Soft Length Controller in VarLenRec

• Soft Length Controller is a differentiable mechanism that adaptively determines the effective length of semantic IDs in generative recommendation.
• It computes continuous layer retention probabilities from hyperbolic residuals and codeword inputs via a lightweight sigmoid gate.
• It is regularized by popularity-derived PIBA priors, enabling end-to-end training that enhances recommendation performance across different popularity groups.

The Soft Length Controller is a differentiable mechanism for learning variable-length tokenization in generative recommendation. In VarLenRec, an item’s semantic ID is produced by multi-layer residual quantization, and the controller determines, per item and per layer, how many residual quantization layers should be retained in the final ID. It does so through continuous layer retention probabilities and cumulative masks, making length decisions trainable end-to-end while remaining guided by popularity-derived priors from Popularity-Weighted Information Budget Allocation (PIBA) (Wang et al., 18 May 2026).

1. Position within generative recommendation

Generative recommendation reformulates recommendation as next-token prediction over discrete semantic identifiers, and each item’s identifier is obtained by quantizing a continuous embedding into a sequence of discrete codebook indices through multi-layer residual quantization. Existing systems such as TIGER, LETTER, and LC-Rec use fixed-length semantic IDs, so every item receives exactly $L$ codewords irrespective of popularity, content complexity, or data sparsity. VarLenRec was proposed against this background after the empirical observation of the Popularity-Length Paradox: for head (popular) items, performance worsens as semantic ID length increases, while for tail (unpopular) items, performance improves as length increases (Wang et al., 18 May 2026).

This observation motivates the controller’s role. Popular items already have strong collaborative information and therefore require only limited semantic detail; tail items must rely more heavily on fine-grained content features and benefit from additional semantic bits. A single global length is therefore a compromise that is suboptimal across popularity strata. VarLenRec addresses this with four coupled components: PIBA, Hyperbolic Residual Quantization (HARQ), the Soft Length Controller, and downstream integration including collision resolution, trie-constrained decoding, and length-normalized scoring.

Within that architecture, the Soft Length Controller is the component that makes length allocation content-adaptive and differentiable. It operates inside the HARQ quantizer, layer by layer, takes hyperbolic residuals and selected codewords as inputs, outputs retention probabilities, and is regularized by PIBA’s popularity-based optimal-length prior.

2. Definition of length and controller mechanics

In this setting, “length” does not denote token width or embedding dimensionality. For item $i$, the semantic ID is a sequence of codewords

$z_i^{(1)}, \dots, z_i^{(K)},$

where $K$ is the maximum number of quantization layers. The effective length $L_i$ is the number of initial layers that are retained: layers $1,\dots,L_i$ are kept, and layers $L_i+1,\dots,K$ are dropped or truncated. The controller therefore solves a specific problem: given an item’s content and quantization state, determine how many residual quantization layers should be kept, but do so in a way that remains compatible with gradient-based optimization (Wang et al., 18 May 2026).

At quantization layer $l$, the controller receives two inputs: the hyperbolic residual before layer $l$,

$\mathbf{r}_i^{(l-1)} \in D_c^d,$

and the selected hyperbolic codebook vector,

$i$0

Both are projected to the tangent space at the origin with the logarithmic map and concatenated: $i$1 A lightweight gate network with sigmoid output then produces the retention probability

$i$2

The quantity $i$3 is the probability of retaining layer $i$4, conditioned on the previous layers. This makes the decision content-aware. If the residual is already small, or if the selected codeword is already well aligned with it, a lower retention probability can be emitted.

To enforce the required prefix structure, the controller defines cumulative masks

$i$5

These masks are non-increasing in $i$6, so they approximate the probability that the first $i$7 layers are retained. They serve simultaneously as soft attention weights on codebook contributions in decoding and as differentiable surrogates for binary activity indicators. A natural soft notion of expected length is then

$i$8

3. Mathematical role in decoding and optimization

The controller is not an external heuristic; it lies on the main computational path of HARQ. The hyperbolic decoder starts from

$i$9

and accumulates mask-scaled code vectors: $z_i^{(1)}, \dots, z_i^{(K)},$0 where $z_i^{(1)}, \dots, z_i^{(K)},$1 is Möbius addition on the Poincaré ball and

$z_i^{(1)}, \dots, z_i^{(K)},$2

If $z_i^{(1)}, \dots, z_i^{(K)},$3 is close to $z_i^{(1)}, \dots, z_i^{(K)},$4, the contribution of layer $z_i^{(1)}, \dots, z_i^{(K)},$5 is suppressed; if it is close to $z_i^{(1)}, \dots, z_i^{(K)},$6, the layer contributes almost fully. The final decoded representation is

$z_i^{(1)}, \dots, z_i^{(K)},$7

Because the computation from inputs to $z_i^{(1)}, \dots, z_i^{(K)},$8, then to $z_i^{(1)}, \dots, z_i^{(K)},$9, then to $K$0, and finally to $K$1, is differentiable, the controller can be trained jointly with the encoder, quantizer, and decoder. The relevant HARQ objective is

$K$2

Two terms are specific to the controller. The first is the length cost

$K$3

which penalizes expected length directly. The second is the length alignment term

$K$4

which is a binary cross-entropy between the continuous mask $K$5 and a binary target mask derived from PIBA. Since

$K$6

all derivatives with respect to the controller outputs remain analytic, and no Gumbel-softmax or straight-through estimator is required (Wang et al., 18 May 2026).

4. PIBA regularization and the popularity prior

The controller is tightly coupled to PIBA, which provides the theoretical prior for optimal semantic-ID length. PIBA assumes that each item requires an information budget $K$7 for effective recommendation, and that this budget is met by two sources: collaborative information from interactions and semantic information from the ID. The collaborative term is modeled as

$K$8

while the semantic term is

$K$9

Equating the semantic contribution to the remaining information gap yields the theorem

$L_i$0

so optimal length scales as a negative power of popularity (Wang et al., 18 May 2026).

