NewWrite together in real time Free · Open source · AGPL-3.0

Write research paperswith less busywork.

Oleafly is a free desktop workspace for writing papers in LaTeX, Typst, and Markdown. Find sources, manage citations, and see your paper take shape in one place. Bring an AI assistant to revise a section, create a figure, or fix a broken build, with changes you can review. Your project stays in ordinary files you own, and you can use the editor with or without AI.

Download for free Read the docs

6We stimulated MCF10A cells expressing the ERK kinase translocation

7reporter~\cite{regot2014ktr} with saturating doses of EGF, HGF or

8IGF-1 and tracked nuclear and cytoplasmic signal every two minutes in 2,184 cellsDaniel Hart

9for four hours. All three ligands produced the same first spike

10within ten minutes (Figure~\ref{fig:overview}b). After that the responses

11diverged: EGF drove pulses about every 19 minutes, HGF about every 41Sofia Marín,

12and IGF-1 rarely more than once an hour (Table~\ref{tab:pulses}).

+2 Recompile 3 co-authors in results.tex

OleaflyTeXFix-Bench: LaTeX Repair Bench…ERK Pulse Frequency Encodes Growth Factor Identity

+2 Compile

Comments1

results.tex3

MCF10A cells

Resolved by Prajwal
1. Elena Petrova2 hours ago
  Should we give the passage number? Reviewer 2 asked about it last time.
2. Daniel Hart1 hour ago
  Passages 12 to 18. I'll put it in Methods.
Write a reply. Type @ to mention someone.

Reply
about every 41

Resolved by Prajwal
1. Sofia Marínnow
  41 here, but 1.5 pulses an hour in Table 1 is every 40 minutes. Which one?
2. Prajwalnow
Write a reply. Type @ to mention someone.

Reply
Pulses/h

Resolved by Prajwal
1. Nina Takeda20 minutes ago
  @Lucas Moreau median or mean?
Write a reply. Type @ to mention someone.

Reply
oncogenic signalling

Resolved by Prajwal
1. Amara Okafor5 minutes ago
  British spelling is fine for this journal.
Write a reply. Type @ to mention someone.

Reply

methods.tex2

introduction.tex1

main.tex results.tex

CodeVisual

results.texResultsEach ligand pulses at its own frequency

1\section{Results}

2\label{sec:results}

4\subsection{Each ligand pulses at its own frequency}

6We stimulated MCF10A cells expressing the ERK kinase translocation

7reporter~\cite{regot2014ktr} with saturating doses of EGF, HGF or

8IGF-1 and tracked nuclear and cytoplasmic signal every two

9minutesDaniel Hart for four hours. All three ligands produced

10the same first spike within ten minutes

11(Figure~\ref{fig:overview}b). After that the responses diverged:

12EGF drove pulses about every 19 minutes, HGF about every 41Sofia Marín, and

13IGF-1 rarely more than once an hour (Table~\ref{tab:pulses}).

15\begin{center}

16 \begin{minipage}{\columnwidth}

17 \centering\small

18 \captionof{table}{Pulse statistics between 15 and 120 minutes

19 after stimulation. Medians across cells.}

20 \label{tab:pulses}

21 \begin{tabular}{lrrrr}

22 \toprule

23 Ligand & Dose & Pulses/h & Width & Cells \\

24 \midrule

25 \textcolor{egf}{EGF} & 20 & 3.1 & 7.8 & 812 \\

26 \textcolor{hgf}{HGF} & 50 & 1.5 & 8.4 & 694 \\

27 \textcolor{igf}{IGF-1} & 100 & 0.8 & 9.6Nina Takeda & 521 \\

28 Vehicle & -- & 0.2 & 11.2 & 157 \\

29 \bottomrule

30 \end{tabular}

31 \end{minipage}

32\end{center}

34\subsection{Frequency, not amplitude, identifies the ligand}

36Pulse amplitude overlapped almost completely between ligands,

37so we trained a decoder on inter-pulse intervalsAI Assistant alone.

38Given the first two hours of a single cell's trace, it named the

39ligand correctly in 91\,percent of held-out cells, compared with

4047\,percent for a decoder that saw only amplitudes. Mutant RAS

41rescales this code rather than erasing it~\cite{gillies2020ras},

42which suggests the frequency channel survives oncogenic signalling.

