Sensory-Aware Sequential Recommendation via Review-Distilled Representations

Yeo Chan Yoon; Chanjun Park; Kyuhan Koh

cs.CLMar 3, 2026

Sensory-Aware Sequential Recommendation via Review-Distilled Representations

Authors: Yeo Chan Yoon, Chanjun Park, Kyuhan Koh

Organizations: Jeju National University, Jeju, Jeju, Korea · Soongsil University, Seoul, Korea · California State University Stanislaus, Turlock, USA

Abstract

Sequential recommenders learn behavioral patterns from item identifiers, while the experiential properties that users describe in reviews, such as how products look, feel, smell, taste, or sound, rarely enter item representations in a controlled, auditable form. We present ASER (Attribute-based Sensory-Enhanced Representation), an offline pipeline that fine-tunes a large language model to extract evidence-grounded sensory attribute-value records, such as color: matte black or scent: vanilla, from review text and distills them into a compact student encoder that produces a frozen five-facet sensory bank for each item catalog. At recommendation time the pretrained backbone stays frozen: a lightweight relational metric between the user history and each candidate is learned over the bank, and its correction is applied within a validation-selected magnitude bound. Across five Amazon domains and four backbones, trained within a common experimental pipeline and evaluated by full-catalog leave-one-out ranking without sampled negatives, this integration improves HR@10 and NDCG@10 in all 20 domain-backbone pairs, with average relative gains of 6.1% and 6.4%. A matched non-sensory control channel, built with the same seed model, schema, and pipeline, separates the sources of the gain: the hit-rate improvement follows from structured, evidence-grounded extraction as such, whereas the sensory vocabulary yields a ranking-quality advantage in eight of nine matched comparisons. An audit of the Beauty evaluation catalog finds that 94.8% of retained records are supported by their cited evidence spans, so the extracted signal remains inspectable against its source text.

Figures & tables

Figure 1: Offline stage of ASER: sensory extraction, facet supervision, and student distillation. Stage 1 applies a sensory teacher to raw item text and produces structured sensory JSON records containing attribute, value, evidence, polarity, negation, and confidence. Stage 2 converts the cleaned teacher outputs into facet-level supervision over five sensory facets, including token evidence masks, facet presence labels, polarity labels, confidence scores, example-level teacher embeddings, and canonical facet targets. Stage 3 trains a multi-facet sensory student. The student encodes text with DeBERTa-v3-Small, applies a per-domain bottleneck adapter, predicts token-level evidence for each facet, derives facet presence with noisy-or pooling, obtains grounded facet representations by evidence-weighted token pooling, and produces example embeddings, canonical embeddings, polarity logits, and confidence scores through facet-specific heads. The trained student is then run offline over catalog texts to construct an item-level facet bank containing canonical facet embeddings, confidence, polarity, and coverage statistics. All three stages run once, before recommender training; neither the teacher nor the student is invoked at recommendation time.

Symbol	Meaning
$\mathcal{F}$	the five sensory facets (Table 4 )
$\mathcal{C}$	content-token mask (item text, excluding prefixes)
$\mathcal{T}_{i}$	set of texts associated with item $i$
$\mathcal{I}$ , $\mathcal{U}$	evaluated item catalog; evaluated user set
$L$	encoder input sequence length; recommender history length
$d_{h}$	text-encoder hidden size ( $768$ )

Table 1: Notation used throughout the paper.

	Pair P	Evid. F	Pol. Acc	Neg. Acc
Mean	0.88	0.90	0.80	0.92

Table 2: Seed label audit on a random 10% sample.

Domain	Raw	Kept	Drop rate (%)
Beauty	21,196	19,612	7.47
Grocery	20,983	17,804	15.15
Sports	22,786	20,435	10.32
Toys	22,174	19,913	10.20
VideoGames	18,813	15,911	15.43
Total	105,952	93,675	11.59

Table 3: Attribute records per domain as returned by the seed model and after evidence-grounded cleaning. Counts are over the 10,882 items of the seed corpus, one packed text per item.

Facet	Fine-grained attributes	#	Conf. thr.
Visual	color , pattern , shape , graphics ,	9	0.60
	brightness , glossiness , transparency ,
	finish , design
Tactile	texture	1	0.60
Auditory	sound	1	0.72
Olfactory	scent	1	0.60

Table 4: Mapping from the 13 fine-grained teacher attributes to the five sensory facets, with the per-facet confidence threshold applied during teacher-label cleaning. The mapping is fixed a priori, is shared by all five domains, and is the only attribute-to-facet routing used anywhere in the pipeline.