For implementation, the continuous optimum is converted to a discrete prior length $L_i$1. Items are ranked by popularity, the normalized “coldness” is computed as $L_i$2, a temperature transform $L_i$3 is applied, and the prior length is mapped into $L_i$4 through

$L_i$5

This prior defines the binary target mask

$L_i$6

A common misconception is that VarLenRec simply enforces PIBA’s popularity-derived lengths. It does not. PIBA acts as a regularizing prior through $L_i$7; the controller is not forced to follow it exactly. The learned gates still adapt to item content and residual quantization state. This distinction is experimentally important because direct PIBA assignment underperforms the full soft, content-aware mechanism (Wang et al., 18 May 2026).

5. Interaction with HARQ, training, and inference

The controller is embedded at each quantization layer of Hyperbolic Residual Quantization. The pipeline is: a hyperbolic encoder maps item features $L_i$8 to

$L_i$9

the hyperbolic residual quantizer selects a codeword by nearest neighbor in hyperbolic distance and computes a residual via Möbius addition, and the Soft Length Controller outputs $1,\dots,L_i$0 at each layer. The decoder then reconstructs with mask-scaled codewords. HARQ supplies the geometric substrate for variable depth: early layers near the origin encode coarse semantics, while deeper layers move toward the boundary, where volume and representational capacity grow exponentially. The controller determines how far along this stratification each item should proceed (Wang et al., 18 May 2026).

During training, there is no hard truncation. All $1,\dots,L_i$1 layers are computed, the decoder uses continuous masks $1,\dots,L_i$2, and the reconstruction, quantization, length cost, and length alignment losses all depend on those masks. Training therefore treats length as a soft, differentiable quantity.

During inference, variable-length semantic IDs must be discrete. The masks are therefore thresholded: $1,\dots,L_i$3 or equivalently layers are retained while $1,\dots,L_i$4 or $1,\dots,L_i$5, with default $1,\dots,L_i$6. The resulting ID is the deterministic variable-length token sequence

$1,\dots,L_i$7

This produces a specific efficiency pattern. Head items receive short IDs, so the generative model often stops after fewer steps; average sequence length and decoding time are reduced. Tail items receive longer IDs when the additional semantic capacity is useful.

6. Empirical behavior, hyperparameters, and limitations

Ablation results isolate the contribution of the controller within VarLenRec-TIGER. The full system includes the Soft Length Controller, $1,\dots,L_i$8, and PIBA alignment $1,\dots,L_i$9. On Beauty, it reports $L_i+1,\dots,K$0 and $L_i+1,\dots,K$1; on Toys, $L_i+1,\dots,K$2 and $L_i+1,\dots,K$3 (Wang et al., 18 May 2026).

Variant	Beauty	Toys
Full VarLenRec-TIGER	R@10 $L_i+1,\dots,K$4, N@10 $L_i+1,\dots,K$5	R@10 $L_i+1,\dots,K$6, N@10 $L_i+1,\dots,K$7
Removing $L_i+1,\dots,K$8	R@10 $L_i+1,\dots,K$9, N@10 $l$0	R@10 $l$1, N@10 $l$2
Direct PIBA assignment	N@10 $l$3	N@10 $l$4

These comparisons establish several points. Removing $l$5 harms performance and leads the controller to keep more layers than necessary. Replacing the learned controller with direct PIBA assignment yields lower $l$6, showing that soft, content-aware length control is superior to hard popularity-based assignment alone. Fixed-length baselines with $l$7 all underperform the learned variable-length scheme.

The learned length distributions are also consistent with the intended behavior. Average learned lengths are 5.31 on Toys and 7.05 on Yelp, with other datasets in between. Lengths 3–6 are most common, and Yelp contains many longer IDs. On the Beauty dataset, VarLenRec with the Soft Length Controller and PIBA achieves simultaneous improvement across popularity groups relative to fixed-length TIGER: +41.3% for head, +48.6% for body, +6.8% for tail, and +40.7% overall. The reported interpretation is that performance scales nearly monotonically with the maximum allowable length because the controller uses longer IDs only where beneficial (Wang et al., 18 May 2026).

The main controller-specific hyperparameters are the maximum length $l$8, the PIBA temperature $l$9, the length-cost weight $l$0, the length-alignment weight $l$1, and the inference threshold $l$2. Performance peaks at $l$3, dataset-dependent. For Beauty, the best settings reported are $l$4 and $l$5; for Toys, $l$6 and $l$7. The design choices emphasized in the paper are the sigmoid gate, residual+codeword inputs, and cumulative product masks.

The main limitations follow directly from the reliance on PIBA and catalog statistics. PIBA requires popularity estimates from interactions; for cold-start items, the method falls back to assigning maximum length. Popularity shift over time requires periodic re-running of HARQ. The prior is also popularity-driven only, so items whose complexity is not aligned with popularity may require other priors such as category, multimodal complexity, or temporal dynamics. Proposed extensions include multimodal items, cross-domain settings, integration with LLM architectures, more sophisticated controllers such as attention-based gating, and alternatives to a fixed power-law prior (Wang et al., 18 May 2026).

In summary, the Soft Length Controller is the mechanism that makes variable-length semantic IDs learnable end-to-end. Its central contribution is to replace discrete truncation decisions with continuous retention probabilities whose cumulative products define a differentiable expected length, a soft decoder mask, and a bridge between hyperbolic quantization and popularity-aware theory. The broader significance of the mechanism lies in showing that adaptive encoding capacity—not merely better token prediction—can be a first-order design variable in generative recommendation.