Identity Anchors in Cross-Domain Systems

• Identity Anchors are persistent reference structures that maintain recognizable identity across transformations, drift, and ambiguity.
• Research shows these anchors stabilize model performance by separating rigid class tokens from adaptable features, achieving metrics like 70.2%-94.5% mF1 in segmentation and HAR tasks.
• Advanced implementations use temporal, geometric, and cryptographic anchoring techniques to ensure continuity and secure identity preservation in diverse applications.

Searching arXiv for papers on “identity anchors” and closely related anchor-based identity-preservation mechanisms across domains. Identity anchors are persistent reference structures used to preserve recognizable identity under transformation, drift, ambiguity, or partial failure. Across recent arXiv literature, the term denotes learnable class tokens that stabilize weakly supervised incremental segmentation, explicit handcrafted feature groups retained inside human-activity-recognition models, exogenous attention projections that preserve token identity in deep Transformers, future keyframes and anchor frames that maintain character consistency in long-form video generation, semantic tokens that bind 3D parts to fixed slots, cryptographic roots that maintain continuity across keys and contexts, and $p$-local invariants attached to irreducible characters (Wang et al., 2 Jun 2026, Yao et al., 28 Apr 2026, Su, 13 Jan 2026, Seo et al., 27 Oct 2025, Yang et al., 31 Mar 2026, Hao et al., 10 Jun 2026, Smith, 2019, Kessar et al., 2015). This suggests that the concept is best treated as a family resemblance across domains: an anchor provides a stable coordinate, provenance trace, or canonical local support, while the surrounding system absorbs variation, performs adaptation, or accumulates evidence.

1. Cross-domain concept and taxonomy

A recurring pattern in the literature is that an anchor is not merely an auxiliary feature. It is the object that remains available when ordinary representations drift, compress, permute, or rotate. In some papers it is learned and persistent; in others it is mobile, cryptographic, or algebraic. The common role is identity preservation under nontrivial state change.

Context	Anchor object	Identity preserved
WILSS (Wang et al., 2 Jun 2026)	Learnable class tokens $\mathbf{A}^t$	Class semantics across incremental steps
HAR (Yao et al., 28 Apr 2026)	Handcrafted TSF groups as feature anchors	Explicit motion statistics
Transformers (Su, 13 Jan 2026)	Exogenous $Q/K/V/G$ projections	Token identity across depth
Animation and video (Seo et al., 27 Oct 2025, Yang et al., 31 Mar 2026)	Future keyframes; global/viewpoint/expression anchor frames	Character identity over long generation
Part-aware 3D generation (Hao et al., 10 Jun 2026)	Semantic identity tokens tied to slots	Part identity across layout and geometry
Localization and alignment (Shokry et al., 2020, Cannistraci et al., 2023)	Human encounters; parallel anchor pairs	Pose correction or shared relative coordinates
Cryptographic systems (Hardjono et al., 2019, Smith, 2019, Wang, 25 Nov 2025)	Anonymous membership credentials, self-certifying identifiers, root entropy $R$	Membership, control provenance, continuity
AI agents and character theory (Menon, 2 Mar 2026, Kessar et al., 2015)	Persistent memory structures; defect groups of $OGe_\chi$	Behavioral continuity or $p$-local character invariant

This taxonomy also corrects a common misconception: anchors are not always static templates. DynamicSLAM’s human anchors are temporary encounter-based observations; Lookahead Anchoring uses keyframes located beyond the current generation window; and soul.py treats identity as distributed across several persistent structures rather than concentrated in one store (Shokry et al., 2020, Seo et al., 27 Oct 2025, Menon, 2 Mar 2026). Another misconception is that anchors are only regularizers. In SASA, the anchor-derived token is the segmentation head itself, and in KERI the anchor is the root of trust for control establishment rather than a secondary constraint (Wang et al., 2 Jun 2026, Smith, 2019).

