Particle-level transformers for 95 GeV Higgs boson searches at future <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"> <mml:msup> <mml:mi>e</mml:mi> <mml:mo>+</mml:mo> </mml:msup> <mml:msup> <mml:mi>e</mml:mi> <mml:mo>−</mml

Motivated by several mild excesses around 95 GeV, we investigate the prospects for a light scalar <a:math xmlns:a="http://www.w3.org/1998/Math/MathML" display="inline"> <a:mi>S</a:mi> </a:math> produced via Higgsstrahlung, <c:math xmlns:c="http://www.w3.org/1998/Math/MathML" display="inline"> <c:msup> <c:mi>e</c:mi> <c:mo>+</c:mo> </c:msup> <c:msup> <c:mi>e</c:mi> <c:mo>−</c:mo> </c:msup> <c:mo stretchy="false">→</c:mo> <c:mi>Z</c:mi> <c:mo stretchy="false">(</c:mo> <c:msup> <c:mi>μ</c:mi> <c:mo>+</c:mo> </c:msup> <c:msup> <c:mi>μ</c:mi> <c:mo>−</c:mo> </c:msup> <c:mo stretchy="false">)</c:mo> <c:mi>S</c:mi> </c:math> , at future <h:math xmlns:h="http://www.w3.org/1998/Math/MathML" display="inline"> <h:msup> <h:mi>e</h:mi> <h:mo>+</h:mo> </h:msup> <h:msup> <h:mi>e</h:mi> <h:mo>−</h:mo> </h:msup> </h:math> Higgs factories. We take the CEPC as a benchmark, with a center-of-mass energy of <j:math xmlns:j="http://www.w3.org/1998/Math/MathML" display="inline"> <j:msqrt> <j:mi>s</j:mi> </j:msqrt> <j:mo>=</j:mo> <j:mn>240</j:mn> <j:mtext> </j:mtext> <j:mtext> </j:mtext> <j:mi>GeV</j:mi> </j:math> and an integrated luminosity of <l:math xmlns:l="http://www.w3.org/1998/Math/MathML" display="inline"> <l:mi>L</l:mi> <l:mo>=</l:mo> <l:mn>20</l:mn> <l:mtext> </l:mtext> <l:mtext> </l:mtext> <l:msup> <l:mi>ab</l:mi> <l:mrow> <l:mo>−</l:mo> <l:mn>1</l:mn> </l:mrow> </l:msup> </l:math> . We focus on the decay modes <n:math xmlns:n="http://www.w3.org/1998/Math/MathML" display="inline"> <n:mi>S</n:mi> <n:mo stretchy="false">→</n:mo> <n:msup> <n:mi>τ</n:mi> <n:mo>+</n:mo> </n:msup> <n:msup> <n:mi>τ</n:mi> <n:mo>−</n:mo> </n:msup> </n:math> and <q:math xmlns:q="http://www.w3.org/1998/Math/MathML" display="inline"> <q:mi>S</q:mi> <q:mo stretchy="false">→</q:mo> <q:mi>b</q:mi> <q:mover accent="true"> <q:mi>b</q:mi> <q:mo stretchy="false">¯</q:mo> </q:mover> </q:math> . To maximize sensitivity, we employ the particle-level transformer networks particle transformer (ParT) and its more-interactive variant MIParT, which exploit the features of all reconstructed objects and their correlations. For a representative signal benchmark, this approach improves the expected statistical precision on the signal strength by factors of 2.3 in the <v:math xmlns:v="http://www.w3.org/1998/Math/MathML" display="inline"> <v:msup> <v:mi>τ</v:mi> <v:mo>+</v:mo> </v:msup> <v:msup> <v:mi>τ</v:mi> <v:mo>−</v:mo> </v:msup> </v:math> channel and 1.4 in