MCMP: Mixed Choice Multiparty Session Types

• MCMP is a multiparty protocol framework that permits a single choice to combine both input and output actions, overcoming classical MPST limitations.
• It employs both synchronous and asynchronous models with techniques like coherence, transient inconsistency, and stale-message handling to guarantee safety and liveness.
• The approach supports modular, interdependent communication and refinement-oriented probabilistic systems while maintaining Subject Reduction, Session Fidelity, and Lock Freedom.

Mixed Choice Multiparty Session Types (MCMP) designate multiparty session disciplines in which a participant may use a single choice construct whose branches are guarded by both input and output actions, rather than being restricted to the unmixed internal and external choices of classical session typing. In the recent literature, MCMP appears both as a synchronous Simple MultiParty Session (SMPS) calculus with global-type assignment, as an asynchronous multiparty framework with explicit mixed-choice state and transient inconsistency, and as the basis for refinement-oriented probabilistic session systems (Barbanera et al., 19 Aug 2025, Bocchi et al., 27 Feb 2026, Blechschmidt, 12 Sep 2025). Across these formulations, the central problem is stable: to increase the expressive power of multiparty protocols without losing Subject Reduction, Session Fidelity, communication safety, deadlock-freedom, Lock Freedom, or progress.

1. Classical MPST limitations and the emergence of mixed choice

Classical multiparty session types describe structured communication protocols among multiple participants by means of global types and their local projections. In this setting, choice is restricted in a characteristic way: a single participant selects a branch and the others follow it, or a participant offers an input-only branch structure. At the level of local behaviour, this yields unmixed choice: a participant may choose among different input branches, or among different output branches, but not both in the same non-deterministic choice (Barbanera et al., 8 Apr 2026).

The restriction is not merely syntactic. Earlier MPST metatheory tied Subject Reduction to pairwise consistency or strong duality conditions on session environments. A representative example is the three-role protocol

$\gtG \;=\; \gtCommRaw{\roleP}{\roleQ}{\!\!\!\! \begin{array}{l} \gtCommChoice{m1}{}{ \gtCommSingle{\roleQ}{\roleR}{m2}{}{ \gtCommSingle{\roleR}{\roleP}{m3}{}{ } } }\,, \gtCommChoice{stop}{}{ \gtCommSingle{\roleQ}{\roleR}{quit}{}{ } } \end{array} \!\!\!\! }$

where $\roleP$ chooses between continuing and stopping, and the later interaction between $\roleR$ and $\roleP$ exists only in the first branch. The abstract on “beyond duality” isolates this pattern as an inter-role dependency that breaks the usual pairwise consistency check between $\gtG \upharpoonright \roleP$ and $\gtG \upharpoonright \roleR$, even though the protocol is intuitively safe (Scalas et al., 2017).

This limitation is closely related to the motivation for MCMP. The asynchronous mixed-choice literature states the same point from a different angle: classical MPST cannot directly express distributed interaction patterns in which several parties resolve choices jointly, especially when a participant must choose among both input and output actions (Barbanera et al., 8 Apr 2026). The expressiveness literature sharpens the claim further. In binary session systems, a surface syntax for mixed choice may still be “rather separate and not mixed” because each choice remains tied to a single endpoint; by contrast, the multiparty setting permits mixed choice to combine input/output variation with variation in communication partners, which changes the expressive landscape (Peters et al., 2022, Peters et al., 2024).

A plausible implication is that MCMP should be read not as a minor extension of branching and selection, but as a response to a specific defect of projection-based, pairwise-duality-centric typing disciplines: they exclude protocols whose correctness depends on globally coordinated, conditionally enabled interactions among more than two roles (Scalas et al., 2017).

2. Core formal idea: mixed choice in process and type syntax

In the SMPS-style synchronous formulation, processes are coinductively defined regular terms

$P ::= \mathbf{0} \mid \Sigma_{i\in I} \pi_i.P_i$

with action prefixes

$\pi ::= p?{\lambda} \mid p!{\lambda}.$

Here $p?{\lambda}$ denotes input from participant $p$ of a message labelled $\roleP$0, and $\roleP$1 denotes output to $\roleP$2 with label $\roleP$3. The summation $\roleP$4 is a nondeterministic choice, and mixed choice is exactly the permission for some $\roleP$5 to be inputs and others outputs in the same summation (Barbanera et al., 19 Aug 2025).

