Decoding Information Aggregation Model

• Decoding Information Aggregation Model is a method that infers uncertainty from fusing internal feature representations across layers rather than relying solely on terminal outputs.
• The approach is implemented in diverse domains such as segmentation with RADMI, ASR via layer aggregation, and Reed–Muller decoding through projection-aggregation techniques.
• Empirical evaluations demonstrate improved prediction metrics and decoding performance while highlighting trade-offs in computational complexity and power consumption.

Decoding Information Aggregation Model denotes a class of decoding formulations in which prediction, uncertainty, or posterior reliability is inferred from how information is fused across intermediate representational stages, rather than from a terminal decision statistic alone. In recent arXiv literature, this notion appears in several technically distinct forms: mutual-information aggregation across decoder layers for dense prediction, top-layer logit aggregation for speech recognition beam search, and projection-aggregation of soft information for Reed–Muller code decoding. Despite domain differences, these formulations share a common premise: decoding performance depends on the internal organization and aggregation of latent evidence across stages, branches, or resolutions, and that internal aggregation can itself be measured, exploited, or analyzed (Stevens et al., 2 May 2026, Wullach et al., 2022, Hashemipour-Nazari et al., 2022, Li et al., 10 Jul 2026).

1. Conceptual scope

The term does not name a single canonical model. Rather, it refers to a family of mechanisms in which decoding is treated as an aggregation problem over internal representations. In encoder–decoder segmentation, the aggregated object is inter-layer dependency structure. In self-supervised ASR, it is a sum of logits produced from several upper transformer layers. In Reed–Muller decoding, it is branch-wise soft evidence obtained from projected subproblems and fused back into updated log-likelihood ratios.

Setting	Aggregated quantity	Decoding role
Segmentation (RADMI)	Mutual information between consecutive decoder layers	Uncertainty proxy
ASR layer aggregation	Logits from the top $M$ layers	Improved beam-search decoding
RM projection-aggregation	Projected branch outputs and LLR contributions	Codeword decoding and soft refinement

Across these settings, the decisive shift is away from output-only decoding. RADMI explicitly differs from predictive entropy and maximum softmax probability because it measures internal feature dependence rather than output distribution spread. The ASR formulation differs from standard top-layer decoding because it aggregates logits from several upper layers after $L2$ normalization. The RM-code formulations differ from terminal hard decision rules because they repeatedly combine evidence across projections, either in IPA hardware architectures or in CPA soft-output decoding (Stevens et al., 2 May 2026, Wullach et al., 2022, Hashemipour-Nazari et al., 2022, Li et al., 10 Jul 2026).

2. Decoder-internal latent aggregation in segmentation

A particularly explicit use of the idea appears in RADMI, “Resolution-Aggregated Decoder Mutual Information,” which can be understood as a decoding information aggregation model because it estimates uncertainty by inspecting how much information must be aggregated and transformed from one decoder stage to the next, rather than by reading the final softmax alone. The central quantity is the inter-layer mutual information

$I(\mathbf{Z}_l; \mathbf{Z}_{l+1}),$

with $\mathbf{Z}_l$ and $\mathbf{Z}_{l+1}$ denoting consecutive decoder feature maps. Homogeneous regions are described as easy to decode because adjacent decoder layers produce similar, tightly clustered activations, whereas boundary or ambiguous regions require reconciliation of competing contextual cues and therefore exhibit stronger dependence between successive layers. High inter-layer MI is accordingly interpreted as evidence that the decoder is actively aggregating latent information to resolve ambiguity (Stevens et al., 2 May 2026).

RADMI is implemented in one forward pass. Feature maps are collected from all consecutive decoder pairs, each pair is aligned by bilinear interpolation to the coarser resolution, and local MI is computed in sliding $p \times p$ windows. Within each patch, the $p^2$ spatial positions are treated as samples, each sample is a channel vector $\mathbf{a}_k \in \mathbb{R}^C$ or $\mathbf{b}_k \in \mathbb{R}^C$, the pair is modeled as jointly Gaussian, and MI is computed in closed form using

$$I(\mathbf{A}; \mathbf{B}) = \frac{1}{2}\log\frac{|\boldsymbol{\Sigma}_{\mathbf{a}}||\boldsymbol{\Sigma}_{\mathbf{b}}|}{|\boldsymbol{\Sigma}|}.$$

This yields a dense layer-wise MI map $L2$0. The maps are then upsampled to the output resolution and combined by resolution-weighted averaging,

$L2$1

Because finer-resolution stages receive larger weights, the resulting uncertainty maps are spatially sharp and boundary-localized (Stevens et al., 2 May 2026).