the <x:math xmlns:x="http://www.w3.org/1998/Math/MathML" display="inline"> <x:mi>b</x:mi> <x:mover accent="true"> <x:mi>b</x:mi> <x:mo stretchy="false">¯</x:mo> </x:mover> </x:math> channel compared to a cut-based analysis. Within the flipped next-to-two-Higgs-doublet model (N2HDM-F), the CEPC can measure the signal strength with a statistical precision down to 1.0% in the <bb:math xmlns:bb="http://www.w3.org/1998/Math/MathML" display="inline"> <bb:msup> <bb:mi>τ</bb:mi> <bb:mo>+</bb:mo> </bb:msup> <bb:msup> <bb:mi>τ</bb:mi> <bb:mo>−</bb:mo> </bb:msup> </bb:math> channel and 0.69% in the <db:math xmlns:db="http://www.w3.org/1998/Math/MathML" display="inline"> <db:mi>b</db:mi> <db:mover accent="true"> <db:mi>b</db:mi> <db:mo stretchy="false">¯</db:mo> </db:mover> </db:math> channel using MIParT. It can achieve a <hb:math xmlns:hb="http://www.w3.org/1998/Math/MathML" display="inline"> <hb:mn>5</hb:mn> <hb:mi>σ</hb:mi> </hb:math> discovery for <jb:math xmlns:jb="http://www.w3.org/1998/Math/MathML" display="inline"> <jb:msubsup> <jb:mi>μ</jb:mi> <jb:mrow> <jb:mi>τ</jb:mi> <jb:mi>τ</jb:mi> </jb:mrow> <jb:mrow> <jb:mi>Z</jb:mi> <jb:mi>S</jb:mi> </jb:mrow> </jb:msubsup> <jb:mo>&gt;</jb:mo> <jb:mn>1.6</jb:mn> <jb:mo>×</jb:mo> <jb:msup> <jb:mn>10</jb:mn> <jb:mrow> <jb:mo>−</jb:mo> <jb:mn>2</jb:mn> </jb:mrow> </jb:msup> </jb:math> or <lb:math xmlns:lb="http://www.w3.org/1998/Math/MathML" display="inline"> <lb:msubsup> <lb:mi>μ</lb:mi> <lb:mrow> <lb:mi>b</lb:mi> <lb:mi>b</lb:mi> </lb:mrow> <lb:mrow> <lb:mi>Z</lb:mi> <lb:mi>S</lb:mi> </lb:mrow> </lb:msubsup> <lb:mo>&gt;</lb:mo> <lb:mn>5.0</lb:mn> <lb:mo>×</lb:mo> <lb:msup> <lb:mn>10</lb:mn> <lb:mrow> <lb:mo>−</lb:mo> <lb:mn>3</lb:mn> </lb:mrow> </lb:msup> </lb:math> , and reach 1% precision for <nb:math xmlns:nb="http://www.w3.org/1998/Math/MathML" display="inline"> <nb:msubsup> <nb:mi>μ</nb:mi> <nb:mrow> <nb:mi>τ</nb:mi> <nb:mi>τ</nb:mi> </nb:mrow> <nb:mrow> <nb:mi>Z</nb:mi> <nb:mi>S</nb:mi> </nb:mrow> </nb:msubsup> <nb:mo>&gt;</nb:mo> <nb:mn>0.93</nb:mn> </nb:math> or <pb:math xmlns:pb="http://www.w3.org/1998/Math/MathML" display="inline"> <pb:msubsup> <pb:mi>μ</pb:mi> <pb:mrow> <pb:mi>b</pb:mi> <pb:mi>b</pb:mi> </pb:mrow> <pb:mrow> <pb:mi>Z</pb:mi> <pb:mi>S</pb:mi> </pb:mrow> </pb:msubsup> <pb:mo>&gt;</pb:mo> <pb:mn>0.14</pb:mn> </pb:math> . These gains are expected to qualitatively carry over to other future lepton colliders such as FCC-ee and the ILC. Our results demonstrate the potential of particle-level machine-learning techniques to strengthen light Higgs searches at future <rb:math xmlns:rb="http://www.w3.org/1998/Math/MathML" display="inline"> <rb:msup> <rb:mi>e</rb:mi> <rb:mo>+</rb:m