44\subsection{Sustained pulsing predicts S-phase entry}

Logs 6.4s 80% 1 of 2

ERK pulse frequency encodes growth factor identity
in mammary epithelial cells

Amara Okafor, Daniel Hart, Sofia Marín, Lucas Moreau, Nina Takeda, Elena Petrova

Department of Systems Biology, Northbridge Institute

Figure 1: Growth factors reach ERK through one cascade but leave different temporal signatures. (a) Receptor tyrosine kinases converge on the RAS–RAF–MEK–ERK module. (b) Single-cell ERK activity from the kinase translocation reporter after stimulation at t = 0.

Abstract

Cells read growth factors through a small number of shared kinases, yet they respond to EGF, HGF and IGF-1 with different fates. We imaged ERK activity in 2,184 MCF10A cells with a kinase translocation reporter and found that each ligand produces pulses at a characteristic frequency rather than a characteristic amplitude. Frequency alone identifies the ligand in 91 percent of cells.

1Introduction

Epidermal growth factor, hepatocyte growth factor and insulin-like growth factor all reach the nucleus through the same RAS–RAF–MEK–ERK module [1], yet epithelial cells answer them with migration, scattering or growth. How a single kinase carries that distinction has been a long-standing question. Early work in PC-12 cells tied fate to whether ERK activity was transient or sustained [2, 3].

Live-cell reporters changed the picture. In single cells ERK does not rise and settle; it fires in discrete pulses, and the pulse frequency tracks the growth factor dose [4, 5]. Here we ask whether frequency also encodes which growth factor a cell has seen.

We find that it does. Each ligand sets its own pulse rate, the rate survives changes in dose, and cells that keep pulsing are the ones that go on to divide. The same code is rescaled, not erased, in cells carrying an activating RAS mutation. Blocking the pulses with a MEK inhibitor at 60 minutes stopped division under all three ligands. Section 2 sets out these results, Section 3 discusses what they mean for signalling by timing, and Section 4 gives the imaging and decoding methods.

2Results

2.1Each ligand pulses at its own frequency

We stimulated MCF10A cells expressing the ERK kinase translocation reporter [6] with saturating doses of EGF, HGF or IGF-1 and tracked nuclear and cytoplasmic signal every two minutes in 2,184 cells for four hours. All three ligands produced the same first spike within ten minutes (Figure 1b). After that the responses diverged: EGF drove pulses about every 19 minutes, HGF about every 41, and IGF-1 rarely more than once an hour (Table 1).

Table 1: Pulse statistics between 15 and 120 minutes after stimulation. Medians across cells.

Ligand	Dose	Pulses/h	Width	Cells
EGF	20	3.1	7.8	812
HGF	50	1.5	8.4	694
IGF-1	100	0.8	9.6	521
Vehicle	–	0.2	11.2	157

2.2Frequency, not amplitude, identifies the ligand

Pulse amplitude overlapped almost completely between ligands, so we trained a decoder on inter-pulse intervalsthe gaps between pulses alone. Given the first two hours of a single cell's trace, it named the ligand correctly in 91 percent of held-out cells, compared with 47 percent for a decoder that saw only amplitudes. Shuffling the intervals within each cell removed the signal, and so did pooling cells across doses.

A decoder trained on amplitudes alone did no better than chance for HGF against IGF-1, while interval-based decoding separated all three ligands after the first hour of imaging.

2.3Sustained pulsing predicts S-phase entry

Cells that kept pulsing past the second hour entered S phase within 18 hours in 74 percent of cases, against 21 percent for cells whose pulses stopped early (Figure 2). Pulse count over the first four hours predicted entry better than mean ERK activity over the same window.

3Discussion

Frequency coding lets one kinase carry several messages without crosstalk at the level of amplitude.

Mutant RAS rescales this code rather than erasing it [7], which suggests the frequency channel survives oncogenic signalling and could be read by downstream targets that integrate over time.

4Methods

MCF10A cells were imaged every two minutes for four hours on a widefield microscope with environmental control. Nuclear and cytoplasmic reporter signal was segmented automatically and checked by hand for a random subset of cells.

Lucas Moreau Editor Editing sections/methods.tex

Daniel Hart Editor Editing sections/results.tex

Nina TakedaEditor · sections/results.texSofia MarínEditor · sections/results.tex

main.tex

CodeVisual

1% TeXFix-Bench technical report, manuscript version 0.2.

