PseudoAct in Weak Action: Category, Monoids, LLMs

• PseudoAct is a multifaceted concept defined by structured weak actions that relax strict laws via coherent isomorphisms, correction data, or explicit pseudocode planning.
• In category and monoid theories, it introduces mechanisms—coherence cells and correction systems—to extend classical action frameworks to weak settings.
• In LLM agent systems, PseudoAct replaces reactive actions with a pseudocode plan and control-flow execution, yielding marked performance and stability improvements.

Searching arXiv for the requested topic and cited papers. {"query":"id:(Shulman, 2010) OR id:(Yihan et al., 27 Feb 2026) OR id:(Martins-Ferreira, 2021) OR id:(Dostál, 2017)","max_results":10} “PseudoAct” is not a single technical notion but a polysemous label used in several research literatures. In higher category theory, it is a convenient symbolic name for the 2-category of pseudoalgebras, $\PsAlg(T)$, associated to a 2-monad $T$; in algebra, it denotes the category of pseudo-actions of monoids, equivalent to the category of semi-biproducts of monoids; and in large-language-model agent systems, it names a framework that performs planning and action control through pseudocode synthesis rather than purely reactive action selection (Shulman, 2010, Martins-Ferreira, 2021, Yihan et al., 27 Feb 2026). The common thread is a shift from strict or fully local behavior toward explicitly structured weak action: algebraic laws may hold up to coherent isomorphism, extensions may require correction data beyond ordinary actions, and agent execution may be constrained by an explicit program-like plan rather than emergent stepwise replanning.

1. PseudoAct as a family of research meanings

The category-theoretic usage arises from the notation $\PsAlg(T)$, described as what one might symbolically call PseudoAct(T), for pseudo $T$-algebras of a strict 2-monad $T$ on a 2-category $\mathcal K$ (Shulman, 2010). In that setting, “PseudoAct” refers to weak algebraic structure: the action laws of the monad hold not strictly, but via invertible 2-cell constraints.

A second usage appears in monoid theory, where “PseudoAct” denotes the category whose objects are pseudo-actions of a monoid $B$ on a monoid $X$, with morphisms given by suitable pairs of monoid homomorphisms (Martins-Ferreira, 2021). Here the term is tied to extension theory: pseudo-actions generalize the classical action-plus-factor-system description of group extensions by adding a correction system that becomes essential in the noncommutative monoid case.

A third, unrelated usage is the agent framework “PseudoAct,” which introduces pseudocode synthesis as the planning representation for LLM agents (Yihan et al., 27 Feb 2026). In this literature, the term names a concrete architecture: an LLM first generates a global pseudocode plan, and a control-flow executor then enforces the plan during tool-using execution.

These meanings are historically independent. This suggests that “PseudoAct” functions less as a stable field-wide term than as a reusable label for structured, weak, or program-mediated action.

2. Higher-category-theoretic PseudoAct: pseudoalgebras of a 2-monad

In the 2-categorical setting, a strict 2-monad $(T,\mu,\eta)$ on $\mathcal K$ consists of a 2-functor $T$0 together with 2-natural transformations $T$1 and $T$2 satisfying the monad axioms strictly (Shulman, 2010). The corresponding strict algebras form $T$3, while the pseudoalgebras form $T$4, the structure symbolically identified with PseudoAct(T).

A pseudo $T$5-algebra is given by an object $T$6, a 1-cell $T$7, and invertible 2-cells

$T$8

subject to coherence axioms (Shulman, 2010). The associated pseudo $T$9-morphisms also carry invertible 2-cell structure. There is always a fully faithful inclusion

$\PsAlg(T)$0

and the strictification problem asks whether every object of PseudoAct(T) is equivalent to some strict algebra (Shulman, 2010).

The paper “Not every pseudoalgebra is equivalent to a strict one” exhibits a finitary 2-monad on a locally finitely presentable 2-category for which this strictification fails (Shulman, 2010). Its counterexample is built from higher category theory. The underlying 2-category is $\PsAlg(T)$1, and the composite 2-monad $\PsAlg(T)$2 has strict algebras precisely the strict 3-categories (Shulman, 2010). The pseudoalgebras of this monad are identified with iconic tricategories, a semi-strict form of tricategory characterized by identity 1-cell components in associativity and unit constraints (Shulman, 2010).

The decisive point is that not every Gray-category is equivalent to a strict 3-category, while every Gray-category, viewed as a tricategory, is iconic (Shulman, 2010). Hence not every object of $\PsAlg(T)$3 is equivalent to a strict $\PsAlg(T)$4-algebra. This shows that accessibility or rank of a 2-monad is not, by itself, sufficient for strictifiability (Shulman, 2010).

