TL;DR for operators

Catalog enrichment systems need to recover attributes explicitly stated in product descriptions alongside values that require interpretation, normalization, or inference. Generating every attribute-value pair sequentially imposes an unnecessary decoding burden when many values already appear verbatim in the input.

Kolluru and colleagues test an architecture that shares one language-model prefill between parallel source-span tagging and conventional generation. On 11,132 held-out Amazon US products, attribute recall increases from 69.8% to 80.2%, while mean inference latency falls from 56.6 to 31.5 milliseconds. LLM-judged precision remains approximately 97%.

The evidence supports dividing extraction according to whether values can be copied, not universally replacing generation. Pure tagging remains faster on fully extractive tasks, while attributes outside the tagging vocabulary remain difficult. Catalog operators should compare all three configurations on representative records, examine recall by attribute frequency and category, and separate copied values from generated values during quality control. Production savings and downstream business improvements remain untested.

The cost of making the model write everything

A merchant product description may already contain its brand, ingredients, dimensions, materials, and other specifications. Extracting those fields should not require the system to compose their values again. Yet the same record may need a normalized product type, dietary classification, or other attribute that cannot be copied directly.

A generative extractor accommodates both cases, but processes its output token by token. As the attribute list grows, the model must spend additional decoding steps producing values already available in the source. Longer output lists can also omit relevant fields.

A conventional tagging system avoids much of that work by locating values in the input. Its limitation is structural: it cannot directly extract values that are absent as contiguous source spans, and a fixed attribute vocabulary restricts what it can recognize.

The resulting architecture decision is more specific than choosing between a fast and a comprehensive model. It concerns which attributes actually require generation.

In Hybrid Generative-Tagging Decoding for Efficient Product Attribute-Value Extraction, Keshav Kolluru and colleagues investigate whether those two workloads can share computation without forcing every attribute through the same decoding process.1

One prefill, two extraction channels

The proposed architecture first runs the product prompt through a causal language model. The resulting hidden states are shared by two channels.

The tagging channel tests a fixed vocabulary of frequent attribute names. Each attribute has a cached query representation that attends to the product’s contextual states. Separate start and end pointers locate a candidate value in the source text.

Crucially, every attribute also has a learned NULL option. The model claims an attribute only when its best source-start score exceeds the abstention score:

$$ m_k=\max_i s_{k,i}-\nu_k,\qquad \mathrm{claim}(k)\iff m_k\gt 0 $$

Here, $s_{k,i}$ scores source position $i$ for attribute $k$, while $\nu_k$ scores the option to abstain.

This makes attribute-presence detection independent across the candidate vocabulary. A field need not survive the model’s sequential construction of a long output list to be recovered.

The generative channel handles the remainder: attributes outside the tagging vocabulary and values that require inference, normalization, or non-contiguous extraction.

During training, an in-vocabulary attribute-value pair is assigned to tagging when its value appears as an exact contiguous token span. Other pairs remain generative. During inference, the selected spans are copied directly, and both channels’ outputs are merged by attribute name. Generation takes precedence when both channels produce the same attribute.

The arrangement is not equivalent to running two complete extraction models independently. Both reuse the same prefill, while routing reduces the tokens that must be generated sequentially.

More attributes recovered with less decoding

The main benchmark uses Qwen3-0.6B, vLLM, matched inference hardware, and 11,132 validation records from an Amazon US catalog.

Metric Generation only Tagging only Hybrid
Attribute recall 69.8% 65.3% 80.2%
LLM-judged precision 97.5% 97.4% 97.1%
Emitted pairs per product 21.3 29.3 34.8
Mean latency 56.6 ms 9.9 ms 31.5 ms
Human quality rating (1–5) 3.93 3.50 4.40

Source: Section 5.1, Table 3. Recall is measured against annotated attributes; precision uses an LLM judge. Emitted-pair counts are not counts of verified correct pairs.

The hybrid improves recall by 10.4 percentage points over generation and emits approximately 63% more attribute-value pairs. Meanwhile, autoregressively generated output falls from 257.8 to 80.0 tokens per product, a reduction of approximately 69%.

The latency reduction is 44.3%, or a 1.80-times speed-up. Bootstrap equivalence testing supports the reported precision equivalence within a one-percentage-point margin. Blinded human ratings across 1,000 products also favor the hybrid.

These results address two different mechanisms. Shorter generated output explains much of the efficiency improvement; independent attribute detection explains why coverage can improve simultaneously.

