Tier Group Encoding Overview

• Tier Group Encoding is a family of hierarchical strategies that embed tiered structures into model representations and aggregations across diverse domains.
• It organizes elements such as spatial attributes, token positions, resolution tiers, and subgroup chains to preserve both coarse and fine details.
• Applications include geospatial federated learning, streaming LLMs, multi-resolution video compression, coding theory constructions, and self-supervised speech recognition.

Tier Group Encoding is a label used in several research literatures for schemes that organize entities into tiers or groups and encode that structure directly into representation, computation, or decoding. In geospatial federated learning, it denotes the joint use of encoded spatial attributes and hierarchical client grouping (Kawnine et al., 10 Jan 2025). In streaming LLMs, it corresponds to Group Position Encoding, which assigns independent, monotonic positions to source and target groups with a constant offset $\phi$ (Tong et al., 22 May 2025). In video compression, it denotes a multi-resolution tier multi-ABR reuse pipeline organized by resolution tiers (Menon, 2023). In coding theory, related constructions appear in two-level fingerprinting, the Schreier matrix form for group codes, and semi-quantitative group testing, where user groups, subgroup chains, or quantizer levels define the operative tiers (0905.0417, Jr, 2011, Emad et al., 2014). In self-supervised speech recognition, a related usage studies how speaker-group information is distributed across representation tiers, namely model layers (Herron et al., 9 Jun 2026). A plausible implication is that Tier Group Encoding is best treated as a family of hierarchical encoding strategies rather than a single formalism.

1. Conceptual scope and recurring structure

Across the cited literature, the term refers to explicit encoding of hierarchical structure into a system’s internal variables. The hierarchy may be spatial, temporal, architectural, representational, or combinatorial. The encoded object may be an input feature vector, a position index, a reusable analysis artifact, a codeword component, or a quantized test outcome.

Domain	Tier/group definition	Encoding mechanism
Geospatial FL	client, intermediate server, global server	spatial attributes $s$ encoded as $e=E(s)$ and used for group assignment
Streaming LLMs	source and target token groups	positions $p(x_i)=i$, $p(y_j)=\phi+j$
Video compression	resolution tiers and bitrate rungs	seed-rung analysis reused within and across tiers
Two-level fingerprinting	user groups and users within groups	centers encode groups, offsets encode users
Group trellis codes	Schreier-matrix columns and factors	coset representatives assigned tier-by-tier
SQGT	quantizer bins induced by thresholds	subset sums mapped to tiered outcomes
Speech models	network layers as tiers	SGI probed in layerwise embeddings

A common structural pattern is the separation of a global or coarse component from a local or fine component. In the federated learning formulation, spatial encoding augments local input vectors while the same spatial attributes assign clients to aggregation tiers. In GPE, source and target retain separate within-group orderings even though they share one Transformer. In two-level fingerprinting, group identity is carried by a random center and user identity by a constant-weight offset. In the Schreier-matrix encoder, macro-tiers are matrix columns and micro-tiers are refined subgroup factors. This suggests that tiering is used to preserve distinctions that would be blurred by a flat encoding.

2. Encoded spatial attributes in multi-tier federated learning

In "Encoded Spatial Attribute in Multi-Tier Federated Learning" (Kawnine et al., 10 Jan 2025), Tier Group Encoding is defined as a two-part process: spatial attributes $s$ are encoded into a numerical representation $e=E(s)$ and concatenated with the original input $x$ to form $x'=[x;e]$, and the same spatial attributes are used to assign clients to spatially meaningful tiers. The paper’s client tiers are provinces or territories for greenhouse gas emissions and stations or city-level partitions for NB power EV charging; tier-1 servers aggregate subsets of clients; tier-2 aggregates tier-1 outputs into a national or provincial/global model.

The spatial attributes include latitude/longitude, administrative units such as provinces and cities, and geohash. The paper states that continuous and discrete spatial representations are combined to capture both fine-grained local variability and coarse administrative trends. Compatible encodings include one-hot or multi-hot vectors for administrative units, learned embeddings for discrete keys, geohash-based embeddings, and sinusoidal positional encodings for latitude and longitude. With multiple components, the final spatial vector is

$$e=[E_{\text{admin}}(s_{\text{admin}});e_g;e_{\text{lat}};e_{\text{lon}}],$$

and the model input becomes $s$0.