A concrete low-dimensional manifestation is obtained by restricting to doubly-degenerate objects. Doubly-degenerate pseudo $\PsAlg(T)$5-algebras correspond to braided monoidal categories, while doubly-degenerate strict $\PsAlg(T)$6-algebras correspond to strictly symmetric strict monoidal categories (Shulman, 2010). Therefore any non-symmetric braided monoidal category yields a pseudoalgebra not equivalent to any strict one (Shulman, 2010).

This usage of PseudoAct belongs to the broader Gray-categorical environment in which pseudoextensions, pseudoliftings, and pseudoadjunctions replace strict universal properties by coherent higher-dimensional ones. A related Gray-categorical result shows that giving a pseudoadjunction $\PsAlg(T)$7 with unit $\PsAlg(T)$8 is equivalent to giving an absolute left Kan pseudoextension of $\PsAlg(T)$9 along $T$0 (Dostál, 2017). This suggests that PseudoAct(T) should be understood as part of a larger weak-coherence program rather than as an isolated notation.

3. PseudoAct in monoid extension theory

In the monoid-theoretic literature, PseudoAct is the category of pseudo-actions, introduced to describe semi-biproducts of monoids (Martins-Ferreira, 2021). The central theorem states that the category of semi-biproducts of monoids is equivalent to the category of pseudo-actions (Martins-Ferreira, 2021).

A semi-biproduct of monoids is a diagram

$T$1

with additional maps $T$2 and $T$3, where $T$4 and $T$5 are monoid homomorphisms, $T$6 and $T$7 are zero-preserving maps, and the identities

$T$8

hold (Martins-Ferreira, 2021). This generalizes biproducts of commutative monoids by allowing some structure maps to be merely identity-preserving rather than multiplicative.

A pseudo-action of $T$9 on $T$0 consists of three ingredients (Martins-Ferreira, 2021):

• a correction system $T$1,
• a pre-action $T$2,
• a factor system $T$3.

These satisfy the normalization conditions

$T$4

together with a master associativity condition and the compatibility laws

$T$5

(Martins-Ferreira, 2021).

The correction system is the genuine novelty. In groups, correction systems are trivial, which is why the notion had not appeared in the classical action-plus-factor-set framework (Martins-Ferreira, 2021). In monoids, however, it measures the failure of a decomposition $T$6 to behave as if the middle monoid were simply $T$7. This is the mechanism that distinguishes pseudo-actions from ordinary actions with cocycles.

The synthetic monoid attached to a pseudo-action is the subset

$T$8

equipped with the operation

$T$9

which is associative precisely because of the master equation (Martins-Ferreira, 2021). This yields a semi-biproduct, and conversely every semi-biproduct yields a pseudo-action via

$\mathcal K$0

(Martins-Ferreira, 2021).

The group case is recovered as a degeneration. When $\mathcal K$1 is a group and $\mathcal K$2 is right cancellable, the correction system collapses to $\mathcal K$3, and pseudo-actions reduce to the familiar combination of action and factor system from group extension theory (Martins-Ferreira, 2021). By contrast, genuinely monoidal examples need not be Schreier extensions, and the middle object need not be in bijection with the full cartesian product $\mathcal K$4 (Martins-Ferreira, 2021).

4. PseudoAct as pseudocode-guided planning for LLM agents

In the LLM-agent literature, PseudoAct is a framework for flexible planning and action control through pseudocode synthesis (Yihan et al., 27 Feb 2026). It is motivated by limitations of reactive agents such as ReAct-style systems, where at each step $\mathcal K$5 an action is selected according to

$\mathcal K$6

with $\mathcal K$7 the query and $\mathcal K$8 the accumulated history (Yihan et al., 27 Feb 2026). The paper identifies several resulting difficulties for long-horizon tasks: redundant tool usage, unstable reasoning paths, high token consumption, and poor handling of conditional branches, loops, and multi-tool workflows (Yihan et al., 27 Feb 2026).

PseudoAct decomposes the problem into a planning policy and an execution policy (Yihan et al., 27 Feb 2026): $\mathcal K$9 The planner first synthesizes a pseudocode plan $B$0, and a deterministic Control-Flow Executor then traverses that plan while invoking an executor LLM for local action selection (Yihan et al., 27 Feb 2026).

The plan representation is

$B$1

where $B$2 is an ordered sequence of subtask steps, $B$3 is the workflow topology, $B$4 is a termination condition, and $B$5 is a hard iteration cap (Yihan et al., 27 Feb 2026). Each step has the form

$B$6

with operational context, natural-language objective, pseudocode logic, input variables, and output variables (Yihan et al., 27 Feb 2026).

