\makeatletter
\@ifundefined{HCode}
{\documentclass[screen,CRBIOL,Unicode,biblatex,published]{cedram}
\addbibresource{crbiol20250691.bib}
\newenvironment{noXML}{}{}
\let\citep\parencite
\let\citet\textcite
\def\xcitealp#1#2{\citeauthor{#1}, \citelink{#1}{#2}}
\def\citealt#1#2{\citeauthor{#1}, \citelink{#1}{#2}}
\newcommand*{\citelink}[2]{\hyperlink{cite.\therefsection @#1}{#2}}
\def\defcitealias#1#2{}
\let\citetalias\textcite
\let\citepalias\parencite
\def\thead{\noalign{\relax}\hline}
\def\endthead{\noalign{\relax}\hline}
\def\tbody{\noalign{\relax}}
\def\tabnote#1{\vskip4pt\parbox{.78\linewidth}{#1}}
\def\xxtabnote#1{\vskip4pt\parbox{.69\linewidth}{#1}}
\def\tsup#1{$^{{#1}}$}
\def\tsub#1{$_{{#1}}$}
\def\ndash{\text{--}}
\def\hyphen{\text{-}}
\def\0{\phantom{0}}
\def\mn{\phantom{$-$}}
\makeatletter
\g@addto@macro{\UrlBreaks}{\UrlOrds}
\gappto{\UrlBreaks}{\UrlOrds}
\RequirePackage{etoolbox}
\usepackage{upgreek}
\def\jobid{crbiol20250691}
%\graphicspath{{/tmp/\jobid_figs/web/}}
\graphicspath{{./figures/}}
\def\xmorerows#1#2{#2}
\newcounter{runlevel}
\usepackage{multirow}
\usepackage{comment}
\def\nrow#1{\@tempcnta #1\relax%
\advance\@tempcnta by 1\relax%
\xdef\lenrow{\the\@tempcnta}}
\def\morerows#1#2{\nrow{#1}\multirow{\lenrow}{*}{#2}}
\let\MakeYrStrItalic\relax
\def\refinput#1{}
\let\ubreak\break
\csdef{Seqnsplit}{\\}
\def\botline{\\\hline}
\def\back#1{}
\newcommand*{\hyperlinkcite}[1]{\hyper@link{cite}{cite.#1}}
\DOI{10.5802/crbiol.187}
\datereceived{2025-08-27}
\daterevised{2025-09-27}
\dateaccepted{2025-09-30}
\ItHasTeXPublished
}
{
\PassOptionsToPackage{authoryear}{natbib}
\documentclass[crgeos]{article}
\usepackage[T1]{fontenc}
\def\CDRdoi{10.5802/crbiol.187}
\def\citelink#1#2{\citeyear{#1}}
\def\xcitealp#1#2{\citealp{#1}}
\def\xmorerows#1#2{\morerows{#1}{#2}}
\let\ubreak\relax
\def\rowcolor#1#2#3#4{#4}
\def\cellcolor#1{}
\makeatletter
\def\CDRsupplementaryTwotypes#1#2{}
\def\hyperlinkcite#1#2{\citeyear{#1}}
\def\xxtabnote#1{\tabnote{#1}}
\def\href#1#2{\url[#1]{#2}}
\let\newline\break
}
\makeatother

\usepackage{upgreek}

\dateposted{2025-10-16}
\begin{document}

\begin{noXML}


\CDRsetmeta{articletype}{review}

\title{Cognitive neuroscience and miniature brains---Dissecting
higher-order learning in the brain of honey bees}

\alttitle{Neurosciences cognitives et cerveau miniature ---
Apprentissages d'ordre sup\'{e}rieur dans le cerveau des abeilles}

\author{\firstname{Martin} \lastname{Giurfa}\CDRorcid{0000-0001-7173-769X}}
\address{Sorbonne Universit\'{e}, CNRS, Inserm, Institut de Biologie Paris-Seine, 
IBPS, F-75005 Paris, France}
\address{Sorbonne Universit\'{e}, CNRS, Inserm, Centre de Neuroscience Neuro-SU, 
F-75005 Paris, France}
\email{martin.giurfa@sorbonne-universite.fr}

\keywords{\kwd{Cognition}
\kwd{Learning}
\kwd{Memory}
\kwd{Honey bee}
\kwd{Elemental learning}
\kwd{Non-elemental learning}
\kwd{Insect brain}}

\altkeywords{\kwd{Cognition}
\kwd{Apprentissage}
\kwd{M\'{e}moire}
\kwd{Abeille}
\kwd{Apprentissage el\'{e}mentaire}
\kwd{Apprentissage non-el\'{e}mentaire}
\kwd{Cerveau d'insecte}}

\thanks{European Union (ERC Advanced Grant COGNIBRAINS; project number
835032).}

\thanks{\textbf{Note.} Martin Giurfa is the recipient of the 2024 Medal
of the Integrative Biology Section of the Acad\'emie des sciences.}

\begin{abstract} 
Despite having a miniature brain---smaller than one cubic millimeter
and comprising roughly one million neurons---honey bees display a rich
behavioral repertoire in which learning and memory play a central role.
This raises the question of whether their adaptive behavior extends
beyond simple forms of learning, and whether the neural mechanisms
underlying complex cognition can be elucidated in this insect model.
Elemental olfactory conditioning, where bees learn to associate an
odorant with a sucrose reward, has provided an unparalleled framework
to dissect the neural circuits underlying conditioned (odor) and
unconditioned (sucrose) stimulus processing. This work revealed how
these pathways converge in the brain---particularly within the antennal
lobes, lateral horn, and mushroom bodies---and how learning reshapes
neural coding, notably at the level of the antennal lobe. Beyond
elemental tasks, bees master non-elemental discriminations such as
negative patterning and biconditional learning, which require
configural processing. Neural interference studies identify the
mushroom bodies as essential for these higher-order functions. Even
more complex capacities have been demonstrated: bees categorize visual
stimuli, learn abstract rules (sameness, difference, above/below),
transfer learning across sensory modalities, and display numerical
competence, including rudimentary arithmetic and an understanding of
zero. Together, these findings reveal a degree of cognitive
sophistication once thought unique to vertebrates and establish the
honey bee as a powerful system for investigating both basic and
advanced cognitive processes, as well as their neural foundations,
within a miniature brain.
\vspace*{-2pt}
\end{abstract}

