Action-On-Item Preference Flow: A Shared Event Schema for Predictive and Generative Personalization

Parthiv Chatterjee; Kashish Kanjaria; Vashisth Purani; Sourish Dasgupta; Tanmoy Chakraborty

cs.LGOct 1, 2026

Action-On-Item Preference Flow: A Shared Event Schema for Predictive and Generative Personalization

Authors: Parthiv Chatterjee, Kashish Kanjaria, Vashisth Purani, Sourish Dasgupta, Tanmoy Chakraborty

Organizations: KDM Lab, Dhirubhai Ambani University · Dhirubhai Ambani University · LCS2 Lab, Indian Institute of Technology Delhi

Abstract

A user's movie, news, and dialogue histories differ in their native actions and outputs, yet each interaction supplies evidence that can update user memory. We study whether these histories can train one reusable update mechanism. An action-on-item schema pairs a mapped interaction role with a content embedding, allowing shared update parameters to operate on separate user states. We establish invariance to native relabeling, bounded state changes under item-embedding perturbations, and a pooled-training bound under explicit compatibility conditions. The Multi-Timescale State Hypothesis (MTSH) specifies how this evidence enters, persists, and is consumed; PerTIDE implements it with action gating, three state-space traces, fusion, and command-conditioned readout. On PENS, the same history encoder supports both next-news prediction and personalized headline generation. In a controlled PENS-to-MovieLens experiment, a frozen source-trained core exceeds an identically structured random core by 15.23 MRR points after fitting the same target consumer. On MIND, PerTIDE retains a 4.12-point MRR advantage over a same-input three-branch state-space control. Action, readout, and trace interventions identify complementary contributions to these gains. Together, the theory and experiments support learning history updates across compatible sources and reusing them through predictive and generative consumers.

Figures & tables

Figure 1 : PerTIDE pipeline : a user’s UIG trajectory is mapped to action-gated b-cells and encoded by a tri-trace SSM with an adaptive memory mixer; a command-conditioned readout produces a task-ready state that conditions task-specific heads for prediction (MLP) and generation (DistilGPT2 decoder). The equations in Section 4 additionally specify the shared eigenbasis, current-event residual, and routing through the preceding task-ready state.

Setting	Comparator	Baseline	PerTIDE	Gain
MovieLens-1M movie ranking	SSD4Rec	14.95 / 20.31	17.31 / 23.78	+2.36 / +3.47
MIND news ranking	MINER	36.60 / 40.20	46.65 / 49.21	+10.05 / +9.01
PENS next-news ranking	LSTUR	9.04 / 10.16	9.94 / 11.58	+0.90 / +1.42
PENS headline generation	DeepSeek-R1-32B (2-shot)	0.263 / 0.174	0.366 / 0.355	+0.103 / +0.181
OpenAI-Reddit TL;DR generation	DeepSeek-14B (2-shot)	0.243 / 0.109	0.285 / 0.294	+0.042 / +0.185

Table 1 : One encoder supports prediction and generation. Predictive scores are MRR/nDCG (0–100), with nDCG@5 for MIND and nDCG@10 for MovieLens and PENS; generative scores are PSE-JSD/PSE-METEOR (0–1). Each row uses its strongest listed task-native comparator. Full scores appear in Appendix J.2 , Tables 13 – 15 ; Appendix J.4 , Table 18 ; and Appendix J.5.4 , Table 23 . Table 2 gives the same-input controls.

State family	Model	MRR	nDCG@5	HR@10
PAH	MeanPool-B	29.43	31.22	33.17
PAH	MaxPool-B	28.32	31.16	32.41
MDH	GRU-B	32.64	35.17	37.92
MDH	LSTM-B	32.17	34.92	37.05
LDH	D-FM Attention-B	32.17	34.56	36.21
LDH	Discretized SSM-B	35.14	38.53	37.27

Table 2 : Same-input state-model control on MIND. All “-B” controls consume the same ordered schema-normalized interaction-block embeddings as PerTIDE . Tri-SSM-B provides a structurally matched three-branch SSM control without the explicit long-, short-, and episodic retention maps used by PerTIDE . Full task-native results appear in Appendix J.2.2 , Table 14 .

