Semantic-Neutral Drift Overview

Updated 14 July 2026

Semantic-neutral drift is a concept where system changes in structure or behavior occur while preserving intended semantic content, defining an invariance condition.
It is operationalized in fields like genetic programming using semantics-preserving mutations and in language dynamics via null models such as random copying.
Empirical studies show that leveraging semantic neutrality enhances system robustness, delineating it from semantically consequential drift observed in other contexts.

Searching arXiv for recent and relevant papers on semantic-neutral drift and closely related semantic drift formulations. Semantic-neutral drift denotes a class of changes in which a system moves in structure, representation, or behavior while semantic content is preserved or intended meaning is held fixed. The expression is explicit in genetic programming as “Semantic Neutral Drift (SND),” where semantics-preserving mutations alter active program structure without changing truth-table behavior or, under a semantics-only fitness function, fitness. In neighboring literatures, related notions appear as neutral evolution or random copying in language dynamics, semantically equivalent prompt rewordings that still induce response instability, and operator-theoretic cases in which embedding displacement occurs without corresponding change in local semantic conditionals. Taken together, these uses suggest that semantic-neutral drift is less a single settled doctrine than a family of neutrality conditions defined relative to a domain-specific semantic model (Atkinson et al., 2018, Blythe, 2011, Li et al., 11 Jun 2025, Russell, 21 Feb 2026).

1. Terminological range and core distinctions

Across current work, neutrality is defined differently in different research programs. In some settings it is an explicit invariance property; in others it is a null hypothesis; in still others it is only a proxy notion inferred from discordant drift signals.

Domain	Neutrality condition	Representative source
Genetic programming	Child has identical semantic meaning to parent	(Atkinson et al., 2018)
Language dynamics	Change arises from stochastic transmission without intrinsic difference between variants	(Blythe, 2011)
Prompt variance in LLMs	Prompts preserve semantic intent while outputs may shift	(Li et al., 11 Jun 2025)
Geometric semantic substrates	Large embedding motion with stable local conditionals and operator behavior	(Russell, 21 Feb 2026)

The strongest explicit definition comes from graph-based genetic programming: “Semantic Neutral Drift (SND) is the augmentation of a GP system with semantics-preserving mutations,” and a semantics-preserving mutation “guarantees that the semantic meaning of a child generated by that mutation is identical to that of its parent, for any choice of parents and a given semantic model” (Atkinson et al., 2018). In language-change work, by contrast, neutrality usually means that frequency change is produced by stochasticity in transmission or random copying rather than by intrinsic preference, prestige, or functional asymmetry (Blythe, 2011).

This plurality matters because many neighboring literatures study the opposite phenomenon. Multilingual semantic drift, factual drift in generation, emoji meaning change, label drift in translation, and semantic drift in incremental segmentation all analyze semantically consequential movement rather than a neutral baseline. Semantic-neutral drift is therefore best understood as a contrast concept: it asks when movement is structurally or statistically real yet semantically non-substantive.

2. Neutrality as a null model in language dynamics

In language dynamics, neutral drift functions primarily as a null model. “Neutral evolution: A null model for language dynamics” argues that rejection of a null model should have theoretical consequences, and proposes neutral evolution, genetic drift, or random copying as a viable null model for language change. Under this view, speakers reproduce variants by unbiased copying, and variants change in prevalence because of stochastic fluctuation rather than because one form is intrinsically better or socially favored (Blythe, 2011).

“Evolutionary forces in language change” sharpens this distinction by defining drift as change caused by stochasticity in transmission and selection as directional change caused by intrinsic differences between variants. The paper models neutral change with the Wright–Fisher diffusion and tests the composite null $H_0: s=0,\ N\ \text{arbitrary}$ with the Frequency Increment Test. It reports that some English changes do not reject drift, while others do. For past-tense regularization and irregularization, 6 of 36 verbs reject neutral drift at nominal $p<0.05$ ; 2 show selection toward the regular form and 4 toward the irregular form. In the rise of periphrastic do, drift is not rejected for affirmative interrogatives ( $p=0.23$ ) or negative interrogatives ( $p=0.77$ ), but is rejected for negative declaratives ( $p=0.01$ ) and negative imperatives ( $p=0.02$ ). The same paper argues that the strength of drift depends on a word’s frequency, providing a neutral explanation for why lower-frequency items are more labile (Ahern et al., 2016).

A spatial challenge to neutrality is provided by “Spatial evidence that language change is not neutral.” Modeling speakers as neurons in a spatially embedded Hopfield network, the paper contrasts the neutral Voter class, where interfaces are driven by noise, with a conformity-driven Ising class, where interfaces are driven by surface tension. Maximum-likelihood analysis and spatial autocorrelation over Survey of English Dialects maps indicate that the maps are more likely to belong to the conformity-driven Ising class: 82% of maps have Moran’s $I$ values more typical of non-neutral models, and about 80% are better fit by conformity-driven rather than neutral matching curves. This does not eliminate neutral drift as a null model, but it narrows its empirical scope in spatial dialect change (Burridge et al., 2020).