The control-flow language is built from seven logic primitives (Yihan et al., 27 Feb 2026):

Primitive	Semantics
EXECUTE	Atomic single-action execution
IF–ELIF–ELSE	Conditional branching
FOR x IN C	Bounded iteration over a collection
WHILE $B$7	Convergence/condition-driven iteration
TRY–ON_FAILURE	Fault-tolerant execution with fallback
PARALLEL	Concurrent independent computations
DATA-FLOW	Explicit inter-step data passing

This representation externalizes control logic that would otherwise remain implicit in chain-of-thought-like reactive traces. The executor enforces dependencies, pauses steps whose required inputs are missing, maintains loop counters and branch state, and stops when either $B$8 is satisfied or the iteration count reaches $B$9 (Yihan et al., 27 Feb 2026). The paper emphasizes that the plan is generated once, while each step receives only compact workflow-level and local-step context rather than the full history (Yihan et al., 27 Feb 2026).

The framework is evaluated on FEVER, HotpotQA, and power-grid applications (Yihan et al., 27 Feb 2026). The reported main results are:

Method	FEVER Acc.	FEVER F1	HotpotQA Acc.
ReAct	60.78	62.63	46.40
DFSDT	67.31	64.16	73.21
PseudoAct	88.24	83.35	82.14

The paper reports a $X$0 absolute gain in FEVER accuracy over DFSDT and states that PseudoAct sets a new state-of-the-art on HotpotQA in its experimental setting (Yihan et al., 27 Feb 2026). It also attributes qualitative improvements to explicit loop handling, termination logic, and data-flow annotations, particularly in cases where reactive baselines perform redundant searches or enter uncontrolled branching regimes (Yihan et al., 27 Feb 2026).

5. Common structural motif across the three usages

Despite their disciplinary separation, the three uses of “PseudoAct” share a structural motif. In each case, a naive strict or local formulation is replaced by a richer action formalism that explicitly encodes coherence.

In the 2-monad setting, the move is from strict equations to invertible coherence 2-cells in pseudoalgebras (Shulman, 2010). In monoid extension theory, ordinary action data is supplemented by a factor system and, crucially, a correction system that captures non-Schreier behavior (Martins-Ferreira, 2021). In LLM agents, stepwise reactive action selection is replaced by an explicit pseudocode object encoding sequencing, branching, iteration, fallback, parallelism, and data flow (Yihan et al., 27 Feb 2026).

A plausible implication is that the repeated appeal of the label “PseudoAct” comes from this shared weak-structure intuition: action is preserved, but strictness is relaxed in favor of an explicit mediating structure. In category theory that mediating structure is coherence isomorphism; in monoid theory it is correction data; in agent systems it is a plan graph expressed as pseudocode.

There are, however, important differences. The category-theoretic and algebraic usages are about mathematical equivalence, decomposition, and coherence laws (Shulman, 2010, Martins-Ferreira, 2021). The agent usage is an algorithmic framework whose concerns are success rate, token efficiency, control flow, and execution stability (Yihan et al., 27 Feb 2026). Any attempt to treat these literatures as a single research program would therefore overstate the connection.

6. Conceptual significance and recurring questions

In higher category theory, the significance of PseudoAct lies in the failure of automatic strictification: PseudoAct(T) may be strictly larger than $X$1, even for finitary 2-monads on locally finitely presentable 2-categories (Shulman, 2010). This has direct implications for coherence theory, because pseudoalgebras are not merely a notational convenience but sometimes the only adequate setting for the structures under study (Shulman, 2010).

In monoid theory, the significance lies in extension classification. The equivalence between semi-biproducts and pseudo-actions provides a noncommutative generalization of the classical group-extension picture, while isolating the correction system as the feature absent in groups (Martins-Ferreira, 2021). This reframes certain monoid extensions as action-like data without forcing the middle object to be a full cartesian product.

In LLM-agent research, the significance lies in control. PseudoAct makes planning inspectable and executable as a structured object, with explicit termination and bounded iteration, rather than relying on purely reactive, history-conditioned inference (Yihan et al., 27 Feb 2026). The framework is positioned for long-horizon, branching, iterative, and safety-sensitive tasks, including power-grid applications (Yihan et al., 27 Feb 2026).

Across these literatures, the recurring question is whether weakly specified action can be reduced to a stricter presentation without loss. The answer is negative in the category-theoretic case (Shulman, 2010), conditionally negative in the monoid case because correction data can be essential (Martins-Ferreira, 2021), and practically negative in agent design because explicit pseudocode plans are introduced precisely to avoid the failure modes of unconstrained reactive execution (Yihan et al., 27 Feb 2026). This suggests that “PseudoAct,” wherever it appears, typically marks a setting in which weak or mediated action is not an accidental relaxation but the central object of study.