Hybrid Generative-Tagging Decoding for Efficient Product Attribute-Value Extraction

Keshav Kolluru; Manas Gupta; Maroof Ahmad; Suchitra Sathyanarayana

cs.CLOct 15, 2026

Hybrid Generative-Tagging Decoding for Efficient Product Attribute-Value Extraction

Authors: Keshav Kolluru, Manas Gupta, Maroof Ahmad, Suchitra Sathyanarayana

Organizations: Amazon

Abstract

Attribute-value extraction is important for building high-quality product catalogs in e-commerce. Existing systems either rely on generative or tagging strategies for extracting the attribute names and their corresponding values. Tagging models directly label spans in the source, making them fast and grounded, but they cannot produce values that require reformatting or generate rare attribute names outside their label space. Generative models avoid these drawbacks by generating both attribute names and values token by token, but this auto-regressive nature makes them slow and limits attribute coverage. We introduce hybrid generative-tagging decoding, a novel model architecture that combines tagging source spans and open-ended generation. For frequent attributes whose values are copyable, a tagging head extracts values directly from the source. This lets the generative head focus on the remainder - rare attributes outside the tagger's label space and non-copyable values, including implicit or normalized values. Integrated into vLLM inference, the hybrid decoder raises attribute coverage by 10% and extracts 63% more correct attribute-value pairs at no precision cost, while reducing latency by 44%.

Figures & tables

Product: “Mrs Ball’s Original Chutney. Ingredients: Sugar, Vinegar, Water, Peaches, Apricots, Modified Maize Starch, Spices. Imported from South Africa.”
Taggable attributes (label source span)
brand	Mrs Ball’s
product_name	Original Chutney
ingredients	Sugar, Vinegar, Water, Peaches, Apricots, Modified Maize Starch, Spices
Generative attributes (reformat / implicit)
product_type	condiment

Table 1: A product listing with both attribute types. Taggable values are contiguous spans of the source and can be copied verbatim; generative values must be reformatted or are only implied by the text.

Figure 1: Hybrid decoder with two decoupled channels over a shared prefill. During prefill the backbone maps the source tokens to hidden states $H$ . The generative channel ( [GEN] ) autoregressively decodes only the remainder—generative, out-of-vocabulary, and non-copyable attributes—as inline name : value pairs. In parallel, the tagging channel ( [TAG] ) runs a span head over $H$ for all $K$ vocabulary attributes at once, claiming an attribute against a learned per-attribute NULL score and copying the chosen source span verbatim. The two channels are merged by attribute name, preferring the generative value on collision.

Split	#Products	Avg. AVE	Extr%	Gen%
Train	100,195	22.3	88.3	11.7
Validation	11,132	22.4	88.2	11.8

Table 2: Statistics of the corpus. Both splits average 22 attribute-value pairs per product, of which about 88% are taggable and 12% generative.

Model	Precision	Recall	Quality ( $\mu\pm\sigma^{2}$ )	Num Extr.	Speed (ms)
Generative	97.5	69.8	3.93 $\pm$ 0.57	21.3	56.6 (1.00 $\times$ )
Tagging	97.4	65.3	3.50 $\pm$ 0.51	29.3	9.9 ( 5.69 $\times$ )
Hybrid	97.1	80.2	4.40 $\pm$ 0.43	34.8	31.5 (1.80 $\times$ )

Table 3: Main Results. The proposed hybrid decoding strategy achieves the best attribute recall and quality among all the covered models. It also achieves 1.8

\times

speed-up over generative model while maintaining precision. Precision : LLM-judged correctness of emitted pairs; Recall : fraction of ground-truth attributes recovered; Quality : blinded

1

–

5

rating (

\mu\pm\sigma^{2}

); Num Extr. : mean attribute-value pairs per record; Speed : wall-clock time (in milli-seconds).

	AE-110K		OA-Mine
System	F1	Speed	F1	Speed
Generative	0.838	130.8 (1.0 $\times$ )	0.810	254.6 (1.0 $\times$ )
Tagging	0.830	8.1 ( 16.2 $\times$ )	0.818	7.4 ( 34.4 $\times$ )
Hybrid	0.841	62.4 (2.1 $\times$ )	0.826	114.7 (2.2 $\times$ )
MVP-RAG	0.836	142.7 (0.9 $\times$ )	0.826	275.2 (0.9 $\times$ )
GenToC	0.705	79.1 (1.7 $\times$ )	0.775	120.7 (2.1 $\times$ )

Table 4: Evaluation on public AE-110K and OA-Mine. We report Slot F1 and Speed in milliseconds. The speed-up over the generative baseline is in parentheses; HPD does not report absolute latency.

Backbone	System	Prec.	Recall	Speed (ms)
Qwen3-0.6B	Generative	97.5	69.8	56.6 (1.00 $\times$ )
	Tagging	97.4	65.3	9.9 ( 5.69 $\times$ )
	Hybrid	97.1	80.2	31.5 (1.80 $\times$ )
Qwen3-1.7B	Generative	98.2	70.4	71.5 (1.00 $\times$ )
	Tagging	97.7	65.1	20.2 ( 3.54 $\times$ )
	Hybrid	97.4	79.8	43.1 (1.66 $\times$ )

Table 5: Evaluation using different backbone LLMs spanning three model sizes and two model families.

