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

Parthiv Chatterjee; Dhiraj Golhar; Ummesalma Diwan; Sourish Dasgupta; Manjunath Joshi; Tanmoy Chakraborty

cs.LGOct 1, 2026

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

Authors: Parthiv Chatterjee, Dhiraj Golhar, Ummesalma Diwan, Sourish Dasgupta, Manjunath Joshi, Tanmoy Chakraborty

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

Abstract

Personalization encoders compress evolving interaction histories into preference states used to rank items or condition text generation. A task head operating only on this state can miss useful evidence that remains in the frozen encoder's cached representations for individual timesteps. We study this recoverability gap and propose REPAIR, which compares cached representations with the current preference state in a compact learned coordinate space. It resolves corrective evidence over extended history, recent interactions, and localized bursts. It then selects which patterns at which timesteps contribute and adds their aggregate correction to the state before the task head. Encoder-host repair reuses representations from the existing forward computation without re-encoding the history. Across MovieLens, PENS, MIND, and Amazon Reviews 2023, training only REPAIR improves MRR and nDCG@10 for all twelve representative recommendation hosts while both encoder and task head remain frozen. Head-only finetuning of the same hosts yields smaller gains. For example, Mamba4Rec on MovieLens gains 3.96 MRR points, compared with 0.19 from head-only finetuning. Rank and temporal diagnostics support a compact, host-dependent corrective structure. In personalized generation, IMPerSumm improves the two reported weighted PerSEval variants, which assess responsiveness to user preference, by up to 25.23%. These results support post-compression state correction and distinguish the availability of preference evidence from its downstream use.

Figures & tables

Figure 1: Overview of REPAIR . Given a frozen preference state and cached per-timestep representations from host computation, REPAIR constructs state-relative corrective evidence, resolves it through long-, short-, and episodic-support branches, selectively retains timestep-coordinate corrective components, and fuses the aggregated correction with the frozen state before the downstream interface consumes the repaired state.

Task	Dataset	Host Coverage	Test Instances
Movie recommendation	MovieLens	9 sequential hosts	15K
News recommendation	MIND	7 history encoders	20K
News recommendation	PENS	History encoders and LLMs	20K
Product recommendation	Amazon Reviews 2023	3 multimodal hosts	2K
Headline generation	PENS	11 encoder interfaces and 2 LLMs	20K

Table 1: Evaluation design. Recommendation ranks one target against

499

negatives. Host architectures and citations are in Appendix E , with protocols in Appendix C .

Dataset	Host	MRR / nDCG@10
		Original Host	Head-only FT	REPAIR (Frozen Head)	REPAIR + Head Alignment
MovieLens	S: TiSASRec	35.51 / 41.14	35.57 / 41.19	36.38 / 42.07	36.78 / 42.91
	M: Mamba4Rec	29.29 / 38.50	29.48 / 38.89	33.25 / 39.20	34.26 / 40.17
	W: SASRec	26.43 / 33.82	27.05 / 34.37	29.57 / 40.08	28.49 / 35.18
PENS	S: LSTUR	10.44 / 12.15	10.98 / 12.32	11.98 / 12.71	12.85 / 15.88
	M: EBNR	2.65 / 2.45	2.93 / 2.86	4.07 / 3.97	11.97 / 14.28

Table 2: RQ1: Post-compression recoverability under frozen-host controls. Columns compare the original host, head-only continuation on the unchanged state, REPAIR with encoder and head frozen, and two-epoch joint REPAIR -head alignment with the host encoder frozen. Values are MRR / nDCG@10. Three-seed means, variability, and bootstrap intervals are available for OpenCLIP and SigLIP2 on Amazon and EBNR on PENS (Appendix Table 10 ). The remaining rows are reported as point estimates. S/M/W denote strongest, median, and weakest original hosts.

Host / Dataset	Original Host	REPAIR Rank $K$
		32	64	128	192
OpenCLIP / Amazon	1.59 / 1.05	9.31 / 9.78	11.17 / 12.22	19.49 / 22.08	11.63 / 12.16
SigLIP2 / Amazon	2.40 / 2.18	10.44 / 11.36	8.42 / 8.77	17.72 / 20.35	13.46 / 13.15
EBNR / PENS	2.65 / 2.45	6.13 / 6.19	8.65 / 8.58	11.97 / 14.28	11.23 / 13.17
Mamba4Rec / MovieLens	29.29 / 38.50	31.47 / 38.85	32.38 / 39.26	34.26 / 40.17	34.02 / 39.96
LSTUR / MIND	51.43 / 59.63	52.67 / 62.14	53.49 / 64.13	55.84 / 66.92	54.73 / 66.12

Table 3: RQ2: Repair-space rank sensitivity after compression. MRR / nDCG@10 for two-epoch downstream alignment across

K

. Bold marks the strongest repair rank for each host.