Panel of Figure 2 — Figure 2 : Correct action roles and complementary traces improve personalization. (a) On fixed PENS trajectories, targets, and candidates, only positive/negative roles change. (b,c) The episodic trace is weak alone, yet adding it to L+S improves MIND MRR by 5.39 points and PENS PSE-METEOR by 0.173. The panels use separate protocols and metric scales. Full interventions are in Tables 21 and 17 .

	500 Target Trajectories			3K Target Trajectories
Method	MRR	nDCG@10	HR@10	MRR	nDCG@10	HR@10
UniSRec	4.31	3.79	8.23	15.18	16.92	33.49
RecGURU	6.03	5.73	11.49	14.21	16.17	29.44
PerTIDE	6.18	5.94	12.96	16.61	19.90	38.75

Table 3 : PENS-trained state reuse with target alignment. MovieLens is withheld from source-encoder training. Methods receive the same target budget and one-positive/699-negative protocol. Scores use the 0–100 scale. The frozen-core control in Table 4 uses a different slice, so absolute scores should be compared within each table.

Frozen core	Head	MRR	nDCG @10	HR@10
Random	Trained	1.115	0.780	1.44
PENS-trained	Untrained	1.371	0.960	2.18
PENS-trained	Trained	16.344	16.002	20.28

Table 4 : Learned updates survive frozen-core transfer. PENS

\rightarrow

MovieLens with PerTIDE cores and the same head architecture. Trained heads use 3K target rows and matched candidates, loss, and optimization. Scores use the 0–100 scale. Table 3 gives the separate transfer-baseline comparison.

Appendix figures & tables27 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset	Event evidence and consumer	Primary metrics
MovieLens-1M	Positive/negative movie events; movie ranking	MRR, nDCG@10, HR@10
MIND	Click/ignore news events; news ranking	MRR, nDCG@5, HR@10
PENS	Click/skip news events; next-news ranking	MRR, nDCG@10
PENS	News history and headline request; personalized generation	PSE-JSD, PSE-SU4, PSE-METEOR
OpenAI-Reddit	Feedback-derived summary events and request; eventized TL;DR generation	PSE-JSD, PSE-SU4, PSE-METEOR

Appendix

Table 5 : One event interface, five predictive and generative settings. Ranking uses one positive target and 699 sampled negatives. Generation appends produced responses to the running trajectory. PSE denotes PerSEval; metric definitions are in Appendices G and I .

Figure 3 : Construction of the User Interaction Graph (UIG) from multi-user, multi-turn dialogue. Turn-level interactions are annotated with behavioral actions and sentiment signals, then projected into a unified UIG over participating users and content nodes.

Figure 4 : UIG Augmentation Pipeline: Chatterjee et al. (2025) proposed framework – (a) Pipeline overview depicting the two-step augmentation , (b) Double Shuffling (DS) to ensure cross-trajectory augmentation and induce diffusion , (c) Stochastic Markovian Perturbation (SMP) to smoothen the response-nodes and modulate random diffusion incorporated in DS stage .

Symbol	Meaning	Symbol	Meaning
UIG and Action-On-Item Schema
$G_{u}=\langle\mathcal{N}_{u},\mathcal{E}_{u}\rangle$	User-specific interaction graph.	$\mathcal{N}_{u},\mathcal{E}_{u}$	User-specific nodes and edges.
$u^{(t_{0})}$	Initial user anchor node.	$d^{(t_{i})}$	Item/content node at timestep $t_{i}$ .
$r_{u}^{(t_{i})}$	Response node at timestep $t_{i}$ .	$\alpha^{(t_{i})}$	Domain-native action label.
$\mathcal{A}$	Shared action-role vocabulary.	$a^{(t_{i})}$	Schema-level action role.
$\rho_{D}(\cdot)$	Domain-specific role mapping.	$\psi_{D}(\cdot)$	Domain-specific target-locus embedding map.

