LZ's 248 keV dark matter candidate event drew five theory papers overnight
Contents
The LZ collaboration announced on September 1, 2026, at the TeVPA 2026 conference in Tendo, Yamagata, that its dark matter search data contains one nuclear recoil event at 248 keV.
The paper is posted as a preprint on LZ’s own site, with an arXiv listing and a Physical Review Letters submission in progress.
And by the early hours of September 2 Japan time, five theory papers interpreting that single event as dark matter had appeared on arXiv.
The collaboration itself keeps repeating that it is not claiming a dark matter discovery.
What LZ saw
LZ (LUX-ZEPLIN) is a liquid xenon detector about 1.5 km underground at the Sanford Underground Research Facility in South Dakota. It looks for the recoil of a xenon nucleus struck by a WIMP (weakly interacting massive particle), the classic dark matter candidate.
A recoiling nucleus emits light and electrons, and the ratio of the prompt light (S1) to the light produced when the electrons are extracted into the gas phase (S2) separates nuclear recoils from electron recoils.
This analysis widened the nuclear recoil energy window from the 55 keV used in earlier WIMP searches to 270 keV.
In non-relativistic effective field theory (NREFT) some models put a non-negligible share of the signal at high recoil energies, and inelastic scattering suppresses the low-energy part, so the old window alone loses sensitivity to those signals.
Inside the widened window there was exactly one event that the known backgrounds do not explain well.
| Item | Value |
|---|---|
| Exposure | 2.84 tonne-years (4.71 tonne fiducial mass × 220 live days) |
| Data period | March 27, 2023 to April 1, 2024 |
| Event timestamp | June 16, 2023, 21:22:39 UTC |
| Recoil energy | 248 ± 23 (stat) ± 23 (sys) keV |
| Position | 26.4 cm above the cathode (the electrode at the bottom), 26.9 cm inward from the wall |
| Distance from the nuclear recoil band | 1.5σ below the median |
| Distance from the electron recoil band | 6.7σ below the median |
| Expected background near the event | 0.0106 ± 0.0008 events (S1c between 500 and 600 phd; phd is the unit for detected photons) |
| Local significance | up to 3.4σ, depending on the model |
| Global significance | 2.6σ |
Across the whole search window about 1710 events were observed against a model prediction of 1713 ± 39, almost all of them electron recoil background.
Only this one event sits inside the nuclear recoil band, and in the region above 500 phd in S1c the expected background was just 0.01 events.
How to read 0.5% and 2.6σ
The press release line “a 0.5% chance the event came from a known source” does not mean the detection is 99.5% accurate. It means that, assuming there is no dark matter, background alone produces an event like this 0.5% of the time.
Particle physics calls 3σ “evidence” and 5σ a “discovery”, and 5σ corresponds to a one-sided background probability of roughly 1 in 3.5 million.
Compared with the roughly 0.5% of 2.6σ, that is a factor of about 17,000, or about four orders of magnitude.
There is also the local versus global distinction.
LZ scans mass or mass splitting across 20 Lagrangians and 2 inelastic operators, which leaves plenty of room for some model to come out significant by chance.
The local significance for a single model reaches 3.4σ, but correcting for how many places were searched (the look-elsewhere effect) brings it down to 2.6σ.
The paper also classifies this analysis as non-blind.
It used salting, injecting fake signal-like events to keep analysts honest, but the salt distribution did not cover the high-energy region well, so bias mitigation rests on keeping the selection cuts unchanged from the previous analysis.
Nothing on the low-energy side
For the simplest elastic-scattering WIMP, the recoil energy distribution piles up at low energies.
If there is one event at 248 keV, there should be many more below a few tens of keV.
But LZ’s low-energy side matches the background prediction, and those events are not there.
Even if this event is dark matter, then, it does not look like the simplest WIMP.
That is where the LZ paper brings up inelastic scattering as an example.
The dark matter particle has two states with slightly different masses, and scattering off a nucleus moves it from the lighter to the heavier state.
The transition needs kinetic energy equal to the mass splitting δ, so slow particles cannot scatter and only fast ones react.
The result is that low-energy events are suppressed and the high-energy share grows.
The LZ paper says outright that this isoscalar operator for inelastic spin-independent scattering (a term that couples to protons and neutrons with equal strength) “resembles a Higgsino model”.
The higgsino is the supersymmetric partner of the Higgs boson.
The paper also notes that inelastic models can produce an annual modulation peaking around June 2, because the Earth’s orbit changes the relative velocity of dark matter.
The event was recorded on June 16, two weeks past that peak but still in June.
The background candidates LZ examined
The paper says of itself that “increased scrutiny of the analysis is warranted” given the 3σ-level local significance, and goes through the candidates one by one.
