Repetition-Enabled Variants (CI-R)

Updated 10 July 2026

CI-R is a framework that treats repetition as an explicit design variable, altering system behavior by preserving key signals or amplifying challenges.
It underpins diverse applications including quantum decision games, grammar compression, language model prompting, neural architectures, fault-tolerant coding, and continual learning.
CI-R reorganizes complexity by shifting focus from raw measurements to repetition-sensitive statistics, enabling efficient design and robust performance.

Repetition-Enabled Variants (CI-R) denotes an umbrella class of constructions in which repetition is treated as an explicit design variable rather than a nuisance: repeated measurements in quantum games, repeated substrings in compressors and indexes, repeated prompt segments in LLM inference, repeated modules in neural architectures, repeated parity checks in coding experiments, repeated classes in continual-learning streams, and replicated processes in reversible calculi all alter the behavior of the system in structurally important ways. Across much of the literature, CI-R is not a standardized author term; several works introduce it only conceptually to organize repetition-aware variants in their own domains (Coladangelo et al., 1 Sep 2025, Furuya, 2019, Belazzougui et al., 2015, Leviathan et al., 17 Dec 2025, Kvalsund et al., 1 Jul 2025).

1. Conceptual scope and recurrent patterns

A common feature of CI-R formulations is that repetition changes the effective object being solved. In some settings repetition preserves a stable signal: the XOR decision variant of a monogamy-of-entanglement game keeps success probability at $\cos^2(\pi/8)$ for all $n$ ; prompt repetition improves non-reasoning LLM accuracy without increasing generated tokens; parameter-shared module repetition keeps parameter count at $P$ instead of $NP$ while repeating the same functional unit across depth or space (Coladangelo et al., 1 Sep 2025, Leviathan et al., 17 Dec 2025, Kvalsund et al., 1 Jul 2025). In other settings repetition amplifies hardness or exposes failure modes: Goldreich–Levin parity in the monogamy setting drives success toward $1/2$ for semi-classical adversaries, aggressive suffix caching in diffusion-based multimodal LLMs induces a “Repeat Curse,” and class-imbalance in repetitive continual-learning streams makes naive replay allocation suboptimal (Coladangelo et al., 1 Sep 2025, Zhao et al., 28 Jan 2026, Hemati et al., 2023).

A second recurring pattern is that CI-R often reorganizes complexity around a repetition-sensitive statistic. Grammar compressors move from digrams to maximal repeats and run-length rules; repetition-aware indexes replace worst-case text length by measures such as the number of BWT runs $r$ , the number of LZ77 factors $z$ , and the number of extensions of maximal repeats $e$ ; repetition-code experiments report cluster-size decay $\rho$ , bias ratio $\beta$ , and correlated-error fractions instead of only logical error rate (Furuya, 2019, Belazzougui et al., 2015, Belazzougui et al., 2016, Liepelt et al., 2023).

A plausible cross-domain synthesis is that CI-R systems fall into two broad operational regimes. One regime uses repetition to preserve or stabilize an informative statistic under scale-up; the other uses repetition either to amplify difficulty or to reveal pathologies that must be controlled. This distinction is stated explicitly for the XOR and Goldreich–Levin decision variants of monogamy-of-entanglement games and is echoed, in different formal languages, by prompt repetition versus repetition degeneration in LLMs, and by beneficial environmental repetition versus imbalance-sensitive replay in continual learning (Coladangelo et al., 1 Sep 2025, Leviathan et al., 17 Dec 2025, Mahaut et al., 1 Apr 2025, Hemati et al., 2023).

Domain	Repetition mechanism	Reported effect
Monogamy-of-entanglement games	Repeated measurements with decision post-processing	Constant XOR bias or exponentially decaying GL advantage
Grammar compression and indexes	Maximal repeats, runs, factor boundaries, CDAWG arcs	Smaller grammars and improved space–time tradeoffs
LLM inference	Prompt repetition or cache reuse	Accuracy gains in one case, degenerate repetition in another
Neural architectures	Repeated parameter-shared modules	Robustness, adaptability, generalization
Quantum coding benchmarks	Repeated parity checks and ARC schedules	Cross-platform diagnostics beyond logical error rate
Continual learning	Repeated class exposure in the stream	Knowledge accumulation and new replay policies
Reversible concurrency	Replication with memory and reversible contexts	New replication semantics and non-congruence results

2. Quantum-information formulation: decision variants of monogamy games

The most explicit CI-R formulation in the provided literature appears in the study of decision variants of the Tomamichel–Fehr–Kaniewski–Wehner monogamy-of-entanglement game (Coladangelo et al., 1 Sep 2025). In the original “search” game, Alice prepares a state on registers $n$ 0, with $n$ 1 consisting of $n$ 2 qubits. The Referee chooses a basis string $n$ 3 uniformly at random, measures each qubit in the computational basis if $n$ 4 and the Hadamard basis if $n$ 5, obtains $n$ 6, and sends $n$ 7 to Bob and Charlie. Bob and Charlie win if both output the full string $n$ 8. The optimal success probability is

$n$ 9

a perfect parallel repetition law (Coladangelo et al., 1 Sep 2025).