2% Compile with: tectonic -X compile main.tex

4\documentclass[journal,10pt]{IEEEtran}

5\usepackage{amsmath,amssymb,booktabs,array,graphicx}

6\usepackage{xcolor,microtype,url,fontspec,newtxsf}

7\usepackage{tikz,pgfplots,balance,hyperref}

8\pgfplotsset{compat=1.18}

9\usetikzlibrary{arrows.meta,positioning,fit,backgrounds}

11\definecolor{accentblue}{HTML}{006D9C}

12\definecolor{accentorange}{HTML}{D55E00}

13\definecolor{accentteal}{HTML}{0F766E}

14\definecolor{slate}{HTML}{64748B}

16\newcommand{\CR}{\textsc{cr}}

17\newcommand{\NEM}{\textsc{nem}}

18\newcommand{\MER}{\textsc{mer}}

20\title{TeXFix-Bench: Engine-Grounded Evaluation of\\

21Single-Fault \LaTeX{} Compilation Repair}

22\author{\textbf{Prajwal~S.~Venkateshmurthy}\\[0.10em]

23\small Independent Researcher\\[0.12em]

24\small San Jose, CA, USA}

26\begin{document}

27\maketitle

29\begin{abstract}

30Compilation repair for \LaTeX{} is easy to score badly. A candidate may

31compile after deleting or hiding content, while exact source equality

32rejects valid alternative repairs. We present TeXFix-Bench, a controlled

33benchmark for repairing complete, broken \LaTeX{} sources under a pinned engine~\cite{tan2024tex}.

34The generator applies 12 deterministic single-fault operators to 13

35hand-authored seed documents. It retains 627 engine-validated rows,

36comprising 300 unique hard-track inputs and 67 unique warning-only

37inputs after hash deduplication.

38\end{abstract}

40\begin{IEEEkeywords}

41automated program repair, \LaTeX{}, benchmark, compilation, large

42language models, reproducibility

43\end{IEEEkeywords}

45\section{Introduction}\label{sec:intro}

46\LaTeX{} compilation failures are a practical target for automated repair.

47A missing delimiter or misspelled control sequence can prevent a complete

48document from building, and the first reported diagnostic may be distant

49from the underlying fault. LLM-assisted writing systems now operate inside

50\LaTeX{} editors~\cite{wen2024overleafcopilot,hou2026paperdebugger}, but their

51publications do not provide a public, engine-pinned evaluation of

52compilation-failure repair.

54Program repair benchmarks pair a defect with an executable oracle and a

55fixed harness~\cite{just2014defects4j,legoues2015manybug}. SWE-bench extends

56the contract to repository-scale issues~\cite{jimenez2024swebench}.

57TeXFix-Bench studies a narrower input-output contract: a system receives

58an existing broken source document and must return a repaired full source

59under a specified engine.

61\section{Benchmark Construction}\label{sec:construction}

62\begin{figure*}[t]

63\centering

64\begin{tikzpicture}[font=\footnotesize\sffamily,node distance=0.58cm,

65 phase/.style={rounded corners=2.5pt,line width=0.7pt,

66 minimum width=2.55cm,minimum height=0.82cm,align=center}]

67\node[source] (seed) {13 known-good\\seed documents};

68\node[transform,right=of seed] (mutate) {12 deterministic\\mutation operators};

69\node[engine,right=of mutate] (gate) {Tectonic 0.17.0\\quality gate};

70\node[artifact,right=of gate] (rows) {627 retained\\rows};

71\draw[flow] (seed) -- (mutate);

72\draw[flow] (mutate) -- (gate);

73\draw[flow] (gate) -- (rows);

74\end{tikzpicture}

75\caption{TeXFix-Bench data and evaluation flow. Construction is

76deterministic. Evaluation reuses saved candidates and reruns only local

77engine and PDF-text checks.}

78\label{fig:pipeline}

79\end{figure*}

81\section{Results}\label{sec:results}

82Table~\ref{tab:main} reports the aggregate outcome.

83Figure~\ref{fig:paired} shows \CR{} and \NEM{} for each system, and

84Figure~\ref{fig:contrasts} shows the published paired contrasts.

86\begin{table*}[t]

87\caption{Version 0.2 results on 86 distinct hard-track inputs. \CR{} is

88intention-to-treat. \MER{} and \NEM{} use the same denominatior.}

89\label{tab:main}

90\input{results/tables/main-results-v02.tex}

91\end{table*}

93Claude Sonnet 5, GPT-5.6 Sol, and Grok 4.5 each recover 85 of 86 inputs.