		Controlled one-view diagnostic			Full REPAIR ( $L+S+E$ jointly)
Dataset	Host	L-Tr	S-Tr	E-Tr	L-Tr	S-Tr	E-Tr
Copy-through prediction		$L$	$S$	$E$	all three views used jointly
MovieLens	Mamba4Rec	$S$ [33.95]	$S$ [33.92]	$S$ [33.88]	[34.87]	[35.52]	[34.39]
MovieLens	TiSASRec	$L$ [36.19]	$S$ [36.20]	$S$ [36.04]	[37.38]	[36.93]	[36.63]
MIND	NAML	$L$ [52.12]	$S$ [52.03]	$L$ [53.70]	[53.31]	[54.55]	[53.80]
MIND	EBNR	$S$ [55.08]	$L$ [53.64]	$E$ [53.70]	[56.79]	[55.69]	[53.19]

Table 4: RQ2 temporal diagnostic. If the temporal module merely copied the controlled input timing, the copy-through prediction would be

L

S

, and

E

for L-Tr, S-Tr, and E-Tr, respectively. This pattern holds in only 6/12 settings. Full REPAIR , which jointly uses

L+S+E

, outperforms the best isolated view in 11/12 settings. Brackets report MRR.

Tasks (Datasets)	Host Model	Base		+ SERAC-Corr		+ Rec-Denoiser-Corr		+ REPAIR
Tasks (Datasets)	Host Model	MRR	nDCG@10	MRR	nDCG@10	MRR	nDCG@10	MRR	nDCG@10
Movie Reco. (MovieLens)	TiSASRec	35.51	41.14	33.17	41.32	34.21	42.08	36.78	42.91
	Mamba4Rec	29.29	38.50	29.81	38.63	31.32	39.22	34.26	40.17
	SIGMA	33.27	38.96	31.95	38.98	29.43	39.61	34.81	40.72
News Reco. (MIND)	LSTUR	51.43	59.63	53.66	62.17	52.45	61.35	55.84	66.92
	NAML	45.65	46.56	46.47	51.33	48.12	48.94	53.22	63.18

Table 5: RQ3: Intervention-locus comparison under the same frozen-host protocol. Pre-head REPAIR is compared with two adapted post-head correctors using the same task-trained host and two-epoch downstream alignment protocol.

Host	RG-L	METEOR	PSE-W-JSD	PSE-W-RG-L
Walk2Pers	+6.58	+13.38	+13.25	+15.32
IMPerSumm	-2.69	+4.76	+20.76	+25.23
DeepSeek-32B	+53.21	+45.49	+9.32	+5.05
Qwen2.5-32B	+41.58	+47.64	+31.52	+12.61

Table 6: RQ3 generation and personalization on PENS. Values are relative gains (%). Full ranking and generation comparisons are in Appendix Table 17 . Absolute scores are in Appendix Tables 15 and 16 .

Appendix figures & tables22 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 2: Encoder-host repair from cached evidence to the task head. Read from left to right. The state and cached sequence enter a shared space, the low-rank reader produces corrective coordinates, and centering precedes temporal resolution and coordinate-wise selection. State gating and weighted aggregation form a correction, which is fused with the original state. The diagram’s raw-residual label denotes $r^{(t_{j};t_{i})}$ in Appendix A.3 . Dashed arrows indicate conditioning, and the original state remains the residual base.

Panel of Figure 3 — Figure 3: Prompt-conditioned repair for a frozen LLM. The prompt computation supplies the terminal state, while a separate interaction construction supplies corrective evidence. Repair maps the fused state into continuous prefix vectors. The frozen decoder then evaluates the prefix-conditioned input to produce the task output. Decoder weights remain fixed, and the attention cache must correspond to this modified input.