The local client objective is supervised classification over datasets $s$1:

$s$2

The reported experiments used simple two-layer neural networks, with encoded features entering at the input layer by concatenation. The paper writes the local gradient step as

$s$3

Aggregation is hierarchical. Within a tier, weighted FedAvg may be written as

$s$4

with $s$5, while the paper’s explicit regional formula is

$s$6

Across tiers, the global model is

$s$7

with $s$8 proportional to total data volume in the tier. The paper also notes that spatial influence can enter aggregation through group assignment and spatially informed weights.

The empirical results are reported for two datasets. For gas emissions, the tier-2/global N-tier encoded FL model achieved $s$9, close to the local NN at $e=E(s)$0, while FedAvg Weighted achieved $e=E(s)$1 at tier-2. For NB power, the tier-2/global N-tier encoded FL model achieved $e=E(s)$2, compared with local NN at $e=E(s)$3. The paper emphasizes that these results were obtained “without having to train the model using the data constituted with the designated tier,” meaning that the tier-2/global model is formed by aggregation of lower-tier models rather than by centralized training on a separate tier-level dataset. It also states that the approach is important in real-time applications and reports edge deployment on reComputer J1010 devices.

The paper is explicit that simple MLPs still have difficulty capturing complex spatial dependencies, and that non-IID and imbalance can harm certain aggregators. Proposed improvements include temporal encodings, spatiotemporal models, hybrid or ensemble models, refined aggregation strategies, and privacy mechanisms such as differential privacy and secure aggregation.

3. Group Position Encoding in streaming LLMs

In "LLM as Effective Streaming Processor: Bridging Streaming-Batch Mismatches with Group Position Encoding" (Tong et al., 22 May 2025), Tier Group Encoding is the paper’s Group Position Encoding, or GPE. It partitions tokens into two tiers, source $e=E(s)$4 and target $e=E(s)$5, and assigns positions independently within each group:

$e=E(s)$6

where $e=E(s)$7 is a constant offset. In batch processing, $e=E(s)$8; in group-streaming, $e=E(s)$9 is fixed and independent of current input length.

The method is motivated by three mismatches between batch-trained LLMs and streaming use: input-attention mismatch, output-attention mismatch, and position-ID mismatch. The paper’s central empirical claim is that input-attention mismatch dominates. Eliminating only that mismatch yielded the largest improvement over interleaved streaming: for Gemma2-2B-Instruct, En–Fr improved by $p(x_i)=i$0 BLEU at wait-1 and En–De by $p(x_i)=i$1; for Phi3-Mini, En–Fr improved by $p(x_i)=i$2 and En–De by $p(x_i)=i$3. By contrast, position-ID re-encoding added at most $p(x_i)=i$4 BLEU, and full KV re-encoding added at most $p(x_i)=i$5 BLEU.

The formal rationale is tied to RoPE. If rotated queries and keys satisfy

$p(x_i)=i$6

then attention depends only on relative distance:

$p(x_i)=i$7

Under GPE, within-source and within-target relative distances are preserved exactly because positions are monotone within each group. Cross-segment distances differ from classic batch concatenation only by the constant shift $p(x_i)=i$8. The paper argues that fine-tuning enables the model to internalize this stable offset.

The streaming algorithm requires no architectural modifications. Source and target KV caches can be maintained separately and merged for attention. New source tokens receive positions $p(x_i)=i$9, new target tokens receive positions $p(y_j)=\phi+j$0, and standard RoPE is applied to these positions. The masks are standard causal or prefix masks augmented so that source tokens cannot attend to target tokens and target tokens can attend only to the inputs available at their emission time.

The reported efficiency gains are substantial. For Phi-3 Mini on En–Fr, wait-5 inference took $p(y_j)=\phi+j$1 hours with re-encoding versus $p(y_j)=\phi+j$2 hours without, a $p(y_j)=\phi+j$3 time reduction, and throughput increased from $p(y_j)=\phi+j$4 to $p(y_j)=\phi+j$5 tokens/s. At wait-9, time decreased from $p(y_j)=\phi+j$6 to $p(y_j)=\phi+j$7 hours and throughput increased from $p(y_j)=\phi+j$8 to $p(y_j)=\phi+j$9 tokens/s. The paper also reports robustness to the offset hyperparameter: varying $s$0 changes BLEU by at most $s$1, and in ASR varying $s$2 changes WER by at most $s$3.