94Their normalized exact-match counts are 52, 66, and 59, respectively.

96\begin{thebibliography}{19}

97\bibitem{just2014defects4j}

98R. Just, D. Jalali, and M. D. Ernst, \emph{Defects4J: A database of existing

99faults to enable controlled testing studies for Java programs}, ISSTA, 2014.

100\bibitem{legoues2015manybugs}

101C. Le Goues, N. Holtschulte, E. K. Smith, Y. Brun, P. Devanbu, S. Forrest,

102and W. Weimer, \emph{The ManyBugs and IntroClass benchmarks}, TSE, 2015.

103\bibitem{jimenez2024swebench}

104C. E. Jimenez, J. Yang, A. Wettig, S. Yao, K. Pei, O. Press, and

105K. Narasimhan, \emph{SWE-bench}, ICLR, 2024.

106\bibitem{chen2021codex}

107M. Chen et al., \emph{Evaluating large language models trained on code},

108arXiv:2107.03374, 2021.

109\bibitem{smith2015overfitting}

110E. K. Smith, E. T. Barr, C. Le Goues, and Y. Brun, \emph{Is the cure worse

111than the disease? Overfitting in automated program repair}, FSE, 2015.

112\bibitem{demillo1978}

113R. A. DeMillo, R. J. Lipton, and F. G. Sayward, \emph{Hints on test data

114selection}, Computer, 1978.

115\bibitem{jia2011mutation}

116Y. Jia and M. Harman, \emph{An analysis and survey of the development of

117mutation testing}, TSE, 2011.

118\bibitem{gupta2017deepfix}

119R. Gupta, S. Pal, A. Kanade, and S. Shevade, \emph{DeepFix}, AAAI, 2017.

120\bibitem{yasunaga2020drrepair}

121M. Yasunaga and P. Liang, \emph{Graph-based, self-supervised program repair

122from diagnostic feedback}, ICML, 2020.

123\bibitem{zhu2022eqfix}

124F. Zhu, Z. Liu, G. Yang, J. Li, and S. Cheung, \emph{EqFix}, ASE, 2022.

125\bibitem{wang2024texocr}

126K. Wang et al., \emph{TexOCR-Bench: Benchmarking page-image-to-LaTeX

127reconstruction}, 2024.

128\bibitem{wen2024overleafcopilot}

129H. Wen, Z. Zhang, and Y. Li, \emph{OverleafCopilot: Empowering academic

130writing in Overleaf with large language models}, 2024.

131\bibitem{hou2026paperdebugger}

132S. Hou et al., \emph{PaperDebugger: An in-editor assistant for manuscript

133repair}, 2026.

134\bibitem{tan2024tex}

135J. Tan, K. Zhang, and X. Chen, \emph{Instrumenting TeX produced diagnostics

136for repair}, 2024.

137\bibitem{zheng2023judging}

138L. Zheng et al., \emph{Judging LLM-as-a-judge with MT-Bench and Chatbot

139Arena}, NeurIPS, 2023.

140\bibitem{panickssery2024selfpref}

141A. Panickssery, S. R. Bowman, and S. Feng, \emph{LLM evaluators recognize

142and favor their own generations}, NeurIPS, 2024.

143\bibitem{hui2024qwen}

144B. Hui et al., \emph{Qwen2.5-Coder technical report}, 2024.

145\bibitem{metaai2024llama}

146Meta AI, \emph{The Llama 3 herd of models}, 2024.

147\bibitem{kassis2024reproduce}

148E. Kassis et al., \emph{On the reproducibility of research artifacts}, 2024.

149\end{thebibliography}

150

151\end{document}

Logs 1.9s 100% 1 of 2

Compiling 1/3 done

1 running · 1 queued

Language analysisCurrent project revision analyzed
CompilingProducing and verifying PDF output
Rendering PDFWaiting for verified compile output

TeXFix-Bench: Engine-Grounded Evaluation of
Single-Fault LATEX Compilation Repair

Prajwal S. Venkateshmurthy
Independent Researcher
San Jose, CA, USA

Abstract—Compilation repair for LaTeX is easy to score badly. A candidate may compile after deleting or hiding content, while exact source equality rejects valid alternative repairs. We present TeXFix-Bench, a controlled benchmark for repairing complete, broken LaTeX sources under a pinned engine. The generator applies 12 deterministic single-fault operators to 13 hand-authored seed documents. It retains 627 engine-validated rows, comprising 300 unique hard-track inputs and 67 unique warning-only inputs after hash deduplication.

