems developers post on their own initiative alongside whether other developers resolve them, providing a decade of revealed demand and community supply.Unlike surveys and interviews organized around a chosen theme, the archive captures questions developers raise unprompted.Source: §I · p. 1
  • Earlier ESE studies covered individual subtopics or platform facets, but none tracked resolution across the platform’s full history; annual question volume fell from 8,316 in 2022 to 873 in 2025.The open question is whether the remaining questions changed or the platform’s resolution of them did.Source: §I · p. 1
  • The study labels all 54,637 questions from January 2016 to December 2025 using TopicGPT seeded with eight Ethereum-specific topics and expanded to 22, then screens and adjudicates labels.The three-provider language-model committee and manual adjudication put label accuracy between 87.9% and 93.5%.Source: §I · pp. 1–2
  • Across five estimators, the answered share fell 9.5–18.6 percentage points and acceptance among answered questions fell 16.2–24.5 points; topic mix explains little of the acceptance decline.At the earlier topic mix, the later acceptance fall is 9.32 rather than 9.40 points.Source: §I · p. 2
  • Acceptance spans 47 points across asker-reputation buckets, versus 17 across topics, and the shift toward newer askers accompanies up to 59% of the acceptance fall.The largest association among seven asker-controlled properties is 5 points.Source: §I · p. 2
  • The paper contributes a reusable decade-scale record of developer needs and a two-stage resolution analysis that separates task mix from resolution within tasks.It also locates the support gap in the toolchain and among newcomers.Source: §I · p. 2

II. MINING AND LABELING ETHEREUM STACK EXCHANGE

The study constructs a labeled corpus of ESE questions and defines the outcomes used to measure resolution. It adapts an open-ended topic-labeling pipeline to Ethereum questions while retaining multi-topic assignments.

II. MINING AND LABELING ETHEREUM STACK EXCHANGE

  • The study needs a topic label for each question and an analysis that separates changes in task mix from changes in resolution on the complete record.Topic-level analyses use the multi-label question-topic pairs; question-level analyses use questions.Source: §II · p. 2, B. Topic Labeling · p. 3

1) Question Outcomes:

  • Each question is accepted, orphaned, or never answered; the answered share combines accepted and orphaned questions, while closure is the accepted share among answered questions.Acceptance is a community-approval signal, not a correctness or security signal, and reputation measures past platform activity rather than topic expertise.Source: 1) Question Outcomes · p. 2

2) Corpus:

  • The corpus contains 54,637 questions posted from January 2016 through December 2025, with question text, tags, timestamps, and asker and accepted-answerer reputation.Bodies are truncated at 15,000 characters for 0.4% of questions, and reputation is recorded as an export-time snapshot rather than at posting.Source: 2) Corpus · p. 2

B. Topic Labeling

  • Questions are classified into named topics with natural-language definitions using a generation, refinement, and assignment pipeline seeded with eight Ethereum-specific topics and left open to new labels.The resulting taxonomy has 22 canonical topics; 12,148 questions have multiple labels, yielding 72,572 question-topic pairs.Source: B. Topic Labeling · pp. 2–3

C. Label Validation

The paper checks topic assignments with a three-model committee and manual review of every flagged case, then analyzes resolution as a two-stage funnel. Its methods separate topic-mix change from within-topic change and compare asker-controlled properties within strata.

1) Committee Screening:

  • Three independent language-model judges audit 660 assignments, flagging a label only when at least two judges vote it absent and cite concrete post evidence.The committee validates 570 assignments and flags 90; pairwise agreement is 86.7%–89.5%.Source: 1) Committee Screening · p. 3

2) Manual Adjudication:

  • Manual review of 90 flagged assignments confirms 43 misclassifications, overturns 10 false alarms, and treats 37 multi-topic cases as ambiguous, bounding verified accuracy at 87.9%–93.5%.Contract Architecture and Upgradeability passes at 50%, while Cryptography, Debugging and Testing, and Solidity Language pass at 100%.Source: 2) Manual Adjudication · p. 3