Appendix

Table 6 : Notation and symbol table.

Symbol	Meaning	Dimensions
Action-Gated Event Cells
$d$	Model / latent width (seed embeddings, memory traces, user state, task embeddings)	$768$
$d_{a}$	Action-role input width	$4$
$\mathbf{e}^{(t_{i})}_{tl}$	Target-locus embedding at timestep $t_{i}$ ; induces item-side abstraction	$\mathbb{R}^{768}$
$\mathbf{a}^{(t_{i})}$	One-hot action-role input to the gate MLP	$\mathbb{R}^{4}$
$W^{(1)}_{\text{gate}},\,\mathbf{b}^{(1)}_{\text{gate}}$	Gate MLP layer-1 parameters ( $4\!\rightarrow\!768$ )	$\mathbb{R}^{768\times 4},\;\mathbb{R}^{768}$

Appendix

Table 7: Event input, traces, and fusion. Architecture dimensions for

d=768

. MLPs have two layers with hidden width 768 unless stated otherwise.

Symbol	Meaning	Dimensions
Command-Conditioned Readout (Realization of $\Pi$ and $G$ )
$\mathbf{g}_{q}$	Task / command seed embedding	$\mathbb{R}^{768}$
$\bm{\pi}^{(t_{i})}_{u}$	Command/task distribution $\bm{\pi}^{(t_{i})}_{u}=\mathrm{SoftMax}(W_{\text{task}}[\mathbf{z}^{(t_{i-1})}_{b_{u}};\mathbf{g}_{q}])$ ; routing uses the preceding task-ready state (Section 4 )	$\mathbb{R}^{\|\mathcal{Q}\|}$
$W_{\text{task}}$	Task-distribution projection	$\mathbb{R}^{\|\mathcal{Q}\|\times 1536}$
$W_{\pi}$	Command-to-state gating projection in $G(\cdot)$	$\mathbb{R}^{768\times\|\mathcal{Q}\|}$
$\mathbf{b}_{G}$	Command-gate bias	$\mathbb{R}^{768}$

Appendix

Table 8: Command readout, supervision, and decoder settings. Architecture dimensions for

d=768

. MLPs have two layers with hidden width 768 unless stated otherwise.

Symbol	Meaning	Dimensions
Decoder-UFI (User Fused Injection)
$W^{(1)}_{\text{UFI}},\,\mathbf{b}^{(1)}_{\text{UFI}}$	UFI MLP layer-1 parameters ( $[\hat{\mathbf{z}}_{q,u};\mathbf{e}_{d_{q}}]\in\mathbb{R}^{1536}\!\rightarrow\!768$ )	$\mathbb{R}^{768\times 1536},\;\mathbb{R}^{768}$
$W^{(2)}_{\text{UFI}},\,\mathbf{b}^{(2)}_{\text{UFI}}$	UFI MLP layer-2 parameters ( $768\!\rightarrow\!768$ )	$\mathbb{R}^{768\times 768},\;\mathbb{R}^{768}$
$\mathbf{c}^{\mathrm{user}}_{\mathrm{UFI}}$	User-fused control, $\mathbf{c}^{\mathrm{user}}_{\mathrm{UFI}}=\mathrm{MLP}_{\mathrm{UFI}}([\hat{\mathbf{z}}_{q,u};\mathbf{e}_{d_{q}}])$	$\mathbb{R}^{768}$
Decoder-UGI (User Gated Injection)
$W^{(1)}_{\gamma},\,\mathbf{b}^{(1)}_{\gamma}$	FiLM-scale MLP layer-1 parameters ( $\mathbf{z}^{(t_{l})}_{u}\in\mathbb{R}^{768}\!\rightarrow\!768$ )	$\mathbb{R}^{768\times 768},\;\mathbb{R}^{768}$

Appendix

Table 9: User fused and user gated injection. Architecture dimensions for

d=768

. MLPs have two layers with hidden width 768 unless stated otherwise.