The evaluation was conducted on the F3 seismic facies segmentation benchmark with 6 lithological classes, using FaciesSegNet, a U-Net variant, trained on 360 inline sections with a 90/10 train/validation split and tested on 200 held-out seismic sections with natural domain shift. Baselines included softmax entropy, $L2$2MSP, MC-Dropout $L2$3, deep ensembles $L2$4, and prediction switches. Since true uncertainty is unavailable, deep ensemble uncertainty was used as the reference standard, and agreement was measured by Pearson correlation, Spearman correlation, cosine similarity, KL, JS, $L2$5, Chamfer, EMD, mIoU, DICE, and histogram intersection. RADMI achieved the best single-pass overall performance against deep ensemble uncertainty, with Pearson $L2$6 and Spearman $L2$7, reported as improvements of 5.5% in Pearson and 10.7% in Spearman over the next-best baselines. It also achieved best mIoU $L2$8, best histogram intersection $L2$9, and lowest $I(\mathbf{Z}_l; \mathbf{Z}_{l+1}),$0 distance $I(\mathbf{Z}_l; \mathbf{Z}_{l+1}),$1. When compared directly to ground-truth error maps, deep ensembles remained best overall, while RADMI ranked as the best single-pass method in Pearson and Spearman and second overall in overlap metrics after the ensemble. A common misconception is that RADMI is merely a confidence heuristic; its stated purpose is instead to use internal model effort during decoding as a proxy for epistemic uncertainty (Stevens et al., 2 May 2026).

3. Layer aggregation in speech recognition decoding

In self-supervised ASR, decoding information aggregation appears as a mechanism for counteracting overconfident top-layer predictions. The setting studied includes fine-tuned wav2vec 2.0 Base/Large and HuBERT Large/X-Large models trained with CTC on LibriSpeech. Standard decoding uses top-layer logits alone,

$I(\mathbf{Z}_l; \mathbf{Z}_{l+1}),$2

either with greedy decoding or with beam search plus an external LLM. The observed problem is that top-layer predictions can be extremely confident, too peaked around a single token, and based largely on local information, thereby limiting the effectiveness of beam search because the beam is deprived of meaningful alternatives (Wullach et al., 2022).

The proposed method aggregates the top $I(\mathbf{Z}_l; \mathbf{Z}_{l+1}),$3 layers: $I(\mathbf{Z}_l; \mathbf{Z}_{l+1}),$4 The $I(\mathbf{Z}_l; \mathbf{Z}_{l+1}),$5 normalization is motivated by the observation that confidence increases with depth, so deeper layers would otherwise dominate the sum. The aggregated logits are then interpolated with the original top-layer logits,

$I(\mathbf{Z}_l; \mathbf{Z}_{l+1}),$6

This is presented as a decoding rule, not a retraining procedure. The associated layer analysis shows that confidence generally increases toward upper layers, that intermediate layers can contain correct token predictions absent from the final layer, and that HuBERT top layers exhibit locally concentrated attention. Label smoothing and temperature scaling also reduced confidence, but did not improve recognition performance in the same way as layer aggregation, reinforcing the claim that the mechanism is not merely probability softening but multi-layer information fusion (Wullach et al., 2022).

Experiments used pyctcdecode, beam search, a 4-gram LLM, and Ax for decoder parameter tuning. Results were reported on LibriSpeech test-clean, test-other, dev-clean, and dev-other in WER and CER. The reported gains were up to 10% WER reduction and 22% CER reduction. Representative test-set improvements include wav2vec 2.0 Base on test-clean from WER 2.5 / CER 0.76 to 2.4 / 0.75 and on test-other from 6.3 / 2.5 to 6.1 / 2.4; HuBERT X-Large on test-other improved from 3.0 / 1.0 to 2.9 / 0.9. Qualitative examples include correcting “break” to “brake,” “tranship” to “transship,” “Phedrus” to “Phaedrus,” “Credius” to “Critias,” and “Chelsey” to “Chelsea.” The stated interpretation is that intermediate layers retain useful acoustic and linguistic information that the final layer may underuse, and aggregation reconstructs a richer decoding distribution more suitable for sequence-level search (Wullach et al., 2022).