D. Analysis Methods

  • Resolution is analyzed in two stages—whether a question receives an answer and whether an answered question is accepted—with each rate using its own denominator.Five decade-change estimators are reported because endpoint contrasts depend heavily on the small 2025 corpus.Source: D. Analysis Methods · p. 3, Table I · p. 4
  • Direct standardization rebuilds later resolution rates at the earlier topic mix, separating the mix contribution from changes within topics; the split is exact arithmetic on the labels.Property comparisons use within-stratum differences across reputation, primary topic, and era, then take a Mantel–Haenszel weighted mean.Source: D. Analysis Methods · pp. 3–4
  • The reputation analysis compares rates across five asker buckets, checks the gradient within each year, and standardizes for changes in the reputation mix.Within-year comparisons address the possibility that askers had unequal time to accumulate reputation.Source: D. Analysis Methods · p. 3

III. RESULTS

The study asks what Ethereum developers seek help with, where resolution changed, whether task mix explains that change, and how outcomes relate to asker reputation and question properties. Its analysis is designed around the complete archive and a distinction between the two resolution stages.

A. Setup

  • The analysis uses the complete 54,637-question corpus, reports rates and percentage-point differences, and splits before and after the end of 2022 for comparisons.A replication package rebuilds the labeled corpus, and an independent verification script recomputes the published numbers.Source: A. Setup · p. 4

B. Research Questions

  • The research asks what developers request help with and whether ESE changed in answering questions or in accepting answers, the two stages of resolution.It also asks whether changes reflect task mix or resolution within tasks, how outcomes vary with asker reputation, and which asker-controlled question properties associate with acceptance.Source: B. Research Questions · p. 4

C. Revealed Demand and the Stage of Decline (RQ1)

ESE demand is concentrated on practical contract work, while some growing topics are served slowly. Resolution weakens at both stages, with acceptance among answered questions falling more than the answered share.

1) What Developers Ask For:

  • Solidity Language and Deployment and Tooling account for 34.9% of question-topic pairs; the top five topics account for 60.6%.Figure 1 shows the 22 canonical topics by question-topic pair count and share of all pairs.Source: 1) What Developers Ask For · p. 4, Figure 1 · p. 4

2) Demand Against Supply:

  • Deployment and Tooling is the second-largest topic and grew by 1.5 points between eras, but it is served slowly, with 33.8% accepted.User Experience and Debugging and Testing also grew and are among the three slowest topics after Networking.Source: 2) Demand Against Supply · pp. 4–5, Table II · p. 5

3) Where Resolution Failed:

  • Across five estimators, the answered share fell by 9.5–18.6 percentage points, while acceptance among answered questions fell by 16.2–24.5 points.All five estimators agree on the direction, and each puts the acceptance fall above the answered-share fall.Source: 3) Where Resolution Failed · p. 5

4) Latency and the End of the Record:

The decline in acceptance is largely within topics rather than due to a changing topic mix, and it is concentrated at the closure stage. First-answer waits also lengthened substantially by 2025.

4) Latency and the End of the Record:

D. Task Mix Versus Resolution of the Same Tasks (RQ2)

1) The Spread the Mix Could Exploit:

  • Acceptance across topics ranged from 26.9% for Wallets to 44.3% for Solidity Language, a 17.3-point spread.Answered share ranged from 73.3% for Networking to 85.6% for Solidity Language.Source: 1) The Spread the Mix Could Exploit · p. 5

3) The Stage That Moved:

  • Within-topic movement was −9.6 points on acceptance, compared with −4.0 points on answering; the mix terms were +0.4 and −0.2 points, respectively.The authors note that labels cannot detect a shift in subject matter within a task.Source: 3) The Stage That Moved · pp. 5–6, Table III · p. 6

2) Acceptance at the Earlier Mix:

E. Asker Reputation and the Answerer Pool (RQ3)

Asker reputation separates acceptance outcomes more sharply than topic, and the gradient persists within years. A shift toward lower-reputation askers accompanies much of the acceptance decline, though snapshot timing may overstate that mix effect.