2. Representation-space identity anchors

In representation learning, identity anchors are used to decouple a stable semantic basis from task-driven adaptation. The clearest instance is SASA for weakly supervised incremental semantic segmentation. SASA introduces semantic anchors as learnable class tokens

$$\mathbf{A}^{t} \in \mathbb{R}^{|\mathbf{C}^{0:t}| \times D},$$

with each row $\mathbf{A}_k^t$ acting as a rigid class-level reference. Instance-specific variation is handled by a residual branch, and the final token is

$\mathbf{Z}^{t} = \mathbf{A}^{t} + \mathbf{R}^{t}.$

The paper explicitly contrasts anchors with conventional prototypes: prototypes are recomputed from noisy, drifting features, whereas anchors are persistent class-level coordinates. SASA further constrains residual magnitude, token-anchor separation, and old-anchor distillation over time to reduce feature drift and semantic corruption under weak supervision (Wang et al., 2 Jun 2026).

TCNet adopts an analogous strategy for time-series sensor-based human activity recognition, but its anchors are handcrafted time-series features rather than class tokens. Differentiable TSF families remain visible inside the model as explicit intermediate representations, and raw-signal context predicts bounded scale, bias, and gating parameters to modulate them directly in feature space. The correction is explicitly interpolative rather than substitutive: $$Z^{(m)}_k = (1-\lambda^{(m)}_k)\odot Z^{(m)}_{\mathrm{raw}} + \lambda^{(m)}_k \odot \widetilde{Z}^{(m)}_k.$$ The loss

$\mathbf{A}^t$0

penalizes excessive deviation from the raw anchors and optionally enforces temporal smoothness. Empirically, the paper argues that the gains come primarily from anchor guidance rather than simple branch fusion, and reports 70.2% mF1 on USC-HAD, 85.1% mF1 on Daphnet, 93.9% mF1 on MHealth, and 94.5% mF1 on PAMAP2 (Yao et al., 28 Apr 2026).

ExoFormer moves the same separation principle into Transformer attention. Instead of forcing the first layer to be both a stable reference and a normal computational block, it learns exogenous anchor projections outside the sequential stack: $\mathbf{A}^t$1 Each layer mixes its current projections with the normalized anchor source: $\mathbf{A}^t$2 The paper interprets the resulting behavior through an Offloading Hypothesis: external anchors preserve essential token identity, allowing the sequential layers to specialize in computational refinement. The dynamic variant reports a 2.13-point downstream accuracy gain over baseline, matches baseline validation loss with 1.84x fewer tokens, and reduces attention sink by 2x relative to standard Gated Attention (Su, 13 Jan 2026).

ISAP-3D applies identity anchoring to structured generation. Its central claim is that part-aware 3D generation is unstable because identity-slot correspondence is not identifiable under weak supervision. The proposed remedy is to anchor each slot with semantic identity tokens $\mathbf{A}^t$3, then maintain one-to-one identity-slot alignment through layout prediction and geometry synthesis. The slot update

$\mathbf{A}^t$4

makes identity injection architectural rather than post hoc. Reported results include CD 0.1410, [email protected] 0.8249, Part-IoU 0.0330, and NMI 0.6157, with NMI serving as the key measure of alignment stability (Hao et al., 10 Jun 2026).

Taken together, these papers support a shared technical recipe: preserve an explicit identity-bearing object, restrict adaptation to local or bounded corrections, and use either loss terms or one-to-one architectural constraints to prevent overwriting.

3. Temporal and generative anchoring

In autoregressive generation, identity drift is a temporal accumulation problem rather than a static representational one. Lookahead Anchoring addresses this in audio-driven human animation by moving keyframes out of the current generation window and treating them as future temporal anchors: $\mathbf{A}^t$5 The keyframe is therefore not a hard boundary inside the current segment but a future target beyond it. This changes its function from exact interpolation constraint to directional beacon. The method also introduces self-keyframing by setting the lookahead target to the reference image itself, eliminating the need for a separate keyframe generation model. The paper reports that smaller $\mathbf{A}^t$6 strengthens identity adherence and larger $\mathbf{A}^t$7 allows greater motion freedom, with lip synchronization peaking around 12 frames. Across Hallo3, HunyuanVideo-Avatar, and OmniAvatar, the method improves lip synchronization, identity preservation, and visual quality (Seo et al., 27 Oct 2025).

