Pith. sign in

REVIEW 4 major objections 7 minor

How Developers Experience Debugging Unfamiliar Codebases with Code Tours Generated and Evaluated by Local LLMs

T0 review · 4 major / 7 minor · reviewed 2026-07-30 · grok-4.5

Pith's one-line read Open-weight LLM code tours help debug unfamiliar codebases only when detail, structure, tone, and trust are right—and LLM judges of those tours are unreliable.

desk verdict Solid empirical DX study on open-weight LLM code tours with real 2025 bugs; the themes are useful and the judge-failure findings land, but every frequency rests on reconstructed notes rather than audio. read the letter →

arxiv 2607.26987 v2 pith:K7PBZEV2 submitted 2026-07-29 cs.SE

classification cs.SE
keywords codetoursonboardingopen-weightLLMsdeveloperexperiencetrustcalibrationLLM-as-a-judgedebuggingsummarization
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper asks how the properties of code-tour components authored and judged by local open-weight LLMs shape what developers experience when they try to fix real bugs in unfamiliar Java codebases. The authors built a pipeline that turns 2025 GitHub bug-fix commits into stack-trace-driven tours, has one model write each tour and two others score it, then sat 26 developers down to think aloud while rating the tours and the machine judgments. What emerges is a concrete preference map: detail should scale with segment length and not merely restate the code; text should be scannable and guiding rather than lazy or theatrical; stack traces alone miss steps people want (constructors, concrete implementations); and people trust prose they read as human-written while discounting prose they read as AI-written. The same study shows that the LLM judges themselves are poor critics—sycophantic, confabulatory, and incoherent across criteria—so automated quality labels cannot yet be trusted. The result is a practical agenda for fine-tuning, personalization, better step selection, trust calibration, and more honest machine evaluators.

What carries the argument

A generation-and-judgment pipeline that authors 26 code tours from real 2025 Java stack traces with three local open-weight models, has the two non-author models produce 52 Likert-plus-rationale annotations on Transparency, Scrutability, and Efficiency, then elicits think-aloud experiences from 26 developers that three coders reduce to 62 first-person experience labels clustered around understanding, acting, navigating, liking, and trusting.

What would settle it

Re-run the same tours and annotations with audio-recorded, independently transcribed sessions (or a second independent interviewer) and check whether the same 62 experience labels, especially the trust and preference clusters, still reach high inter-coder agreement and the same frequency rankings.

Watch

Extended reading notes

Core claim

Properties of open-weight LLM-authored code-tour components systematically shape developer experience when debugging unfamiliar Java codebases: developers generally prefer detail scaled to segment length, non-restating scannable structure, and a guiding tone; some preferences (such as imperative mood) conflict; stack traces are often insufficient for step selection; perceived human versus AI authorship shifts trust toward misuse or disuse; and LLM judges of tour quality exhibit pervasive sycophancy, confabulation, and incoherence.

Load-bearing premise

That interviewer notes taken on the fly, later translated from French and rewritten as first-person excerpts, faithfully capture what participants experienced rather than what the interviewer reconstructed.

Editorial extensions

If this is right

  • Fine-tuning open-weight models on the shared preferences (scaled detail, scannable structure, guiding tone, no pure restatement) should raise baseline tour usefulness.
  • Personalization is required for mutually exclusive tastes such as imperative mood and deliberately incomplete descriptions.
  • Step selection must reach beyond the raw stack trace to constructors, concrete implementations, and other missing frames while staying under roughly five steps.
  • Trust calibration needs explicit disclosure of AI authorship and gradual exposure so users neither over-trust human-sounding text nor under-trust capable AI text.
  • LLM-as-judge pipelines for code tours cannot be used as quality gates until sycophancy, confabulation, and cross-criterion incoherence are reduced.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The same preference map (scannable structure, task-scoped suggestions, calibrated detail) likely transfers to other forms of in-IDE onboarding documentation, not only debugging tours.
  • If stack-trace-only selection systematically omits constructors and implementations, hybrid static-analysis plus LLM step pickers become a natural next experiment.
  • The observed misuse/disuse split suggests that any production onboarding tool should surface provenance and AI-literacy cues rather than hide them.
  • Because judges praised duplicate lambda frames that humans rejected, automated evaluators may need an explicit structural checklist before free-form rationale generation.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 7 minor

Summary. This paper studies how properties of code-tour components authored and judged by open-weight LLMs shape developer experience when debugging unfamiliar Java codebases. The authors mine non-flaky 2025 GitHub bugs via Gitbug-Actions, generate 26 stack-trace-based tours with three local SE models (Qwen2.5-Coder 14B, DeepSeek-Coder-V2, Devstral Small 1.1), obtain 52 author≠judge annotations on Transparency/Scrutability/Efficiency, and run a BIBD user study with 26 developers who think aloud while rating tours and annotations. Three coders produce 62 first-person experience labels (α≥0.8) clustered around understanding, acting, navigating, liking, and trusting. Main findings: developers prefer detail scaled to segment length, non-restating scannable structure, and a guiding tone (with some mutually exclusive preferences); stack traces miss steps developers want; perceived human vs AI authorship shifts trust toward misuse or disuse; and LLM judges show sycophancy, confabulation, and incoherence.