Index Terms—automated program repair, LaTeX, benchmark, compilation, large language models, reproducibility

I. INTRODUCTION

LaTeX compilation failures are a practical target for automated repair. A missing delimiter or misspelled control sequence can prevent a complete document from building, and the first reported diagnostic may be distant from the underlying fault. LLM-assisted writing systems now operate inside LaTeX editors , but their publications do not provide a public, engine-pinned evaluation of compilation-failure repair.

Program repair benchmarks pair a defect with an executable oracle and a fixed harness . TeXFix-Bench studies a narrower contract: a broken source document in, a repaired full source out, under a specified engine.

II. RELATED WORK

TexOCR-Bench evaluates page-image-to-LaTeX reconstruction , and EqFix repairs errors in individual equations by synthesizing rules from examples . Overfitting to a weak oracle is a known hazard .

III. TASK AND SCOPE

A system receives one broken source file and returns one repaired source file. The engine, its version, and the build command are fixed for every attempt.

CR = compile(c) ∧ LCS_tok(T_c, T_r) ≥ 0.80(1)

IV. BENCHMARK CONSTRUCTION

Fig. 1. TeXFix-Bench data and evaluation flow. Construction is deterministic.

Every generated row passes an engine gate before it is retained, and duplicate broken sources are removed by SHA-256 so that no input is scored twice .

V. EVALUATION METHOD

Compile Recovery requires a successful Tectonic 0.17.0 run and at least 0.80 token-level LCS similarity between text extracted from the candidate and reference PDFs. Normalized exact match sits beside it, because compilation alone cannot separate a faithful repair from a deletion .

TABLE I
Version 0.2 results on 86 distinct hard-track inputs. CR is intention-to-treat.

System	CR	CR %	NEM	Fail
Identity	0/86	0.0	0/86	0
Rule	34/86	39.5	28/86	0
Qwen-7B + fb	57/86	66.3	17/86	0
MiniMax M3	75/86	87.2	50/86	2
DeepSeek V4	81/86	94.2	38/86	2
Gemini 3.5	82/86	95.3	26/86	0
Claude S5	85/86	98.8	52/86	0
GPT-5.6 Sol	85/86	98.8	66/86	1
Grok 4.5	85/86	98.8	59/86	0

Fig. 2. Compile Recovery and normalized exact match on the fixed cohort of n = 86 hard-track inputs. Circles denote CR, squares NEM.

VI. RESULTS

Claude Sonnet 5, GPT-5.6 Sol, and Grok 4.5 each recover 85 of 86 inputs, yet their normalized exact-match counts are 52, 66, and 59 . Compilation and edit precision must be reported together.

VII. CONCLUSION

The contribution is an auditable task definition, an executable oracle, a fixed cohort, and a replayable candidate ledger.

REFERENCES

[1]R. Just, D. Jalali, and M. D. Ernst, “Defects4J: A database of existing faults to enable controlled testing studies for Java programs,” in Proc. ISSTA, 2014, pp. 437–440.

[2]C. Le Goues, N. Holtschulte, E. K. Smith, Y. Brun, P. Devanbu, S. Forrest, and W. Weimer, “The ManyBugs and IntroClass benchmarks for automated repair of C programs,” IEEE Trans. Softw. Eng., vol. 41, no. 12, pp. 1236–1256, 2015.

[3]C. E. Jimenez, J. Yang, A. Wettig, S. Yao, K. Pei, O. Press, and K. Narasimhan, “SWE-bench: Can language models resolve real-world GitHub issues?” in Proc. ICLR, 2024.

[4]M. Chen et al., “Evaluating large language models trained on code,” arXiv:2107.03374, 2021.

[5]E. K. Smith, E. T. Barr, C. Le Goues, and Y. Brun, “Is the cure worse than the disease? Overfitting in automated program repair,” in Proc. ESEC/FSE, 2015, pp. 532–543.

[6]R. A. DeMillo, R. J. Lipton, and F. G. Sayward, “Hints on test data selection: Help for the practicing programmer,” Computer, vol. 11, no. 4, pp. 34–41, 1978.

[7]Y. Jia and M. Harman, “An analysis and survey of the development of mutation testing,” IEEE Trans. Softw. Eng., vol. 37, no. 5, pp. 649–678, 2011.