\begin{altabstract}
Malgr\'{e} leur cerveau miniature --- plus petit qu'un millim\`{e}tre
cube et contenant environ un million de neurones --- les abeilles
pr\'{e}sentent un riche r\'{e}pertoire comportemental dans lequel
l'apprentissage et la m\'{e}moire jouent un r\^{o}le central. Cela
soul\`{e}ve la question de savoir si leur comportement adaptatif va
au-del\`{a} des simples formes d'apprentissage, et si les
m\'{e}canismes neuronaux sous-jacents \`{a} la cognition complexe
peuvent \^{e}tre \'{e}lucid\'{e}s dans cet insecte mod\`{e}le. Le
conditionnement olfactif \'{e}l\'{e}mentaire, dans lequel les abeilles
apprennent \`{a} associer un odorant \`{a} une r\'{e}compense de
saccharose, a permis de diss\'{e}quer les circuits neuronaux
sous-jacents au traitement d'un stimulus conditionn\'{e} (l'odeur) ou
inconditionn\'{e} (saccharose) . Ces travaux ont r\'{e}v\'{e}l\'{e}
comment ces voies convergent dans le cerveau --- en particulier dans les
lobes antennaires, les cornes lat\'{e}rales et les corps en champignon
--- et comment l'apprentissage remod\`{e}le le codage neuronal,
notamment au niveau du lobe antennaire. Au-del\`{a} des t\^{a}ches
\'{e}l\'{e}mentaires, les abeilles ma\^{i}trisent des discriminations
non-lin\'{e}aires telles que le patterning n\'{e}gatif et
l'apprentissage biconditionnel, qui n\'{e}cessitent un traitement
configural. Des \'{e}tudes sur les interf\'{e}rences neuronales
identifient les corps en champignon comme essentiels \`{a} ces
fonctions sup\'{e}rieures. Des capacit\'{e}s encore plus complexes ont
\'{e}t\'{e} d\'{e}montr\'{e}es : les abeilles cat\'{e}gorisent les
stimuli visuels, apprennent des r\`{e}gles abstraites (similitude,
diff\'{e}rence, au-dessus/en dessous), transf\`{e}rent l'apprentissage
entre les modalit\'{e}s sensorielles et font preuve de comp\'{e}tences
num\'{e}riques, notamment en arithm\'{e}tique rudimentaire et en
compr\'{e}hension du z\'{e}ro. Ces r\'{e}sultats r\'{e}v\`{e}lent un
degr\'{e} de sophistication cognitive autrefois consid\'{e}r\'{e} comme
propre aux vert\'{e}br\'{e}s, et font de l'abeille un mod\`{e}le
puissant pour \'{e}tudier les processus cognitifs fondamentaux et
avanc\'{e}s, ainsi que leurs fondements neuronaux, au sein d'un cerveau
miniature.
\end{altabstract}

\maketitle

\vspace*{-2pt}

\twocolumngrid

\end{noXML}

\defcitealias{1}{ibid.}
\defcitealias{72}{ibid.}
\defcitealias{75}{ibid.}
\defcitealias{77}{ibid.}
\defcitealias{84}{ibid.}
\defcitealias{85}{ibid.}
\defcitealias{81}{ibid.}

\section{Introduction}\label{sec1}
Karl von~Frisch (1886--1982), who dedicated his life to the study 
of honey bees, became renowned for his discovery of the bee dance---a
ritualized behavior that allows a successful forager to inform
nestmates about the distance and direction of a profitable food source 
\citep[e.g.][]{1}. Beyond his groundbreaking work on dance
communication, von~Frisch left behind a remarkably rich body of
evidence on honey bee behavior, encompassing studies on navigation,
vision, olfaction, taste, magnetic sensing, and  more \citepalias{1}. He
often described honey bees as a ``magic well'' for biological
discoveries---the more one draws from it, the more there\break is to draw.

Surprisingly, however, this fascination did not 
\mbox{extend} to the cognitive
abilities of bees. Reflecting on communication behavior, von~Frisch
wrote: ``The brain of a bee is the size of a grass seed and is not made
for thinking. The actions of bees are mainly governed by instinct'' 
\citep{2}. It is striking that such a dismissive view of bee cognition
came from the very scientist most captivated by their behavioral
complexity.

In recent decades, however, honey bees have emerged as a valuable model
for the study of learning and memory  \citep{3,4,5,6}. More recently,
they have also gained prominence in research on higher-order cognitive
capacities---abilities that were long considered the exclusive domain
of certain vertebrates known for their advanced learning skills 
\citep{7,8}.\looseness=1

In this review, I will examine the key contributions of honey bee
research to the fields of learning and memory, and how this work has
shaped our understanding of cognition. I will highlight both
established findings and open questions that illustrate the extent to
which honey bees have advanced our knowledge of cognitive
processing---at both the behavioral and cellular levels. In doing so, I
aim to emphasize the power and potential of the honey bee as a model in
cognitive neuroscience.

\section{Experimental access to learning and\newline memory in honey
bees}\label{sec2}

Honey bees can be individually trained to solve a wide variety of
discrimination tasks. Several experimental paradigms have been
developed to study learning and memory in single honey bees. This
individualized approach is critical because learning and memory are the
products of individual experience. It also enables a neurobiological
analysis that can be directly correlated with individual performance
scores.\looseness=-1

I will describe two primary protocols widely used to investigate
learning and memory in honey bees, selected for their experimental
robustness and impact:  (1)~Conditioning of approach flights to visual
targets in free-flying bees (in part 3); and  (2)~Olfactory
conditioning of the proboscis extension reflex (PER) in harnessed bees
(in part 4).

Both rely on the appetitive context of food search, using sucrose
solution as a reward mimicking nectar. In these paradigms---and in
various modified 
versions tailored to specific experimental goals---the
basic design includes an acquisition (or training) phase, during which
bees are exposed to a stimulus or perform a task that is reinforced,
followed by a test (or retrieval) phase, in which the same stimulus is
presented without reinforcement to assess memory retention. To explore
generalization and discrimination capabilities, novel stimuli may also
be introduced during the test phase alongside the trained stimulus.
Additionally, testing in the absence of the original stimulus allows
for assessment of stimulus transfer and cognitive flexibility\break (see
below).