The asynchronous formulation adopts essentially the same prefix discipline, but places it in a calculus with a single logical queue of messages, where outputs are non-blocking and inputs pull messages from the queue. It explicitly characterizes mixed choice as allowing nondeterministic choices whose branches are guarded by arbitrary prefixes, including both inputs and outputs, for instance

$\roleP$6

with some prefixes $\roleP$7 and others $\roleP$8 arbitrarily mixed (Barbanera et al., 8 Apr 2026).

At the local-type level, the technical report on separation and encodability gives the mixed-choice constructor directly as

$\roleP$9

where each branch is either a send action or a receive action, and the peer participants $\roleR$0 may differ across branches (Peters et al., 2024). This distinguishes MCMP from classical MP, where local choice is either a pure input sum or a pure output structure, and from binary mixed-session calculi where each mixed choice remains attached to a single endpoint (Peters et al., 2022).

Several later developments keep this mixed-choice core while changing the ambient typing framework. The asynchronous 2026 framework types sessions directly with global types equipped with a coinductively defined labelled transition system and a notion of coherence of communication label sets (Barbanera et al., 8 Apr 2026). The modular 2025 framework uses a single-layered judgement

$\roleR$1

assigning global types directly to sessions under a participant partition $\roleR$2, with no explicit local types (Barbanera et al., 19 Aug 2025). The probabilistic extension enriches the same mixed-choice shape with probabilistic output sums:

$\roleR$3

where the summands may be input types or probabilistic output types (Blechschmidt, 12 Sep 2025).

The common invariant is that MCMP treats the local control state of a participant as a genuinely mixed frontier of communication possibilities, not as a choice already polarised into sender-only or receiver-only behaviour.

3. Typing disciplines: from duality to coherence, liveness, and direct global typing

One line of development replaces restrictive pairwise duality conditions with semantic criteria over richer typing objects. The “beyond duality” abstract proposes a judgement

$\roleR$4

with guarantee and rely typing contexts. In that system, the crucial hypothesis is not consistency of a single session environment but liveness of the pair $\roleR$5, ensuring along context reductions that each output can synchronise with a compatible input and vice versa (Scalas et al., 2017). Although that work does not use the term MCMP, it isolates the semantic move that later mixed-choice systems exploit: replace pairwise syntactic duality by a whole-context notion capable of supporting complex inter-role dependencies.

The asynchronous 2026 framework takes a different route. It is “orthogonal to the well-established, projection-based approaches of the MultiParty Session Type framework” and types sessions directly with global types. The key technical notion is coherence of communication label sets: roughly, a coherent set contains either all the communications enabled in the session, or all the actions currently exhibited by a participant, provided that at least one input is enabled for each expected sender (Barbanera et al., 8 Apr 2026). From this basis it proves Subject Reduction and Session Fidelity, which in turn imply Lock-Free and Orphan-Message-Free behaviour, and it further studies Eventual Reception for an extension of the type system (Barbanera et al., 8 Apr 2026).

The modular synchronous system also uses coherence, but now relative to a partition of the participants into modules. Its typing rule TComm types a whole coherent set of labels at once, rather than a single local send or receive step. The labels must be $\roleR$6-coherent for the current session and have a witness module in the unique $\roleR$7-modularisation (Barbanera et al., 19 Aug 2025). This is a strong departure from standard local-rule MPST typing: an entire “frontier” of interactions for one module is checked in a single coinductive step.

The probabilistic framework extends the same trend toward richer typing judgements. Its subtyping relation is not limited to one endpoint refining another; it relates local contexts with active channels,

$\roleR$8

and allows “one channel (the interface) to be substituted by several channels (the refinement)” (Blechschmidt, 12 Sep 2025). The interface-existence theorem states that for any safe, deadlock-free local context $\roleR$9 there exists a single interface channel $\roleP$0 such that

$\roleP$1

This makes MCMP not only a vehicle for richer control flow, but also a substrate for compositional refinement (Blechschmidt, 12 Sep 2025).

These systems differ in presentation—rely/guarantee liveness, coherent label sets, modular global typing, multi-channel subtyping—but converge on one point: mixed-choice multiparty protocols cannot be handled adequately by the classical discipline of branch-by-branch projection plus pairwise duality checks.