The XOR decision variant replaces full-string reconstruction by parity prediction. Bob and Charlie now output a single bit, required to equal

$P$ 0

The striking result is that the optimal success probability does not decay with $P$ 1:

$P$ 2

for every $P$ 3, and this value is optimal even against fully quantum adversaries (Coladangelo et al., 1 Sep 2025). The explicit strategy uses no entanglement. Let

$P$ 4

and for $P$ 5 define

$P$ 6

With $P$ 7 the bitwise complement of $P$ 8, the states

$P$ 9

form two orthonormal bases, and measurement in any $NP$ 0 basis yields a parity bit that is a deterministic function of $NP$ 1 with success probability $NP$ 2 (Coladangelo et al., 1 Sep 2025). For $NP$ 3, for example,

$NP$ 4

The structural reason is that the parity observable

$NP$ 5

admits a two-dimensional invariant subspace

$NP$ 6

on which $NP$ 7 acts as either single-qubit $NP$ 8 or $NP$ 9 up to sign and global phase. Under the induced isomorphism $1/2$0, the $1/2$1-qubit parity task collapses to the one-qubit optimum. This is the canonical example of a repetition-enabled variant that preserves adversarial advantage under parallel repetition (Coladangelo et al., 1 Sep 2025).

The Goldreich–Levin variant changes the decision predicate by introducing an additional random mask $1/2$2 and asking Bob and Charlie to predict $1/2$3 over $1/2$4. Here the reported behavior is the opposite. For semi-classical adversaries, meaning $1/2$5 is unentangled from $1/2$6 and the answers are deterministic functions of $1/2$7, the winning probability satisfies

$1/2$8

The proof decomposes the average winning projector into a diagonal term $1/2$9, bounded by $r$ 0, and an off-diagonal interference term $r$ 1, bounded by Parseval-based Fourier analysis (Coladangelo et al., 1 Sep 2025). In this case the random mask destroys the fixed two-dimensional invariant-subspace trick used by XOR: the relevant observable depends on both $r$ 2 and $r$ 3, and there is no single $r$ 4D structure invariant under all $r$ 5.

This contrast is central to the CI-R idea in the quantum setting. XOR repetition is repetition-enabled for preserving constant bias, whereas Goldreich–Levin repetition is repetition-enabled for amplifying hardness. The paper further formulates a conjecture equivalent to exponential decay against fully quantum adversaries in the GL game; if proved, it would imply an information-theoretically secure “unclonable bit” (Coladangelo et al., 1 Sep 2025).

3. Repetition as structure in compression and indexing

In grammar compression, CI-R appears as a move from local replacement heuristics to repetition-aware replacement units. “Practical Repetition-Aware Grammar Compression” studies MR-RePair and RL-MR-RePair, two variants that explicitly exploit repetitiveness beyond standard RePair (Furuya, 2019). The basic grammar size model assigns cost $r$ 6 to terminal rules $r$ 7, cost $r$ 8 to concatenation rules $r$ 9, and cost $z$ 0 to run-length rules $z$ 1, with total size

$z$ 2

MR-RePair replaces the most frequent conditional maximal repeat rather than the most frequent digram, while RL-MR-RePair extends this with run-length rules when the most frequent maximal repeat is $z$ 3 (Furuya, 2019).

The asymptotic guarantees are explicit. MR-RePair runs in $z$ 4 expected time and uses

$z$ 5

words of space, while RL-MR-RePair retains $z$ 6 expected time and uses

$z$ 7

words (Furuya, 2019). The practical encoding contribution is equally central: general CFG and RLCFG outputs are serialized by converting run-length rules to the symbol sequence “0 $z$ 8 $z$ 9” and then encoded with POPPT plus Increasing Bit Length Encoding or Packed Gamma Encoding, with $e$ 0 often outperforming $e$ 1 (Furuya, 2019). Empirically, on repetitive datasets such as dna.001.1, RL-MR-RePair constructs the smallest grammar among the compared methods, and POPPT+PGE yields the best or near-best compression ratios across most real datasets (Furuya, 2019).

