Daily Summary for 2026-11-14

daily

Transcript

Introduction to the show: ident: High Energy Physics: Experiment. The day in high-energy experiment: detectors, searches and limits.

Kai: It's the fourteenth of November, twenty twenty-six, and this is the day's research.

Mira: 5 new papers came out today.

Kai: I'm Kai.

Mira: And I'm Mira, here with Lev.

Kai: We'll take the day in one pass, then pull out the papers we're staying with.

Mira: Start us off with DrSai: an agentic AI system toward autonomous physics analysis workflows at BESIII.

The day's research: Kai: We are automating complex workflows due to slow manual methods with petabyte scale data.

Mira: The agentic AI system DrSai translates natural language into physics workflows.

Lev: DrSai performed large-scale re-measurements of ten J/psi decay branching fractions.

Kai: It navigated the BESIII environment to produce benchmark-matching results without manual coding.

Mira: Improvements in particle track reconstruction are seen at high collision rates.

Lev: A fast Kalman filter significantly improved momentum resolution for D0 and J/psi hadrons.

Kai: This used analytical parameterizations instead of expensive lookups for efficiency gain.

Mira: The ATLAS collaboration searched for boosted diphoton resonances in the 10 to 70 GeV mass range.

Lev: That search set strong limits on axion-like particles in a specific mass range.

Kai: We are also constraining long-lived particle decays through data scouting at the LHC.

Mira: CMS performed a search for long-lived particles decaying into muons using data scouting.

Kai: DrSai an agentic AI system toward autonomous physics analysis workflows at BESIII High Energy Physics experiments can use AI agents to automatically handle complex data analysis tasks.

Mira: Track fitting at the full LHC collision rate The LHCb experiment uses a fast, simplified Kalman filter to track charged particles at the very high collision rates of the LHC.

Lev: Higgs Physics at a sqrt s =3 TeV Muon Collider with detailed detector simulation This paper explores how well current detector designs can measure Higgs boson production cross sections at a future 3 TeV muon collider.

Kai: Search for boosted diphoton resonances in the 10 to 70 GeV mass range using 138 fb-1 of 13 TeV pp collisions with the ATLAS detector Researchers searched for new particles decaying into two photons in a specific mass range at the LHC and found no evidence.

Mira: Search for long-lived particles decaying into muons in proton-proton collisions at sqrt s = 13.6 TeV using data scouting This study looked for particles that decay into muons, setting limits on their properties based on CMS data from the LHC.

Lucky paper: 2604.22541: Kai: So we're looking at DrSai an agentic AI system toward autonomous physics analysis workflows at BESIII, which is about using AI agents to automate complex data analysis tasks for high-energy physics experiments.

Mira: It’s built on a Multi-Agent System architecture with six specialized agents: Host, Planner, Coder, Tester, Calculator, and Reflector.

Lev: The system uses a hierarchical task management structure inspired by fractal-like selfsimilar structures to break down large objectives into subtasks.

Kai: And the Coder agent generates code blocks based on a high-level JSON configuration that specifies things like decay channels and selection criteria.

Mira: They also developed HepScript, which is a domain specific language designed to formalize human analytical logic and translate that intent into standardized executable sequences.

Lev: When it comes to performance validation, they tested the system with ten distinct J-psi decay channels to re-measure branching fractions.

Kai: And one of the key findings is that the primary bottleneck across all models was functional calling capability rather than information extraction.

Mira: DeepSeek R1 showed diminishing returns with iterative reflection because it had a fundamental weakness in tool call initiation, while GPT 4o did well at tool call initiation but struggled with synthesizing the schema compliant parameters needed for execution.

Lev: This suggests that even deep reasoning depth isn't enough if the underlying agent can't precisely initiate the right tools.

Kai: It seems like they are pointing towards a need for more specialized synthesis layers or fine tuned models that have a syntax level mastery of professional HEP tools.

Mira: The paper also mentions future work focusing on how to represent domain specific knowledge to turn static information into executable logic that mimics human expert decision making.

Lev: That's the big challenge, moving from a chat based assistant to something truly autonomous that understands the physics ontology deeply enough.

Kai: So DrSai can maintain physical consistency and handle the demands of real world experimental frameworks at BESIII, which is what they confirmed with their experimental validation.