Figure 4: Task prompts used for the LLM headline-generation and recommendation settings. The history supplies preference evidence, while the task block specifies the requested output and, for generation, the current source document. The examples illustrate prompt structure rather than additional evaluation instances. The repair-specific continuous prefix is constructed separately as described in Appendix A.7 .

Symbol	Meaning	Symbol	Meaning
Interaction and host interface
$\mathcal{G}_{u}=\langle\mathcal{N}_{u},\mathcal{E}_{u}\rangle$	User interaction graph	$\mathcal{N}_{u}$	User-specific typed node set
$\mathcal{E}_{u}$	Action-labeled edge set	$u^{(t_{0})}$	Initial user node
$v^{(t_{i})}$	Content or response node	$a^{(t_{i})}$	Observed action label
$\{\mathrm{pos},\mathrm{neg},\mathrm{cmd}\}$	Normalized action types	$b^{(t_{i})}=\langle a^{(t_{i})},v^{(t_{i})}\rangle$	Interaction unit at timestep
$\tau^{t_{1:i}}_{b}$	Model-side prefix trajectory	$t_{i}$	Current prefix endpoint

Appendix

Table 7: Notation for the interaction sequence, frozen host interface, repair operators, and supervision. Hats on

U_{\mathrm{rep}}

and

V_{\mathrm{rep}}

denote column normalization. The family weights denoted by

\omega

in the main text are denoted by

\lambda

here, with

\omega

reserved for within-family coefficients.

Symbol	Meaning	Symbol	Meaning
Temporal resolution
$\rho\in\{L,S,E\}$	Temporal basis index	$\kappa^{L}(\ell_{j})$	Long-support temporal basis
$\kappa^{S}(\ell_{j})$	Short-support temporal basis	$\kappa^{E}(\ell_{j})$	Episodic-support temporal basis
$\beta$	Learned long-support decay vector	$\xi$	Learned episodic-support rate vector
$W_{L},W_{S},W_{E}$	Temporal branch-mixing matrices	$G_{\mathrm{rep}}$	Multi-basis resolution operator
$\bm{\eta}^{(t_{j})}$	Temporally resolved corrective vector

Appendix

Table 8: Notation continued from Table 7 . Temporal resolution, selective retention, aggregation, and training objective.

Component	Symbol / Configuration	Dimension / Value
Core repair-interface dimensions
Repair-interface state dimension	$d_{z}$	$768$
Repair-interface cached-representation dimension	$d_{e}$	$768$
Shared repair space	$m$	$192$
Repair rank	$K$	$128$
Maximum history length	$L$	$50$

Appendix

Table 9: Architecture and fixed hyperparameters for the encoder-host repair interface. The dimensions describe the implemented common interface, not every host’s native hidden width. The default

K=128

is varied only in the rank sweep. Appendix C.4.4 defines the loss families and within-family coefficients.

Setting	$H_{\rm best}$	$R$	$RH_{2}$
OpenCLIP / Amazon	$2.21\pm 0.18$ [2.04, 2.36]	$4.43\pm 0.25$ [4.17, 4.64]	$19.49\pm 1.43$ [18.97, 20.06]
SigLIP2 / Amazon	$2.67\pm 0.14$ [2.33, 2.91]	$4.16\pm 0.17$ [3.97, 4.33]	$17.72\pm 2.34$ [17.37, 18.04]
EBNR / PENS	$2.93\pm 0.31$ [2.73, 3.05]	$4.07\pm 0.19$ [3.98, 4.24]	$11.97\pm 0.62$ [11.76, 12.25]

Appendix

Table 10: Repeated-run reliability for three RQ1 settings. Entries are MRR mean

\pm

standard deviation across three training seeds, followed by 95% user-clustered bootstrap intervals from 10,000 resamples. MRR uses the

0

100

scale.

Setting	Host	MRR			nDCG@5
		Base	+ REPAIR	$\Delta$	Base	+ REPAIR	$\Delta$
Direct	Mamba4Rec	29.29	34.26	+4.97	30.07	35.56	+5.49
	TiSASRec	35.51	36.78	+1.27	36.87	38.41	+1.54
	HSTU	29.00	30.62	+1.62	29.69	31.55	+1.86
	FEARec	28.03	30.62	+2.59	28.20	31.55	+3.35
	SIGMA	33.27	34.81	+1.54	34.43	36.26	+1.83

Appendix

Table 11: MovieLens recommendation under Stage-3 repair and head alignment, with the host encoder frozen. Base is the original task-trained host. Direct rows train repair for that host. Transfer rows use the module trained on Mamba4Rec. Scores are on a

0

100

scale and higher is better. Each

\Delta

is an absolute score-point change. All paired conditions rank the same target against

499

negatives. Appendix G.1 distinguishes these aligned results from frozen-head attribution.