Significance. The work fills a clear gap: prior code-tour work has not jointly examined fully AI-generated tours with open-weight models, developer experience, trust calibration, a debugging focus, and LLM-as-a-judge in one pipeline (Table 1). Strengths include use of post-cutoff 2025 reproducible bugs (mitigating leakage), stratified sampling, BIBD assignment, author≠judge separation, multi-coder coding with reported Krippendorff α, an explicit threats section (§5.2), and a public replication package. The trust-calibration and LLM-judge unreliability results are timely for SE tooling. If the qualitative substrate is accepted as adequate, the preference and step-selection findings give concrete, actionable directions for fine-tuning, personalization, and tour design.

major comments (4)
  1. [§3.7, §5.2.1, Table 4, Figures 6 and 8] §3.7 and §5.2.1: All 62 experience labels (Table 4) and the edge weights in Figures 6 and 8 rest on interviewer on-the-fly notes with no audio, then French→English translation and reformulation into first-person Lejeune-style experiences. Krippendorff α≥0.8 after iterative discussion measures coder agreement on these reconstructions, not fidelity to participants’ utterances. Concurrent think-aloud plus interviewer presence is also flagged (P19). This is load-bearing for the central claim that tour-component properties systematically shape goals. The manuscript should either (a) strengthen the chain of evidence (e.g., second-pass member checks, dual note-takers on a subset, or limited audio where consent allows) or (b) systematically bound every frequency claim as “reported in interviewer reconstructions,” and avoid language that treats n’s as direct counts of participant utterances witho
  2. [§4.5.2, abstract, Conclusion] §4.5.2 and abstract: The claim that sycophancy, confabulation, and incoherence are “pervasive” in LLM judges is supported by participant reports (e.g., E23 n=13, E27 n=9, E26 n=3) and vivid examples, but there is no systematic audit of the 52 annotations (error rates per criterion/model, inter-judge agreement, rate of self-contradiction, factual checks against tour content). A compact quantitative table over the full annotation set would make the judge-unreliability result falsifiable and proportionate to how strongly it is stated in the abstract and conclusion.
  3. [§3.4] §3.4: Generation and judging use temperature=1.0 and top-p=1.0 “to explore freely.” For a study that treats specific tour texts and judge rationales as the objects of human evaluation, this maximizes sample variance and weakens reproducibility of the exact artifacts participants saw. The paper should justify this choice against lower-temperature or multi-sample settings, report whether tours/annotations were regenerated or fixed once, and discuss sensitivity: would the same preference and judge-failure themes be expected under more deterministic decoding?
  4. [Abstract, §4 synthesis, §5.2.3, Conclusion] §3.6 and §5.2.3: The study is scoped to Java, mostly junior developers (13 junior / 9 medior / 4 senior), a web interface rather than an IDE with full codebase access, and three mid-size 2025 open-weight models. That scope is acceptable for an exploratory qualitative study, but several result statements (e.g., “developers generally preferred…”) read as broader than the sample. Tighten claim scope in the abstract, §4 synthesis box, and conclusion to “in this Java debugging onboarding setting with predominantly junior participants,” and separate design implications that are sample-bound from those offered as general SE guidance.
minor comments (7)
  1. [Figure 2] Figure 2 / Graphical Abstract: several labels render with missing characters (e.g., “bug-xes”, “Ofine”), which hurts readability of an otherwise helpful pipeline figure.
  2. [Table 4, E17] Table 4 E17 name text says “I can locate the fault when it is not highlighted” while valence is Negative; this appears inconsistent with E16/E15 and with the prose in §4.2.1 (nine could not identify the fault). Please correct the label wording.
  3. [§2.1, Table 1] §2 and related work: Kara et al. is cited as 2026 in places and Lacy interchangeably; normalize naming and year for the reader.
  4. [Figure 2] Criteria spelling: “Scrutatibility” appears in Figure 2 caption while “Scrutability” is used elsewhere; standardize.
  5. [§3.2, §3.3] §3.2–3.3: The path from 110 bugs → 243 traces → 26 stratified samples is clear, but a one-row summary table (projects, error types, tour lengths) would help readers assess diversity without mining Table 5.
  6. [§3.5.2] Rokeman [34] is cited for the 7-point scale; the more common reference is typically Rokeman/Likert scale methodology literature—double-check the bibliographic key and spelling (Rokeman vs common variants).
  7. [Throughout] Minor prose/typos: “mediors” (Table 3), “prematurely” misspelled as “prematuraly” (§4.1.5), “strack trace” in Figure 3 caption, “forgetAllTotalMethodsPath” spacing in Figure 9.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: inductive qualitative themes from a user study, not definitional or fitted predictions.

