PaperlayerView exact revision on arXiv

Source-linked AI summary

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

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

arXiv:2610.01375v1cs.LG

TL;DR

Movie, news, and dialogue histories differ in their native actions and outputs, raising whether they can train one reusable history-update mechanism. The paper formalizes action-on-item sharing and implements it in PerTIDE; experiments report reuse across predictive and generative consumers, including a 15.23-point transfer comparison.

  • Problem

    Movie, news, and dialogue histories differ in action vocabularies, item types, temporal structure, and objectives, leaving open whether their interactions can train reusable memory updates.

    Source: §1 · p. 1

  • Method

    The paper maps interactions to action roles and content embeddings, formalizes invariance, conditional stability, and compatible pooled learning, and implements MTSH with PerTIDE's gated multi-trace encoder and command-conditioned readout.

    Source: §1 · p. 2

  • Results

    15.23 MRR points separate a frozen PENS-trained core from an identically structured random core after fitting the same MovieLens consumer, alongside PENS prediction and generation and a 4.12-point MIND advantage over a same-input three-branch control.

    Source: §6.1 · p. 6, §6.2 · p. 7, §6.3 · p. 9

  • Takeaways & Limitations

    The theory and experiments support learning history updates across compatible sources and reusing them through predictive and generative consumers.

    Source: §8 · p. 9

    Reuse depends on supplied action mappings, informative content embeddings, and compatible downstream semantics; reported resource figures are estimates rather than matched-hardware measurements.

    Source: §7 · p. 9

Abstract

from arXiv · show

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.

1 Introduction

Personalization histories differ in actions, items, and objectives, but each interaction can update user memory; the paper asks whether shared action-on-item updates can be reused across predictive and generative tasks. It formalizes this sharing, proposes PerTIDE, and reports cross-task and transfer results.

Motivation

  • Movie, news, and dialogue histories differ in action vocabularies, item types, temporal structures, and downstream objectives, yet each interaction provides evidence about an evolving user.The motivating question is whether their common action-on-item form supports one encoder reusable across predictive and generative tasks.Source: §1 · p. 1

Prior work and gap

  • Prior systems share tasks, representations, or retrieved histories, but those forms of reuse alone do not establish how a changing action or item affects evolving state.The paper studies that response, its stability, and whether learned update operations remain useful through different consumers.Source: §1 · p. 2

Formal account

  • An action-on-item interface maps native interactions to roles and content embeddings, enabling shared updates with invariance, conditional stability, and pooled-learning guarantees under compatibility conditions.The stability bound concerns the supplied embedding geometry; pooled learning explicitly accounts for distribution and label-semantic discrepancies.Source: §1 · p. 2

MTSH and PerTIDE

  • MTSH separates evidence entry, persistence, and task use; PerTIDE implements these with action gating, long-, short-, and episodic traces, adaptive fusion, and command-conditioned readout.Users retain separate history-specific states while sharing update operations across pooled event trajectories.Source: §1 · p. 2

Contributions and results

  • The frozen PENS-trained core exceeds an identically structured random core by 15.23 MRR points after both use the same fitted MovieLens consumer.On MIND, PerTIDE also leads a same-input three-branch SSM control by 4.12 MRR points; trace interventions show complementary evidence from combining traces.Source: §1 · p. 2

2 Related Work

Prior personalization work shares tasks, representations, temporal histories, or generation context. This paper focuses on how an action-dependent event update changes user state and is retained and consumed across tasks.

Sharing tasks and representations

  • Task and representation sharing includes language-interface systems, multi-task methods, transferable item or sequence representations, and universal-plus-domain-specific representations.The paper's question is how each event's action-dependent update behaves under item substitution and through retention and readout.Source: §2 · p. 2

Temporal history encoders

  • Temporal history encoders model profiles, sequences, interests, and memory, while MTSH separates event entry, persistence, fusion, and task consumption on a shared interface.A same-input three-branch control tests the additional value of PerTIDE's retention parameterization.Source: §2 · pp. 2–3

Personalized generation and event construction

  • Personalized generation uses retrieved evidence, user embeddings, or user-specific parameters; PerTIDE instead supplies a command-adapted history state to predictive heads or a prefix-conditioned decoder.Reuse of the learned core is evaluated separately from training a consumer for a new output space.Source: §2 · p. 3

3 Action-On-Item Preference Flow

The action-on-item interface represents interactions as mapped roles paired with item or response embeddings, then asks how evidence should enter, persist, and be consumed. MTSH organizes those operations into complementary temporal traces and task-specific readout.