Symbol	Meaning	Dimensions
Decoder Document Grounding
$W^{(1)}_{\text{doc}},\,\mathbf{b}^{(1)}_{\text{doc}}$	Document-grounding MLP layer-1 parameters ( $[\mathbf{c}^{\mathrm{user}};\mathbf{e}_{d_{q}}]\in\mathbb{R}^{1536}\!\rightarrow\!768$ )	$\mathbb{R}^{768\times 1536},\;\mathbb{R}^{768}$
$W^{(2)}_{\text{doc}},\,\mathbf{b}^{(2)}_{\text{doc}}$	Document-grounding MLP layer-2 parameters ( $768\!\rightarrow\!768$ )	$\mathbb{R}^{768\times 768},\;\mathbb{R}^{768}$
$\mathbf{c}^{\mathrm{doc}}$	Document-grounding control, $\mathbf{c}^{\mathrm{doc}}=\mathrm{MLP}_{\mathrm{doc}}([\mathbf{c}^{\mathrm{user}};\mathbf{e}_{d_{q}}])$	$\mathbb{R}^{768}$
$\mathbf{c}$	Final normalized decoder control, $\mathbf{c}=\mathrm{LN}(\mathbf{c}^{\mathrm{user}}+\mathbf{c}^{\mathrm{doc}})$	$\mathbb{R}^{768}$
Learned Prefix Injection

Appendix

Table 10: Document grounding, prefix, and generation loss. Architecture dimensions for

d=768

. MLPs have two layers with hidden width 768 unless stated otherwise.

Metric	PerTIDE (Ours)	Single-scale encoder	Prompted LLM (ICL)
A. Parameter and Memory Footprint
Trainable parameters (encoder only)	$\sim$ 154.36M	$\sim$ 130–145M	0
Decoder contextualization parameters	$\sim$ 8.27M (UFI) / $\sim$ 10.63M (UGI)	task-dependent	0
Encoder + contextualization parameters	$\sim$ 162.63M (UFI) / $\sim$ 164.99M (UGI)	$\sim$ 130–145M + context	0
Decoder parameters fine-tuned for generation	$\sim$ 14.18M (last 2 blocks + final LN)	same protocol if DistilGPT2 is used	0 (ICL)
Pretrained decoder backbone deployed	$\sim$ 82M (DistilGPT2)	$\sim$ 82M (if used)	7B–32B

Appendix

Table 11 : Runtime and deployment costs for PerTIDE . We report approximate resource figures for PerTIDE and comparison systems. The action-on-item encoder supports online updates from cached user state without replaying the full history, whereas prompted LLM personalization conditions on history through the inference context. Numbers correspond to

d{=}768

|\mathcal{Q}|{=}4

, and, for generation, a pretrained DistilGPT2 decoder conditioned by a learned

K{=}8

prefix. The final two decoder blocks and final layer normalization are fine-tuned during generation training; the remaining decoder parameters stay fixed. Latency, throughput, VRAM, FLOPs, and training time are source estimates, not a matched hardware benchmark; the per-user cache is computed directly from the state dimensions.

Metric	PerTIDE	Single-scale encoder	Prompted LLM
C. Predictive Serving Cost
Latency per sample (end-to-end)	$\sim$ 2–6 ms	$\sim$ 2–5 ms	$\sim$ 20–60 ms
Throughput (samples/sec)	$\sim$ 160–420	$\sim$ 180–450	$\sim$ 20–60
FLOPs per sample	$\sim$ 3.8 $\times 10^{9}$	$\sim$ 2.5–3.2 $\times 10^{9}$	$\geq$ 1 $\times 10^{11}$
D. Generative Serving Cost
Contextualization	UFI / UGI learned prefix	concat / none	ICL prompt

Appendix

Table 12 : Approximate serving and training resources. Continuation of the resource estimates in Appendix J.1 , Table 11 . These estimates are not a matched hardware benchmark.

