HARP: Training-Free Dual-Profile Agentic Communication for LLM-Based Recommendation

Zhen Tao; Xun Zhou; Xuhui Chen; Ziyue Qiao; Qingqiang Sun

cs.IRNov 3, 2026

HARP: Training-Free Dual-Profile Agentic Communication for LLM-Based Recommendation

Authors: Zhen Tao, Xun Zhou, Xuhui Chen, Ziyue Qiao, Qingqiang Sun

Organizations: Harbin Institute of Technology, Shenzhen · Great Bay University · Shenzhen Loop Area Institute

Abstract

In LLM-based recommendation, memory acts as a user-level agent in a memory-augmented backbone. Yet users and items remain asymmetric: only the user is given a structured profile; it enters a bandwidth-limited reranker prompt; and it is frozen after warmup. We recast LLM-based recommendation as an agentic-communication process in four stages: BUILD a self-description, TRANSMIT it along the collaborative graph, UNDERSTAND the other side, and SYNC it from the next interaction. This formulation reveals three unresolved communication challenges: the item side never produces a comparable self-description (one-sided BUILD); profile-level signals enter only the bandwidth-limited reranker prompt rather than retrieval, where they would not compete for context tokens (prompt-trapped UNDERSTAND); and the user's description is frozen after warmup with no principled update rule (absent SYNC). We instantiate the missing stages as HARP (Homophily-aware Agentic Recommendation via dual Profiles), a training-free, three-operator addition to a frozen LLM backbone: a rule-extracted symmetric item profile (zero extra LLM calls), a channel-decomposable profile-level homophily H(phi_u, phi_i) consumed at retrieval time rather than inside the reranker prompt, and a single-pass reflective update realising in-context profile evolution without any gradient step. On four public benchmarks (Yelp, Books, MovieTV, Goodreads), HARP lifts Hit@1 over the strongest published baseline by 6.9%-24.0%, with family-wise Holm-corrected p < 1e-9 on every dataset. Ablations and a per-case attribution study show that the operators contribute complementary gains, with the retrieval-side gate as the main source; we also report the trade-off introduced by the reflective pass. We formulate agentic communication and provide a training-free instantiation. Code is available at https://github.com/Jove-Zhen/HARP.

Figures & tables

Figure 1: Harp at a glance. A user with a niche query in a half-dialogue session, compared with the full dialogue restored by Harp . HARP integrates three training-free innovations: Dual Profile (§ 3.2 ), Homophily Gate (§ 3.3 ), and Reflection (§ 3.4 ).

Figure 2: Harp pipeline overview. Upper : story strip illustrating the flow of a user’s query across the entire pipeline. Lower : block-exploded technical view with input/output annotations and equation references.

Channel $c$	User field $\phi_{u}^{(c)}$	Item field $\phi_{i}^{(c)}$
preference	interests	serves
quality	quality_bar	quality_signal
context	context	context
audience	behavior	audience

Table 1: Four-channel correspondence between user and item profiles.

		Books					Goodreads
Paradigm	Model	H@1	H@3	N@3	H@5	N@5	H@1	H@3	N@3	H@5	N@5
Trad.	LightGCN †	.175	.326	.260	.570	.359	.250	.588	.443	.790	.526
	SASRec †	.091	.283	.200	.485	.282	.132	.352	.258	.541	.335
	P5 †	.219	.361	.299	.527	.367	.157	.323	.251	.506	.326
Agent (match)	Vanilla LLM †	.314	.562	.453	.727	.523	.286	.466	.395	.739	.504
	iAgent †	.393	.556	.486	.691	.541	.262	.495	.395	.659	.463

Table 2: Main results on four benchmarks. Bold

=

best, underline

=

second best; † = from original paper. “Rel. Improv.” = relative improvement over the strongest published baseline. All H@1 improvements

p<10^{-9}

(Holm–Bonferroni corrected).

Config	$\phi_{u}$	$\phi_{i}$	CH	RI	Sync	H@1	H@3	H@5	N@3	N@5
Memory						.580	.761	.855	.683	.722
Monologue	✓			✓		.578	.738	.824	.671	.707
Consult	✓	✓		✓		.580	.750	.842	.678	.716
Dialogue	✓	✓	✓			.635	.804	.888	.728	.763
Harp	✓	✓	✓	✓	✓	.643	.808	.889	.734	.767

Table 3: Harp ablation (Yelp subsplit). CH = retrieval-side homophily gate; RI = reranker prompt injection; Sync = reflection.

Figure 3: Five-rung ablation ladder (four benchmarks, matched

1{,}000

-user subsplits). (a) Hit@1; (b) Hit@3; (c) NDCG@3. Red numbers = total gain (pp) over Memory .