3.1 A Shared Event Interface

  • The event pairs a mapped action role with the item or response acted upon, whose embedding supplies continuous content; a separate command seed specifies the task.Roles come from preprocessing, while the encoder learns how each role modulates content evidence.Source: §3.1 · p. 3
  • For equal initial states and commands, identical mapped trajectories produce identical states; changing an item affects its later trace contribution according to embedding distance when subsequent inputs agree.The complete-state bound also accounts for fusion, the current-input residual, and recurrent readout; halving the embedding perturbation halves the upper bound.Source: §3.1 · p. 3
  • Pooled learning has an estimation term plus schema-distribution and label-semantic discrepancies; compatible distributions and conditional label semantics yield a shared-model estimation allowance scaling as 1/√N.Examples from participating domains can contribute to estimating one shared update rule when the compatibility conditions hold.Source: §3.1 · p. 3

3.2 Multi-Timescale State Hypothesis

  • MTSH retains action-conditioned evidence through long-term, short-term, and episodic traces, whose intended retention biases must be tested rather than assumed to differ in practice.The decomposition includes evidence entry, trace persistence, fusion, routing, and readout.Source: §3.2 · p. 4
  • A residual path carries the current event into fusion, while routing uses the preceding task-ready state; shared operators still produce user-specific histories and command-adapted outputs.The task determines how the retained history is consumed.Source: §3.2 · p. 4

4 PerTIDE: Instantiating MTSH

PerTIDE instantiates MTSH as a shared history encoder that gates event content by action, maintains three state-space traces, fuses them, and reads them out for a command. Its user-specific history can condition predictive or generative consumers.

Architecture

  • Each schema-normalized event becomes action-conditioned evidence, passes through long-, short-, and episodic traces, and is fused into a user-specific preference-history state.The active command converts that history into a task-ready representation for a predictive or generative consumer.Source: §4 · p. 4

Action entry

  • The supplied action role controls how item or response content enters the state, but a negative-role gate has nonnegative entries and does not itself subtract embeddings or impose monotone dislike.Thus clicked and ignored content can share an embedding space while contributing differently through their roles.Source: §4 · p. 4

Trace persistence

  • Three lag-biased SSM traces share an orthogonal eigenbasis; their slow and fast modal bases produce different retention, while trace names indicate intended biases rather than guaranteed ordering or semantic separation.The trace kernels may overlap, and event inputs leave decaying tails.Source: §4 · p. 4

Trace fusion

  • Adaptive fusion combines the updated traces, with a residual path carrying the current event directly rather than adding another recurrent trace.This is the implemented fusion operation.Source: §4 · p. 4

Command consumption

  • A command-conditioned readout converts the user-specific history state into a task-ready representation, allowing one news history to support next-news prediction or headline generation.The command seed and downstream consumer determine the required representation.Source: §4 · p. 5

Recurrent user memory

  • The recurrent cache holds three 768-dimensional traces and the preceding task-ready state in 12 KiB of FP32 memory, independent of history length.This excludes model weights and runtime estimates.Source: §4 · p. 5

Training and consumers

  • The encoder is jointly trained on MovieLens, MIND, PENS, and OpenAI trajectories using temporal-position, next-event, and command-classification objectives with globally fixed weights.For generation, a pretrained DistilGPT2 decoder uses a learned prefix; contextualization modules and selected decoder components are optimized while the history encoder remains fixed.Source: §4 · pp. 5–6

5 Evaluation

The evaluation spans predictive histories and personalized generation, with task-native and schema-matched controls. It also tests source-trained transfer and specifies metrics and evaluation protocols.

Datasets and tasks

  • Five main settings cover MovieLens-1M, MIND, PENS, OpenAI-Reddit, and MPChat, with OpenAI-Reddit used only as an eventization stress test.PENS also supports personalized headline generation; MPChat tests dialogue event-role inference as a boundary case.Source: §5 · p. 6

Baselines

  • Baselines include task-native systems, prompted instruction-tuned LLMs, and schema-matched controls that isolate state design from input construction.The MIND controls use the same ordered interaction-block trajectories; Tri-SSM-B matches three branches but omits explicit L/S/E retention maps.Source: §5 · p. 6

Metrics and protocol

  • Prediction uses one positive target and 699 sampled negatives with MRR, nDCG, and HR@10; generation uses PerSEval, reference metrics, and human similarity ratings.Generation metrics measure reference agreement rather than article factuality; main comparisons use paired bootstrap evaluation with 10,000 resamples.Source: §5 · p. 6

6 Results And Analysis

PerTIDE supports predictive and generative tasks and outperforms listed task-native comparators across the five main settings. Controlled comparisons and interventions examine its retention structure, update stages, and reuse across consumers and domains.

6.1 RQ1: Can The Shared Encoder Serve Prediction And Generation?

  • Across five main settings, PerTIDE improves primary metrics over listed task-native comparators, including 2.36/3.47 MovieLens MRR/nDCG points and 10.05/9.01 MIND MRR/nDCG@5 points.On PENS, prediction gains are 0.90/1.42 MRR/nDCG points over LSTUR; the same encoder also supports headline generation.Source: §6.1 · p. 6, Table 1 · p. 7