Statistic	Avg Tokens	Median	P95	Max
Packed item text	497.6	542	632	1,011

Table 5: Token-length statistics of the packed item texts that form the student training corpus, measured with the DeBERTa tokenizer before the 512-token input limit is applied. Each item contributes one such text.

Facet	Beauty	Toys	Sports	Games	Groc.
Visual	85.45	97.19	97.34	87.91	56.56
Tactile	89.71	88.24	95.40	70.59	83.84
Auditory	10.29	42.86	23.13	49.64	3.00
Olfactory	66.12	19.79	29.90	4.06	52.06
Gustatory	8.51	1.50	10.89	0.05	94.53

Table 6: Facet coverage across domains in the student training corpus, given as the percentage of supervision examples in that domain for which the facet is present. Each example is one item.

Figure 2: Runtime stage of ASER (Stage 4): bounded trust-region reranking with the frozen facet bank. The sequential backbone is frozen and scores the catalog exactly as its ID-only baseline does. In parallel, the precomputed item-level facet bank entry $B_{i}$ is retrieved by item lookup for each history item, each stored canonical facet embedding is mapped by the learned per-facet projection $P_{f}$ (Eq. 30 ), and one attention pass per facet, biased by evidence quality and recency (Eqs. 29 – 32 ), produces a per-facet user prototype. The relational score of a candidate is the facet-weighted sum of cosine agreements between prototype and candidate (Eq. 33 ), which is bounded a priori by the learned facet weights. Base score and relational score are standardized per query and combined as a residual that is applied only inside the base top- $N$ trust region (Eq. 36 ), so the residual is expressed in base-score standard deviations and cannot promote an item the backbone ranks below $N$ . No language model or text encoder is invoked at recommendation time.

Domain	Users	Items	Interactions	Avg.Len
Beauty	22,363	12,101	198,502	8.9
Sports	35,598	18,357	296,337	8.3
Toys	19,412	11,924	167,597	8.6
Video Games	24,303	10,672	231,780	9.5
Grocery	14,681	8,713	151,254	10.3

Table 7: Dataset statistics for Amazon 2014 5-core (Beauty, Sports & Outdoors, Toys & Games, Video Games, Grocery & Gourmet Food). Avg. Len is the average number of interactions per user.

Domain	Model	Setting	HR@5	HR@10	HR@20	NDCG@5	NDCG@10	NDCG@20	Mean
Beauty	SASRec	Base	6.40	8.76	11.86	4.66	5.42	6.20	7.21
	SASRec	ASER	6.61 $\uparrow$ 3.3%	9.16 $\uparrow$ 4.5%	12.59 $\uparrow$ 6.1%	4.74 $\uparrow$ 1.6%	5.56 $\uparrow$ 2.6%	6.42 $\uparrow$ 3.6%	7.51 $\uparrow$ 4.1%
	BERT4Rec	Base	4.09	6.28	9.24	2.69	3.39	4.14	4.97
	BERT4Rec	ASER	4.36 $\uparrow$ 6.4%	6.69 $\uparrow$ 6.5%	9.99 $\uparrow$ 8.1%	2.88 $\uparrow$ 7.1%	3.63 $\uparrow$ 7.1%	4.46 $\uparrow$ 7.8%	5.33 $\uparrow$ 7.3%
	BSARec	Base	6.81	9.31	12.51	4.89	5.70	6.50	7.62
	BSARec	ASER	7.00 $\uparrow$ 2.8%	9.76 $\uparrow$ 4.8%	13.09 $\uparrow$ 4.6%	5.03 $\uparrow$ 2.9%	5.93 $\uparrow$ 4.0%	6.77 $\uparrow$ 4.1%	7.93 $\uparrow$ 4.1%

Table 8: Full-ranking recommendation performance (%) on five domains. All 20 domain–backbone cells are evaluated within a common data and evaluation pipeline: the four backbones use the same preprocessing, leave-one-out split, full-catalog evaluator, and matched model capacity where permitted by their implementations; absolute values obtained under different preprocessing, capacity, or evaluator configurations are not comparable to this table. Base is the ID-only backbone; the DIFF side-information channels are disabled, so all four backbones are ID-only. ASER is the bounded relational integration of the frozen sensory facet bank: the backbone is kept frozen, only the sensory metric is trained, and the trust-region strength

\lambda

and neighborhood size are selected on the validation split. Every cell is one run at seed 42; relative changes are computed from the unrounded values and the rightmost Mean column is the arithmetic average of the six reported metrics for that row. ASER improves HR@10 and NDCG@10 in all 20 domain–backbone cells, and, in fact, all 120 reported metric values increase. Seed-to-seed variability for the three representative domains (Beauty, Toys, Games) is reported in Table 9 .