Gloria develops a broader anchor memory for long-duration character video generation. It represents identity through three content-anchor types: a global anchor $\mathbf{A}^t$8 for overall scene and subject-background correlation, viewpoint anchors $\mathbf{A}^t$9 for up to four views, and expression anchors $Q/K/V/G$0 for up to eight expression categories. These anchors are encoded by the same 3D VAE as the target video and injected directly into the DiT self-attention stream. To prevent copy-paste behavior, Superset Content Anchoring samples anchors from the full source long video rather than only from the current 5-second training clip. To disambiguate multiple anchors, RoPE as Weak Condition assigns distinct temporal offsets: $Q/K/V/G$1 The system is reported to generate character videos exceeding 10 minutes and to improve long-term identity, multi-view appearance, and expressive consistency (Yang et al., 31 Mar 2026).

A notable controversy in this area concerns rigidity. Traditional keyframe methods can enforce exact poses and timestamps, which restricts natural motion dynamics. Both Lookahead Anchoring and Gloria argue for softer anchor usage: future beacons rather than hard boundaries in the former, and weak positional disambiguation rather than hard semantic selection in the latter (Seo et al., 27 Oct 2025, Yang et al., 31 Mar 2026). This suggests that temporal identity anchoring is strongest when persistence is directional and structured, not when it is simply immutable.

4. Anchors as observations, correspondences, and geometric references

A second major family of anchor methods treats anchors as external observations or cross-domain correspondences rather than internal identity tokens. In DynamicSLAM, anchors are identifiable reference points used to reset unbounded dead-reckoning error. The system generalizes anchors from static building landmarks to human mobile anchors: encounters between users provide relative-distance observations that are fused in a FastSLAM-style probabilistic framework,

$Q/K/V/G$2

The paper proves that human-anchor encounters reduce accumulated error in expectation, while known building anchors guarantee convergence even for $Q/K/V/G$3 particle if at least one building anchor is known in advance. Empirically, DynamicSLAM reports a median localization error of 1.1 m, 55% improvement over SemanticSLAM, and 29% reduction in worst-case error (Shokry et al., 2020).

In relative representation learning, anchors define a shared coordinate system across domains. “Bootstrapping Parallel Anchors for Relative Representations” formalizes parallel anchors as semantically corresponding pairs $Q/K/V/G$4, with relative representations given by cosine similarities to the anchor sets. Its Anchor Optimization procedure begins from a tiny seed $Q/K/V/G$5 of known correspondences and alternates between Sinkhorn-based soft matching and optimization of unknown target-side anchors. The reported setup expands 15 seed pairs toward 300 total anchors, enabling latent space communication and zero-shot model stitching in word embeddings, vision, and cross-lingual RoBERTa-based rating prediction (Cannistraci et al., 2023).

Social network alignment introduces yet another variant: pseudo anchors implanted around real anchor users to reshape local embedding geometry. The paper argues that structural-proximity objectives create “overly-close” neighborhoods that obscure the correct match. Pseudo anchors are therefore added to the anchor’s context, and PSML meta-learns their update directions so that they widen the local neighborhood without drifting into fuzzy regions. Gains are especially large in low-anchor regimes; for example, on Twitter–Foursquare at 3% training anchors, IONE rises from 5.24 to 9.67 and SNNA from 2.65 to 7.54 under PSML (Yan et al., 2021).

These uses differ from representation anchors in one important respect. Here the anchor is an observation, matched pair, or injected structural reference that constrains geometry from outside the current latent state. This suggests a broader distinction between intrinsic anchors, which persist inside the model, and relational anchors, which define how states in different times, users, or domains are made commensurate.

5. Cryptographic and infrastructural identity anchoring

In cryptographic systems, an identity anchor is not primarily a semantic coordinate but a root of trust or continuity. ChainAnchor addresses identity and access control in shared permissioned blockchains by separating enrollment-time real identity from later anonymous yet verifiable blockchain participation. Membership is established through EPID-based zero-knowledge proofs, after which the user registers a transaction public key $Q/K/V/G$6 with the Permissions Verifier. Consensus nodes enforce access control by a simple lookup against a read-only permissions database of anonymous members’ public keys. The paper emphasizes unlinkability across multiple transaction keys and optional disclosure of a challenged key without linking unrelated activity (Hardjono et al., 2019).