Host	MRR			nDCG@5
	Base	+ REPAIR	$\Delta$	Base	+ REPAIR	$\Delta$
EBNR	2.65	11.97	+9.32	1.82	12.68	+10.86
NAML	1.29	12.38	+11.09	0.43	13.16	+12.73
NRMS	1.18	12.11	+10.93	0.39	12.89	+12.50
TrRMIo	8.02	12.86	+4.84	7.85	13.97	+6.12
SCAPE	1.57	8.34	+6.77	0.76	7.91	+7.15

Appendix

Table 12: PENS recommendation with paired target and

499

-negative candidate sets. Base denotes the task-trained host. Repaired encoder rows use Stage-3 repair and head alignment with frozen encoders. LLM rows use the frozen-decoder soft-prefix interface in Appendix A.7 . Scores are on a

0

100

scale, higher is better, and

\Delta

denotes absolute score-point change.

Host	MRR			nDCG@5
	Base	+ REPAIR	$\Delta$	Base	+ REPAIR	$\Delta$
EBNR	47.08	53.22	+6.14	48.65	57.21	+8.56
NAML	45.65	53.22	+7.57	43.13	57.21	+14.08
NRMS	44.91	52.74	+7.83	42.38	56.48	+14.10
TrRMIo	47.43	55.21	+7.78	46.27	52.83	+6.56
SCAPE-GRU	42.53	45.32	+2.79	37.53	44.15	+6.62

Appendix

Table 13: MIND recommendation under Stage-3 repair and head alignment with frozen encoders. Base denotes the original task-trained host. Paired conditions rank the same target against

499

negatives. Scores are on a

0

100

scale, higher is better, and

\Delta

denotes absolute score-point change. Appendix G.1 discusses the seven hosts.

Host	MRR			nDCG@5
	Base	+ REPAIR	$\Delta$	Base	+ REPAIR	$\Delta$
SigLIP2	2.40	17.72	+15.33	1.53	17.71	+16.18
OpenCLIP	1.59	19.49	+17.91	0.64	19.46	+18.83
DINOv2 E5	1.19	16.15	+14.96	0.43	16.30	+15.87

Appendix

Table 14: Amazon Reviews 2023 All_Beauty recommendation with frozen multimodal encoders and Stage-3 repair and head alignment. Each instance ranks one target against the same

499

negatives in paired conditions. Scores are on a

0

100

scale, higher is better, and

\Delta

is an absolute score-point change.

Host	RG-1			RG-L			RG-SU4
	Base	+ REPAIR	Gain (%)	Base	+ REPAIR	Gain (%)	Base	+ REPAIR	Gain (%)
DeepSeek-32B	20.62	26.84	+30.16%	13.85	21.22	+53.21%	4.42	11.99	+171.27%
Qwen2.5-32B	18.51	23.47	+26.80%	12.53	17.74	+41.58%	3.86	8.46	+119.17%
FPG	16.48	19.51	+18.39%	13.08	14.12	+7.95%	2.51	3.11	+23.90%
SCAPE-GRU	19.02	22.47	+18.14%	15.26	16.41	+7.54%	2.88	3.75	+30.21%
GTP	15.97	18.28	+14.46%	12.75	13.67	+7.22%	2.45	2.92	+19.18%

Appendix

Table 15: Full PENS headline-generation quality results. Base and repaired values equal

100

times the native metric score. Higher is better. Gain is the reported relative percentage change from Base. RG abbreviates ROUGE, W denotes weighted overlap, and P denotes paraphrase-aware overlap. Both panels use the same generated outputs. Encoder hosts use the aligned repair condition, while LLMs use the frozen soft-prefix interface. Definitions are in Appendix D.3 .