Regime	Model	MRR	nDCG@10	HR@10
PAH	BPR-MF	6.50	6.12	12.81
MDH	GRU4Rec	12.76	16.42	29.01
	Caser	13.54	19.21	28.92
	Diff4Rec	12.02	15.72	22.38
LDH	S 3 Rec	10.62	13.67	19.21
	SASRec	13.20	19.96	30.23

Appendix

Table 13 : MovieLens movie recommendation under action-on-item preference flow. We compare PerTIDE against task-native recommendation baselines on next-item ranking. Baselines expose profile, recurrent, or long-horizon history states, while PerTIDE uses schema-normalized action evidence with multi-timescale persistence. Point estimates are means under the paired-bootstrap evaluation in Appendix E ;

\pm

denotes standard deviation over three uniformly sampled 25% evaluation subsets for PerTIDE variants.

Regime	Model	MRR	nDCG@5	HR@10
PAH	NAML	32.75	35.66	41.40
	PLM-NR	35.39	38.71	44.38
	Mean Pooling-B (ours)	29.43	31.22	33.17
	Max Pooling-B (ours)	28.32	31.16	32.41
MDH	EBNR	31.26	32.18	39.04
MDH	GRU-B (ours)	32.64	35.17	37.92

Appendix

Table 14 : Action-on-item encoder performance on MIND. We compare task-native history encoders and schema-matched controls with PerTIDE on sequential ranking. AC denotes action-conditioned evidence entry, and CR denotes command-conditioned readout. Point estimates are means under the paired-bootstrap evaluation in Appendix E ;

\pm

denotes standard deviation over three uniformly sampled 25% evaluation subsets for PerTIDE variants.

Regime	Model	MRR	nDCG@5	nDCG@10	HR@10
MDH	EBNR	2.65	1.82	2.45	4.87
PAH	NAML	1.29	0.43	0.81	2.01
LDH	NRMS	1.18	0.39	0.74	1.92
LDH	TrRMIo	8.02	7.85	8.53	12.77
LDH	LSTUR	9.04	8.21	10.16	12.93
LDH	MINER	5.17	4.71	5.61	9.77

Appendix

Table 15 : Structured news prediction on PENS-Structured. We report same-user next-news ranking under the action-on-item schema. Baselines are grouped by the exposed history-state assumption in Section 3 . All values are on a 0–100 scale.

Dataset	Family	Model / Variant	L-Tr	S-Tr	E-Tr
MovieLens	Baseline	Mamba4Rec	26.43/29.38	26.12/28.87	25.72/28.31
		GRU4Rec	24.17/27.32	23.65/26.58	22.13/24.71
		SASRec	25.82/28.16	25.85/28.45	25.17/27.91
	PerTIDE	PerTIDE-L	27.31/29.96	27.18/29.35	25.83/28.77
		PerTIDE-S	25.21/28.64	27.13/29.22	22.16/24.84
		PerTIDE-E	22.19/24.08	22.75/25.04	27.14/30.36

Appendix

Table 16 : Temporal regime-specific trace specialization. Controlled L-Tr, S-Tr, and E-Tr probes emphasize persistent, recent, and event-local evidence. Scores are MRR/nDCG@10 on the 0–100 scale. Bold denotes the overall best result; underlining marks the strongest isolated PerTIDE trace in each regime. The full model leads every reported probe; the long trace narrowly leads the short trace on MovieLens S-Tr.