A frequent assumption in prior streaming adaptation was that output re-encoding is necessary. The paper directly disputes this. Its position is that preserving relative order within source and within target contexts is more critical than preserving absolute order, and that frequent re-encoding is not justified by the measured gains.

4. Tiered reuse in multi-resolution video compression

In "Multi-resolution encoding and optimization for next generation video compression" (Menon, 2023), Tier Group Encoding denotes a multi-encode strategy for adaptive bitrate ladders in which the ladder is partitioned into spatial-resolution tiers such as $s$4p, $s$5p, and $s$6p, and each tier contains several bitrate rungs. Instead of encoding every rung independently, the method encodes a representative seed rung, chosen as the median bitrate in a tier, and reuses that analysis for the other rungs in the same tier. The refined seed is then extrapolated to the next higher-resolution tier.

The reported seed rungs are $s$7p@$s$8 Mbps, $s$9p@$e=E(s)$0 Mbps, and $e=E(s)$1p@$e=E(s)$2 Mbps. The rationale is that encoder decisions such as the CTU/CU quadtree, prediction modes, motion vectors, and frame-level structure exhibit strong intra-tier and significant inter-tier correlation, while the median rung avoids the performance bottleneck of seeding from the slowest, highest-quality rung.

Reusable x265 analysis metadata include lookahead and slice-type decisions, scene cuts, reference lists, CU quadtree structure, intra and inter modes, motion information, RD pruning outcomes, early-termination flags, and rate-control hints. Cross-resolution reuse is dyadic. If the source and target resolutions are $e=E(s)$3 and $e=E(s)$4, then scaled motion vectors satisfy

$e=E(s)$5

with $e=E(s)$6 and $e=E(s)$7.

The pipeline is bottom-up. The lowest tier is seeded first, other rungs in the tier load that seed with configurable refinement, the next tier’s seed is extrapolated from the lower-tier seed using --scale-factor=2, and the process repeats. Quality-efficiency tradeoffs are reported using Bjøntegaard metrics. Speed-up is defined as

$e=E(s)$8

and time reduction as

$e=E(s)$9

Within a single resolution tier, the proposed median-seed pattern at reuse-level $x$0 improved both speed and coding efficiency relative to a highest-rung seed. At $x$1p it gave $x$2, BD-Rate $x$3, and BD-PSNR $x$4 dB; at $x$5p, $x$6, BD-Rate $x$7, and BD-PSNR $x$8 dB; at $x$9p, $x'=[x;e]$0, BD-Rate $x'=[x;e]$1, and BD-PSNR $x'=[x;e]$2 dB. Cross-resolution reuse without refinement was too aggressive, producing BD-Rate penalties of about $x'=[x;e]$3 for $x'=[x;e]$4p$x'=[x;e]$5p and $x'=[x;e]$6 for $x'=[x;e]$7p$x'=[x;e]$8p. With refinement, those penalties fell to about $x'=[x;e]$9 and $$e=[E_{\text{admin}}(s_{\text{admin}});e_g;e_{\text{lat}};e_{\text{lon}}],$$0, respectively, with time savings around $$e=[E_{\text{admin}}(s_{\text{admin}});e_g;e_{\text{lat}};e_{\text{lon}}],$$1 per higher tier. Across complete ladders, the scheme achieved overall speed-ups up to about $$e=[E_{\text{admin}}(s_{\text{admin}});e_g;e_{\text{lat}};e_{\text{lon}}],$$2 versus conventional single-instance encoding.

The paper also discusses hardware-aware SIMD optimization, evaluated with the Arowana XVC encoder. Kernel-level gains of $$e=[E_{\text{admin}}(s_{\text{admin}});e_g;e_{\text{lat}};e_{\text{lon}}],$$3–$$e=[E_{\text{admin}}(s_{\text{admin}});e_g;e_{\text{lat}};e_{\text{lon}}],$$4 for SATD kernels and $$e=[E_{\text{admin}}(s_{\text{admin}});e_g;e_{\text{lat}};e_{\text{lon}}],$$5–$$e=[E_{\text{admin}}(s_{\text{admin}});e_g;e_{\text{lat}};e_{\text{lon}}],$$6 for SAD kernels are reported, with encoder-level speed-ups up to about $$e=[E_{\text{admin}}(s_{\text{admin}});e_g;e_{\text{lat}};e_{\text{lon}}],$$7–$$e=[E_{\text{admin}}(s_{\text{admin}});e_g;e_{\text{lat}};e_{\text{lon}}],$$8 over scalar baselines at $$e=[E_{\text{admin}}(s_{\text{admin}});e_g;e_{\text{lat}};e_{\text{lon}}],$$9p. The stated principle is that tiered reuse prunes expensive searches, while SIMD accelerates the evaluations that remain.