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset	$n$	hits	Harp H@1	Wilson 95% CI	MemRec	$p$ -value
Books	7,377	4,034	0.5468	$[0.5355,\,0.5582]$	0.5117	$7.8\!\times\!10^{-10}$
Goodreads	11,734	5,498	0.4686	$[0.4595,\,0.4776]$	0.3997	$<10^{-16}$
MovieTV	5,649	3,641	0.6445	$[0.6320,\,0.6569]$	0.5882	$<10^{-16}$
Yelp	2,950	1,782	0.6041	$[0.5863,\,0.6216]$	0.487	$<10^{-16}$

Appendix

Table 4: Harp vs. MemRec, per-dataset binomial test on the full test split.

n

and hits are computed from our evaluation on the full InstructRec test splits; MemRec numbers are from the MemRec paper’s Tables 2–3.

Figure 4: Per-metric radar visualisation of the main results. Each subplot is one metric (H@1, H@3, N@3, H@5, N@5); the four vertices correspond to the four benchmarks; each polygon is one method. Harp (red, bold) is the outermost polygon at every vertex of every metric; MemRec (dark blue, up-triangle) is the closest competitor per benchmark but lies strictly inside Harp .

Figure 5: Cross-dataset five-rung ladder on four metrics (H@1, H@3, N@3, N@5). Each subplot is one metric; each line within a subplot is one dataset (Yelp in red diamonds). The Monologue

\to

Dialogue transition is the largest single-rung jump on Yelp, MovieTV and Books; the Consult rung falls back near Memory ; Harp adds a small residual from prompt injection and Sync (positive on Yelp/MovieTV/Goodreads, slightly negative on Books). The ordering is stable across all four metrics.

Figure 6: Repair source decomposition: each Harp -over- Memory repair is attributed to Pillar 2 (Dial-rep) or Pillar 3 (Sync-rep). Dark green = Pillar 2, light green = Pillar 3; the right-hand annotation gives the total number of repairs and Pillar 2’s share.

Transition	H@1 A	H@1 B	only $A\checkmark$	only $B\checkmark$	Gap (pp)	$p$ -value
Memory $\to$ Monologue	$0.580$	$0.578$	$79$	$77$	$-0.20$	$0.94$
Monologue $\to$ Dialogue	$0.578$	$0.635$	$51$	$108$	$+5.70$	$7.2\!\times\!10^{-6}$
Dialogue $\to$ Consult	$0.635$	$0.580$	$104$	$49$	$-5.50$	$1.0\!\times\!10^{-5}$
Consult $\to$ Harp	$0.580$	$0.643$	$43$	$106$	$+6.30$	$2.5\!\times\!10^{-7}$
Memory $\to$ Harp (overall)	$0.580$	$0.643$	$62$	$125$	$+6.30$	$4.8\!\times\!10^{-6}$

Appendix

Table 5: Pairwise McNemar tests across adjacent configurations on the deep-analysis split (

1{,}000

-user Yelp). “only

A\checkmark

” = users for whom

A

hits and

B

misses; “only

B\checkmark

” = reverse; Gap

=\text{H@1}_{B}-\text{H@1}_{A}

in pp;

p

-value is the exact two-sided binomial under

b=c

. The non-significant row Memory

\to

Monologue serves as a designed negative control : injecting the profile into the reranker prompt without changing retrieval scoring has no measurable effect on Hit@1, which isolates the Monologue

\to

Dialogue step (moving the same profile to the retrieval-side homophily gate) as the locus of gain. Dialogue

\to

Consult is the corresponding reverse-ablation : switching the retrieval-side gate back to prompt-side injection incurs a significant drop of

-5.5

pp (

p=1.0\times 10^{-5}

), providing directional confirmation that Dialogue is the component responsible for the gain.

Symbol	Meaning
$\mathcal{V}_{U},\mathcal{V}_{I}$	User and item sets; $N_{U},N_{I}$ their sizes
$H_{u}$ , $\mathcal{I}_{u}$	User $u$ ’s interaction history and evaluation instruction
$\mathcal{N}_{U}(u),\mathcal{N}_{I}(u)$	$k$ -nearest user/item neighbours of $u$ from a fixed collaborative embedding
$\mathcal{C}$	Channel set $\{\mathrm{preference},\mathrm{quality},\mathrm{context},\mathrm{audience}\}$ , $\|\mathcal{C}\|=4$
$\phi_{v},\phi_{v}^{(c)}$	Agent $v$ ’s JSON profile (four channel fields) and its channel- $c$ field
$\mathcal{H}(\phi_{u},\phi_{i})$	Profile-level homophily (Eq. 6 ); $\mathcal{H}_{c}$ its channel- $c$ component

Appendix

Table 6: Notation used throughout § 3 . All symbols are introduced at first use; this table is an index.

Parameter	Value
LLM backbone (all operators)	gpt-4o-mini-2024-07-18
LLM temperature (all inference / retrieval calls)	$0.0$
LLM max tokens (reranker etc. main calls)	$4000$
Neighbors $k$	$16$
Facets $N_{f}$	$7$
Candidate list $\|C\|$	$10$

Appendix

Table 7: Complete hyperparameter settings. LLM calls use top_p

=1

(OpenAI default, not explicitly overridden).

