Home / Curated / Architectures / List-augmented recurrent neural network

List-augmented recurrent neural network

Architectures AIMech:ListAugmentedRecurrentNeuralNetwork DRAFT

Aliases: List RNN

A recurrent sequence architecture whose neural controller accesses an external list through a movable read/write head, learning continuous mixtures of left movement, right movement, insertion and optional no-operation updates.

Identity Notes & Concept Scope Compared all 442 pinned AIO and 574 native IDs, names, aliases and scoped synonyms; no normalized identity match. Ignored- and hidden-inclusive fixed-string searches of the recovery, older discovery and original Documents checkouts returned 30, 0, 0 lines. Matches were inspected as shared paper citations, Stack RNN comparison text, prior review evidence and discovery leads rather than an existing List RNN record. Semantic matching inspected 54 AIO and135 native definitions, with unchanged preceding rows verified and new matches read. Four full annotated AIO node/edge/generated contexts and four full native records were inspected. StackedRNNCellsLayer describes layer composition, not movable-head memory; MemoryAugmentedLLM requires large-language-model scope not inherent to List RNN; the pinned NeuralTuringMachineNetwork feedforward/spiking description does not establish this list topology. None supplies exact equivalence or a suitable broader mapping. RecurrentNeuralNetwork supplies the broader architecture mapping. Stack RNN defines top push/pop updates with a leading-entry readout, whereas List RNN defines bidirectional adjacent movement and head insertion. These are explicitly named separate source models. Ensemble stacking, beam-stack search and linear SCRN context have different identities. Soft versus one-hot action configurations stay within List RNN.

Claims & Grounded Evidence

6 Claims

Mechanism #head-centered-list-transition

The paper expresses memory in coordinates with a fixed head index H. A RIGHT action contributes the old value at i+1 and a LEFT action the old value at i-1 at every index i. INSERT contributes sigmoid(D*h_t) at H, the old value at i+1 for i<H, and the old value at i for i>H. Action weights blend these branches, and optional NO-OP adds its weight times the old value at i. Thus the fixed index is a coordinate convention for head movement, not a claim that the controller can never move along the logical list.

supports Source: Inferring Algorithmic Patterns with Stack-Augmented Recurrent Nets · Locator: Doubly-linked lists, displayed piecewise update and no-op paragraph, p. 5

Scope #insertion-topology

In the ideal one-hot INSERT branch, a new value occupies the head; the former head and the side with indices below H shift one position toward lower indices, while the side above H retains its values. LEFT and RIGHT permit adjacent movement in both directions. This differs from updating a stack through top push/pop operations and from treating a stored sequence as a fixed vector. These symbolic descriptions assume the required neighboring memory positions exist.

supports Source: Inferring Algorithmic Patterns with Stack-Augmented Recurrent Nets · Locator: Figure 1, p. 4; Doubly-linked lists and piecewise update, p. 5; curator indexing consequence

Scope #soft-state-is-not-discrete-history

Continuous action weights produce deterministic weighted memory updates, not sampling of one list operation. A nontrivial mixture need not encode one discrete head position and symbolic list. A single affine update can be read as a weighted mean of candidate results, but feeding its values through the nonlinear controller is not generally equivalent to averaging separate discrete trajectories. One-hot operations describe a limiting or selected configuration, not an automatic consequence of gradient-based training.

supports Source: Inferring Algorithmic Patterns with Stack-Augmented Recurrent Nets · Locator: Figure 1 and continuous stack controller, p. 4; list update and Optimization, p. 5; curator affine/nonlinear consequence

Mechanism #controller-and-memory-scope

Figure 1(b) couples list memory to a neural hidden state that produces action weights and inserted values; its caption includes optional NO-OP. The text says the recurrent hidden-state matrix R is omitted from the figure for clarity. The absence of a drawn hidden-to-hidden arrow therefore does not define List RNN as universally lacking hidden recurrence. The list controller and its local access topology define this record, rather than a particular trained model or task-specific capacity setting.

supports Source: Inferring Algorithmic Patterns with Stack-Augmented Recurrent Nets · Locator: Figure 1(b), p. 4; Doubly-linked lists, p. 5

Limitation #memory-and-implementation-boundaries

The source proposes the list as expandable external memory, but this does not supply infinite physical capacity, exact numerical operations or unrestricted generalization. A finite window with truncating boundaries can lose information during a shift; in that constructed setting, opposite movements need not undo one another. No list implementation, allocator, boundary convention or backward routine was inspected, so this record does not transfer the separately inspected Stack-RNN code behavior to List RNN.

supports Source: Inferring Algorithmic Patterns with Stack-Augmented Recurrent Nets · Locator: Doubly-linked lists and Optimization, p. 5; discussion/conclusion, p. 8; curator finite-window counterexample

Limitation #training-and-evaluation-scope

The source describes continuous Stack and List RNN training with stochastic gradient descent and truncated backpropagation through time. It reports difficulty learning more complex memory controllers and uses random restarts. List RNN appears in selected counting and memorization comparisons with specified capacities, supervision and sequence-length ranges. Those results do not establish universal algorithm learning, exact behavior at arbitrary lengths, or performance of an independently reproduced implementation. No numerical benchmark is curated here.

supports Source: Inferring Algorithmic Patterns with Stack-Augmented Recurrent Nets · Locator: Optimization and implementation details, p. 5; Table 2, p. 5; Figure 2, p. 6; Section 5.1, pp. 5–7

⚠️ Curated Limitations & Negative Results

Soft action mixtures need not preserve a discrete list or head interpretation; exact nonlinear propagation of an ensemble is not asserted.
The paper ideal memory and the curator finite-window example are distinguished. No particular list capacity, sentinel, allocator or implementation is certified.
Stack-RNN historical code is evidence about a different memory topology, not an inspected List RNN implementation.
All nine primary conference-paper pages were read; no supplement, List RNN source execution, training, backward audit or empirical replication was performed.
Published capacities, task settings, sequence lengths and tuning limit the evaluation claims. No universal algorithm-learning or unlimited exact-memory guarantee follows.

Artificial Intelligence Ontology (AIO) Grounding

Mapped AIO Classes

AIO Target Class	Mapping Relation	Mapping Rationale
`AIO:RecurrentNeuralNetwork` Recurrent Neural Network	broad_match	The architecture is a recurrent sequence network with a particular external-list access topology; the AIO target is the broader recurrent architecture category.

Evidence Sources & Primary Literature

1 Source

Source ID	Publication / Document	Type / Access	Reference & Checksum
`joulin2015`	Inferring Algorithmic Patterns with Stack-Augmented Recurrent Nets Armand Joulin, Tomas Mikolov	primary paper full_text	sha256:8c9bbbbf4e91…

Curation Provenance

Curator: Codex

Method: ai_assisted

Model: GPT-6

Created: 2026-10-13T05:41:54Z

Updated: 2026-10-13T05:41:54Z

Source File: data/curated/ListAugmentedRecurrentNeuralNetwork.yaml