\section{Conditioning of approach flights to visual targets in
free-flying bees}\label{sec3}
Free-flying honey bees can be conditioned to respond to various visual
cues, including color, shape, pattern, motion, and depth 
\citep{9,10,11}. In this protocol, each bee is individually marked
(typically with a colored spot on the thorax or abdomen) and displaced
by the experimenter to the training site, where it receives a sucrose
reward to encourage repeated visits  (Figure~\ref{fig1}a). This
pre-training phase occurs in the absence of training stimuli to prevent
uncontrolled associative learning. Once the bee begins visiting the
training site autonomously, visual stimuli are introduced, and correct
choices are reinforced with sucrose. The resulting associations may be
operant, classical, or a combination of both: visual stimuli (CS) may
become linked to the reward (US), to the motor response (e.g.,
landing), or both. Nevertheless, the task is predominantly operant, as
reinforcement depends on the bee's\break behavior.

\begin{figure*}
\vspace*{-1pt}
\includegraphics{fig01}
\vspace*{-2pt}
\caption{\label{fig1}Experimental protocols for the study of learning
and memory in honey bees. (a)~\textit{Visual appetitive conditioning of
free-flying bees}. A bee marked with a green spot on the abdomen is
trained to collect sugar solution in the middle of a ring pattern. 
(b)~\textit{Visual appetitive conditioning in a virtual reality (VR)
environment}. \textit{Left}: Global view of the VR system.
1:~Semicircular projection screen made of tracing paper. 2:~Holding
frame to place the tethered bee on the treadmill. 3:~Tethered bee.
4:~The treadmill is a Styrofoam ball positioned within a cylindrical
support (not visible) floating on an air cushion. 5:~Infrared mouse
optic sensors allow to record the displacement of the ball and to
reconstruct the bee's trajectory. The video projector displaying images
(not visible) is placed behind the screen. \textit{Right}. Color
discrimination learning in the VR setup. The bee had to learn to
discriminate two vertical stimuli based on their different color and
their association with reward and punishment. Stimuli were green and
blue on a black-and-white grating background.  (c)~\textit{Olfactory
appetitive conditioning of harnessed bees}. \textit{Left}: A bee
immobilized in a tube displays the proboscis extension response (PER).
\textit{Right}: Schematic of a conditioning protocol in which a
rewarded odorant (CS${+}$) is paired with a sucrose solution
(unconditioned stimulus, US) delivered via a toothpick, while a
non-rewarded odorant (CS${-}$) is also presented. In a retention test,
the CS${+}$ is presented without reward, and a PER to it indicates learning
(conditioned response). Adapted from \citet{3}.}
\vspace*{-2pt}
\end{figure*}

A recent variant of this protocol involves virtual reality setups,
where bees walk or fly in a stationary position while being exposed to
a dynamically projected visual environment that responds to their
movements (a closed-loop setup)  \citep{12,13,14,15,16}. In this
setting  (Figure~\ref{fig1}b), bees efficiently learn to discriminate
between virtual objects differing in color and shape  \citep{12,17},
offering new opportunities to study visual learning under highly
controlled experimental conditions.

\section{Olfactory conditioning of the proboscis\newline 
extension reflex (PER)
in  harnessed bees}\label{sec4}
Harnessed honey bees can be conditioned to 
\mbox{respond} to odorants in the
laboratory setting  \citep{18}. Each bee is restrained in an individual
holder that immobilizes its body while leaving the antennae and
mouthparts (mandibles and proboscis) free to move 
(Figure~\ref{fig1}c). When a hungry bee's antennae are touched with
sucrose solution, it reflexively extends its proboscis to drink---a
behavior known as the proboscis extension response (PER). In na\"{i}ve
bees, odorants alone do not elicit this response. However, if an odor
is presented just before sucrose (a forward pairing), an association is
formed such that the odor alone can later elicit PER 
(Figure~\ref{fig1}c). In classical-conditioning terms  \citep{19}, the
odor functions as the conditioned stimulus (CS), and sucrose as the
unconditioned stimulus (US). The PER to the odor then becomes the
conditioned response~(CR).

Of the two protocols, it is olfactory conditioning of the proboscis
extension reflex (PER) that has provided access to the neural and
molecular bases of learning and memory, thanks to the possibility of
recording from the bee brain in an immobilized individual that can
still learn and memorize odors.

\section{Cellular bases of appetitive olfactory PER
conditioning}\label{sec5}
A major advance enabled by olfactory PER conditioning has been the
opportunity to trace the \textbf{conditioned stimulus (CS)} and
\textbf{unconditioned stimulus (US)} pathways in the honey bee brain
and to study the underlying neural circuits of elemental associative
learning. 

Odorants are processed via a well-defined \textbf{CS processing
pathway} comprising multiple stages  
{(Figure~\ref{fig2}a)}. Olfactory
perception begins at the antennae, where olfactory receptor neurons
(ORNs) are located within sensilla. These neurons transmit information
to the antennal lobes, the primary olfactory centers of the insect
brain. Each antennal lobe contains approximately
165~glomeruli---synaptic integration units where ORNs, inhibitory local
interneurons, and projection neurons (PNs) interact. PNs relay
processed olfactory information to higher-order brain centers including
the lateral horn and the mushroom bodies. The mushroom bodies serve as
multimodal integration centers, receiving input from olfactory, visual,
gustatory, and mechanosensory modalities.\looseness=-1

\begin{figure*}
\includegraphics{fig02}
\caption{\label{fig2}CS-US associations in the honey bee brain.
(a)~Scheme of a frontal view of the bee brain showing the olfactory
(CS; in blue on the left) and sucrose (US, in red on the right) central
pathways. \textit{The CS pathway}: Olfactory sensory neurons send
information to the brain via the antennal nerve (AN). In the antennal
lobe (AL), these neurons synapse at the level of glomeruli (Gl) onto
local interneurons (not shown) and projection neurons (Pn) conveying
the olfactory information to higher-order centers, the lateral horn
(LH) and the mushroom bodies (MB). MBs are interconnected through
commissural tracts (in violet). \textit{The US pathway}: this circuit
is partially represented by the VUM$_{\mathrm{mx}1}$ neuron, which has
its soma in the subesophagic zone (SEZ) and converges with the CS
pathway at three main sites: the AL, the LH and the MB. CC: central
complex. Adapted from \citet{97}.  (b)~Scheme of the
localization and distribution of CS-US associations in the bee brain.
ORNs: olfactory receptor neurons; GRNs: gustatory receptor neurons. The
dashed line between GRNs and VUM$_{\mathrm{mx}1}$ indicates that this
part of the circuit is actually unknown.}
\end{figure*}

Neural activity along this pathway has been studied using
electrophysiology and calcium imaging  \citep{20,21,22,23,24,25}. In
experiments where proboscis movement is restricted, myogram recordings
of muscle~M17 (which controls proboscis extension) have been used as a
proxy for the conditioned response  \citep{26}.