1) Who Is Served:

  • Acceptance rises from 19.8% for askers below 10 reputation to 66.8% at 10k and above, a 47.0-point spread—2.7 times the spread across topics.The advantage is chiefly in whether a question receives an answer, not in how quickly answered questions get one.Source: 1) Who Is Served · p. 6

2) The Gradient Inside Each Year:

  • The reputation gradient persists within every year: the top-minus-bottom acceptance spread ranges from 17.5 to 57.9 points and exceeds 30 points in nine years.The spread is 37.7 points in 2016 and 56.8 in 2025, inconsistent with accumulation time alone explaining it.Source: 2) The Gradient Inside Each Year · p. 6

3) The Asker Mix:

  • Below-100-reputation askers’ share rose from 41.0% before 2023 to 65.8% after; standardizing to the earlier reputation mix accounts for 5.6 of the 9.5-point acceptance fall.Acceptance also fell within all five reputation buckets, by 3.9 points on average; the export-time snapshot may overstate the mix effect because recent askers have had less time to accumulate reputation.Source: 3) The Asker Mix · p. 6

4) Who Answers:

The accepted-answer pool lost seniority over the decade, while matched accepted questions show answerers typically had substantially higher reputation than askers. The paper then finds that asker-controlled question properties have comparatively small adjusted associations with acceptance.

4) Who Answers:

  • The share of accepted answers from users with at least 10k reputation fell from 48.5% in 2016 to 9.0% in 2024, then recovered to 23.6% in 2025.Median accepted-answerer reputation fell from 9,693 to 1,641; the 2025 recovery is less certain because it rests on 212 accepted answers versus 2,039 in 2016.Source: 4) Who Answers · pp. 6–7

5) The Expertise Gap on Matched Populations:

  • Within 20,337 accepted questions, answerers had a median 13.3× the asker’s reputation and higher reputation in 75.3% of cases.Comparing separate populations instead yields 30.2×, but accepted-question askers have a higher median reputation than askers overall.Source: 5) The Expertise Gap on Matched… · p. 7

1) Raw and Adjusted Associations:

  • After adjustment for reputation bucket, primary topic, and era, code snippets accompany 5.0 points higher acceptance; the other reported properties range from −0.3 to 2.6 points.The seven properties include post length, tags, snippets, and text markers for versions, frameworks, and errors.Source: 1) Raw and Adjusted Associations · p. 7

2) Direction Across the Reputation Scale:

Most asker-controlled properties do not point uniformly across reputation groups, limiting their value as general advice. The discussion connects the record to different support needs across Ethereum work areas and to declining newcomer support.

2) Direction Across the Reputation Scale:

  • Only tag count points in the same direction across all five reputation buckets; its association is about 2 points, while the largest adjusted association is 5.0 points.Code snippets, version markers, framework names, and error traces reverse direction among the highest-reputation askers; the reported differences are associations, not effects of changing a question.Source: 2) Direction Across the Reputation Scale · p. 8

A. What the Record Means for the Ethereum Community

  • Toolchain and integration accounts for 40.9% of topic pairs, gained 6.0 points after 2022, and is slowest to answer at 4.35 hours; Wallets had the largest main-topic acceptance loss.Wallet acceptance fell from 29.2% to 15.7%; the discussion points to environment-aware diagnostics and wallet and front-end SDK documentation as targets.Source: A. What the Record Means for the… · p. 8
  • Newcomers now post most questions and are served worst, while the senior answerer pool has thinned; the paper frames deliberate support for the venue’s live layer as needed.One newcomer question in five is never answered, compared with one in ten from established askers.Source: A. What the Record Means for the… · p. 8

B. What Changed When Assistants Arrived

The post-2022 archive is a filtered record of needs assistants did not meet, and the paper cannot attribute concurrent changes to assistants alone. Its proposed responses emphasize capturing context, routing questions to experts, and validating labels and time trends in Q&A archives.

B. What Changed When Assistants Arrived

  • The post-2022 volume fall coincides with a cryptomarket downturn, and without an external activity index the corpus cannot separate those explanations; assistants may explain lower volume and longer waits, but not the earlier decline in closure.The senior-answerer share had already fallen from 53.5% in 2020 to 10.9% in 2022.Source: B. What Changed When Assistants Arrived · pp. 8–9
  • After 2022, the archive is a filtered sample of needs assistants did not meet, so falling question volume need not indicate falling developer need and an ESE-orphaned question may be resolved elsewhere.Across 22 tasks, the later topic mix alone raises the answered share by 0.4 points and lowers acceptance by 0.08, while closure kept falling for reasons the archive cannot see.Source: B. What Changed When Assistants Arrived · p. 9