3. Explicit Semantic Neutral Drift in genetic programming

The most formalized and operational use of the term occurs in “Evolving Graphs with Semantic Neutral Drift” (Atkinson et al., 2018). In that paper, semantics is exact: “Two circuits are semantically equivalent if they describe identical truth tables.” SND is introduced to exploit equivalence laws so that mutations change active graph structure while preserving meaning. If $S$ is the search space, $M$ a set of semantics-preserving mutations, and $f$ a fitness function that depends only on semantics, then

$p<0.05$ 0

Neutrality is therefore guaranteed by construction, not discovered post hoc.

The implementation is graph-transformational. Individuals are directed graphs over the function set $p<0.05$ 1, encoded in P-GP2 and evolved with EGGP under a $p<0.05$ 2 strategy with $p<0.05$ 3, mutation rate $p<0.05$ 4, and fixed representation size of 100 function nodes. Semantics-preserving rewrites are derived from Boolean equivalences, including De Morgan laws,

$p<0.05$ 5

idempotence-style duplication and removal,

$p<0.05$ 6

double negation,

$p<0.05$ 7

and graph-level copy/collapse rules that duplicate or merge semantically identical subgraphs (Atkinson et al., 2018).

Empirically, the paper evaluates 12 digital circuit benchmark problems and reports that designed neutrality improves evolutionary performance on most of them. DMN and ID each improve all but one problem; the combined DMID rule set significantly outperforms baseline EGGP on 11 of 12 problems, while not improving the Comparator benchmark. The analysis argues that the benefit is not reducible to neutral growth or bloat: DMN and ID increase active size, but size alone does not explain gains, and the weaker performance of the generic copy/collapse rule set indicates that richer semantic rewrites matter more than mere neutral expansion. In this literature, semantic-neutral drift is thus a deliberately engineered traversal of neutral networks of equivalent programs rather than an observational regularity.

4. Diagnostics and proxy formalisms in NLP and representation spaces

In NLP, semantic-neutral drift is usually approached indirectly through decomposition or stability diagnostics rather than through exact equivalence laws. “Characterizing and Measuring Linguistic Dataset Drift” decomposes dataset drift into vocabulary, structural, and semantic components. Vocabulary drift is content-word frequency cross-entropy, structural drift is POS 5-gram cross-entropy, and semantic drift is defined as meaning change not captured by word frequencies, operationalized through lexical semantic change scores derived from contextualized embeddings of shared content tokens. This decomposition makes it possible to isolate cases with high vocabulary or structural drift but comparatively low semantic drift; such cases are the paper’s closest operational analogue of semantically neutral drift (Chang et al., 2023).

“Prompt-Based Semantic Shift” moves the problem to input paraphrase stability. It studies prompt variance, defined as behavioral shifts in LLM outputs triggered by surface-level paraphrases that preserve intent. For semantically equivalent prompts $p<0.05$ 8, with outputs $p<0.05$ 9 and $p=0.23$ 0, PBSS measures output drift by cosine distance between sentence embeddings: $p=0.23$ 1 The paper shows that benign paraphrases can induce consistent, model-specific output shifts, reframing semantic-neutrality as a quality-of-service stability question: intended meaning is fixed, but behavior is not (Li et al., 11 Jun 2025).

“Semantic Substrate Theory” offers the most systematic formal proxy. It defines a time-indexed substrate

$p=0.23$ 2

where $p=0.23$ 3 is a set of semantic objects, $p=0.23$ 4 an embedding-induced metric, and $p=0.23$ 5 a one-step Markov diffusion kernel. It then distinguishes translational drift, rewiring drift,

$p=0.23$ 6

coarse Ricci curvature, bridge mass, and recursive instability. The paper does not define semantic-neutral drift explicitly, but its closest formal proxy is precisely the case where translational displacement is nonzero while local neighborhood measures, induced diffusion, bridge mass, and recursive stability remain approximately unchanged. In that setting, representational motion occurs without meaningful change in local semantic conditionals (Russell, 21 Feb 2026).

5. Non-neutral semantic drift as the contrast case

Several recent literatures clarify semantic-neutral drift by showing what it is not. “Semantic Drift in Multilingual Representations” defines drift as stronger within-family than cross-family similarity of concept neighborhoods across languages, measured through representational similarity analysis and a family-conditioned semantic drift score. The paper explicitly states that this drift is not semantically neutral: high scores mean that semantic organization depends on language-family membership, while some observed drift may also be model-driven, for example when nearest-neighbor translation picks loanwords or semantically narrowed forms (Beinborn et al., 2019).