full rationale

This is an empirical software-engineering user study. The load-bearing chain is: mine 2025 reproducible Java bugs → generate code tours with open-weight LLMs from stack traces → have other LLMs annotate tours on Transparency/Scrutability/Efficiency → 26 developers think aloud while exploring tours and rating annotations → three coders produce 62 first-person experience labels (α≥0.8) clustered around understand/act/navigate/like/trust. Those labels and frequencies (Table 4, Figs. 6/8) are inductive reports of participant reactions, not quantities derived from parameters fitted to the same outcomes. The three evaluation criteria are adapted from Balfroid et al. [4] and used as discussion prompts and LLM-judge axes; they do not define or force the qualitative preferences (detail scaling, scannability, guiding tone, authorship-trust effects, judge sycophancy). Self-positioning in Table 1 and Background against the authors’ prior code-tour papers is normal related-work framing, not a uniqueness theorem or ansatz that closes the result. There is no fitted-input-called-prediction, no self-definitional X⇔Y, and no renaming of a known law as a first-principles derivation. Methodological threats (reconstructed notes, no audio) affect construct fidelity, not circularity of the claim chain. Score 0 is the proportionate finding.

Assumptions & free parameters 3 free parameters · 6 assumptions · 2 invented entities

Load-bearing commitments are methodological domain assumptions of qualitative HCI/SE research and engineering choices in the tour pipeline, not free physical parameters. No invented physical entities. The central experiential claims rest on validity of reconstructed think-aloud coding, representativeness of stack-trace-selected Java bugs and junior-leaning recruits, and the three evaluation criteria borrowed from prior work.

free parameters (3)
  • Temperature=1.0 and top-p=1.0 for generation and judging = temperature 1.0, top-p 1.0
    Chosen to 'explore freely' (§3.4); different decoding would change tour and annotation text and thus participant reactions.
  • BIBD block size r=6 annotations per developer = r=6, k=3, λ=1, v=13 per round
    Set from pilot timing and attention limits (§3.6); shapes exposure breadth and fatigue.
  • Stratified sample of 26 stack traces from 243 = 26 tours / 52 annotations
    One trace per project-error type per round to counter jhy-jsoup imbalance; different sample could shift themes.
assumptions (6)
  • domain assumption Think-aloud concurrent verbalization plus interviewer note-taking yields usable evidence of developer experience after translation and first-person reformulation.
    Core of §3.7 and Results; authors acknowledge reconstruction and presence effects in §5.2.1.
  • domain assumption Stack-trace frames are a reasonable primary skeleton for debugging-focused code-tour steps.
    Inherited from Balfroid et al. and §2.5; the paper partly falsifies sufficiency (§4.3.4) but still builds the pipeline on it.
  • domain assumption Transparency, Scrutability, and Efficiency (adapted from Balfroid et al. 2024) are adequate discussion anchors for tour quality without subcriteria.
    §3.5.2; chosen to reduce anchoring while stimulating talk.
  • standard math Krippendorff's α≥0.8 after discussion indicates stable shared meaning of experience labels.
    Standard qualitative reliability practice applied in §3.7–§4.
  • domain assumption Open-weight models with pre-2025 cutoffs on 2025 bugs sufficiently mitigate training-data leakage for this study.
    §3.2–3.4 motivation for Gitbug-Actions 2025 mining and model choice.
  • standard math Balanced incomplete block design equalizes annotation load and pairwise overlap enough for fair qualitative coverage.
    §3.6 BIBD relations used for assignment.
invented entities (2)
  • 62 first-person 'Experience' labels (E1–E62) decomposed into goal/valence/component/property
    purpose: Unit of analysis for clustering developer reactions to tour components
    Coding scheme invented for this study following Lejeune-style experience formulation; not an external validated instrument.
  • Author–judge multi-LLM code-tour pipeline with three fixed open-weight SE models independent evidence
    purpose: Generate and annotate tours without proprietary APIs
    Engineering assembly of existing models/tools; findings about judge failure modes are about this class of setup.

how reviews work

0 comments
Cite this review

Pith. "Pith review of How Developers Experience Debugging Unfamiliar Codebases with Code Tours Generated and Evaluated by Local LLMs." pith.science (2026). https://pith.science/paper/K7PBZEV2

@misc{pith2026260726987,
  author       = {Pith},
  title        = {Pith review of: How Developers Experience Debugging Unfamiliar Codebases with Code Tours Generated and Evaluated by Local LLMs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/K7PBZEV2}},
  note         = {Machine review of arXiv:2607.26987}
}
read the original abstract