It also states that neutrons are generally considered the most likely cause of nuclear recoils, so they are in the table too.
| Candidate | Finding |
|---|---|
| Neutrons | Giving a xenon nucleus 250 keV takes a neutron of at least 8 MeV, and higher-energy neutrons scatter more forward and deposit less, so several lower-energy events would be expected too. For (α,n) neutrons from the detector’s own materials, the veto systems (which catch neutrons and gamma rays around the TPC to reject them) tag 92 ± 4%, leaving a bit under 10% untagged |
| Atmospheric neutrinos | The recoil spectrum from coherent elastic scattering (where the whole nucleus recoils as one) falls off exponentially, and 248 keV is outside the range where coherence holds |
| Accidental coincidences | Random pairings of an isolated S1 with an isolated S2. They favor low energies, and the validated model expects 2.7 events across the whole window. They can reach this high-energy region as the tail of the distribution, but the absence of a large low-energy population constrains that |
| MSSI (multiple-scintillation single-ionization) | An electron recoil that scatters more than once with some deposits in charge-dead regions, mimicking a nuclear recoil. Expectation 0.0049 ± 0.0049. Validated with a 5.4 tonne volume and a high-energy sideband, 12-bin p-value 0.7 |
| Calibration source | A 57Co source had been deployed on the opposite side of the detector until 25 minutes before the event. Its gamma rays have a mean free path under 4 mm in liquid xenon, and no anomalous population appeared in the data around it |
| Muons | The previous muon passed through the outer detector 41 minutes earlier, and the TPC (the main xenon volume) 127 minutes earlier |
| Double-vacancy decays of 124Xe and 125I | Read as an electron recoil they land at 64 to 67 keV, close to the event’s reconstructed energy. The paper notes a systematic uncertainty in projecting into the tail that is not included in the statistical inference |
The conclusion’s wording is that “several rare background processes and detector effects were studied in detail but none could be identified as a likely explanation”.
The radon tag (a method for identifying 214Pb decays) could not be applied because the detector was in its mixed-flow state at the time.
Five papers in one night
LZ’s announcement was on September 1, and all five submissions fall between 16:10 and 17:55 UTC the same day.
That is 1 to 3 a.m. on September 2 in Japan, and all five appeared together in the next day’s arXiv listing.
| arXiv | Authors | Interpretation |
|---|---|---|
| 2609.01475 | Su, Yang, Yang | Endothermic inelastic dark matter. Mass above 500 GeV, splitting of order 300 keV. Representative point from a two-bin likelihood at 1.105 TeV and 350 keV |
| 2609.01504 | Fan, Reece | A 1.1 TeV thermal higgsino with a splitting of about 350 keV. Gauginos (superpartners of gauge bosons) and heavy Higgs scalars must sit at 10^3 to 10^4 TeV |
| 2609.01583 | Freese, Theodosopoulos | A nearly pure higgsino near 1 TeV. The electroweak interaction fixes the scattering cross section, which approaches LZ’s 90% confidence interval at δ of about 350 keV |
| 2609.01590 | Wu, Zhang, Zhu | A 1.1 TeV higgsino explaining both the LZ event and a mild excess in Fermi-LAT Galactic center gamma rays. A gamma-ray line testable with H.E.S.S. and CTAO |
| 2609.01592 | Lou, Lu | Absorption of a 247 MeV fermionic dark matter particle by the nucleus, rather than scattering, giving a monoenergetic 248 keV recoil. Already excluded by KamLAND data |
Three of the five land on a higgsino at roughly 1 to 1.1 TeV.
Two put the mass splitting at about 350 keV, and Wu and colleagues say a few hundred keV.
The remaining inelastic dark matter paper (2609.01475) names no particle, and its representative point of 1.105 TeV and 350 keV is almost the same spot.
The reason the answers come out so close is in the abstracts.
For a higgsino left over from thermal equilibrium in the early universe to account for today’s dark matter density, its mass has to be about 1.1 TeV.
Separately, the splitting needed to produce a 248 keV recoil comes out at a few hundred keV for the standard dark matter velocity distribution, about 350 keV in Fan and Reece’s calculation.
The precise value apparently depends on the high-speed tail of that distribution, and Fan and Reece say a better understanding of that tail is still needed.
Because two conditions pin the mass and the splitting to narrow ranges, the three higgsino papers arrive at similar answers.
The higgsino has two electrically neutral states with almost the same mass, and the Z boson connects them, so scattering off a nucleus just ends up inelastic on its own.
Fan and Reece describe the model as a “classic, but still viable” WIMP scenario, and in the Brown University press release Fan names inelastic scattering and momentum-dependent elastic scattering as the candidates.
The three higgsino papers and the one generic inelastic dark matter paper fit the LZ event into the existing inelastic scattering framework.
The LZ paper itself had already used an inelastic operator resembling a higgsino model as its example.
Only Lou and Lu try a different mechanism, absorption of the dark matter particle by the nucleus, and they write themselves that the parameters it needs are already excluded by KamLAND’s neutron-emission channel.
What happens next
Wu, Zhang, and Zhu predict a Galactic center gamma-ray line and endpoint signal near the current H.E.S.S. sensitivity and within CTAO’s projected reach, and Su and coauthors say the planned CRESST upgrade can test their region.
Whether either holds up looks checkable with a different experiment.
As precedent, Science News points to the excess XENON1T reported in 2020, which its successor XENONnT later ruled out.
LZ’s Rick Gaitskell says “we have seen something interesting that we want to share with the scientific community for their input”. Science News quotes Dan Hooper of the University of Wisconsin–Madison: “It’s only one event. So who knows what’s really going on here,” he says. “That said, it’s intriguing.”
The Imperial College release lists extending LZ’s run beyond 2028, and XLZD, holding about ten times LZ’s liquid xenon, coming online in the mid-2030s as the next steps.
Before that, according to Science News, LZ already has at least three times as much additional data as it has analyzed, and LZ itself has kept taking data under the same electric field conditions since April 1, 2024.