“Know When To Stop: A Study of Semantic Drift in Text Generation” studies a correct-then-incorrect trajectory in long-form generation. It defines semantic drift as generated text diverging from the subject matter designated by the prompt, with deterioration in relevance, coherence, or truthfulness, and introduces an SD score over ordered atomic facts. For LLaMa2-70B biographies, the mean SD score is 0.78 over all 500 examples and 0.80 after filtering out completely correct and completely incorrect samples, indicating strong left-correct/right-incorrect separation. This is an explicitly non-neutral drift: later text is systematically less faithful than earlier text (Spataru et al., 2024).

“How to Do Things without Words” extends the same contrast to emoji. It treats semantic drift as changes in contextual distribution over time and shows that graphical redesigns can trigger semantic change. The pistol emoji, for example, shifts from anger, death, failure, and gun-rights discourse toward games and playful online usage after platform redesigns, illustrating drift driven by visual realization rather than neutral geometry (Arviv et al., 2021).

“Semantic Label Drift in Cross-Cultural Translation” shows that even nominally neutral domains are not drift-free. It operationalizes drift as a mismatch between original labels and labels assigned to translated text, using Label Preservation Rate, KL divergence, and MCC. In culturally neutral Amazon product reviews, negative and positive classes are preserved at about 0.90–0.96, but the neutral class is much less stable, roughly 0.40–0.48, with low KL divergence of 0.046–0.072. The paper concludes that drift in neutral domains is comparatively small and concentrated in ambiguous middle categories, whereas culturally sensitive domains such as irony and mental-health severity exhibit larger, semantically consequential drift (Kabir et al., 29 Oct 2025).

A related suppression result appears in “Multitasking Inhibits Semantic Drift.” In latent language policies, semantic drift means that messages come to be used in ways inconsistent with their original natural-language meanings. The paper proves that multitask training eliminates drift in a simple family of signaling games under stated assumptions, and in MiniRTS shows that a shared executor improves interoperability and lowers off-diagonal message–action mismatch. Here, semantic neutrality is maintained not by exact invariance but by population pressure across tasks and communicative partners (Jacob et al., 2021).

Outside language alone, drift-robust systems often aim at semantics preservation rather than neutrality. In class-incremental semantic segmentation, “IPSeg” argues that semantic drift arises both from evolving semantic content and from separate optimization of different heads at different phases, which produces misaligned probability scales. Its response is image posterior guidance plus permanent-temporary semantics decoupling. On Pascal VOC 2012 with a Swin-B backbone, IPSeg reaches 78.6 mIoU on VOC 10-1 and 72.4 on VOC 2-2, improving over CoinSeg-M by 4.9 and 6.4 respectively. The method is best described as drift-robust rather than semantically neutral, because it compensates for cross-phase misalignment rather than defining an invariant semantic substrate (Yu et al., 7 Feb 2025).

In long-horizon tracking, “QueST” defines semantic drift as silently ceasing to track the same interaction-relevant entity while remaining visually confident. It replaces local propagation with persistent semantic queries and 3D physical grounding. On articulated PartNet-Mobility sequences, it reports APE $p=0.23$ 7, Drift@100 $p=0.23$ 8, and Identity Accuracy $p=0.23$ 9, versus TAP-Net’s APE $p=0.77$ 0 and identity accuracy $p=0.77$ 1. The framework is explicitly semantics-preserving rather than semantics-neutral: its persistent queries encode affordance-level priors such as handle, hinge, or rim (Anand et al., 10 May 2026).

“TokenDrift” extends drifting objectives to discrete diffusion LLMs by applying anti-symmetric drifting in a frozen semantic space. It constructs soft-token features, computes a drift vector in that fixed geometry, and trains the DDLM toward a stop-gradient target. The paper emphasizes that this is not strict semantics preservation: it guarantees no drift at feature-space equilibrium, but the objective intentionally moves representations in semantic feature space. In controlled continual-training experiments, it reduces Gen.-PPL at 4 NFEs by 89% on MDLM and 86% on DUO, showing that a frozen semantic space can stabilize a correction geometry without making the correction itself semantically neutral (Oba et al., 19 May 2026).

A recurring theoretical gap is that no single invariance principle yet unifies these uses. The operator-theoretic substrate paper explicitly notes the absence of a formal equivalence relation, invariance group, or theorem of semantic identity across representations, while PBSS and linguistic dataset-drift work show that fixed intent, stable meaning, lexical stability, and behavioral stability are distinct diagnostics rather than interchangeable ones. This suggests that semantic-neutral drift remains a comparative research concept: exact in some domains, null-model based in others, and only proximally formalized where semantics must be inferred from neighborhoods, operators, or downstream behavior (Russell, 21 Feb 2026, Li et al., 11 Jun 2025, Chang et al., 2023).