Host	PSE-W-JSD			PSE-W-RG-L
	Base	+ REPAIR	Gain (%)	Base	+ REPAIR	Gain (%)
DeepSeek-32B	118	129	+9.32%	148.5	156	+5.05%
Qwen2.5-32B	92	121	+31.52%	125.3	141.1	+12.61%
SCAPE	103	112.5	+9.22%	128	135.2	+5.62%
FPG	74.8	83.9	+12.17%	104.2	110.4	+5.95%
GTP	89.5	98.7	+10.28%	113.8	120.8	+6.15%

Appendix

Table 16: Full PENS personalization-sensitive results on the same outputs as Table 15 . Base and repaired scores are displayed in units of

10^{-3}

, so a displayed value of

118

denotes a native score of

0.118

. Higher is better, and Gain is the reported relative percentage change. PSE denotes PerSEval. JSD, W, P, and RG are defined in Appendix D.4 . Encoder and LLM repair interfaces follow the same distinction as in the generation-quality table.

Host	Interface	Ranking		Generation Quality		Personalization
Host	Interface	MRR	nDCG@10	RG-L	METEOR	PSE-W-JSD	PSE-W-RG-L
NAML	T1	+859.69	+1767.90	+8.20	+25.24	+32.66	+9.94
NAML	T2	+859.69	+1767.90	+7.96	+23.28	+32.30	+9.49
EBNR	T1	+351.70	+482.86	+7.76	+27.51	+31.09	+10.33
EBNR	T2	+351.70	+482.86	+7.48	+26.87	+30.25	+9.53
NRMS	T1	+926.27	+1916.22	+8.21	+24.14	+35.15	+10.20

Appendix

Table 17: Full RQ3 comparison of task-specific downstream use on PENS. Normalized relative gains (%) after preference-state repair across ranking, generation-quality, and personalization-sensitive metrics. The original results in absolute values are in Appendix Tables 12 , 15 , and 16 .

Dataset	Host	Base	NoTS	L	S	E	L+S	S+E	L+E	L+S+E
MovieLens	Mamba4Rec	29.29	28.43	29.92	29.63	29.45	32.17	31.35	31.04	34.26
	TiSASRec	35.51	35.52	35.54	35.78	35.83	35.91	35.66	35.41	36.78
	SIGMA	33.27	33.29	33.66	33.32	33.29	33.88	33.73	33.43	34.81
MIND	NAML	45.65	45.95	47.65	48.32	47.13	49.31	49.18	48.79	53.22
	EBNR	47.08	47.74	48.88	48.03	47.62	50.13	49.38	48.76	53.22
	NRMS	44.91	45.06	46.43	45.13	47.73	47.58	46.44	46.15	52.74

Appendix

Table 18: Temporal-pathway ablation on MovieLens and MIND. All values are MRR on a

0

100

scale, with higher values better. Base is the original host, NoTS removes temporal resolution, and

L

S

, and

E

retain the indicated long-support, short-support, and episodic-support branches. The full configuration is

L+S+E

. Mamba4Rec and SIGMA are state-space hosts, TiSASRec and NRMS use self-attention, NAML uses attentive aggregation, and EBNR is recurrent. These aligned configurations are distinct from the controlled input-timing diagnostic in Table 4 .

Dataset	Host	Base	$\mathcal{L}_{\mathrm{task}}$	$\mathcal{L}_{\mathrm{task}}+\mathcal{L}_{\mathrm{pos}}$	$\mathcal{L}_{\mathrm{task}}+\mathcal{L}_{\mathrm{pos}}+\mathcal{L}_{\mathrm{coh}}$
MovieLens	TiSASRec	35.51	35.71	36.14	36.78
MovieLens	Mamba4Rec	29.29	31.43	32.16	34.26
MIND	LSTUR	51.43	52.17	54.65	55.84
MIND	NRMS	44.91	46.39	51.29	52.74

Appendix

Table 19: Progressive supervision ablation on MovieLens and MIND. Values are MRR on a

0

100

scale, with higher values better. Successive columns add task utility, timestep provenance, and source coherence. This fixed-order comparison measures the evaluated combinations and does not isolate each family under every possible combination. Appendix G.4 connects it to the conditional local theorem.