4. Modularity, safety, and liveness in synchronous MCMP

The modular synchronous theory gives one of the most explicit and technically constrained formulations of MCMP. A modular multiparty session is a parallel composition of participant processes partitioned into modules, or subsessions, within which mixed choice may be used without restriction. Communication between modules is performed by designated connectors, whose processes satisfy a connecting discipline: if some guard in a choice involves a participant outside the module, then all guards in that choice must involve that same external participant (Barbanera et al., 19 Aug 2025).

This connector discipline is the mechanism by which unrestricted internal mixed choice coexists with analyzable global behaviour. It is coupled to a notion of $\roleP$2-modularisation of a session and to the fact that modularisability is preserved by reduction (Barbanera et al., 19 Aug 2025). The paper also proves the existence of a unique minimal modular partition with respect to refinement. This suggests that modularity is not an optional presentation device but a structural invariant of the typed semantics.

The metatheory is built around three theorems. Subject Reduction states that if

$\roleP$3

then there exists $\roleP$4 such that

$\roleP$5

Session Fidelity states the converse correspondence: if a global type steps, the session can perform the corresponding communication and remain typable. Lock Freedom states that every typable modular session is lock free in the sense that any participant still present after a finite trace can engage in a future communication (Barbanera et al., 19 Aug 2025).

The paper’s examples are instructive because they are not merely toy branch/selection variations. A local election protocol over five participants and a station uses mixed choice such as

$\roleP$6

with cyclically renamed variants for the other participants (Barbanera et al., 19 Aug 2025). A larger modular example composes three local elections with one global election; the coherent sets chosen by successive TComm applications isolate module-specific frontiers while preserving the overall protocol structure (Barbanera et al., 19 Aug 2025).

The article’s central design slogan is stated plainly: mixed choice is “fully exploited only inside loosely coupled modules” (Barbanera et al., 19 Aug 2025). This is not simply an implementation heuristic. It marks a precise safety tradeoff: unrestricted mixed choice is admitted where intensive internal coordination can be analysed as a unit, and cross-module interactions are restricted enough to preserve Subject Reduction, Session Fidelity, and Lock Freedom.

5. Asynchronous MCMP: races, transient inconsistency, and stale messages

The 2026 asynchronous framework pushes MCMP into the setting where mixed choice is hardest: buffered, distributed, asynchronous communication (Bocchi et al., 27 Feb 2026). Its starting point is that many realistic patterns—timeouts, interrupts, heartbeat-based failure handling, cancellation—require a role to be ready both to send and to receive across different channels. Classical directed choice excludes such states because it assumes a single participant determines the branch and all local views stay aligned (Bocchi et al., 27 Feb 2026).

The framework introduces a core global construct for mixed choice with an asymmetric observer role. Informally, a left-hand side models one communication regime, a right-hand side another, and the observer on the right-hand side can overrule the left-hand side by sending on the RHS. The key innovation is to permit transient inconsistencies in protocol state between distributed participants, while ensuring that all participants can always eventually reach a mutually consistent state (Bocchi et al., 27 Feb 2026).

A running timeout protocol with roles A, B, and C makes the point concrete. On the LHS, A and B begin the normal interaction and C participates later; on the RHS, B sends timeout notifications TOa and TOc. Because communication is asynchronous, a1 and TOa may cross in flight, so A may have already advanced on the LHS when B commits to the RHS. The system addresses this by tracking commitment and by associating messages with paths through nested mixed choices. Messages that become semantically obsolete are identified as stale and can be discarded locally by a garbage-collection step (Bocchi et al., 27 Feb 2026).

Three structural notions control this semantics. Committing sets identify the labels whose receipt commits a role to one side of a mixed choice. Awareness requires, roughly, that on the RHS only the observer can initiate the branch, while on the LHS any role that terminates must have received enough information to know which side the protocol has taken. Balance generalises the standard requirement that branches align their participating roles; for active mixed choices it requires that the roles present on one side together with those already committed match the roles present on the other side together with the opposite commitments (Bocchi et al., 27 Feb 2026).

Under these conditions the paper proves a Global Progress theorem for initial, aware, balanced global types, and an Operational Correspondence theorem between global types and their distributed local projections (Bocchi et al., 27 Feb 2026). The projected systems also satisfy local progress. A further theorem states that every queued message can eventually either be received or become stale and be garbage-collected; this yields the paper’s strong asynchronous form of orphan-message freedom (Bocchi et al., 27 Feb 2026).