Category	# misses	% of 357	Signature
Sync-reg (near-miss)	167	$46.8\%$	GT $\in$ Dialogue top-3 but dropped out of Harp top-1
B1 (vocabulary mismatch)	18	$~5.0\%$	instruction token $\cap$ user preference phrases $=\varnothing$
B2 (reranker bandwidth)	21	$~5.9\%$	$\mathrm{score}_{\text{top-1}}-\mathrm{score}_{\text{GT}}\leq 0.2$
LLM-prior	151	$42.3\%$	$\mathrm{score}$ gap $>0.3$ with no preference-overlap evidence
Other	0	$~0.0\%$	residual
Total misses	357	$100\%$	(Hit@1 $=643/1000=0.643$ )

Appendix

Table 8: Per-case attribution of the

357

Harp misses on the

1{,}000

-user Yelp split. Near-misses (“Sync-reg”) dominate: nearly half of all Harp misses are users whose ground truth was in the Dialogue top-3 but is not ranked first by Harp ; this bucket includes the Sync-hurts pattern of Case 1 (User 157) in Appendix E.5 but is not by itself a regression count. The LLM-prior residual (

42.3\%

) sets a natural ceiling for what any three-operator addition to a memory-augmented backbone can achieve without touching the reranker-level prior. “Sync-reg” is an audit category specific to this per-miss attribution and should not be conflated with the B3 Communicative Drift bottleneck of § 3.5 , which Sync is designed to address rather than to cause.

Dataset	deepseek-v3.2	gemini-3-flash	qwen3-30b
Yelp	.510 / .586	.424 / .471	.479 / .647
MovieTV	.545 / .651	.474 / .527	.481 / .590
Books	.393 / .584	.578 / .596	.507 / .596
Goodreads	.347 / .418	.283 / .392	.318 / .389

Appendix

Table 9: Hit@1 of MemRec / Harp with other backbones.

Recommender Systems

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

See all 540 papers

On the Memorization Behavior of LLMs in Generative Recommendation: Observations, Implications, and Training Strategies
Jun 15, 2026Sunwoo Kim, Sunkyung Lee, Clark Mingxuan Ju +5Sequential Recommendation Memorization in Language Models
1KAIST · 2Sungkyunkwan University · 3Snap Inc.
Personalized Recommendation Tool Learning via Autonomous Language Agents
Jul 22, 2026Mingdai Yang, Zhiwei Liu, Weizhi Zhang +3LLM Tool Use Recommender Systems
Univ. of Illinois Chicago Chicago, IL, USA · Microsoft Mountain View, USA · Beihang University, & Hangzhou Innovation Institute of BUAA Beijing & Hangzhou, China
RRCM: Ranking-Driven Retrieval over Collaborative and Meta Memories for LLM Recommendation
May 8, 2026Shijun Li, Pranav Belligundu, Tianxin Wei +3Agentic RAG Recommender Systems
The University of Texas at Austin · University of Illinois at Chicago · Capital One AI Foundations +1

Jun 15, 2026J/K move · Enter open · S save

On the Memorization Behavior of LLMs in Generative Recommendation: Observations, Implications, and Training Strategies

Sunwoo Kim, Sunkyung Lee, Clark Mingxuan Ju, Donald Loveland+4

1KAIST · 2Sungkyunkwan University · 3Snap Inc.

Sequential Recommendation Memorization in Language Models Large Language Model-Based Recommendation Generative Recommendation

Open paper AskPDF arXiv

Generative recommendation (GR) has emerged as a promising direction for recommender systems. Recently, large language models (LLMs) have been increasingly adopted for GR, as their rich pretrained knowledge is expected to help them generalize beyond common user behavior patterns that traditional memorization-oriented baselines can capture. However, existing LLM-based GR works largely ignore LLMs' well-known tendency to memorize, which, if present in LLMs fine-tuned for GR, would restrict their utilization of pretrained knowledge. In this work, we investigate this concern by examining one-hop memorization, where a model recommends items that are direct successors of items in the training data. We show that LLMs do this more than non-LLM-based GR models-in fact, the vast majority of their gains over GR baselines are actually on users whose target items can be predicted through one-hop memorization. We intuit that improving performance on the remaining users requires LLMs to learn richer item-item relations beyond one-hop transitions. To achieve this, we propose IIRG, a novel training strategy that teaches LLMs to capture: (1) collaborative relations derived from item co-occurrences across multiple hops in user sequences, and (2) semantic relations among items with similar themes, both of which can serve as useful recommendation signals. We show that IIRG significantly improves over LLMs trained solely with standard next-item prediction, with especially large gains for users whose test items are not covered by train-time one-hop transitions.

Abstract

Recommender Systems

Explore similar work

On the Memorization Behavior of LLMs in Generative Recommendation: Observations, Implications, and Training Strategies

Personalized Recommendation Tool Learning via Autonomous Language Agents

RRCM: Ranking-Driven Retrieval over Collaborative and Meta Memories for LLM Recommendation