Host	$i$	MRR			nDCG@10			SelSteps
		Base	NS	SEL	Base	NS	SEL
Mamba4Rec	10	17.43	19.32	19.45	22.62	27.15	27.94	7
	20	21.65	23.22	23.43	27.85	30.61	31.24	16
	30	26.11	28.45	28.29	33.32	35.40	35.36	22
	40	27.32	29.89	29.45	35.53	36.41	36.21	28
	50	29.29	33.95	34.26	38.50	40.03	40.17	34

Appendix

Table 20: Length-controlled selective-retention ablation. At each history length

i

, Base is the original host, NS uses non-selective repair, and SEL uses adaptive timestep-coordinate retention. MRR and nDCG@10 use a

0

100

scale, with higher values better. SelSteps counts positions retained by at least one corrective coordinate. Mamba4Rec and TiSASRec use MovieLens. NAML and EBNR use MIND. Comparisons between NS and SEL hold the prefix length fixed. Appendix G.5 discusses both gains and reversals.

Dataset	Host Encoder	Noisy Base		+ REPAIR
		MRR	nDCG@10	MRR	nDCG@10
MovieLens	Mamba4Rec	18.21	23.94	21.43	25.74
MovieLens	TiSASRec	22.34	25.92	24.65	28.17
MIND	NAML	28.61	29.23	32.85	35.81
MIND	EBNR	29.63	32.44	31.65	35.17

Appendix

Table 21: Noisy-history stress test with approximately

35

40\%

of valid positions replaced with random training-pool content. The supervised target and the relative order of non-corrupted interactions are preserved. Noisy Base and repaired scores use the same corrupted input without retraining. MRR and nDCG@10 are on a

0

100

scale, with higher values better. Appendix G.6 states the scope of the test.

Host	Base MRR	+ REPAIR	$\Delta$	Gain (%)
Mamba4Rec	29.29	34.26	+4.97	16.96
TiSASRec	35.51	36.78	+1.27	3.57
EBNR	47.08	53.22	+6.14	13.04
NAML	45.65	53.22	+7.57	16.58
DeepSeek-32B	1.31	1.83	+0.52	39.69

Appendix

Table 22: Serving overhead relative to the same host in the same execution environment. Base, repaired MRR, and

\Delta

use a

0

100

scale. Gain is a relative percentage. Memory overhead is derived from parameter counts and excludes activation and cache residency. Throughput change is derived from reciprocal latency. The lower panel divides multiplicative score gain by multiplicative resource cost. Ratios above one favor score gain relative to cost. Ratios below one favor cost growth. The LLM row uses its separate soft-prefix serving protocol.

Model	Work Unit	Base	Added Repair	Total
Mamba4Rec	GFLOPs/query	0.75	0.18	0.93
TiSASRec	GFLOPs/query	1.45	0.18	1.63
NAML	GFLOPs/query	0.03	0.02	0.05
DeepSeek-32B	GFLOPs/token	64.0	unreported	64.0

Appendix

Table 23: Arithmetic work for representative serving paths at history length

T=50

. GFLOPs denotes billions of floating-point operations. Budget columns are required arithmetic throughput, computed as work divided by

50

25

milliseconds. They are not measured latencies. Scaling describes the host path. Encoder work is per query, whereas DeepSeek work is per generated token. For DeepSeek, added repair work is unreported, and the total is therefore a decoder reference rather than a complete repaired-system estimate.

Host And Serving Mode	Resident Artifacts	Online Operations And Cost Interpretation
Mamba4Rec . end-to-end encoder serving	Frozen : item embedding table + 4-layer selective SSM encoder. Added : repair module attached to exposed terminal state.	one frozen encoder pass + residual repair fusion. Sequence evidence is cached within the existing forward pass and released after the step.
TiSASRec . end-to-end encoder serving	Frozen : item table + time-interval-aware self-attention encoder. Added : repair module attached to exposed terminal state.	attention score/value passes + repair fusion. Latency is more sensitive to sequence interactions because the host is attention-based rather than SSM-based.
NAML . cached news-table serving	Frozen : cached NEWS_TABLE + frozen user encoder. Added : repair module attached after cached-history composition.	table lookup / memory bandwidth + attentive user aggregation. Serving shifts a large fraction of work out of online text encoding and into cached vector retrieval.
DeepSeek-32B . context-conditioned autoregressive decoding	Frozen : frozen decoder weights + KV cache. Added : repair module + soft-prefix projection.	token-by-token decoding. Prefix-conditioned decoding builds the appropriate attention cache for the modified input.

Appendix

Table 24: Resident artifacts and online operations for the evaluated serving paths. A within-forward sequence cache, a persistent news-vector table, and an attention key-value (KV) cache serve different roles. The LLM decoder is evaluated with the inserted soft prefix, so a KV cache for an unmodified prompt cannot be reused unchanged. Appendix G.7 relates these paths to the measured and estimated costs.

