LLM Physics Plotting Benchmark

How reliably and cheaply can LLMs turn ATLAS analysis requests into validated Python plots?

Benchmark setup
ServiceX + Awkward Array + Hist · Combined current + historical contexts

The default combines current and historical contexts for one workflow. Select an individual context below when you need to inspect it alone; different workflows remain separate.

3138trials
20models
18questions
overall pass rate
Pass
Generated code completed and satisfied the configured validation checks.
Pass rate
Successful trials divided by trials.
No data
That model/question combination has no trials recorded.
Filter results
Benchmark setupView profile configuration
Models
Questions

Cost versus success

Each point shows the average estimated API cost per trial, including successful and failed trials. Only trials with cost data contribute. The horizontal axis is pass rate. Cost is not cost per successful trial.

Closed circle: complete question coverage Open circle: partial question coverage

Model summary

Model performance for the selected setup, models, and questions

3138 trial(s) across 20 model(s) and 18 question(s).

Pass fraction by question

Each cell shows successful trials as a percentage and the number of trials. The question glossary below provides the full plotting request for every Q-number.

Question glossary
Plotting requests used in the benchmark
QuestionRequestExpected plotsFurther details
Q1Plot the ETmiss of all events in the rucio dataset user.zmarshal:user.zmarshal.364702_OpenData_v1_p6026_2024-04-23.1Repository definition
Q2Plot the pT of all jets in the rucio dataset user.zmarshal:user.zmarshal.364702_OpenData_v1_p6026_2024-04-23.1Repository definition
Q3Plot the pT of jets with |η| < 1 in the rucio dataset user.zmarshal:user.zmarshal.364702_OpenData_v1_p6026_2024-04-23.1Repository definition
Q4Plot the ETmiss of events that have at least two jets with pT > 40 GeV in the rucio dataset user.zmarshal:user.zmarshal.364702_OpenData_v1_p6026_2024-04-23.1Repository definition
Q5Plot the ETmiss of events that have an opposite-charge muon pair with an invariant mass between 60 and 120 GeV in the rucio dataset opendata:mc20_13TeV.700325.Sh_2211_Zmumu_maxHTpTV2_CVetoBVeto.deriv.DAOD_PHYSLITE.e8351_s3681_r13167_p6026.1Repository definition
Q6For events with at least three jets, plot the pT of the trijet four-momentum that has the invariant mass closest to 172.5 GeV in each event and plot the maximum b-tagging discriminant value among the jets in this trijet in the rucio dataset opendata:mc20_13TeV.410471.PhPy8EG_A14_ttbar_hdamp258p75_allhad.deriv.DAOD_PHYSLITE.e6337_s3681_r13167_p6026.2Repository definition
Q7Plot the scalar sum in each event of the pT of jets with pT > 30 GeV that are not within 0.4 in ΔR of any light lepton with pT > 10 GeV in the rucio dataset opendata:mc20_13TeV.410471.PhPy8EG_A14_ttbar_hdamp258p75_allhad.deriv.DAOD_PHYSLITE.e6337_s3681_r13167_p6026.1Repository definition
Q8For events with at least three light leptons and a same-flavor opposite-charge light lepton pair, find such a pair that has the invariant mass closest to 91.2 GeV in each event and plot the transverse mass of the system consisting of the missing transverse momentum and the highest-pT light lepton not in this pair in the rucio dataset user.zmarshal:user.zmarshal.700588_OpenData_v0_p6026_2024-04-16.1Repository definition
Q9Plot the number of jets per event in the rucio dataset user.zmarshal:user.zmarshal.364702_OpenData_v1_p6026_2024-04-23.1Repository definition
Q10Plot the electron transverse momentum for electrons with p_T > 25 GeV in the first 20000 events of the rucio dataset opendata:mc20_13TeV.410471.PhPy8EG_A14_ttbar_hdamp258p75_allhad.deriv.DAOD_PHYSLITE.e6337_s3681_r13167_p6026.1Repository definition
Q11Plot the leading jet p_T (highest p_T jet per event) in the rucio dataset user.zmarshal:user.zmarshal.364702_OpenData_v1_p6026_2024-04-23.1Repository definition
Q12Plot the invariant mass of all muon pairs with the same charge in the opendata:mc20_13TeV.700325.Sh_2211_Zmumu_maxHTpTV2_CVetoBVeto.deriv.DAOD_PHYSLITE.e8351_s3681_r13167_p6026.1Repository definition
Q13Plot the missing transverse energy for events with no jets in the opendata:mc20_13TeV.700325.Sh_2211_Zmumu_maxHTpTV2_CVetoBVeto.deriv.DAOD_PHYSLITE.e8351_s3681_r13167_p6026.1Repository definition
Q14Plot delta-R between the two leading jets for events with at least two jets in the rucio dataset user.zmarshal:user.zmarshal.364702_OpenData_v1_p6026_2024-04-23.1Repository definition
Q15Make stacked histogram of the pT of the highest pT jet in each event in user.zmarshal:user.zmarshal.364702_OpenData_v1_p6026_2024-04-23 and user.zmarshal:user.zmarshal.364703_OpenData_v1_p6026_2024-04-23 between 100 and 200 GeV.1Repository definition
Q16Plot the reconstructed top quark pair mass near 3 TeV in events with one charged lepton in the rucio dataset user.zmarshal:user.zmarshal.301333_OpenData_v1_p6026_2024-04-23. Baseline object cuts: jets pT > 30 GeV, electrons pT > 25 GeV, muons pT > 25 GeV. Require exactly one charged lepton and at least four jets. Do not add extra shaping cuts.1Repository definition
Q17Plot the b-tagging discriminant value for all jets the dataset opendata:mc20_13TeV.410471.PhPy8EG_A14_ttbar_hdamp258p75_allhad.deriv.DAOD_PHYSLITE.e6337_s3681_r13167_p6026.1Repository definition
Q18Plot the b-tagging discriminant value for all jets with pT>45 and eta<2.4 for events that have at least 2 of those jets in the dataset opendata:mc20_13TeV.410471.PhPy8EG_A14_ttbar_hdamp258p75_allhad.deriv.DAOD_PHYSLITE.e6337_s3681_r13167_p6026.1Repository definition

Additional attempts

An additional attempt is a code-generation or execution attempt after a previous attempt failed. Each question/model cell shows two numbers: trials needing a new attempt, followed by the total number of additional attempts. Legacy records derive the total from their recorded attempt count.

Darker cells indicate more additional attempts; this scale is independent of pass rate.

Diagnostics

Failure types

Final failures for the selected results
Failure typeCountShare of final failures

Failure type definitions

Trial explorer

Trial details
QuestionModelTrialOutcomeFailureValidationDetails