KERI pushes the anchor deeper into identifier formation itself. A self-certifying identifier is cryptographically bound at issuance to a signing key-pair, and control is maintained through an append-only chained key-event log. Identity continuity is therefore established by replaying inception and rotation events rather than trusting a central registry. The architecture distinguishes direct one-to-one trust from indirect one-to-any trust with witnesses and Key Event Receipt Logs, and key pre-rotation commits to the next key set in advance. The paper’s stated goal is ambient verifiability: a control history that is independently checkable anywhere and at any time (Smith, 2019).

MSCIKDF generalizes the same idea into a single-root, multi-context identity primitive. It defines a secret long-term root entropy $Q/K/V/G$7, derives context-specific identities via

$Q/K/V/G$8

and introduces a stateless protective transformation

$Q/K/V/G$9

for secret rotation without changing the underlying identity root. The paper frames this as a response to the lack of context isolation, multi-curve independence, and secure non-destructive rotation in BIP-39/BIP-32-style schemes. Its claimed properties include zero-linkability, resistance to cross-context correlation, and PQC-pluggable algorithm agility (Wang, 25 Nov 2025).

Across these systems, the anchor is largely invisible to ordinary operation. It is a continuity substrate rather than a user-facing identifier: a membership credential, self-certifying prefix, event history, or secret root. A common misunderstanding is to equate these anchors with public labels. The cited work instead treats them as cryptographic origins whose preservation enables later verification, revocation, and controlled disclosure.

6. Distributed identity, resilience, and formal invariants

The most explicit generalization of the term appears in work on persistent identity for AI agents. That paper defines an identity anchor as “a persistent data structure that contributes to an agent's behavioral continuity across sessions,” and defines anchor resilience of degree $R$0 as survival of identity after complete loss of up to $R$1 anchors. Behavioral continuity is modeled by a KL-bound on response distributions,

$R$2

and anchor contributions are written as

$R$3

The implemented architecture, soul.py, separates SOUL.md from MEMORY.md, then proposes additional anchors such as PROCEDURES.md, SALIENCE.md, RELATIONS.md, and IDENTITY_HASH.md, together with a hybrid RAG+RLM retrieval system. The paper is explicit that several claims remain hypotheses rather than fully validated empirical results, including the assumed 90%/10% split between focused and exhaustive queries and the effectiveness of the full multi-anchor system (Menon, 2 Mar 2026).

At the opposite end of abstraction, character theory defines anchors as local algebraic invariants. For a prime $R$4, an irreducible character $R$5 of a finite group $R$6 has an anchor defined as a defect group of the primitive $R$7-interior $R$8-algebra $R$9, equivalently a minimal $OGe_\chi$0-subgroup $OGe_\chi$1 such that

$OGe_\chi$2

The paper proves that anchors form a single conjugacy class of $OGe_\chi$3-subgroups, lie inside block defect groups, contain a vertex of a lattice affording $OGe_\chi$4, and contain $OGe_\chi$5. In favorable cases they coincide with more familiar invariants, but not always: anchors may differ from defect groups, Green vertices, or Navarro vertices, and the relationship can involve strict inclusion rather than equality (Kessar et al., 2015).

These two literatures show how far the notion can be stretched without losing its core. In one case an anchor is a persistent behavioral support; in the other it is a uniquely determined $OGe_\chi$6-local conjugacy class. The common thread is minimality with persistence: the anchor is the smallest or most stable structure that still supports recognizable identity. This also clarifies an unresolved tension visible across the broader literature. Stronger anchoring often improves continuity, but it can also introduce costs or constraints: expressivity may fall when temporal anchors are too near, efficiency decreases as anchor redundancy grows, and algebraic anchors need not coincide with larger or more familiar invariants (Seo et al., 27 Oct 2025, Menon, 2 Mar 2026, Kessar et al., 2015).