Early studies combining PER conditioning with neural interference
techniques used cooling-induced retrograde amnesia to explore the role
of specific brain structures in memory formation. Cooling the antennal
lobes within one minute after a single {conditioning} trial induced
memory loss, while \mbox{cooling} the mushroom bodies produced amnesia when
done 5--7 minutes post-conditioning  \citep{27,28}. In contrast,
chilling the lateral horn had no effect. These results led to the
conclusion that the mushroom bodies are essential for late short-term
memory consolidation, while the antennal lobes are involved in early
short-term memory  \citep{29}. These findings were foundational and
influenced subsequent research confirming the role of mushroom bodies
in memory in other insects, such as \textit{Drosophila melanogaster} 
\citep{30,31}.

Calcium imaging studies have shown that in na\"{i}ve bees, odors evoke
reproducible glomerular activation patterns in the antennal lobes 
\citep{20,32,33}  (Figure~\ref{fig2}a). These patterns are bilaterally
symmetric and conserved across individuals  \citep{34,35}. Each odor is
thus coded by a characteristic spatial pattern, and when odors are
mixed, their neural representation can reflect additive responses or
dominance by a single component, depending on mixture complexity 
\citep{36}. As more components are added, inhibitory interactions
become apparent  \citep{20,36}. While this across-fiber pattern coding
persists in higher brain areas  \citep{37}, the mushroom bodies exhibit
sparser coding, particularly in the calyces, their input regions 
\citep{21}.

Learning modifies these neural representations. Calcium imaging shortly
after differential conditioning (A${+}$ vs.\ B${-}$) revealed increased
glomerular activation for the rewarded odor (A), but not for the
non-rewarded one (B), as well as decorrelation of odor representations,
improving discrimination  \citep{38}. This result was confirmed in a
later study examining neural activity 2--5 hours post-training 
\citep{39}, where successful learners showed increased pattern
separation between A and B, while non-learners did not.

At the level of the mushroom bodies  (Figure~\ref{fig2}a), recordings
from Kenyon cells (KCs) showed that their responses are sparse and
temporally sharp, shaped by pre- and postsynaptic mechanisms and likely
by inhibitory feedback  \citep{21}. \mbox{Associative} learning alters these
responses: while repeated odor presentation alone reduces KC activity
\mbox{(non-associative} adaptation), pairing an odor with sucrose prolongs the
KC response  \citep{40}. After conditioning, responses to the CS${+}$
recovered, while CS${-}$ responses decreased further, and the
spatiotemporal patterns of KC activation changed more for CS${-}$ than
for CS${+}$.

Molecular interference studies have further clarified mechanisms
underlying olfactory learning. For example, RNAi-mediated silencing of
the NR1 subunit of the NMDA receptor in mushroom bodies impaired
mid-term and early long-term memory, but not late long-term memory,
indicating a time-dependent role of NMDA signaling  \citep{41} (for
further analyses coupling PER conditioning and molecular interferences
see \xcitealp{42}{2006}).

Despite these advances, the impact of learning on olfactory processing
across the brain remains an open question. Future studies should
address how different conditioning protocols and memory phases affect
neural coding at multiple levels of the olfactory circuit.

Knowledge of the \textbf{US processing pathway} remains more
fragmentary. Currently, only one neuron---the VUMmx1 neuron---is known
to mediate sucrose reinforcement  (Figure~\ref{fig2}a). Located in the
subesophageal zone (SEZ), the first relay of the gustatory system 
\citep{43}, VUMmx1 responds with sustained spike activity to sucrose
stimulation of the antennae or proboscis  \citep{44}. Its~axonal
projections arborize bilaterally in three key brain regions: the
antennal lobes, the mushroom body calyces, and the lateral horns,
providing a clear anatomical basis for convergence with the olfactory
CS pathway (Figures~\ref{fig2}a,b).

VUMmx1's role as a neural representation of the US was elegantly
demonstrated by substituting sucrose with artificial depolarization of
this neuron: when stimulation of VUMmx1 followed odor presentation
(forward pairing), bees learned the association; backward pairing did
not induce learning  \citep{44}. This mirrored results with actual
sucrose reinforcement and confirmed that VUMmx1 encodes the instructive
value of the US.

VUMmx1 belongs to a class of octopamine-immunoreactive neurons 
\citep{45}. Octopamine is a biogenic amine known to promote arousal and
behavioral activation in invertebrates  \citep{46,47}. In bees, it
enhances responsiveness to both sucrose  \citep{48} and odors 
\citep{49}. When octopamine was injected into the antennal lobes or
mushroom bodies (but not the lateral horn) paired with an odor, it
substituted for the sucrose reward and induced a lasting PER to the
odor  \citep{50}, demonstrating that octopamine acts as an instructive
signal, , i.e.\ as a system allowing ordering, prioritizing and
assigning a ``good'' label to odorants  \citep{51}.

Altogether, these findings demonstrate that elemental associative
olfactory learning can be dissected at both behavioral and cellular
levels in honey bees. The bee brain enables the mapping of distributed
but localized interactions between CS and US pathways
(Figure~\ref{fig2}b). These interactions occur in at least three
regions---antennal lobes, mushroom bodies, and lateral
horns---illustrating a combination of distribution (multiple sites) and
localization (precise synaptic convergence). While these sites may
appear redundant, their contributions may differ, suggesting that
distinct types of learning and memory may be mediated by specific brain
regions---a hypothesis to be explored in the following sections.

\section{Non-elemental learning in bees}\label{sec6}
Elemental appetitive learning, as discussed above, relies on the
establishment of direct associative links between two specific and
unambiguous events in the bee's environment (e.g.\ CS
${\rightarrow}$ US). In contrast, the forms of associative learning
discussed here 
\mbox{involve} events that are ambiguous in terms of their
outcomes, rendering simple one-to-one associative links ineffective.
They thus represent greater cognitive challenges and provide a means to
assess the capacity of the bee's miniature brain to solve higher-order
cognitive tasks.