Repetition-aware indexing makes a parallel shift from text length to repetition statistics. “Composite repetition-aware data structures” combines the run-length encoded BWT with LZ77-based components and with the compact directed acyclic word graph (Belazzougui et al., 2015). The relevant measures are the number of BWT runs $e$ 2, the number of LZ77 factors $e$ 3, and the number of CDAWG extensions of maximal repeats $e$ 4. For highly repetitive strings these quantities all grow sublinearly in $e$ 5, and the composite designs obtain bounds that depend on several of them simultaneously. The RLBWT+LZ77 composite occupies

$e$ 6

words, supports counting in

$e$ 7

and reports all primary and secondary occurrences in

$e$ 8

time. The RLBWT+CDAWG composite occupies $e$ 9 words and answers exact matching in

$\rho$ 0

time (Belazzougui et al., 2015).

“Practical combinations of repetition-aware data structures” refines the same theme with implementable variants that mix RLBWT with LZ77 boundaries, bidirectional search, unidirectional extraction, sparsified parses, and CDAWG traversal (Belazzougui et al., 2016). In the full RLBWT+LZ configuration, the space bound is

$\rho$ 1

bits, with counting in $\rho$ 2 and locating in $\rho$ 3 time (Belazzougui et al., 2016). The CDAWG+RLBWT variant uses $\rho$ 4 words and gives $\rho$ 5 counting and $\rho$ 6 locating (Belazzougui et al., 2016). In both the theoretical and practical indexing literatures, CI-R therefore denotes structures whose efficiency is tied to repetition-aware measures rather than to raw text length.

4. Language-model regimes: deliberate repetition, pathological repetition, and internal mechanisms

The LLM literature represented here treats repetition as both an intervention and a failure mode. “Prompt Repetition Improves Non-Reasoning LLMs” defines a family of prompt-construction strategies in which a base prompt $\rho$ 7 is repeated $\rho$ 8 times, optionally with separators, while response length and decoding behavior remain unchanged (Leviathan et al., 17 Dec 2025). The core empirical claim is that, without reasoning, prompt repetition wins 47 out of 70 model–dataset configurations with 0 losses, and in a reasoning setting the effects are neutral to slightly positive, with 5 wins, 1 loss, and 22 ties (Leviathan et al., 17 Dec 2025). The reported extreme example is Gemini 2.0 Flash Lite on NameIndex, where accuracy improves from $\rho$ 9 to $\beta$ 0 under $\beta$ 1, an absolute increase of $\beta$ 2 percentage points and a relative improvement of approximately $\beta$ 3 (Leviathan et al., 17 Dec 2025). Padding controls do not help, and the measurements show no increase in generated tokens and similar end-to-end latency when reasoning is disabled (Leviathan et al., 17 Dec 2025).

The explanatory account is mechanistic rather than purely empirical. Because causal LLMs do not allow earlier tokens to attend to later ones, a second copy of the prompt exposes “future” information inside the same input sequence. Repeating an instruction-bearing embedding $\beta$ 4 $\beta$ 5 times yields an attention amplification term of the form

$\beta$ 6

which concentrates attention on instructions and constraints (Leviathan et al., 17 Dec 2025). This is a CI-R regime in which repetition is explicitly injected at the input level to improve accuracy.

A different body of work studies undesired repetition. “Repetitions are not all alike” shows that repetition in autoregressive LMs is not unitary: naturally emergent repetition after human-written text and ICL-induced repetition have different confidence profiles, different head-level circuitry, and different responses to top- $\beta$ 7 perturbations (Mahaut et al., 1 Apr 2025). In the Natural condition, the model enters a cycle with higher confidence and more diffuse pro-cycle head support. In the ICL condition, confidence is initially lower and then rises sharply around cycle 2, consistent with a copy loop becoming active once two matching context spans are available (Mahaut et al., 1 Apr 2025). The implicated heads differ by condition: L4H4 is strongly pro-cycle in ICL, L8H12 is Natural-specific, and L10H2 is ICL-specific (Mahaut et al., 1 Apr 2025).