4. Projection-aggregation decoding for Reed–Muller codes

In coding theory, the aggregation idea takes a more explicit algorithmic form. For Reed–Muller codes $I(\mathbf{Z}_l; \mathbf{Z}_{l+1}),$7, the recursive projection-aggregation (RPA) decoder operates by projection, recursive decoding, and aggregation. The received vector is represented in LLR form as $I(\mathbf{Z}_l; \mathbf{Z}_{l+1}),$8, and the recursive structure exploited by the decoder comes from the Plotkin construction. The base case is the first-order code $I(\mathbf{Z}_l; \mathbf{Z}_{l+1}),$9, which is decoded optimally using the fast Hadamard transform (Hashemipour-Nazari et al., 2022).

The projection step maps $\mathbf{Z}_l$0 to shorter vectors $\mathbf{Z}_l$1 of length $\mathbf{Z}_l$2. For coordinate $\mathbf{Z}_l$3,

$\mathbf{Z}_l$4

with $\mathbf{Z}_l$5. A common approximation is the min-sum rule,

$\mathbf{Z}_l$6

Each projected vector is then decoded recursively, and the aggregation stage combines all projected branches back into updated soft information. The aggregation rule is a sign-corrected average: $\mathbf{Z}_l$7

$\mathbf{Z}_l$8

Iterative projection-aggregation (IPA) modifies RPA by removing internal recursive iterations and replacing them with a more hardware-friendly iterative structure, with minimal degradation in error-correcting performance according to the reported simulations (Hashemipour-Nazari et al., 2022).

The hardware contribution is a flexible pipelined architecture configurable from fully sequential to fully parallel. Its main components are Processing Units, an FOD block based on FHT, a PreAggregation block, a tree divider or Voting block, a register array, and a control unit. Projection hardware uses ROC and a min-sum block; PreAggregation uses Extension, ReArrangement, and TwosComp submodules. The tree-like divider averages $\mathbf{Z}_l$9 pre-aggregated vectors and inserts a dummy all-zero vector to simplify the divider and control logic with negligible performance impact. For the second-order decoder, the architecture can process a new codeword every $\mathbf{Z}_{l+1}$0 cycles in the fully sequential case and every $\mathbf{Z}_{l+1}$1 cycles with $\mathbf{Z}_{l+1}$2 Processing Units. ASIC synthesis in STM $\mathbf{Z}_{l+1}$3 FD-SOI reported, for RM$\mathbf{Z}_{l+1}$4, representative configurations $\mathbf{Z}_{l+1}$5: 294 cycles, 357 Mbps, 0.38 mm$\mathbf{Z}_{l+1}$6; $\mathbf{Z}_{l+1}$7: 168 cycles, 714 Mbps, 0.61 mm$\mathbf{Z}_{l+1}$8; $\mathbf{Z}_{l+1}$9: 106 cycles, 1428 Mbps, 1.21 mm$p \times p$0. For a code with block length 128 and information length 29, the IPA decoder was reported to have 41% lower area consumption, 44% lower latency, four times higher throughput, but seven times higher power consumption relative to a state-of-the-art polar SCL decoder with comparable decoding performance (Hashemipour-Nazari et al., 2022).

5. Collapsed projection-aggregation and density evolution

A subsequent development analyzes soft-decision collapsed projection-aggregation (CPA) decoding over the BIAWGN channel. Here the decoding information aggregation model is treated probabilistically: each projected subproblem returns a soft piece of evidence, and CPA aggregates these pieces across subspaces. The projection rule for a coset $p \times p$1 of a $p \times p$2-dimensional subspace $p \times p$3 is

$p \times p$4

Each projected vector is decoded by ML FHT decoding as an $p \times p$5 subcode, and aggregation lifts the branch decisions back to original coordinates: $p \times p$6 The average aggregated LLR is then

$p \times p$7

This formulation supports both the original extrinsic update and a broadcast update that reuses the same averaged LLR across subspaces in later iterations (Li et al., 10 Jul 2026).