Mira: It’s interesting how they validated this system across five mainstream LLMs to see how success rate heatmaps changed under varying retry budgets and token expenditure.

Lev: The systematic evaluation showed that a moderate reflection allowance significantly enhanced sub task reliability for most of those models.

Kai: What about the limitations they found? They clearly stated that the current bottlenecks are precision of execution and the depth of domain integration.

Mira: They also flagged a lack of a formal representation of expert heuristics to transform static information into executable logic that mirrors strategic decision making.

Lev: So, even with this agentic structure, getting it to act like an expert still requires figuring out how to embed those human decision-making rules correctly.

Lucky paper: 2607.14793: Kai: We're looking at Track fitting at the full LHC collision rate today. It shows a fast implementation of a Kalman filter being used in the first trigger stage since two thousand twenty-five data taking period.

Mira: That's the paper we started with, and it deals with how they replaced expensive magnetic field map lookups and numerical integration methods with analytical parameterizations.

Lev: The core of this work is a fast implementation of a Kalman filter used in the first trigger stage since two thousand twenty-five data taking period, which replaces the previous partial track fitting algorithm in the real time trigger environment at the cost of a two percent increase in processing time.

Kai: A two percent increase sounds like a small trade-off for what they are achieving here.

Mira: They show that this parameterised Kalman filter results in a factor of two improvement in invariant mass resolutions for reconstructed D0 and J psi hadrons compared to the VELO Only Kalman Filter.

Lev: They also demonstrate greater robustness against detector misalignment effects and substantially sharpen the discrimination between genuine particle trajectories and accidental background, more than doubling the rejection of the latter at no cost to genuine track efficiency for a standard selection working point.

Kai: So, they are getting better mass resolution and better fake track rejection simultaneously in that trigger stage.

Mira: The paper also details how this algorithm extends the track fitting to all three tracking detectors: VELO, UT and SciFi.

Lev: They state that the Parameterised Kalman Filter makes up roughly three point five percent of the total runtime of the default HLT1 sequence, while the VELO Only Kalman Filter made up one point four percent.

Kai: That shows a real efficiency gain when you look at how much processing time it consumes overall.

Mira: The authors compare this to a standard implementation that accounts for approximately fourteen percent of the total processing time in HLT2 reconstruction, which is where the previous method was used more heavily.

Lev: The primary benefit of fitting the full track comes from incorporating hits upstream and downstream of the LHCb magnet, which enables precise measurements of particle momenta.

Kai: So those extra hits give them better momentum measurements than just pattern recognition algorithms alone.

Mira: In invariant mass distributions, they find that J psi and D0 peak widths are almost halved compared to the VELO Only Kalman Filter.

Lev: They also show it demonstrates robustness against detector misalignment effects that degrade the performance of momentum estimates based solely on pattern recognition algorithms.

Kai: It seems like this technique is a significant step towards offline quality reconstruction being available in real time on GPUs, which is what they conclude.

Mira: The conclusion points to further developments including the implementation of outlier removal and the addition of state estimates for a future particle identification using reconstruction in the RICH detectors.

Lev: I'm interested in how this fits into your work on agentic AI systems, Kai; do you see parallels between translating natural language into physics workflows and this kind of parameterization?

Lucky paper: 2405.19314: Kai: We're moving into paper number three of four today, which is about Higgs Physics at a sqrt s = three TeV Muon Collider with detailed detector simulation. It covers how to measure Higgs boson production cross sections using a simulated environment.

Mira: It dives into the experimental setup for this MuC complex, describing the muon injector, acceleration ring, and collision ring.

Lev: The paper discusses ionization cooling as the proposed technique for cooling those muon beams because of their short lifetime.

Kai: Then it gets into how they generate signal and background processes in section three point one, which is important because that determines the kinematic properties we measure later.

Mira: They detail the detector used in section three point two and the simulation framework in section three point three, explaining the digitization procedure for events when particles hit sensors simultaneously.

Lev: For measuring Higgs boson production cross sections at a sqrt s = three TeV collider, they estimate the cross section as NS over epsilon times L, where N is signal yield and epsilon is resolution.

Kai: The statistical uncertainty on that cross section estimate is dominated by the uncertainty on the signal yield itself.