[8]R. Gupta, S. Pal, A. Kanade, and S. Shevade, “DeepFix: Fixing common C language errors by deep learning,” in Proc. AAAI, vol. 31, no. 1, pp. 1345–1351, 2017.

[9]M. Yasunaga and P. Liang, “Graph-based, self-supervised program repair from diagnostic feedback,” in Proc. ICML, PMLR, vol. 119, pp. 10799–10808, 2020.

[10]F. Zhu, Z. Liu, G. Yang, J. Li, and S. Cheung, “EqFix: Fixing LaTeX equation errors by examples,” in SETTA 2022, LNCS 13649, pp. 89–109.

[11]K. Wang et al., “TexOCR-Bench: Benchmarking page-image-to-LaTeX reconstruction,” in Proc. 64th ACL, Vol. 1, 2026.

[12]H. Wen, Z. Zhang, and Y. Li, “OverleafCopilot: Empowering academic writing in Overleaf with large language models,” arXiv:2403.09733, 2024.

[13]S. Hou et al., “PaperDebugger: Assisting scientific writing, review, and editing,” in Companion Proc. ACM Web Conf., 2026.

[14]J. Tan, K. Zhang, and X. Chen, “Inconsistencies in TeX-produced documents,” in Proc. ISSTA, 2024.

[15]L. Zheng et al., “Judging LLM-as-a-judge with MT-Bench and Chatbot Arena,” Adv. Neural Inf. Process. Syst., vol. 36, pp. 46595–46623, 2023.

[16]A. Panickssery, S. R. Bowman, and S. Feng, “LLM evaluators recognize and favor their own generations,” Adv. Neural Inf. Process. Syst., vol. 37, 2024.

[17]B. Hui et al., “Qwen2.5-Coder technical report,” arXiv:2409.12186, 2024.

[18]Meta AI, “Llama 3.2 model card,” 2024.

[19]E. Kassis et al., “Agent skills for reproducible research agents,” arXiv:2609.00065, 2026.

Zoom in

Zoom out

Fit to width

Fit to height

Reset to 100%

25%

50%

75%

100%

150%

200%

400%

Undefined reference

bibentry · legoues2015manybug

@article
ibitem{legoues2015manybugs}
main.tex:956

Oleafly AgentCLI Agent BETA

Connect an AI provider to continue

Bring your own API key (OpenAI, Anthropic, Groq, and more) or run a model locally with Ollama. The assistant can read and edit files, compile your project, and verify the PDF.

Connect a providerRun a local model with Ollama

What would you like to do today?

Working on "TeXFix-Bench: LaTeX Repair Bench..."

ResearchAuthoringFiguresReviewSubmissionCommunicationToolingYour skills

citation-management

Comprehensive citation management for academic research. Search OpenAlex, PubMed, and Google Scholar for papers, extract accurate metadata, validate citations, and generate properly formatted BibTeX entries. This skill should be used when you need to find papers, verify citation information, convert DOIs to BibTeX, or ensure reference accuracy in scientific writing.

exploratory-data-analysis

Perform bounded, local exploratory analysis of explicitly supported scientific files. Use for redacted CSV/TSV/JSON profiles; optional NumPy, HDF5, FASTA/FASTQ, and basic image metadata inspection; missingness/leakage audits; outlier and transformation sensitivity; and rigorous EDA report scaffolds. Other domain formats are reference-only and unknown formats fail closed.

hypothesis-generation

Formulate evidence-bounded scientific questions, candidate hypotheses, rival explanations, causal or associational claims, discriminating predictions, measurements, and preregistration-ready analysis plans. Use when turning observations or preliminary findings into transparent, testable research plans without treating hypotheses as facts.

oleafly-literature-sweep oleafly-research-loop paper-lookup research-lookup oleafly-literature-sweep

RECENT CHATS2

What would you like to do today?

Working on "TeXFix-Bench: LaTeX Repair Bench..."

Check the introduction for unsupported claims. Do not edit files.

Claude Code

Read main.tex

Read references.bib

The energy-use comparison needs a supporting source. I found the claim in the introduction, but its citation covers detection accuracy. Your files are unchanged.

Sweep the literature for this project

Working...

Thinking...

Thought for 10s

update todos

Explored 3 files

Read file

I'll inspect the current citations first, then search for papers that actually support the claims.

Working...

Thinking...