Code tours are interactive, onboarding documentation to guide developers through a codebase. Large Language Models (LLMs) can automatically synthesize code tours. Prior work on code tour generation has not studied developer experience or trust calibration when debugging unfamiliar codebases with code tours generated and evaluated by open-weight LLMs. This study surveys how the properties of components in open-weight LLM-authored code tours influence developers' experiences when debugging unfamiliar codebases. We built a pipeline that generated and evaluated code tours from real reproducible bugs. 26 developers with varying backgrounds participated in a user study. In total, 26 code tours were authored from real Java bugs mined from 2025 GitHub commits, with each tour independently judged by two different LLMs, resulting in 52 evaluated configurations. Participants thought aloud as they explored each tour. Three authors qualitatively coded the interviews to identify recurring themes. Developers generally preferred tours that scaled detail with the code length, avoided merely restating code, were easily scannable, and adopted a guiding tone. However, some preferences were mutually exclusive, such as the use of imperative mood. Stack traces were often insufficient to identify all steps developers found relevant. Developers also trusted descriptions they perceived as human-written more than those they believed were AI-generated. Finally, LLM-generated annotations of tour quality were unreliable: sycophancy, confabulation, and incoherence were pervasive. This work lays a basis for future research on fine-tuning open-weight models for code tour generation, personalizing generation to accommodate diverging preferences, selecting relevant steps beyond stack traces, calibrating users' trust to avoid both disuse and misuse, and improving open-weight LLMs' ability to be more trustworthy evaluators

Figures

Figures reproduced from arXiv: 2607.26987 by the authors.

Figure 1
Figure 1. The tutorial code tour. It is designed to help a developer understand a NullPointerException in Apache Commons Lang, a popular Java library for string manipulation, including formatting. It is an interactive document that guides the developer from the failing test case through the rendering logic to the uninitialized variable responsible for the crash. 2. Background Onboarding is the process through which new employ… view at source ↗
Figure 2
Figure 2. Pipeline: (1) collecting reproducible bugs using Gitbug-Actions pipeline [35], (2) executing failing tests to produce candidate stack traces, (3) generating code tour files directly from the stack trace using one of the LLMs as an author, (4) generating annotations — a chain-of-thoughts rationale (text) and a rating (Likert-7) — of code tours for each criterion (Transparency, Efficiency, and Scrutatibility) by each … view at source ↗
Figure 4
Figure 4. Prompt components for the code tour generation task: (a) the prompt and (a) the schema defining the structure in Pydantic. the synthetic judge as an experienced software engineer. This serves as a cultural anchor, implicitly conveying ex￾pectations for tone and technical depth that would otherwise require lengthy, explicit instructions [33], thereby steering the distribution of the next likely tokens. The task secti… view at source ↗
Figures from the paper (6 more)
Figure 5
Figure 5. Figure 5: Prompt components for the code tour evaluation task: (a) the template, (b) the quality criteria, and (c) the rating scale. where 𝑏 represents the total number of annotations. In this context, 𝑏 is the total number of rating-rationale pairs to be evaluated, 𝑣 is the tot…
Figure 6
Figure 6. Figure 6: This diagram illustrates how the properties of the Description component (white rectangles) impact the functional goals (blue ovals). The connecting edges represent participant experiences, marked by label IDs (Ex). Edge color indicates whether the experience was posit…
Figure 7
Figure 7. Figure 7: Code Tour 18 (traccar-traccar, commit de97c50, trace 1/1) generated by Qwen2.5-Coder 14B. Project. Traccar is an open-source GPS tracking server that supports many GPS device protocols. Commit. This commit adds support for a new device variant (a magnetic card reader c…
Figure 8
Figure 8. Figure 8: This diagram illustrates how the properties of the Description component (white rectangles) impact the Liking goals (blue ovals). The connecting edges represent participant experiences, marked by label IDs (Ex). Edge color indicates whether the experience was positive …
Figure 9
Figure 9. Figure 9: Code Tour 9 (joular-joularjx, commit 9940cfc, trace 116/150) generated by Devstral Small 1.1. Project. JoularX is a plugin that monitors the power consumption of Java code. Commit. Refactoring of the class generating the path for the results, which are now UnixPath ins…
Figure 10
Figure 10. Figure 10: Code Tour 32 (mwilliamson-java-mammoth, commit b65b...) generated by Qwen2.5-Coder 14B. The tour covers the parsing order of XML styles (supposedly for a Word document, since Mammoth is a Word-to-HTML converter). The test fails because of an AssertionFailedError, expe…

Discussion (0). Continue with ORCID to comment.

Pith tools

Reviewed July 30, 2026 · model on record in the stance chip above.