A typical example is ``negative patterning'', in which bees must learn
to discriminate a non-reinforced compound from its individually
reinforced components (A${+}$, B${+}$ vs.\ AB${-}$). This problem resists
elemental solutions because bees must recognize that the compound AB
differs fundamentally from the linear sum of A and B. Animals learning
that both A and B are individually reinforced should inhibit the summed
expectation that the compound AB is twice as rewarding. Accordingly,
the task is described as non-linear. A related case is biconditional
discrimination, where bees are required to respond to the compounds AB
and CD, but not to AC and BD (AB${+}$, CD${+}$, AC${-}$, BD${-}$). In this
scenario, each individual element (A, B, C, D) is equally paired with
reinforcement and non-reinforcement, precluding a solution based on
individual associative strength. These examples illustrate that solving
such tasks requires more complex computational strategies.

One influential approach to these problems is configural learning
theory, which posits that compound stimuli are processed as unique
configurations distinct from their elements (e.g.,  AB ${=}$ X $\neq$ A
${+}$ B)  \citep{52}. According to this view, animals trained with AB
respond minimally to A or B alone. Another account is the unique-cue
theory, which proposes that the compound AB is processed as the sum of
its components plus an additional, configural cue $(u)$ specific to the
combination (AB  ${=}$ A ${+}$ B ${+}$\ $u$)  \citep{53}. This theory allows
for stronger responses to individual components and suggests that the
compound carries a unique signature.

Due to their complexity, such tasks have rarely been studied in
invertebrates. Nonetheless, a number of studies have explored elemental
vs.\ non-elemental learning in honey bees using both visual conditioning
of free-flying bees and olfactory PER conditioning. In both modalities,
bees successfully learned biconditional discriminations (AB${+}$, CD${+}$,
AC${-}$, BD${-}$). In the visual domain, bees \mbox{discriminated} complex
patterns conforming to this {structure}  \citep{54}, while in olfaction,
bees trained with odorant mixtures  \citep{55} learned to respond to
compounds independently of the ambiguity of their components. These
results demonstrate that under specific conditions, both visual and
olfactory compounds are learned as configurations, distinct from the
simple sum of their elements.

This conclusion is further supported by studies showing that bees can
solve negative patterning discriminations (A${+}$, B${+}$, AB${-}$) in both
the visual  \citep{54} and olfactory domains  \citep{56,57,58,59}.
Solving this task requires that bees treat the compound AB as distinct
from A and B. Experiments designed to distinguish between configural
and unique-cue theories showed that bees' performance was consistent
with the latter: bees perceive the components of a compound but also
assign a unique identity to the mixture based on the interaction of
these components  \citep{58}.

\section{Cellular bases of non-elemental learning}\label{sec7}
The investigation of the neural substrates of non-elemental learning
took advantage of the olfactory conditioning of PER. The use of a
negative patterning protocol allowed asking the question of the neural
substrates underlying this form of non-linear problem solving in the
olfactory domain  \citep{60}. Control bees that received saline
solution injections at the level of the mushroom bodies 
\mbox{(Figure~\ref{fig3}a)} were able to solve a negative patterning
discrimination (A${+}$, B${+}$, AB${-}$). Yet, when mushroom body activity
was blocked locally using injections of procaine \citep{61}---a sodium
and potassium channel blocker---, bees were unable to learn the negative
patterning task  (Figure~\ref{fig3}b). A critical control experiment
involved again mushroom body blockade using procaine, but this time,
bees were subjected to an elemental discrimination, which nevertheless
resembled a negative patterning problem: A${+}$, B${+}$, CD${-}$ 
(Figure~\ref{fig3}c); in this case, each odorant is associated
univocally with the presence or absence of reinforcement so that there
is no ambiguity in the problem. Controls and bees with mushroom bodies
blocked by procaine could solve equally well this discrimination 
(Figure~\ref{fig3}c), thus showing that the incapacity to solve the
negative patterning discrimination in the absence of mushroom bodies
was inherent to the nature of the problem trained. These 
\mbox{results} thus
show a specific role for the mushroom bodies in non-elemental learning 
\citep{60}. 

\begin{figure*}
\vspace*{-1pt}
\includegraphics{fig03}
\vspace*{-1pt}
\caption{\label{fig3}Mushroom body blockade impairs negative patterning
(NP) discrimination. (a)~Left: harnessed honey bee injected in the
mushroom bodies (MBs). Right: Experimental protocol. After feeding and
rest, bees were injected bilaterally with either the anesthetic
procaine or saline solution for control bees (black arrow at time~0).
The white arrows indicate the sites (VL) of bilateral injections
(scale bar: 250~$\upmu$m). Twenty min after injection, bees were
conditioned following a NP regime.  (b)~Percentage of conditioned PER
of a group of bees injected with saline solution (controls) in response
to rewarded pure odorants (A${+}$/B${+}$, black bar) and to an unrewarded
compound (AB${-}$, white bar) at the end of conditioning (las
conditioning block). Saline-injected bees learned to respond
significantly more to the odorants A${+}$ and B${+}$ than to its unrewarded
compound (AB${-}$). Procaine-injected bees did not learn the NP
discrimination.  (c)~When bees were conditioned with an elemental,
non-ambiguous discrimination (CD${-}$ vs.\ A${+}$/B${+}$), both
saline-injected and procaine-injected bees learned the discrimination,
thus showing that MBs are required for non-elemental learning but are
dispensable for elemental learning. Adapted from 
\citet{60}.}
\vspace*{-2pt}
\end{figure*}

Furthermore, local injections of picrotoxin, a GABAergic inhibitor,
into mushroom bodies were used to dissect the contributions of two
feedback tracks of GABAergic neurons, the A3v neurons---which provide
GABAergic input from the lobes to the calyces of the mushroom bodies
(i.e.\ output ${\rightarrow}$ input)---and the A3d neurons---which provide
GABAergic input from the lobes to the lobes of the mushroom bodies
(i.e.\ output ${\rightarrow}$ output) \citep{62}. Injections of
picrotoxin into the calyces---but not into the lobes---disrupted negative
patterning performance by preventing suppression of responses to the
non-reinforced compound AB  \citep{60}. This finding highlights the
importance of GABAergic feedback inhibition from Av3 neurons at the
mushroom body input for solving such non-linear tasks.