The same paper describes a practical toolchain. Protocols are written in Scribble with mixed { ... } or { ... } blocks, validated against well-formedness, awareness, balance, and projectability conditions, projected to local types, translated into event-driven finite state machines, and compiled into compliant gen_statem processes in Erlang/OTP (Bocchi et al., 27 Feb 2026). The framework is evaluated by specifying and reimplementing part of the amqp_client of the RabbitMQ broker for Erlang (Bocchi et al., 27 Feb 2026).

This asynchronous theory changes the meaning of MCMP in an important way. In the synchronous modular setting, the main difficulty is controlling expressive power. In the asynchronous setting, the main difficulty is tolerating disagreement long enough for it to be resolved soundly. The notion of stale-message purging is the operational counterpart of that shift.

6. Expressiveness, related constructions, and subsequent extensions

The technical report on separation and encodability gives the sharpest expressiveness account. It studies nine subcalculi and proves “8 new encodablity results and 20 new separation results,” concluding that “MCMP is strictly more expressive than classical multiparty sessions (MP) and mixed choice in mixed sessions” (Peters et al., 2024). The key contrast is with the binary case: earlier results showed that binary mixed sessions do not gain expressive power over non-mixed binary sessions under good encodings, but this collapse does not survive the move to multiparty mixed choice (Peters et al., 2022, Peters et al., 2022).

The report attributes the gain to the combination of two dimensions of mixing: input/output polarity and communication partner. In MCMP, a single choice can mix sends and receives to different participants, and this supports synchronisation patterns unavailable in weaker calculi (Peters et al., 2024). In particular, the top layer of the reported hierarchy consists of MSMP and MCMP, which can realise the stronger $\roleP$7 synchronisation pattern, whereas classical MP and several restricted mixed-choice variants cannot (Peters et al., 2024).

A different but related line is the study of symmetric sum types, which model multiparty agreement-based choice rather than observer-driven mixed choice. The symmetric synchronisation construct

$\roleP$8

reduces when all $\roleP$9 participants can agree on a common label in the intersection of their offered label sets (Nielsen et al., 2010). Symmetric sums are not the same as MCMP: they eliminate race conditions by making the choice collective. Still, they represent a multiparty generalisation of richer choice beyond ordinary branching, and the conductor-process encoding shows how such global choice can be compiled into ordinary branching while preserving typability and operational correspondence (Nielsen et al., 2010). This suggests a broader taxonomy in which MCMP occupies the branch of genuinely distributed, potentially conflicting mixed choices, while symmetric sums occupy the agreement-based branch.

Two later extensions indicate how MCMP serves as a foundation rather than an endpoint. The probabilistic framework introduces a “probabilistic mixed choice multiparty session pi-calculus” with a subtyping system “which allows one channel (the interface) to be substituted by several channels (the refinement)” and proves subject reduction, error-freedom and deadlock-freedom (Blechschmidt, 12 Sep 2025). The fairness literature, while not using the name MCMP, studies multiparty sessions whose progress depends on the choices of other participants and provides a type system ensuring fair termination under a fairness assumption, precisely for protocols with such interdependent choice behaviour (Ciccone et al., 2022).

The timed asynchronous binary theory TOAST is also relevant by analogy. It argues that mixed choice has long been barred from asynchronous models because it compromises the decidability assumptions of communicating finite-state machines, and then restores safe asynchronous mixed choice by regulating it with timing constraints (Pears et al., 2024). This suggests one plausible trajectory for future MCMP work: asynchronous mixed choice can be made tractable either by explicit commitment and stale-message handling, as in the 2026 multiparty framework, or by additional regulators such as time, fairness, or modular structure (Bocchi et al., 27 Feb 2026, Pears et al., 2024, Ciccone et al., 2022).

Taken together, these works present MCMP not as a single formalism but as a research programme. Its unifying aim is to admit multiparty protocols in which branch structure is not owned by one participant and does not reduce to pure branching or pure selection, while recovering enough global structure—coherence, liveness, awareness, fairness, balance, or refinement simulation—to make the resulting protocols type-safe and analyzable (Barbanera et al., 19 Aug 2025, Barbanera et al., 8 Apr 2026, Bocchi et al., 27 Feb 2026).