Domain	Metric	Base	ASER (mean $\pm$ sd)	$\Delta$ (mean $\pm$ sd)	$\Delta$ relative
Beauty	HR@10	9.31	9.708 $\pm$ 0.054	+0.398 $\pm$ 0.054	+4.27% $\pm$ 0.58
Beauty	NDCG@10	5.70	5.915 $\pm$ 0.011	+0.218 $\pm$ 0.011	+3.82% $\pm$ 0.20
Toys	HR@10	9.59	10.078 $\pm$ 0.020	+0.491 $\pm$ 0.020	+5.12% $\pm$ 0.20
Toys	NDCG@10	6.06	6.294 $\pm$ 0.012	+0.232 $\pm$ 0.012	+3.83% $\pm$ 0.20
Games	HR@10	13.74	14.023 $\pm$ 0.096	+0.280 $\pm$ 0.096	+2.04% $\pm$ 0.70
Games	NDCG@10	7.62	7.847 $\pm$ 0.034	+0.223 $\pm$ 0.034	+2.92% $\pm$ 0.44

Table 9: Seed stability of the bounded relational integration on the three representative domains with the BSARec backbone, under the same pipeline, the same leave-one-out split and the same full-catalog ranking protocol as Table 8 . Values are percentages; absolute scores are given to two decimals and the deltas, which are in percentage points, to three. Mean and sample standard deviation are taken over three seeds, 42, 43 and 44, of the integration stage. The backbone checkpoint is frozen and shared across the three seeds, so the Base column is identical for all of them and the reported spread measures the variability of the integration stage rather than of backbone training; the trust-region strength and neighborhood size are reselected on validation independently for each seed. All nine runs, three domains by three seeds, improve both HR@10 and NDCG@10 over the base backbone. The standard deviation is between 4% and 35% of the corresponding mean gain; it is smallest for Toys HR@10 and largest for Games HR@10, which is the least stable of the six entries.

Metric	Precision	Recall	F1
Exact match	0.4165	0.4398	0.4278
Attribute match	0.7557	0.7609	0.7583
Semantic value match, 0.8	0.5157	0.5444	0.5297
Semantic value match, 0.7	0.5728	0.6047	0.5883
Semantic value match, 0.6	0.6120	0.6461	0.6286
Facet-class match	0.8523	0.8379	0.8450

Table 10: Teacher alignment between Qwen and GPT-5 Mini on the validation split. Exact match requires exact attribute and value equality after normalization. Attribute match evaluates agreement on the fine-grained extracted attribute only. Semantic value match evaluates value agreement using cosine similarity thresholds of 0.8, 0.7, and 0.6. Facet-class match evaluates agreement after mapping each extracted record to one of the five coarse sensory facets.

	HR@10			NDCG@10
Domain	Base	ASER	Temporal	Base	ASER	Temporal
Beauty	0.0931	0.0976	0.0971	0.0570	0.0593	0.0585
Grocery	0.0975	0.0999	0.1028	0.0565	0.0579	0.0588
Video Games	0.1374	0.1413	0.1412	0.0762	0.0789	0.0789

Table 15: Temporal-control evaluation with BSARec. Temporal ASER rebuilds the sensory channel from supervision items that have at least one review dated on or before the 2014-07-23 evaluation cutoff (51.1% of the supervision training split); the backbone checkpoints are identical to the main table. The improvement over the ID-only base persists in all three domains.

Facet	Support	Precision	Recall	F1
visual	447	0.9032	0.9597	0.9306
tactile	413	0.8692	0.9492	0.9074
auditory	64	0.7647	0.4062	0.5306
olfactory	61	0.8222	0.6066	0.6981
gustatory	91	0.8713	0.9670	0.9167

Table 16: Per-facet presence scores of the released facet-aware student on the supervision validation split (545 examples, presence threshold 0.5). Support is the number of positive examples per facet. Macro F1 is 0.7967 and micro F1 is 0.8909.

Variant	HR@10	$\Delta$	NDCG@10	$\Delta$
A. Full integration	0.0976	–	0.0593	–
B. No quality bias	0.0976	+0.0000	0.0594	+0.0001
C. No recency bias	0.0970	$-$ 0.0006	0.0587	$-$ 0.0005
D. Mean history pooling	0.0964	$-$ 0.0012	0.0590	$-$ 0.0002
E. Uniform facet weights	0.0970	$-$ 0.0006	0.0584	$-$ 0.0009
F. No trust region	0.0976	+0.0000	0.0593	+0.0000

Table 22: Component ablation of the bounded relational integration on Beauty with BSARec (seed 42). Deltas are against the full integration (A). Differences smaller than 0.001 are within the reproduction noise measured across seeds and are not interpreted.