5. Coding-theoretic formulations

In coding theory, Tier Group Encoding appears as explicit hierarchical structuring of codewords or decoder outputs. The cited papers give three distinct constructions: two-level fingerprinting and traceability codes, Schreier-matrix encoders for group trellises, and semi-quantitative group testing with nonuniform thresholds (0905.0417, Jr, 2011, Emad et al., 2014).

In "Two-Level Fingerprinting Codes" (0905.0417), users are indexed by $s$00, where $s$01 is the group index and $s$02 is the within-group user index. A two-level deterministic code is a triple

$s$03

with

$s$04

The code is $s$05-traceability if, for coalitions of size at most $s$06, $s$07 identifies a guilty user, and for coalitions of size at most $s$08, $s$09 identifies a guilty group. A sufficient condition is

$s$10

where $s$11 measures inter-group minimum distance and $s$12 measures within-group minimum distance. In the randomized construction, group identity is encoded by random centers $s$13 and user identity by constant-weight offsets $s$14, with fingerprints

$s$15

This is a literal hierarchical encoding of group and user information.

In "Group Codes and the Schreier matrix form" (Jr, 2011), the tiers are columns of the Schreier refinement of two normal chains in a group trellis. The matrix entries are

$s$16

and column $s$17 forms the chain

$s$18

Shortest-length generator sequences align with these factors, and each generator component is chosen as a coset representative of the corresponding factor. The output branch at time $s$19 is an ordered product of tier components,

$s$20

with order fixed column-by-column. The encoder is implemented as a shove or shift register under the next-branch map $s$21, which moves tier components one column to the right at the next time step. If the branch group $s$22 is solvable, the refinement can be pushed to composition factors of prime cyclic order, so that each micro-tier reduces to a prime-order symbol.

In "Code Construction and Decoding Algorithms for Semi-Quantitative Group Testing with Nonuniform Thresholds" (Emad et al., 2014), the tiers are the output bins of a nonuniform quantizer with thresholds

$s$23

For a defective set $s$24, each test outcome is

$s$25

The construction horizontally concatenates weighted copies of a binary $s$26-disjunct base matrix,

$s$27

where the weights $s$28 are drawn from quantized $s$29, SQLO$s$30, or SQLO$s$31 sequences designed so that all subset sums of up to $s$32 elements fall into distinct tiers. The resulting codes are $s$33-SQ-separable, guaranteeing zero-error identification with up to $s$34 adversarial test errors. The paper provides an exponential-time decoder for quantized $s$35 constructions and linear-time knapsack-style decoders for SQLO$s$36 and SQLO$s$37.

These three formulations differ sharply in application, but each uses tier structure to separate coarse identification from fine identification. In two-level fingerprinting, the coarse output is a guilty group and the fine output is a guilty user. In the Schreier encoder, coarse structure is a subgroup column and fine structure is the coset representative inside that column. In SQGT, the coarse object is the quantizer bin and the fine object is the subset of weights that produced it.

6. Layerwise speaker-group encoding in self-supervised speech models

In "Speaker Group Encoding in Self-supervised Speech Recognition Models" (Herron et al., 9 Jun 2026), tier structure is identified with network depth. The paper studies what self-supervised speech recognition models encode about speaker groups across representation tiers, that is, across layers. The speaker group categories are gender, age, dialect, ethnicity, and native versus non-native status. The models include WavLM-Large, WavLM-Base, Wav2vec 2.0 lv60, and XLSR-53, examined in four states: pretrained, fine-tuned on speaker identification, fine-tuned on ASR with CTC, and ASR-finetuned with fairness-enhancing DET or DAT objectives.