Host	MRR			Hit@10
	Base	+ REPAIR	$\Delta$	Base	+ REPAIR	$\Delta$
SASRec	4.11	4.56	+0.45	7.82	8.76	+0.94
TextCNN	2.57	2.78	+0.21	5.08	5.04	-0.04
GRU4Rec	4.14	4.56	+0.42	8.80	9.13	+0.33

Appendix

Table 25: TG-ReDial recommendation and response generation from dialogue-derived interactions. Base and repaired conditions use the same held-out task instances. Recommendation ranks one target against

499

negatives. Reported scores are on a

0

100

scale, with higher values better.

\Delta

is an absolute score-point change. The first two panels evaluate recommendation, and the third evaluates generation. RG denotes ROUGE. The panels use distinct task interfaces. We discuss the mixed TextCNN response and the difference between the two generators in Appendix G.8 .

Explore similar work

Oct 1, 2026cs.LG

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

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.

Parthiv Chatterjee, Kashish Kanjaria, Vashisth Purani +2

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

Personalization

Ask PDF

Jul 12, 2026cs.IR

RecRec: Recursive Refinement for Sequential Recommendation

Sequential recommender systems typically infer user preferences through single-pass encoding of interaction histories without iterative refinement, relying on increasingly deep architectures to capture complex patterns. In this work, we revisit sequential recommendation from a recursive inference perspective: can user preferences be modeled as a persistent latent state that is recursively refined? We propose RecRec (Recursive Recommendation), a lightweight model that maintains a compact latent state and updates it through a shared recursive module conditioned on interaction evidence. Unlike prior recursive models, RecRec introduces an evidence-anchored correction mechanism that stabilizes refinement by grounding each update in the original interaction context, preventing semantic drift during deep recursive reasoning. Experiments on three benchmark datasets under standard evaluation protocols show that RecRec matches or outperforms state-of-the-art sequential, graph-based, and reasoning-enhanced recommenders while using only 3.9M to 14M parameters. Ablation studies demonstrate that both recursive refinement and the evidence-anchored correction gate contribute significantly to performance, highlighting the effectiveness of recursive latent inference as a scalable alternative to deeper or language-based architectures. Code is available at https://anonymous.4open.science/r/RecRec-6B67/README.md.

Pervez Shaik, Prosenjit Biswas, Abhinav Thorat +2

Sony Research India India

Multimodal Recommendation Model Sequential Recommendation Recommendation Refinement Iterative Self-Correction

Ask PDF

May 26, 2026cs.CL

LATTE: Forecasting Peer Anchored Preference Trajectories for Personalized LLM Generation

Personalized generation with frozen large language models requires a conditioning signal that is both compact and current. Existing personalization methods typically retrieve or summarize user histories in text, or compress them into static latent profiles and soft prompts. These approaches are efficient, but they treat a user's past behavior as an aggregate profile and therefore mix stable identity, recent drift, and item content in the same representation. We propose LAtent Trajectory Tracking and Extrapolation (LATTE), a framework that represents personalization as forecasting a peer anchored relative preference state. For each historical session, LATTE subtracts a time masked baseline formed from comparable users who responded to the same item, producing a state that measures how the target user differs from peers under a shared item context. A lightweight sequence predictor then forecasts the next state in this trajectory, and a State to Token Bridge injects the forecast into a frozen instruction tuned LLM through a single anchored soft token. We provide a latent factor analysis showing when peer anchoring cancels shared item variation and why temporal forecasting trades off stale averages against noisy recent states. Experiments on Amazon Reviews 2023 and MemoryCD show that LATTE consistently outperforms retrieval, summary memory, static latent profiles, difference aware latent profiles, and soft prompt compression baselines. On Amazon Reviews 2023, LATTE improves average ROUGE-L from 0.219 for a static latent profile and 0.245 for the strongest added latent compression baseline to 0.259. Additional pairwise comparisons and diagnostic analyses suggest that the improvement is mainly due to forecasting user-specific trajectory information, rather than merely adding a soft prompt interface.

Jinze Li, Xiaoyan Yang, Shuo Yang +5

The University of Hong Kong · Ant Healthcare, Ant Group

Large Language Model Personalization Frozen Language Model Large Language Model Generation Latent Variable Peers

Ask PDF