I need sources that match the paper's actual claims:
- satellite fire detection
- low-power sensor networks
- TinyML and quantized inference

I'll keep only results with verifiable DOI metadata.

Search literature

Search papers

Quick audit: three citation keys do not resolve, and two existing entries point to unrelated papers. I'm searching for better matches on wildfire sensing, TinyML, and low-power networks.

Working...

Search literature

Check citation

Four solid matches so far: MODIS fire detection, a field-tested sensor-network detector, an energy-harvesting survey, and a TinyML review. I'm checking the DOI metadata before I touch the bibliography.

Working...

Check citation

update todos

The citations check out. I'll add nine verified entries to references.bib, then place them beside the claims they support.

Working...

write fileApproved

The bibliography is in place. Next I'm replacing the unresolved keys in main.tex with the verified sources.

Working...

replace in fileApproved

Compile document

Check PDF pages

Done. I added nine verified references, replaced the three unresolved keys, and recompiled the paper. The two-page PDF now has a complete IEEE bibliography with no missing citations.

Compiled and checked the rendered PDF

STEP 1/6

Plan

Inventory citations used in main.tex and references.bib

Search literature for papers matching each cited claim

Verify every candidate citation is real

Add verified BibTeX entries to references.bib

Replace unresolved cites in main.tex

Recompile and confirm all citations resolve

The assistant wants to change your files

write_filereferences.bibline 1

references.bibfirst change · L1

1@article{open_science2015,

2 title = {Estimating the Reproducibility of Psychological Science},

1@article{giglio2003,

2 author = {Giglio, Louis and Descloitres, Jacques and Justice, Christopher O.},

3 title = {An Enhanced Contextual Fire Detection Algorithm for {MODIS}},

4 journal = {Remote Sensing of Environment},

5 year = {2003}, doi = {10.1016/S0034-4257(03)00184-6}

6}

Reject Approve

The assistant wants to change your files

replace_in_filemain.texline 21

main.texfirst change · L21

54Program repair benchmarks pair a defect with an executable oracle and a

55fixed harness~\cite{just2014defects4j,legoues2015manybug}. SWE-bench extends

55fixed harness~\cite{just2014defects4j,legoues2015manybugs}. SWE-bench extends

56the contract to repository-scale issues~\cite{jimenez2024swebench}.

57TeXFix-Bench studies a narrower contract: a system receives an existing

57TeXFix-Bench studies a narrower contract~\cite{zhu2022eqfix}: a system receives

Reject Approve

Claude Code

Start or reconnect a conversation to send a message Start a conversation with Claude Code

Ask AI to help with your document...

Ask for approval Prompts Persona gpt-6.1-sol

Logs 4.3s 105% 1of 2

TeXFix-Bench: Engine-Grounded Evaluation of
Single-Fault LATEX Compilation Repair

Prajwal S. Venkateshmurthy
Independent Researcher
San Jose, CA, USA

Index Terms—automated program repair, LaTeX, benchmark, compilation, large language models, reproducibility

I. INTRODUCTION

II. RELATED WORK

TexOCR-Bench evaluates page-image-to-LaTeX reconstruction , and EqFix repairs errors in individual equations by synthesizing rules from examples . Overfitting to a weak oracle is a known hazard .

III. TASK AND SCOPE

A system receives one broken source file and returns one repaired source file. The engine, its version, and the build command are fixed for every attempt.

CR = compile(c) ∧ LCS_tok(T_c, T_r) ≥ 0.80(1)

IV. BENCHMARK CONSTRUCTION

Fig. 1. TeXFix-Bench data and evaluation flow. Construction is deterministic.

Every generated row passes an engine gate before it is retained, and duplicate broken sources are removed by SHA-256 so that no input is scored twice .

V. EVALUATION METHOD

TABLE I
Version 0.2 results on 86 distinct hard-track inputs. CR is intention-to-treat.

System	CR	CR %	NEM	Fail
Identity	0/86	0.0	0/86	0
Rule	34/86	39.5	28/86	0
Qwen-7B + fb	57/86	66.3	17/86	0
MiniMax M3	75/86	87.2	50/86	2
DeepSeek V4	81/86	94.2	38/86	2
Gemini 3.5	82/86	95.3	26/86	0
Claude S5	85/86	98.8	52/86	0
GPT-5.6 Sol	85/86	98.8	66/86	1
Grok 4.5	85/86	98.8	59/86	0