\section{Positive transfer of learning}\label{sec8}
In this section, I focus on problem solving in which animals respond
adaptively to novel stimuli they have never encountered
before---stimuli that do not predict a specific outcome based solely on
the animals' past experience. Such positive transfer of learning 
\citep{63} differs fundamentally from elemental forms of learning,
which link known stimuli or actions to specific reinforcers. In the
cases considered here, responses may become independent of the physical
nature of the presented stimuli and are instead guided by abstract
rules (e.g., relational rules such as ``on top of'' or ``larger
than''), which can be applied regardless of stimulus similarity. Most
of these experiments have been conducted in the visual modality, i.e.\ 
with free-flying bees trained to solve visual discriminations 
\citep{5,64}.

\subsection{Categorization of visual stimuli}\label{sec8.1} 
Positive transfer of learning is a hallmark of categorization
performance. Categorization refers to the classification of perceptual
inputs into functional groups  \citep{65}. It is the ability to group
distinguishable objects or events based on a common feature or set of
features, and to respond similarly to them  \citep{65,66,67}.
Categorization thus involves extracting defining features from the
environment. A typical categorization experiment trains an animal to
extract a category's basic attributes and tests its performance using
novel stimuli that either do or do not share these attributes. If the
animal chooses the novel stimuli based on these defining features, it
demonstrates category learning and positive transfer.

Several studies have demonstrated visual categorization in free-flying
honey bees trained to discriminate patterns and shapes. For example, 
\citet{68} trained bees to distinguish between vertical gratings with
different orientations (e.g., 45\textdegree\ vs.\ 135\textdegree),
rewarding only one orientation with sucrose solution. Each bee was
trained with a changing succession of pairs of different gratings, one
of which was always rewarded and the other not. Although the specific
gratings changed across trials, the rewarded and non-rewarded gratings
maintained consistent orientations. Bees learned to extract and respond
to the common orientation among rewarded stimuli and transferred this
knowledge to novel, 
\mbox{non-rewarded} patterns that shared the trained
orientation.

Bees can also categorize visual patterns based on bilateral symmetry.
When trained to discriminate symmetrical from asymmetrical patterns,
they generalize this knowledge to novel patterns  \citep{69}. Similar
capacities apply to radial symmetry, concentric organization, pattern
disruption 
\citep[see][for review]{70}, and even photographs
belonging to a given class, such as flowers or landscapes  \citep{71}.

To explain how bees categorize highly variable photographs of, for
example, radial flowers,  \citet{72} proposed that bees integrate
multiple coexisting orientations into a global, multi-feature
representation. A category such as ``radial flower'' might be defined
by five or more radiating edges. Bees trained with complex patterns
sharing such a layout transferred their responses to novel patterns
that preserved the trained orientation configuration. They even
responded to patterns with fewer correct orientations based on how
closely these matched the trained template \citepalias{72}. Thus, honey bees
extract regularities and generate generalized object representations
from finite feature sets.

These findings show that honey bees exhibit positive transfer from
trained to novel stimuli in a manner consistent with categorization.
However, such results may admit an elemental interpretation. If
categorization is based on specific features such as orientation, its
neural implementation may be straightforward. Stimuli sharing a feature
may activate the same orientation detectors in the bee optic lobes. The
orientation and tuning of these detectors have been already
characterized by means of electrophysiological recordings in the honey
bee optic lobes  \citep{73}. Category learning could thus involve
reinforcing associations between symmetry detectors and reward
pathways, akin to Pavlovian conditioning. From this view, although
categorization involves positive transfer, it could be based on
elemental stimulus-reinforcer\break links.

\subsection{Rule learning}\label{sec8.2} 
In contrast, rule learning involves learning relationships between
objects, not the objects themselves. Therefore, positive transfer
occurs independently of~the stimuli's physical nature   \citep{65,74}.
Classic  examples  include the rules of sameness and difference,
assessed through  delayed matching-to-sample (DMTS) and delayed
non-matching-to-sample (DNMTS) tasks,  \mbox{respectively.}

In DMTS, animals are shown a sample, followed by a choice between
stimuli where one matches the sample. Reinforcement follows selection
of the matching stimulus. Since samples vary, animals must learn the
abstract rule: ``always choose what you were shown.'' In DNMTS, the
rule is reversed: ``always choose the different stimulus.''

Honey bees trained in a Y-maze learned both rules  \citep{75}. Bees
were trained in a DMTS problem in which they were presented with a
changing non-rewarded sample (i.e.\ one of two different color disks or
one of two different black-and-white gratings, vertical or horizontal)
at the entrance of a maze  (Figure~\ref{fig4}). The bees were rewarded
only if they chose the stimulus identical to the sample once within the
maze. Bees trained with colors and presented in transfer tests with
black-and-white gratings that they did not experience before solved the
problem and chose the grating identical to the sample at the entrance
of the maze. Similarly, bees trained with the gratings and tested with
colors in transfer tests also solved the problem and chose the novel
color corresponding to that of the sample grating at the maze entrance.
Transfer was not limited to different kinds of modalities (pattern vs.\ 
color) within the visual domain, but could also operate between
drastically different domains such as olfaction and vision,
demonstrating cross-modal transfer. This transfer even spanned vision
and olfaction. Bees also learned a rule of difference in DNMTS tasks 
\citepalias{75}. Working memory underlying DMTS tasks in bees lasts about 5
seconds  \citep{76}, consistent with short-term memory durations
observed in simpler associative tasks  \citep{6}. 

\begin{figure*}
\includegraphics{fig04}
\caption{\label{fig4}Rule learning in honey bees. Honey bees trained to
collect sugar solution in a Y-maze  (a)~on a series of different
patterns or two different colors  (b)~learn a rule of sameness.
Learning and transfer performance of bees in a delayed
matching-to-sample task in which they were trained to colors
(Experiment~1) or to black-and-white, vertical and horizontal gratings
(Experiment~2).  (c,d)~Transfer tests with novel stimuli. (c)~In
Experiment~1, bees trained on the colors were tested on the gratings. 
(d)~In Experiment~2, bees trained on the gratings were tested on the
colors. In both cases bees chose the novel stimuli corresponding to the
sample although they had no experience with such test stimuli. $n$
denotes number of choices evaluated. Adapted from \citet{75}.}
\end{figure*}