Variant	MIND Recommendation			ML-1M Recommendation			PENS Summarization
Variant	MRR	nDCG@5	HR@10	MRR	nDCG@10	HR@10	PSE-JSD	PSE-SU4	PSE-METEOR
L only	40.41	43.75	46.13	12.31	16.15	20.34	0.231	0.095	0.107
S only	37.65	41.32	44.24	12.47	16.62	20.81	0.238	0.106	0.111
E only	24.18	27.73	31.23	10.18	13.17	18.54	0.076	0.042	0.045
L + S	41.26	43.95	46.72	14.17	18.44	22.11	0.264	0.143	0.182
L + E	40.45	43.82	46.44	14.03	18.07	22.74	0.246	0.103	0.108

Appendix

Table 17 : Trace-subset ablation for schema-normalized evidence persistence. We ablate long-term (L), short-term (S), and episodic (E) persistence channels individually and in combination. Recommendation results are reported on MIND and MovieLens-1M, and personalized summarization is reported on PENS. Best results are bolded.

Category	Model	PSE-JSD	PSE-SU4	PSE-METEOR
LLMs (2-shot)	LLaMA-13B	0.227	0.078	0.081
	DeepSeek-14B	0.248	0.094	0.097
	Gemini-2.5-Flash	0.222	0.104	0.124
	DeepSeek-R1-32B	0.263	0.125	0.174
	Qwen-2.5-32B	0.162	0.103	0.115
Prompt-Chaining	Mistral-7B	0.072	0.026	0.023

Appendix

Table 18 : Command-conditioned personalized headline generation on PENS. We compare prompted LLMs, task-native personalized generators, schema-matched history controls, and PerTIDE . AC denotes action-conditioned evidence entry, CR denotes command-conditioned readout, and UFI/UGI denote User Fused/Gated Injection. Point estimates are means under the paired-bootstrap evaluation in Appendix E ;

\pm

denotes standard deviation over three uniformly sampled 25% evaluation subsets for PerTIDE variants.

Category	Model	RG-SU4	BLEU-4	BScore	HJ
Specialized (Personalized)	PENS-NRMS-T2	13.64	4.48	86.13	2.95
	GTP-TrRMIo	21.91	10.31	88.53	2.44
	SP-Individual	19.54	8.90	86.61	2.86
LLMs (2-shot history)	LLaMA-13B	18.31	11.85	88.76	3.05
	DeepSeek-14B	19.57	12.68	89.43	3.03
	DeepSeek-32B	29.42	19.31	90.87	3.12

Appendix

Table 19 : Reference agreement for command-driven headline generation. Automatic metrics measure reference overlap or semantic similarity; HJ is the six-point reference-similarity rating described in Appendix K .

Relation	Spearman $\rho$	User-clustered 95% CI
$d_{\mathrm{sch}}$ vs. frozen risk	+0.109	[+0.002,+0.207]
$d_{\mathrm{sch}}$ vs. adapted risk	+0.160	[+0.060,+0.258]
$d_{\mathrm{sch}}$ vs. frozen RR gain	-0.089	[-0.169,-0.004]
$d_{\mathrm{sch}}$ vs. adapted RR gain	-0.114	[-0.195,-0.029]

Appendix

Table 20 : Schema-distance diagnostic for conditional reuse on 500 MovieLens trajectories from 99 users. Larger distance is associated with higher risk and smaller reciprocal-rank gain.

Role flips	MRR	nDCG@10	HR@10
0%	12.14	16.02	32.0
10%	10.13	13.02	25.0
25%	9.17	11.10	22.0
50%	7.71	8.93	15.0
100%	2.81	4.90	10.0

Appendix

Table 21 : Action-role corruption on fixed PENS trajectories. Only positive/negative roles are flipped; trajectories, targets, and candidate sets are held fixed.

Events	MRR	nDCG@10	HR@10
10	11.874	14.904	27.8
20	11.881	15.011	28.2
30	11.909	15.091	28.4
40	11.926	15.157	28.6

Appendix

Table 22 : History-length sensitivity on the same 500 PENS trajectories. The target and 1+699 candidate sets are fixed while the visible history grows from 10 to 40 events.