Utterance-level embeddings are formed by mean pooling over one frame, $s$38 frames, or the entire sequence. The paper reports that whole-utterance pooling generally provides the most speaker-group information, and that the first second often carries more SGI than the last second in command-style utterances. Probing uses a low-dimensional linear projection $s$39 with $s$40, followed by a classifier $s$41:

$s$42

A speaker classifier with gradient reversal is trained on the same projection to discourage speaker overfitting. The probe loss is cross-entropy,

$s$43

The layerwise findings are strongly tier-dependent. Except for gender, very early layers encode little SGI; encoding generally increases toward middle layers, approximately layer $s$44 in a $s$45-layer stack, and then remains relatively stable in late layers for WavLM. W2V2 tends to diminish SGI in the very last layers even when pretrained. WavLM-Large encodes the most SGI of the pretrained models, while W2V2 multilingual and English-only variants encode similar amounts for the cohorts considered.

Task fine-tuning alters what each tier retains. Fine-tuned SID amplifies phonetically variant categories such as gender and age and keeps them recoverable through later layers. Fine-tuned ASR suppresses phonetic SGI toward late layers but retains semantically variant SGI such as dialect, ethnicity-related dialectal features, and native/non-native status. DET, which attaches a speaker-ID head at a middle layer, preserves or increases phonetic SGI at middle tiers relative to vanilla ASR. DAT, which attaches a gradient-reversal speaker-ID head at a late layer, further suppresses late-layer phonetic SGI but has limited effect on semantic SGI.

The geometric analysis uses PCA on class centroids and cosine similarity between principal directions, as well as variance explained

$s$46

and subspace removal

$s$47

The paper concludes that SGI is concentrated in a few directions, that DET aligns SID and SGC directions more strongly at the chosen middle layer, and that DAT can make some SGC subspaces more orthogonal to SID near the top of the stack rather than fully eliminating them.

Although this paper does not use Tier Group Encoding as a positional or aggregation mechanism, it provides a related and technically important interpretation of tiered group encoding: group information is distributed nonuniformly across architectural layers, and training objectives can relocate, amplify, suppress, or orthogonalize that information depending on the tier at which they are applied.

7. Common design principles, misconceptions, and open directions

The surveyed literature supports several recurring design principles. First, tier assignment is domain-specific but explicit: spatial granularity in federated learning, source-versus-target role in streaming LLMs, resolution and bitrate ladders in video compression, user groups or subgroup chains in coding theory, nonuniform thresholds in SQGT, and model layers in speech representation analysis. Second, the tier variable is not merely descriptive; it changes the computational graph, the admissible reuse pattern, the decoder, or the optimization target. Third, many constructions preserve within-tier structure while controlling cross-tier interaction. This is exact in GPE, where within-group relative order is preserved, and in the federated learning hierarchy, where local spatial encoding is combined with aggregation only among assigned spatial groups.

A common misconception arises in the streaming LLM setting: that output re-encoding is intrinsically required for high-quality streaming. The GPE paper argues the opposite, reporting that input-attention mismatch is the dominant problem and that output re-encoding yields only marginal gains (Tong et al., 22 May 2025). A different misconception is that hierarchical grouping alone resolves heterogeneity. The federated learning paper does not claim this; it reports variability across servers and datasets, notes that weighted FedAvg can degrade sharply under imbalance, and explicitly calls for further work on model generalization and richer architectures (Kawnine et al., 10 Jan 2025).

The literature also identifies domain-specific limitations. In geospatial FL, simple MLPs are insufficient for complex spatial dependencies and privacy mechanisms were not implemented. In GPE, very large $s$48 and extremely long contexts can move cross-segment distances away from the pretraining distribution. In video compression, blind cross-resolution reuse produces prohibitive BD-Rate penalties unless refinement is used. In two-level fingerprinting, guarantees are asymptotic and tied to the marking assumption. In Schreier-matrix encoders, exact ordering matters in nonabelian settings. In SQGT, arbitrary thresholds still leave test complexity at $s$49 and alphabet size can grow with the number of weight tiers. In speech models, SGI measurements are based on linear probes and domain mismatch remains a caveat.

A plausible implication of these results is that Tier Group Encoding is most effective when the tier variable corresponds to a stable structural asymmetry already present in the problem: geography, streaming causality, resolution hierarchy, decoder traceability levels, subgroup decomposition, quantizer bins, or layerwise specialization. Where that asymmetry is weak or mis-specified, the cited works consistently report the need for refinement, stronger inductive bias, or more selective regularization.