\subsection{Spatial and simultaneous concept learning}\label{sec8.3} 
Bees can also learn spatial concepts, such as ``above'', ``below'',
``left of'', and ``right of'', which are \mbox{essential} for orientation and
displacement. The capacity to learn an above/below relationship between
visual stimuli and to transfer it to novel stimuli that are
perceptually different from those used during the training was shown by
training free-flying bees to choose visual stimuli presented above or
below a horizontal bar  \citep{77}. Training followed a differential
conditioning procedure in which one spatial relation (e.g.\ ``target
above bar'') was associated with sucrose solution whilst the other
relation (e.g.\ ``target below bar'') was associated with 
\mbox{quinine}
solution. One group of bees was rewarded on the ``target above bar''
relation while another group was rewarded on the ``target below bar''
relation. After completing the training, bees were subjected to a
non-rewarded transfer test in which a novel target stimulus (not used
during the training) was presented above or below the bar. Despite the
novelty of the test situation, bees responded appropriately: if trained
for the above relationship they chose the novel stimulus above the bar,
and if trained for the below relationship they chose the novel stimulus
below the bar  \citepalias{77}. 

Bees can also acquire multiple concepts simultaneously. In one study,
they learned both spatial (e.g., ``above/below'', ``left/right'') and
difference concepts  \citep{78}. Stimuli featured two images in
specific spatial relations, and bees had to select configurations that
both obeyed a spatial rule and involved different images. Bees
generalized these concepts to novel images. In conflict tests, where
one concept was satisfied but not the other, bees showed no preference,
suggesting equal weighting of both concepts.

\section{Numerical cognition}\label{sec9}
A new research field on bee numerosity has revealed a remarkable
capacity for numerical cognition  
\citep[see reviews in][]{80,79}. Spatial arrays of items have been used in
experiments employing a delayed matching-to-sample protocol, in which
bees were trained to fly into a Y-maze and choose the stimulus that
matched the numerosity of a sample array presented at the maze entrance
\citep{81}. Bees trained to match sample arrays containing two or three
items successfully learned the task and transferred their choice to
novel arrays that maintained the same number of items but differed in
color, shape, and configuration. However, performance declined when the
sample contained four items, and higher numerosities resulted in
increasingly unsuccessful outcomes. These results suggest that the
upper limit of numerical discrimination in these conditions may be
close to four  \citepalias{81}.

Like some primates and birds, bees also appear to possess a concept of
zero, understood as the lower bound of a continuous numerical scale 
\citep{82}. This was demonstrated in 
\mbox{experiments} where bees were
trained to follow the constant rule of 
\mbox{selecting} the smaller of two
numerosities  (Figure~\ref{fig5}a), varying between 1 and 4. When
tested with an unfamiliar comparison---one item versus an empty
set---they reliably chose the empty set, indicating that they treated
it as a quantity smaller than one, two, or more items  \citep{83} 
(Figure~\ref{fig5}b). Furthermore, their performance improved with
increasing numerical distance (e.g., 0 vs.\ 6 was easier than 0 vs.\ 1; 
\mbox{Figure~\ref{fig5}b}), replicating the \textit{numerical distance effect}
observed in vertebrates, where discrimination between numbers improves
as the numerical difference between them increases.

\begin{figure*}
\includegraphics{fig05} 
\vspace*{2pt}
\caption{\label{fig5}Zero representation in bee numerosity. 
(a)~\textit{Left}.  Experimental setup for studying the presence of
zero in bee numerosity. Bees were trained to fly to a vertical screen
displaying two numerosities via four stimuli made of black dots (here 3
vs.\ 4). Stimuli were controlled to discard the use of low-level cues.
One numerosity (3) was rewarded with sucrose solution (red ${+}$ sign).
\textit{Right}. Training and tests. Bees were trained with varying
pairs of numerosities; they were rewarded for choosing always the
\textit{smaller} numerosity (indicated by a red ${+}$ sign). When they
reached a \% of correct choices ${>}$ 80\%, they were tested with an
empty background (``zero'') vs.\ a single item. Further transfer tests
opposing the empty background to two, three, four, five and six items
were also performed.  (b)~Results of the transfer tests. Bees trained
to choose the smaller numerosity preferred the zero stimulus to any
other numerosity. Their performance is consistent with the
\textit{numerical distance effect}' as the ability to discriminate
between numbers improves as the numerical distance between them
increases.}
\vspace*{4pt}
\end{figure*}

Because this study relied on the bees' capacity to use relative
numerosity (``choose the smaller set''), another experiment was
designed to test whether bees could also use absolute numerosity 
\citep{84}. One group of bees (``larger'') was trained to choose 3 over
2, while another group (``smaller'') was trained to choose 3 over 4. In
subsequent tests, both groups were presented with the previously
rewarded numerosity (3) and a novel one (4 for ``larger'', 2 for
``smaller''). Bees in both groups preferred the three-item array,
suggesting a tendency to rely on absolute numerosity---unlike many
vertebrates, which typically favor relative rules under similar
conditions. Nonetheless, a \textit{numerical size effect}, consistent
with Weber's law, was observed: discrimination performance declined as
overall numerical magnitude increased (e.g., 3 vs.\ 4 was harder than 2
vs.\ 3)  \citepalias{84}.

Bees have also been shown to perform rudimentary arithmetic operations.
In a recent study  \citep{85}, the color of a sample array (blue or
yellow) indicated which arithmetic operation to perform 
(Figure~\ref{fig1}b). If the sample was blue, bees had to choose the
array that contained one more item (addition); if yellow, they had to
select the array with one fewer item (subtraction). Bees learned the
task and successfully transferred this rule to novel numerosities,
adjusting their response based on the indicated operation. These
findings demonstrate that bees can master simple arithmetic operations
under controlled conditions  \citepalias{85}.

Symbolic matching has also been explored in bees. In one experiment 
\citep{86}, \mbox{different} groups of bees were trained to match a visual
sign (e.g., ``N'' or ``${\bot}$'') to a specific numerosity (2 or 3), or
vice versa. Both groups learned their respective associations and
generalized them to novel stimuli with varying colors, shapes, and
\mbox{configurations.} However, they failed to reverse the direction of the
learned association (i.e., from number-to-sign if trained on
sign-to-number, and vice versa). Given that bees are capable of
learning other reversal tasks, this failure may reflect a specific
limitation related to the numerical aspect of the task, suggesting
boundaries to their symbolic numerical competences.