Category	Model	PSE-JSD	PSE-SU4	PSE-METEOR
LLMs (2-shot)	LLaMA-13B	0.232	0.093	0.107
	Zephyr-7B	0.214	0.087	0.104
	Mistral-7B	0.226	0.088	0.103
	DeepSeek-14B	0.243	0.095	0.109
MTSH	PerTIDE -Full	0.285	0.261	0.294

Appendix

Table 23 : Eventized TL;DR personalization on OpenAI-Reddit. We treat OpenAI-Reddit as an eventization stress test under surrogate trajectory construction, not as evidence of unconstrained cross-domain user-history transfer.

Model	PSE-JSD	PSE-SU4	PSE-METEOR
DeepSeek-R1-32B (2-shot)	$0.263\rightarrow 0.152$ (−42.21%)	$0.125\rightarrow 0.087$ (−30.4%)	$0.174\rightarrow 0.092$ (−47.2%)
Gemini-2.5-Flash (2-shot)	$0.222\rightarrow 0.122$ (−45.0%)	$0.104\rightarrow 0.061$ (−41.3%)	$0.124\rightarrow 0.070$ (−43.5%)
Best Baseline (GTP)	$0.024\rightarrow 0.016$ (−33.3%)	$0.017\rightarrow 0.009$ (−47.1%)	$0.019\rightarrow 0.011$ (−42.1%)
PerTIDE	$0.366\rightarrow 0.211$ (−42.35%)	$0.341\rightarrow 0.244$ (−28.45%)	$0.355\rightarrow 0.262$ (−26.2%)

Appendix

Table 24 : Robustness under sparse positive-action evidence. We evaluate each model on a click-only sparse subset and report the original score (left) and sparse-action score (right), with the relative drop in parentheses. Smaller drops indicate higher robustness when positive evidence is limited in the action-on-item trajectory.

Category	Model	RG-2	RG-L	BLEU-4	BScore
LLMs (2-shot)	DeepSeek-32B	9.71	8.46	5.61	71.42
	Qwen-2.5-32B	9.43	8.21	4.75	68.83
	MS/Phi-Instruct	9.49	8.05	3.21	69.52
PAH	Meanpool-B	$8.54\pm 1.33$	$6.31\pm 1.21$	$3.43\pm 2.15$	$62.18\pm 1.45$
PAH	Maxpool-B	$7.21\pm 1.12$	$4.12\pm 1.43$	$2.55\pm 2.11$	$61.42\pm 1.65$
MDH	GRU-B	$11.11\pm 1.31$	$10.82\pm 1.43$	$4.17\pm 0.78$	$63.18\pm 2.43$

Appendix

Table 25 : MPChat broader-applicability stress test: dialogue generation. AC denotes action-conditioned evidence injection, CR denotes command-conditioned readout, UFI denotes user fused injection, and Full denotes PerTIDE encoder with UGI decoder. These results test dialogue eventization under imperfect preference logs rather than serving as primary evidence for schema-level generalizability;

\mu\pm\sigma

denotes mean and standard deviation over three evaluations on uniformly random test-set subsamples.

Model	R@1	MRR
SBERT (zero-shot)	35.67	45.75
SBERT	$51.32\pm 1.32$	$64.76\pm 0.92$
SBERT+ViT ( $c$ )	$57.70\pm 0.71$	$69.39\pm 0.40$
SBERT+ViT ( $c,p^{i}$ )	$58.55\pm 0.70$	$70.17\pm 0.45$
SBERT+ViT ( $c,p^{t}$ )	$64.32\pm 0.64$	$74.31\pm 0.45$
SBERT+ViT ( $c,p^{i+t}$ , Full)	${65.29}\pm{0.66}$	${75.08}\pm{0.43}$

Appendix

Table 26 : MPChat broader-applicability stress test: response prediction. Dialogue eventization is compared with text-only and multimodal consumers. Scores use the 0–100 scale;

\mu\pm\sigma

denotes the reported mean and standard deviation over three uniformly sampled 25% test subsets, shared across methods.