Delimitation	Prec	Rec	Num. Extr.
Per-token tagging	87.7	74.1	36.3
Anchored pointer	93.1	71.5	34.3
Independent pointer	97.2	70.5	34.8

Table 6: Span head ablations using three variations. Prec : LLM-judged precision; Rec : recall; Num. Extr. : mean pairs per record. We use Independent pointer in our final Hybrid decoding strategy due to its balance between precision and recall.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

System	Substring-flag	Genuine
Generative	3.0%	$\approx$ 0.12%
Hybrid	1.3%	$\approx$ 0.04%

Appendix

Table 7: Hallucination on extractive-type attributes,

n{=}1000

records. Substring-flag : fraction of emitted extractive-type values not found verbatim in the source under a format-insensitive match (lowercase, strip punctuation, normalize decimals). Genuine : rate of true fabrication after a 250-case manual audit removes formatting and unit-conversion artifacts. The substring metric is structurally biased toward the hybrid because its tagging channel copies values verbatim while the generative channel reformats them; the genuine rates are the audited, artifact-corrected figures.

System	Precision	F1
Generative	54.5	69.9
Tagging	56.1	71.2
Hybrid	63.8	77.0

Appendix

Table 8: Precision using exact match against gold,

n{=}1000

records. Precision : fraction of emitted pairs whose value matches the gold value string exactly; F1 : harmonic mean of per-pair precision and per-pair recall. Unlike the LLM-judged precision in Table 3 , exact match penalizes any surface-form rephrasing.

$K$	Vocab cov.	Recall	Speed (ms)
100	70.0	69.2	37.5
250	82.0	76.2	33.3
500	87.8	80.2	31.5
1000	92.0	80.9	29.0

Appendix

Table 9: Effect of the tagging vocabulary size

K

on the hybrid. Vocab cov. : fraction of gold attribute–value pairs whose attribute is in the top-

K

vocabulary; Recall and Speed as in Table 3 .

Category	Vocab cov.	Recall gain
grocery	90.9%	$+11.5$
drugstore	88.2%	$+10.4$
wireless	92.3%	$+13.4$
home	89.2%	$+9.6$
pet products	88.5%	$+8.5$

Appendix

Table 10: Category-level breakdown for five representative categories. Vocab cov. : fraction of the category’s gold pairs covered by the

K{=}500

vocabulary; Recall gain : hybrid minus generative recall.

Frequency band	% pairs	Gen	Tag	Hybrid
rank 1–100	73.7	79.5	75.5	90.4
rank 101–500	15.5	54.9	62.2	72.2
rank 501–1000	3.5	41.0	0.0	39.0
rank $>$ 1000	6.2	20.3	0.0	16.8

Appendix

Table 11: Recall by attribute frequency band. % pairs : share of gold pairs in the band; remaining columns are per-band recall for each system.

Product Attribute Extraction

See all 15 papers

Breaking the Autoregressive Chain: Hyper-Parallel Decoding for Efficient LLM-Based Attribute Value Extraction
Apr 29, 2026Theodore Glavas, Nikhita Vedula, Dushyanta Dhyani +2LLM Inference Acceleration Parallel Decoding
Amazon.com, Inc. · McGill University · Mila - Quebec AI Institute +1
Scaling E-Commerce Attribute Extraction with Parallel Decoding
Sep 9, 2026Nikhita Vedula, Dushyanta Dhyani, Bryan Wang +1Document Information Extraction Product Attribute Extraction
Amazon.com, Inc.
BEATS: Bootstrapping E-commerce Attribute Taxonomies for Search through Iterative Human-AI Collaboration
Jun 3, 2026Yung-Yu Shih, Shang-Yu Su, Tzu-I Ho +2Semantic Search LLM-Assisted Annotation
National Taiwan University · Rakuten Group, Inc. · Taiwan Rakuten Ichiba, Inc. +1

Apr 29, 2026J/K move · Enter open · S save

Breaking the Autoregressive Chain: Hyper-Parallel Decoding for Efficient LLM-Based Attribute Value Extraction

Theodore Glavas, Nikhita Vedula, Dushyanta Dhyani, Yilun Zhu+1

Amazon.com, Inc. · McGill University · Mila - Quebec AI Institute+1

LLM Inference Acceleration Parallel Decoding LLM Information Extraction

Open paper AskPDF arXiv

Some text generation tasks, such as Attribute Value Extraction (AVE), require decoding multiple independent sequences from the same document context. While standard autoregressive decoding is slow due to its sequential nature, the independence between output sequences offers an opportunity for parallelism. We present Hyper-Parallel Decoding, a novel decoding algorithm that accelerates offline decoding by leveraging both shared memory and computation across batches. HPD enables out-of-order token generation through position ID manipulation, significantly improving efficiency. Experiments on AVE show that attribute-value pairs are conditionally independent, enabling us to parallelize value generation within each prompt. By further stacking multiple documents within a single prompt, we can decode in parallel up to 96 tokens per prompt. HPD works with all LLMs, and reduces both inference costs and total inference time by up to 13.8X without compromising output quality, potentially saving hundreds of thousands of dollars on industry AVE tasks. Although designed for attribute extraction, HPD makes no assumptions unique to the AVE domain and can in theory be applied to other scenarios with independent output structures.

Abstract

Product Attribute Extraction

Explore similar work

Breaking the Autoregressive Chain: Hyper-Parallel Decoding for Efficient LLM-Based Attribute Value Extraction

Scaling E-Commerce Attribute Extraction with Parallel Decoding

BEATS: Bootstrapping E-commerce Attribute Taxonomies for Search through Iterative Human-AI Collaboration