The frequency breakdown is particularly informative. For attributes ranked 101–500, recall rises from 54.9% to 72.2%. These moderately frequent attributes account for 15.5% of annotated pairs and receive the largest reported frequency-band gain.

For attributes ranked beyond 1,000, however, hybrid recall is 16.8%, below generation’s 20.3%. The parallel channel does not automatically solve long-tail extraction.

The additional tests identify where the advantage stops

The backbone comparisons are robustness checks, not independent evidence of production benefits. Across four tested Qwen and Gemma configurations, hybrid recall exceeds generation by 9.4–13.0 percentage points, with speed-ups ranging from 1.30 to 1.80 times. The gain is therefore not confined to the smallest model.

Two design ablations explain the tradeoffs.

Increasing the tagging vocabulary from 100 to 1,000 raises recall from 69.2% to 80.9% and reduces latency from 37.5 to 29.0 milliseconds. Yet recall barely improves beyond the 500-attribute setting. Vocabulary expansion has diminishing returns even when it transfers additional work away from generation.

Separately, independent start/end pointers deliver 97.2% precision in the span-head ablation, compared with 93.1% for anchored pointers and 87.7% for per-token tagging. Their recall is lower, at 70.5%. This supports a deliberate span-precision tradeoff rather than an across-the-board improvement.

Public benchmark comparisons expose a different workload boundary. AE-110K and OA-Mine use exclusively extractive target values. On AE-110K, hybrid Slot F1 reaches 0.841 versus 0.838 for generation. On OA-Mine, it reaches 0.826 versus 0.810, but trails HPD’s reported 0.876.

Pure tagging is substantially faster on both datasets. These public latency measurements also use HuggingFace Accelerate rather than the main benchmark’s vLLM serving setup, so timings should not be compared directly across the two protocols.

For an entirely copyable schema, a generative channel may add complexity without a corresponding operational benefit.

What catalog operators should validate

The business inference follows from the measured decomposition, not from a reported deployment trial.

An operator processing heterogeneous merchant feeds could first identify frequent attributes whose values ordinarily appear as source spans. These become tagging candidates. Normalized specifications, inferred classifications, and rare attributes remain generative.

Copied and generated results should retain distinct provenance. That allows a catalog-quality workflow to examine source-span errors separately from unsupported inferences, and route conflicting or consequential fields to appropriate exception handling.

Evaluation needs comparable coverage, latency, and correctness measurements across all three configurations. Frequency-band and category-level recall matter because overall recall can conceal weak performance on rare or commercially critical fields. The paper reports hybrid recall improvements in 28 of 30 categories, but this does not establish performance on another merchant’s attribute distribution.

The distinction between textual and semantic correctness is equally consequential. On a 1,000-record exact-match evaluation, hybrid precision is 63.8%, far below its 97.1% LLM-judged precision. Exact matching penalizes valid normalization and formatting differences, while semantic judging introduces its own evaluation uncertainty.

Reliability is not established by the precision headline

The primary training and validation annotations were generated automatically, and the main precision measure relies on an LLM judge, although the authors audited agreement against human judgments.

A separate grounding audit estimates genuine fabrication at approximately 0.12% for generation and 0.04% for the hybrid. The authors explicitly regard this difference as within sampling noise. It is not evidence of a statistically established reduction in hallucination.

The research also excludes image-derived attributes and does not test multilingual merchant feeds, cross-domain transfer, live catalog correction costs, search conversion, or production reliability. Reported proportions of extractive and taggable values differ across the corpus description and statistics table without a supplied reconciliation, reinforcing the need to measure workload composition locally.

Extraction completeness is partly an architecture decision

The central finding is that broader extraction does not necessarily require longer generation. A model can recover more attributes by independently checking copyable fields while reserving sequential decoding for values that genuinely require it.

For high-volume catalogs containing both explicit and interpreted attributes, that division deserves testing before investing exclusively in larger generative models. The relevant comparison is not model size alone, but coverage, semantic correctness, latency, and the share of attributes each channel can reliably handle.

The benchmark establishes a compelling technical case for hybrid decoding in the tested setting. Whether it creates operational value depends on the catalog’s actual attribute distribution, error costs, and downstream requirements.


Cognaptus: Automate the Present, Incubate the Future.


  1. Keshav Kolluru and Manas Gupta and Maroof Ahmad and Suchitra Sathyanarayana (2026). Hybrid Generative-Tagging Decoding for Efficient Product Attribute-Value Extraction. arXiv:2610.21329. https://arxiv.org/abs/2610.21329 ↩︎