Category	Model	R@1	MRR
Text Only	SBERT ( $c^{t},r,\mathcal{P}^{t}_{c}$ )	$56.47\pm 0.58$	$67.92\pm 0.52$
SBERT+ViT	$c,r,\mathcal{P}^{i}_{c}$	$19.56\pm 0.64$	$35.84\pm 0.45$
	$c,r,\mathcal{P}^{t}_{c}$	$56.87\pm 0.60$	$68.33\pm 0.37$
	$c,r,\mathcal{P}^{i+t}_{c}$ (Full)	${57.28\pm 0.44}$	${68.86\pm 0.33}$
SBERT+CLIP	$c,r,\mathcal{P}^{i}_{c}$	$25.71\pm 0.49$	$42.47\pm 0.34$
	$c,r,\mathcal{P}^{t}_{c}$	$56.63\pm 0.66$	$68.15\pm 0.42$

Appendix

Table 27 : MPChat broader-applicability stress test: speaker identification. We report speaker identification under dialogue eventization and compare against systems using persona and multimodal context;

c

: dialogue context,

c^{t}

: text-only context,

r

: response,

\mathcal{P}_{c}

: candidate set with profile information

Figure 6 : Prompt Templates for LLMs for (a)Personalized Summarization, (b)Dialogue Generation

Explore similar work

HorizonBench: Long-Horizon Personalization with Evolving Preferences
Apr 19, 2026Shuyue Stella Li, Bhargavi Paranjape, Kerem Oktar +9Preference Alignment Learning Interaction History
University of Washington · Meta · OpenAI +1
Not All Is Lost: Repairing Lossy User Preference States of Personalization Encoders
Oct 1, 2026Parthiv Chatterjee, Dhiraj Golhar, Ummesalma Diwan +3Preference Alignment Learning Recommendation
KDM Lab, Dhirubhai Ambani University · Dhirubhai Ambani University · LCS2 Lab, Indian Institute of Technology Delhi
Know You Before You Speak: User-State Modeling for LLM Personalization in Multi-Turn Conversation
May 23, 2026Jiani Luo, Xiaoyan Zhao, Yang Zhang +4Large Language Model Personalization Personalization
School of Computing, National University of Singapore, Singapore · School of Artificial Intelligence, Beihang University, Beijing, China · Institute of Computing Technology, Chinese Academy of Sciences, Beijing, China +1

Apr 19, 2026, cs.CLJ/K move · Enter open · S save

HorizonBench: Long-Horizon Personalization with Evolving Preferences

Shuyue Stella Li, Bhargavi Paranjape, Kerem Oktar, Zhongyao Ma+8

University of Washington · Meta · OpenAI+1

Preference Alignment Learning Interaction History Dialogue Benchmarks Horizon

Open paper AskPDF arXiv

User preferences evolve across months of interaction, and tracking them requires inferring when a stated preference has been changed by a subsequent life event. We define this problem as long-horizon personalization and observe that progress on it is limited by data availability and measurement, with no existing resource providing both naturalistic long-horizon interactions and the ground-truth provenance needed to diagnose why models fail. We introduce a data generator that produces conversations from a structured mental state graph, yielding ground-truth provenance for every preference change across 6-month timelines, and from it construct HorizonBench, a benchmark of 4,245 items from 360 simulated users with 6-month conversation histories averaging ~4,300 turns and ~163K tokens. HorizonBench provides a testbed for long-context modeling, memory-augmented architectures, theory-of-mind reasoning, and user modeling. Across 25 frontier models, the best model reaches 52.8% and most score at or below the 20% chance baseline. When these models err on evolved preferences, over a third of the time they select the user's originally stated value without tracking the updated user state. This belief-update failure persists across context lengths and expression explicitness levels, identifying state-tracking capability as the primary bottleneck for long-horizon personalization.

Abstract

Explore similar work

HorizonBench: Long-Horizon Personalization with Evolving Preferences

Not All Is Lost: Repairing Lossy User Preference States of Personalization Encoders

Know You Before You Speak: User-State Modeling for LLM Personalization in Multi-Turn Conversation