“Understanding and Controlling Repetition Neurons and Induction Heads in In-Context Learning” shifts the locus from attention heads to feed-forward units (Doan et al., 10 Jul 2025). Repetition neurons are identified by the activation-difference score

$\beta$ 8

computed around repetition onset. The main finding is strongly layer-dependent: ablating late-layer repetition neurons causes large drops in Pattern recall, up to about $\beta$ 9 in 10-shot settings, while middle-layer ablation reduces repetitive generation with much smaller ICL damage (Doan et al., 10 Jul 2025). Induction-head ablation is more brittle: removing the top $n$ 00 by prefix-matching score causes Pattern recall drops ranging from $n$ 01 to $n$ 02 across the reported tasks, and joint ablation of late repetition neurons with top induction heads produces at least $n$ 03 degradation in most cases (Doan et al., 10 Jul 2025). The paper therefore identifies a CI-R control regime in which only some repetitions should be suppressed, and only at selected depths.

In diffusion-based multimodal LLMs, repetition is linked to cache design. “Context Tokens are Anchors” defines the “Repeat Curse” as the excessive generation of repeated tokens under cache-based acceleration (Zhao et al., 28 Jan 2026). The paper measures repetition with ARR, SRR, MRL, ARL, and 95pRL, and reports that on LLaDA-V with dLLM-Cache, long generations of 512 tokens increase ARR from $n$ 04 to $n$ 05 and SRR from $n$ 06 to $n$ 07 (Zhao et al., 28 Jan 2026). The information-flow account identifies context tokens as semantic anchors whose entropy should converge in deeper layers:

$n$ 08

Under suffix or output-token caching, anchor structure is disrupted and deep-layer entropy fails to converge (Zhao et al., 28 Jan 2026). The proposed mitigation, CoTA, combines Context Tokens Attention Enhancement with Context Tokens Entropy-Guided Voting:

$n$ 09

$n$ 10

On COCO captioning with cache, CoTA reduces long-generation ARR from $n$ 11 to $n$ 12 and SRR from $n$ 13 to $n$ 14 (Zhao et al., 28 Jan 2026).

Taken together, these works imply that CI-R in LLMs is not a single phenomenon. Input repetition can improve non-reasoning accuracy, internal repetition circuits can support ICL while also driving degenerate outputs, and cache-induced repetition can arise from a failure of information-flow regularities rather than from prompt semantics alone (Leviathan et al., 17 Dec 2025, Mahaut et al., 1 Apr 2025, Doan et al., 10 Jul 2025, Zhao et al., 28 Jan 2026).

5. Architectural, coding-theoretic, and continual-learning variants

In neural architecture design, CI-R is formulated as module repetition. “The Generalist Brain Module” distinguishes architectural repetition from parameter-shared module repetition and argues that the latter is the more direct carrier of collective-intelligence properties (Kvalsund et al., 1 Jul 2025). The formal core is a repeated shared module

$n$ 15

or in residual form

$n$ 16

as well as spatial or graph-based repetition with shared parameters (Kvalsund et al., 1 Jul 2025). Parameter count stays at

$n$ 17

instead of

$n$ 18

and the paper associates this design with robustness, adaptability, and generalization, together with an emergent “generalist module” that can assume different roles inside an ensemble (Kvalsund et al., 1 Jul 2025). This is an architectural CI-R in which repetition is not a decoding trick but the substrate of the model itself.

In quantum fault-tolerance benchmarking, CI-R appears as a family of repetition-code experiments designed for cross-platform comparability. “Enhanced repetition codes for the cross-platform comparison of progress towards fault-tolerance” studies repetition codes, Alternating Repetition Codes, interleaved $n$ 19 sequences, generalized link graphs, space-time decoding, and syndrome-derived diagnostics (Liepelt et al., 2023). The basic stabilizers are

$n$ 20

and the paper advocates metrics beyond logical error rate, including cluster-size decay

$n$ 21

On ibm_sherbrooke, the implementation spans 125 qubits with 54 data qubits, 71 ancillas, and $n$ 22 rounds; no logical failures were observed over $n$ 23 shots in any of the four ARC instances (Liepelt et al., 2023). The fitted decay parameter is $n$ 24 with resets, $n$ 25 without resets, and $n$ 26 for $n$ 27 (Liepelt et al., 2023). In this setting, repetition is the basis of a benchmarking protocol rather than a source of compression or a model-input intervention.