The principal theoretical claim is that soft-decision CPA returns the exact marginal probability. Projection to a coset computes the LLR of the parity over that coset, and after substitution of the FHT-decoded sign, the aggregation algebra reconstructs the exact bit marginal LLR. The decoder is also proved symmetric on any binary-input memoryless symmetric channel: if one input LLR vector is sign-flipped by a codeword, the output LLRs transform by the same sign pattern. This symmetry enables the standard density-evolution reduction to the all-zero codeword (Li et al., 10 Jul 2026).

The density-evolution model then approximates projection by a BSC sign error probability and replaces exact FHT/ML analysis by a hard-decision decoding proxy. For the BIAWGN channel with BPSK, if the channel LLR density for $p \times p$8 is $p \times p$9 and the sign error probability is $p^2$0, the projected sign error probability is approximated by

$p^2$1

The subspace contribution $p^2$2 is assigned a closed-form mean and variance, and the averaged output LLR is modeled by the CLT as approximately Gaussian. The analysis establishes positive mean, bounded second moment, and rapid decrease of the bit error probability as the number of subspaces grows. Simulations on $p^2$3 and $p^2$4 show approximately Gaussian post-iteration LLR histograms and fast reduction in mean and variance, while also noting that the independence approximation underestimates variance at finite length because subspace outputs are correlated. The asymptotic conclusion is that CPA can achieve vanishing error probability for RM codes with vanishing code rate (Li et al., 10 Jul 2026).

6. Comparative interpretation and limitations

Across the main formulations, the aggregated object changes, but the governing principle remains stable. RADMI aggregates normalized inter-layer information flow across decoder resolutions; ASR layer aggregation sums normalized logits from multiple upper layers; IPA and CPA aggregate branch-consistent soft evidence across projected subproblems. In each case, the decoder is treated as a structured inference pipeline whose internal representations contain exploitable information not captured by a final-layer statistic alone (Stevens et al., 2 May 2026, Wullach et al., 2022, Hashemipour-Nazari et al., 2022, Li et al., 10 Jul 2026).

Several recurrent distinctions are important. First, decoding information aggregation is not identical to output confidence estimation. RADMI contrasts with predictive entropy and $p^2$5MSP because it measures internal feature dependence. Second, it is not equivalent to indiscriminate probability smoothing. In ASR, label smoothing and temperature scaling did reduce confidence, but did not reproduce the gains from top-layer aggregation. Third, it is not mere averaging in the unstructured sense. In RM decoding, aggregation is a sign-corrected fusion rule tied to code symmetries, and in CPA it is analyzed as exact marginal recovery rather than heuristic consensus (Stevens et al., 2 May 2026, Wullach et al., 2022, Hashemipour-Nazari et al., 2022, Li et al., 10 Jul 2026).

The limitations are equally domain-specific. RADMI was validated on a seismic facies segmentation benchmark and still used deep ensembles as the reference standard because true uncertainty is unavailable. ASR layer aggregation produced consistent but bounded improvements and still depends on beam search, a LLM, and tuned interpolation. IPA achieves favorable area, latency, and throughput trade-offs, but at substantially higher power consumption than the compared polar SCL decoder. CPA density evolution depends on approximations: projection is treated through a BSC surrogate, FHT decoding through a hard-decision proxy, and subspace outputs through an independence assumption that the authors explicitly identify as imperfect at finite length (Stevens et al., 2 May 2026, Wullach et al., 2022, Hashemipour-Nazari et al., 2022, Li et al., 10 Jul 2026).

A plausible implication is that the phrase “decoding information aggregation model” is best understood as a methodological orientation rather than a single architecture. It denotes a shift from terminal-output decoding to decoder-analytic inference: ambiguity, alternative hypotheses, or posterior certainty are inferred from how representations are fused across layers, resolutions, or projected branches. In current arXiv usage, that orientation supports uncertainty estimation in segmentation, sequence-level search in ASR, and near-ML or analytically tractable decoding in Reed–Muller codes (Stevens et al., 2 May 2026, Wullach et al., 2022, Hashemipour-Nazari et al., 2022, Li et al., 10 Jul 2026).