Mira: They also look at how to evaluate the uncertainty on the trilinear Higgs boson self-coupling using a simulated sample involving µ plus minus decaying into HHX, which is described in section six point one.

Lev: The total double-Higgs production cross section is sensitive to the value of the trilinear self-coupling, and they find that this uncertainty ranges between zero point eight one and one point four four at a sixty-eight percent confidence level.

Kai: That range for the trilinear coupling sounds quite wide, so what does that imply for how precisely we can measure it?

Mira: The paper also sets detector requirements based on this sqrt s = three TeV case to inform the design of a future sqrt s = ten TeV detector.

Lev: They state that the average transverse momentum of particles from Higgs decays is about one hundred GeV at both sqrt s = three TeV and sqrt s = ten TeV, which is chosen because it's the peak of the Higgs daughters pT distribution in two body decays.

Kai: So, even though they are simulating a lower energy collider now, what does this simulation tell us about the performance needed for that much higher energy machine?

Mira: They mention that running a MuC at sqrt s = ten TeV for five years is expected to achieve an uncertainty of approximately four percent with ten ab minus one, which they say is the best precision attainable by any future collider.

Lev: They also mention that the estimated precision on the Higgs trilinear self-coupling is currently affected by the performance of simplified algorithms used in event reconstruction.

Kai: That's a big caveat, meaning if their reconstruction isn't perfect, their coupling measurement uncertainty goes up too.

Mira: The object requirements they set for muons, photons, jets, and b-jets are taken from the performance of the reconstruction described in section three and evaluated for objects with pT equals one hundred GeV.

Lev: For b-jets specifically, they quote identification efficiencies and the c mistag rate as part of those requirements.

Kai: So it’s not just about building a better detector; it’s about knowing exactly how well the algorithms can separate those Higgs decays from everything else at that energy scale.

Mira: The paper also includes figures showing projections of the expected precision for the ATLAS and CMS combination on the Higgs production cross sections and kappa i coupling modifiers including HL LHC data.

Lev: That figure shows a lot about how these different experiments combine their results to get a better picture of things at that energy.

Lucky paper: 2606.25840: Kai: We're looking at the search for long-lived particles decaying into muons in proton-proton collisions at sqrt s equals thirteen point six TeV using data scouting.

Mira: So the core idea is looking for a narrow, low-mass resonance where the decay vertex is within seventy centimeters of the primary vertex.

Lev: It uses a dimuon invariant mass spectrum across different transverse displacement ranges to find localized excesses.

Kai: That means they are filtering data based on how far away from the main collision point the decay happens.

Mira: They're looking for signals in specific models like the hidden Abelian Higgs model, dark-shower models, and a long-lived scalar resonance from a b hadron decay.

Lev: The search methodology involves selecting events with at least one pair of OS muons associated with an SV that satisfies several quality criteria.

Kai: Those quality criteria are very specific: the x coordinate uncertainty is less than zero point zero five centimeters, and the y coordinate uncertainty is also less than zero point zero five centimeters.

Mira: So they're filtering based on spatial resolution at the primary vertex to isolate these potential signals.

Lev: The constraints set in this paper extend previous limits on the mean proper decay length of an A prime particle in Scenario A, specifically in the range from ten centimeters to one hundred centimeters.

Kai: That means they are using data from two years of CMS data, which is six hundred twenty point four femtobarns integrated luminosity.

Mira: The paper states that no significant peak-like structures were observed in the data.

Lev: They did set upper limits at ninety five percent confidence level on the branching fraction B(H to Z D Z D) across a wide range of mass and lifetime hypotheses in several beyond the standard model frameworks.

Kai: The most stringent limits they found are for dark-photon proper decay lengths larger than one centimeter and dark-photon masses less than five gigaelectron volts in the hidden Abelian Higgs model.

Mira: That is quite a tight constraint, especially for the dark-shower models where they set limits on proper decay lengths larger than ten centimeters.

Lev: The paper also mentions that systematic uncertainties in trigger efficiency are assessed to be between one and fifteen percent depending on the transverse momentum and displacement of the muons.

Kai: And there's another systematic uncertainty considered, ranging from one to thirty percent, to account for potential mismodeling of the signal simulation affecting event selection acceptance.

Mira: So even with those uncertainties, they still found constraints that are most stringent in a large fraction of the explored parameter space.

More episodes

← Home