A separate coding-theoretic formulation appears in “Hybrid Non-Binary Repeated Polar Codes” (Abbasi et al., 2021). The outer code is a hybrid polar construction with a binary stage and a symbol-level stage over $n$ 28, while the inner code is a non-binary multiplicative repetition code of rate $n$ 29. The total rate is

$n$ 30

Decoding uses CRC-aided SCL, with symbol-level repetition combining and non-binary max-log recursions (Abbasi et al., 2021). The empirical message is that repetition can be made coding-theoretically stronger by moving from binary repetition to multiplicative non-binary repetition. For $n$ 31, $n$ 32, and $n$ 33, hybrid GF(16) with list size $n$ 34 nearly matches hybrid GF(4) with $n$ 35 and polar-repetition with $n$ 36; for $n$ 37, hybrid GF(16) yields the best performance among the compared repetition-enabled schemes at comparable complexity (Abbasi et al., 2021).

Continual learning introduces yet another variant. “Class-Incremental Learning with Repetition” defines Class-Incremental with Repetition (CIR), a family of streams in which repetition is a property of the environment rather than an agent-controlled replay mechanism (Hemati et al., 2023). CIR interpolates between CI and DI and is generated either by a slot-based generator, which revisits classes without repeating instances, or by a sampling-based generator based on first-occurrence distributions and per-class repetition probabilities $n$ 38 (Hemati et al., 2023). The paper emphasizes metrics such as Average Test Accuracy, Seen Classes Accuracy, and Missing Classes Accuracy, and finds that natural environmental repetition leads to knowledge accumulation even for naive finetuning (Hemati et al., 2023). Its main algorithmic contribution is ER-FA, a replay strategy that allocates memory inversely to observed class frequency:

$n$ 39

In imbalanced streams this improves retention of infrequent classes; on CIFAR-100 with a bi-modal stream and Fraction $n$ 40, ER-FA reaches ACA $n$ 41 versus $n$ 42 for ER-CB and MCA $n$ 43 versus $n$ 44 (Hemati et al., 2023).

6. Replication in reversible concurrency and a cross-domain synthesis

The most literal sense of “repetition-enabled variants” appears in the reversible-process-algebra literature. “Enabling Replications and Contexts in Reversible Concurrent Calculus” develops a reversible CCS-style calculus with explicit identifier structures, seeds, splitters, and memories, precisely to make replication and contexts workable in a reversible setting (Aubert et al., 2021). Local identifier generation is handled by patterns $n$ 45 and a generator $n$ 46, with pairing $n$ 47 used for synchronization identifiers. Freshness is local rather than global: an action consumes $n$ 48 and advances the seed to $n$ 49 (Aubert et al., 2021).

The calculus introduces replication $n$ 50 at the identified-process level with two forward rules. In schematic form, one rule spawns a continuation of $n$ 51 together with one active copy of $n$ 52, and the other allows synchronization between two replicated copies through a paired identifier $n$ 53 (Aubert et al., 2021). In the reversible extension, the paper distinguishes four forward replication variants for processes with memory, differing in how base memory, memory deltas, and replicated branches are tagged and distributed:

$n$ 54

The reported conclusion is that unmarked duplication of memory events breaks causal consistency; some form of memory marking is required so that backward rules can recognize replicated lineage (Aubert et al., 2021).

The same work also generalizes contexts to “reversible contexts” that can themselves carry a past:

$n$ 55

This enables “hot-plugging” of a process into an ongoing context with memory, a capability that standard reversible bisimulations do not handle compositionally (Aubert et al., 2021). The paper proves the expected reversible meta-theory—Loop Lemma, Square Property, backward concurrency, well-foundedness, and causal consistency—but also observes that none of the reversible bisimulations defined thus far are congruences under the proposed reversible contexts (Aubert et al., 2021).

A cross-domain synthesis suggested by the collected literature is that CI-R is best understood not as a single technique but as a structural principle. Repetition can preserve a single-qubit bias across arbitrarily many quantum measurements, encode repeated substrings more aggressively than pairwise compressors, reduce prompt-order sensitivity in non-reasoning LLMs, regularize or destabilize generative information flow, support robustness through repeated shared modules, provide device-agnostic fault-tolerance diagnostics, compensate low-SNR channels by inner repetition, turn environmental recurrence into a first-class continual-learning variable, or force a redefinition of causality and context in reversible process algebra (Coladangelo et al., 1 Sep 2025, Furuya, 2019, Leviathan et al., 17 Dec 2025, Zhao et al., 28 Jan 2026, Kvalsund et al., 1 Jul 2025, Liepelt et al., 2023, Abbasi et al., 2021, Hemati et al., 2023, Aubert et al., 2021). What unifies these otherwise heterogeneous constructions is that repetition changes the effective state space, the relevant statistics, or the admissible transformations of the system.