C. What to Build for Newcomers

  • For newcomers, the paper proposes capturing versions, network, and error traces at the source, routing for reach rather than speed, and expert verification of archive-based assistant answers.These proposals are motivated by the findings, not tested interventions; the paper notes that only controlled deployment can test whether context-capturing templates change outcomes.Source: C. What to Build for Newcomers · p. 9

D. Implications for Mining Software Repositories

  • For Q&A archive studies, the paper recommends staged resolution measures, standardizing shifts in composition, category-level label validation, and marking the post-assistant filtered period.Here, topic mix explains under 1% of the acceptance fall and asker mix up to 59%; overall label accuracy is 87.9 to 93.5%, though one broad category reaches only 50%.Source: D. Implications for Mining Software… · p. 9

V. THREATS TO VALIDITY

The study’s conclusions are bounded by label-audit and measurement choices, ESE’s limited reach, and the meanings of acceptance and reputation. Several checks narrow the reach of its estimates, while leaving some interpretation unresolved.

A. Internal Validity

  • The label audit measures precision, not missed topics, and assumes 570 committee-validated assignments are correct; Architecture and Upgradeability passed only 50% on 30 assignments.The study did not re-run topic standardization under perturbed labels or another era boundary; the mix term is 0.08 of the 9.40-point change.Source: A. Internal Validity · p. 9
  • Topic-level rates count each of the 22.2% multi-topic questions once per topic, while primary-topic analyses use an unfixed label order.Topic standardization cannot distinguish changes in what askers accept from changes in difficulty within topics; reputation standardization uses an export-time snapshot that inflates its mix term.Source: A. Internal Validity · p. 9
  • Without sampling intervals, the smallest text-property contrasts are read as no association, and text markers indicate mentions rather than report completeness.The markers record whether a version, framework, or error string is mentioned, not whether the report is complete.Source: A. Internal Validity · p. 9

B. External Validity

  • ESE users are English-speaking public posters, and the archive omits problems developers solve alone; late acceptances can raise the 2025 rate by at most 0.70 points.The export carries no user identifiers, so the study cannot distinguish answers that no longer merit acceptance from askers who left or solved problems elsewhere.Source: B. External Validity · p. 9, A. Internal Validity · p. 9

C. Construct Validity

  • Acceptance signals community approval, not correctness, while reputation records past activity rather than expertise and may partly follow accepted outcomes.Security labels cover only questions framed as security-related, and one judge shares TopicGPT’s model family, although it was not the most lenient judge.Source: C. Construct Validity · p. 10

A. Stated Needs: Surveys and Interviews

Surveys and interviews capture stated needs and can ask developers why; the ESE archive complements them with unprompted questions and resolution outcomes. Across the related work and conclusion, the paper presents a decade-scale view of demand and support, while noting that ESE reflects only needs posted there.

A. Stated Needs: Surveys and Interviews

  • Surveys and interviews identify stated pain points and can ask why; the ESE archive adds unprompted demand, resolution outcomes, and how outcomes vary with who asks.The two views overlap on toolchain pain points, which appear in the largest and slowest-answered cluster.Source: A. Stated Needs: Surveys and Interviews · p. 10

B. Mining Ethereum Stack Exchange

  • Prior ESE studies examine defect taxonomies,, upgradeable-contract posts, or answer-code vulnerabilities, while this study supplies a decade-scale question-side baseline.The corpus also surfaces categories such as Consensus Mechanisms and Networking that sampled studies did not report.Source: B. Mining Ethereum Stack Exchange · p. 10

C. Mining Developer Q&A Platforms

  • Work on developer Q&A platforms studies security questions, scoring, question quality, and activity; this study measures ESE outcomes and asker properties without classifying why questions go unanswered.The paper also notes that ESE’s closure decline predates assistants, but does not separate the volume collapse from the market cycle.Source: C. Mining Developer Q&A Platforms · p. 10

VII. CONCLUSION

  • Toolchain and integration questions form the largest and slowest-answered cluster, while support has weakened unevenly and the decline is not explained by task mix.The conclusion places documentation and tooling effort with the toolchain and newcomer support with context capture and routing.Source: §VII · p. 10
Loading 2610.30829v1…