Similarities with vertebrate numerosity extend to the existence of a
``mental number line'' (MNL), the representation of numbers by which
humans organize numbers spatially from left to right according to their
magnitude. A recent study explored how bees trained to specific
numerosities (e.g., 3) associated with a reward of sucrose solution
respond to novel numbers (i.e.\ 1 or 5) when identical options (1 vs.\ 1
or 5 vs.\ 5) are shown on the left and the right, in the absence of the
trained numerosity  \citep{87}. Results showed that bees order numbers
from left to right according to their magnitude (i.e.\ after being
trained to 3, they preferred 1 on the left and 5 on the right) and that
the location of a number on that line varies with the reference number
previously trained. Thus, the MNL is a form of numeric 
\mbox{representation}
that is evolutionary conserved across nervous systems endowed with a
sense of number, irrespective of their neural complexity.

Overall, these results indicate that bees and vertebrates share
similarities in their numeric competences, thus suggesting that
numerosity is evolutionary conserved and can be implemented in
miniature brains.

\section{An ecological context for honey bee\newline cognition}\label{sec10}
While laboratory paradigms have revealed the extent and richness of
honey bee cognition, it remains essential to understand how these
capacities are expressed in natural behavior. Honey bees are
central-place foragers, meaning that their foraging trips \mbox{always} begin
and end at the hive. To succeed in this task, they rely on flexible
strategies that allow them to recognize and generalize visual patterns
both at flowers and at the nest. Bees extract salient image features
and combine them into specific configural 
\mbox{representations} that support
flower recognition  \citep{72,88}. They can then generalize these
representations to novel 
visual images that share the same
configurations, even when other spatial details or positions within the
visual field vary substantially. Such capacities are likely involved in
extracting and recognizing relational concepts, where the critical
information lies in the relationships among visual features.

During foraging bouts, bees also rely on celestial cues for compass
navigation, while prominent landmarks and landscape features help
define routes and support orientation  \citep{89,90,91,92}. In this
context, the ability to represent spatial relationships in a
generalized form around the hive or food sources is particularly
advantageous. Extracting relations such as ``same'', ``different'',
``to the right (or left) of'', or ``above (below)'' may help bees
maintain reliable routes in a changing environment, where landmarks
themselves can vary in appearance  \citep{93}.

In addition, numerical information contributes to efficient foraging
and navigation. Estimating the number of landmarks can indicate where
to land, while assessing the number of flowers within a patch provides
an estimate of its richness. Such numerical competencies likely enhance
foraging efficiency and, ultimately, colony survival. Selective
pressures may therefore have favored the evolution of numerical
abilities in bees, much as they have in\break vertebrates.

Finally, non-elemental learning capabilities are especially
advantageous in the complex floral market. Honey bees typically exploit
one floral species at a time---a behavior known as flower constancy 
\citep{94,95,96}---which maximizes foraging efficiency. Yet many floral
species share odor components, and if bees were to generalize
indiscriminately across species with overlapping odor cues, flower
constancy would break down. By learning specific odor combinations as
unique configurations, distinct from the mere sum of their components,
bees can discriminate among species and thus sustain both flower
constancy and foraging 
\mbox{efficiency.}

These arguments help bridge the gap between experimental paradigms and
natural behavior, showing that findings from controlled laboratory
setups reveal capacities with genuine adaptive value in the wild. They
also reinforce the status of honey bees as a
\mbox{powerful} model for
studying cognition beyond simple associative learning.

\section{Conclusion}\label{sec11}
This review highlights the remarkable richness and flexibility of
experience-dependent behavior in honey bees, and the fact that various
forms of cognitive processing based on associative learning---ranging
from simple to more complex---can be formalized and studied under
controlled laboratory conditions. The adoption of rigorous definitions
from elemental and non-elemental learning frameworks provides a
valuable foundation for assessing the extent to which honey bees can
transcend basic associative processes. As demonstrated by the numerous
examples reviewed here, such an experimental approach has revealed a
level of cognitive sophistication in bees that challenges traditional
views of insect intelligence as inherently limited.

While specific neural circuits have been identified for elemental forms
of associative learning, the neural basis of more complex forms of
problem solving remains poorly understood. Existing evidence
consistently implicates the mushroom bodies---a central structure in
the insect brain---in learning and memory. Although certain elemental
discriminations can be accomplished without mushroom bodies 
\citep[e.g.,][]{60}, this does not appear to be the case for
higher-order, non-elemental tasks. Although the precise substrates and
circuits underlying complex cognition in the bee brain remain to be
identified, there is reason for optimism. What is now required is a
conceptual shift that promotes deeper investigation into complex
cognitive processing in insect brains.

A key question for future research concerns the specific limitations of
the bee brain compared to larger brains, and what structural or
functional constraints might underlie them. Addressing this question
requires a better understanding of the deficits or boundaries of bee
cognition---an area that remains largely unexplored. Due to space
limitations, we have not addressed how various forms of learning
function in ecologically relevant contexts such as navigation and
communication. These domains offer additional and promising frameworks
for studying cognitive processing. Questions such as how bees represent
space or flexibly adjust their 
\mbox{communication} strategies remain crucial
for evaluating the cognitive potential of the bee brain. Ultimately,
such questions should be linked to specific neural circuits and
structures---a goal that remains elusive, but 
\mbox{attainable.} 

Research on honey bee behavior invites an optimistic perspective on
these challenges. Because honey bee learning shares important features
with that of vertebrates, this insect may serve as a powerful model for
investigating intermediate levels of cognitive complexity and their
neural underpinnings.

\section*{Declaration of interests}
Views and opinions expressed are those of the author only and do not
necessarily reflect those of the European Union or the European
Research Council. Neither the European Union nor the granting authority
can be held responsible for them.

The author does not work for, advise, own shares in, or receive funds
from any organization that could benefit from this article, and has
declared no affiliations other than their research organization.

\section*{Funding}
The work of MG is currently funded by the European Union (ERC Advanced
Grant COGNIBRAINS; project number 835032).

\CDRGrant[EU]{835032}

\back{}

\printbibliography
\refinput{crbiol20250691-reference.tex}

\end{document}