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Component	Parameters
Facet-aware student encoder	154,715,417
Encoder	141,304,320
Adapters and prediction heads	13,411,097

Appendix

Table 23: Measured parameter counts of the facet-aware student.

Hyperparameter	Value	Description
$\lambda_{\mathrm{visual}}$	1.0	Facet loss weight
$\lambda_{\mathrm{tactile}}$	1.0	Facet loss weight
$\lambda_{\mathrm{auditory}}$	0.4	Facet loss weight
$\lambda_{\mathrm{olfactory}}$	0.8	Facet loss weight
$\lambda_{\mathrm{gustatory}}$	0.4	Facet loss weight
$\tau_{\mathrm{pres}}$	0.5	Presence threshold

Appendix

Table 24: Facet-aware student loss and optimization hyperparameters. The facet weights

\lambda_{f}

are global loss-mixing coefficients for all-domain student training, not inverse-frequency class weights.

Hyperparameter	Value	Description
LoRA rank $r$	16	Low-rank adaptation dimension
LoRA alpha	32	LoRA scaling factor
LoRA dropout	0.05	LoRA dropout rate
Target modules	q_proj , k_proj , v_proj o_proj , gate_proj up_proj , down_proj	Attention and feed-forward projection modules

Appendix

Table 25: Teacher LoRA fine-tuning configuration.

Recommender Systems

Recommender systems predict and rank items for people using interaction histories, content, context, and feedback.

See all 514 papers

Is Serving-Time Language-Model Inference Necessary for Recommendation? Evidence from Generative and Agentic Settings
Aug 2, 2026Xiang Chen, Zhongke He, Hongwei Zeng +4Sequential Recommendation Algorithmic Auditing
Tongji University · Dalian University of Technology · University of Chinese Academy of Sciences +3
CALMRec: Causally Aligned Language Memory for Long-Horizon Recommendation
Jul 26, 2026Gengyu ZhanMemory-Augmented Language Models Recommender Systems
Shenzhen University
Beyond Ranking Accuracy: Evaluating LLM-Cited Feature Rationales for Next Basket Repurchase Recommendation
Aug 31, 2026Yanan Cao, Anay Dombe, Murali Mohana Krishna Dandu +5Sequential Recommendation LLM Evaluation
Walmart Global Tech, Sunnyvale, CA, United States

Aug 2, 2026J/K move · Enter open · S save

Is Serving-Time Language-Model Inference Necessary for Recommendation? Evidence from Generative and Agentic Settings

Xiang Chen, Zhongke He, Hongwei Zeng, Kai Deng+3

Tongji University · Dalian University of Technology · University of Chinese Academy of Sciences+3

Sequential Recommendation Algorithmic Auditing Recommender Systems

Open paper AskPDF arXiv

Language models (LMs) are increasingly used in recommender systems to generate recommendations or coordinate recommendation tools. However, does their strong performance imply that serving-time LM inference is necessary? We challenge this seemingly reasonable but unverified assumption. To investigate this question, we introduce recovery-based mechanism auditing, which tests whether independent alternatives can achieve the accuracy of LM-based recommender systems without the audited inference mechanisms under controlled information and evaluation conditions. For generative recommendation, we propose LIME-Rec, which combines sequential, collaborative, and offline semantic information without autoregressive LM decoding. On three Amazon Reviews datasets (Beauty, Toys, and Sports), LIME-Rec achieves the highest reported mean in all 12 dataset-metric combinations, with up to a 15.2% relative improvement in Recall@10 over GRAM, an LM-based generative recommender. For agentic recommendation, we propose DART-Rec, which replaces LM-based tool control with deterministic fusion and fixed expert selection. On 1,000 Beauty users, given the same candidate items and expert scores, deterministic fusion achieves a 16.7% relative improvement in Recall@10 over a fine-tuned LM controller. With at most two tool calls, fixed expert selection matches the LM controllers' Hit@10 outcomes for every user in 17 of 18 comparisons. These findings demonstrate that strong recommendation performance does not necessarily require serving-time LM inference: under the evaluated task and information conditions, alternatives without the specified inference mechanisms can match or exceed the accuracy of their LM-based counterparts. To support reproducibility, code is available at https://github.com/Double-wk/LIME-Rec.

Abstract

Recommender Systems

Explore similar work

Is Serving-Time Language-Model Inference Necessary for Recommendation? Evidence from Generative and Agentic Settings

CALMRec: Causally Aligned Language Memory for Long-Horizon Recommendation

Beyond Ranking Accuracy: Evaluating LLM-Cited Feature Rationales for Next Basket Repurchase Recommendation