6.2 RQ2: Do Entry, Retention, And Readout Explain The Gains?

  • 4.12 MRR, 5.37 nDCG@5, and 9.15 HR@10 points separate PerTIDE from Tri-SSM-B on MIND under the same-input, three-branch control.PerTIDE reaches 46.65 MRR; the comparison supports an advantage for its additional retention parameterization in this setting.Source: Table 2 · p. 7, §6.2 · p. 7
  • Removing action conditioning lowers MIND MRR/HR@10 from 46.65/55.48 to 40.13/46.35, while removing command-conditioned readout gives 42.51/51.06.A fixed-trajectory role-corruption experiment also shows MRR falling from 12.14 to 2.81 as positive/negative labels move from 0% to 100% corruption.Source: §6.2 · p. 7
  • Each isolated trace is strongest in its corresponding MIND probe regime, while the fused state leads every reported probe, including overlapping specialization.MovieLens is an exception to isolated specialization: L slightly exceeds S on S-Tr, at 27.18/29.35 versus 27.13/29.22 MRR/nDCG@10.Source: §6.2 · p. 7
  • Adding the episodic trace to L+S raises MIND MRR by 5.39 points and PENS PSE-METEOR by 0.173, although the episodic trace is weak alone.On MIND, E alone reaches 24.18 MRR; adding it to L+S raises MRR from 41.26 to 46.65.Source: Figure 2 · p. 8, §6.2 · p. 8

6.3 RQ3: Do Learned Updates Survive A New Consumer Or Domain?

  • On PENS, the same encoded histories support next-news prediction and personalized headlines, and UGI improves on UFI by 0.034/0.052/0.052 across the three PSE metrics.Source-trained transfer to MovieLens beats the strongest baseline by 1.43/2.98/5.26 MRR/nDCG@10/HR@10 points with 3,000 target trajectories.Source: §6.3 · p. 8, Table 3 · p. 8
  • A frozen PENS-trained core reaches 16.344 MovieLens MRR with a trained head, versus 1.115 for a frozen random core with the same trained head.The 15.23-point advantage isolates useful source-trained structure; the separate protocol's scores should not be combined with Table 3.Source: Table 4 · p. 9, §6.3 · p. 9
  • Schema distance has modest correlations with frozen/adapted target risk (0.11/0.16) and reciprocal-rank gain (−0.09/−0.11), with all four confidence intervals excluding zero.These associations do not establish that the theorem’s stability or conditional-label assumptions hold in the experiments.Source: §6.3 · p. 9

7 Discussion And Limitations

The discussion frames the learned history update as the reusable object, while noting limits in generation evidence and the conditions required for reuse.

Results and interpretation

  • PENS generation improves over DeepSeek-32B on RG-SU4 and BLEU-4, but human similarity ratings are close and BERTScore is slightly lower.The authors frame the findings as supporting reference agreement rather than article factuality.Source: §7 · p. 9

Limitations

  • Reuse depends on supplied action mappings, informative content embeddings, and compatible downstream semantics.The paper also notes constructed chronology for OpenAI evaluation and deficits in MPChat response prediction and speaker identification.Source: §7 · p. 9

8 Conclusion

The conclusion presents shared history updates as reusable across compatible sources and task-specific consumers.

Conclusion

  • Action-on-item histories support shared update learning while preserving separate user states.The schema provides exact invariance, a conditional continuity bound under item changes, and a pooled-training bound with explicit compatibility conditions.Source: §8 · p. 9
  • MTSH specifies how evidence enters, persists, and is consumed, and PerTIDE implements those operations in a shared encoder.Prediction and generation results demonstrate multiple uses of the learned history state.Source: §8 · p. 9
  • A controlled frozen-core experiment provides evidence that source-trained updates remain useful after a domain change and consumer alignment.Same-input controls and trace interventions connect that reuse to the proposed update structure.Source: §8 · p. 9

AI Use Statement

Generative AI tools assisted with technical and manuscript work, while the authors retained responsibility for the research and its claims.

AI Use Statement

  • Generative AI tools were used to cross-check technical reasoning and assist with literature retrieval, mathematical exposition, and verification.Source: AI Use Statement · p. 10
  • The tools also assisted with manuscript drafting and editing, and code cleaning and documentation.Source: AI Use Statement · p. 10
  • The authors determined the research questions, experimental decisions, scientific claims, and final interpretations, and retained responsibility for the paper’s veracity, originality, and integrity.Source: AI Use Statement · p. 10

A Datasets

The datasets preserve distinct interaction evidence and provenance while supplying events to a shared interface; the schema supports invariant state updates and conditional stability under item changes.

A Datasets

  • Table 5 spans five predictive and generative settings; ranking uses one positive target plus 699 sampled negatives, while generation appends produced responses to trajectories.Source: Table 5 · p. 15

A.1 MIND-Large Dataset

  • MIND-Large is the full-scale Microsoft News recommendation dataset, with anonymized user logs and over 15 million impressions.It contains 1,000,000 users and 161,013 English news articles; MIND-small is a randomly sampled 50,000-user variant.Source: §A.1 · p. 15

A.2 PENS Dataset

  • PENS combines news articles and interaction logs with personalized headline annotations, and its offline test set uses participant-written headlines for unseen articles.The training pool samples 500,000 impressions; test articles have four participant-written references on average, with editorial checks for factual correctness.Source: §A.2 · p. 15

A.3 OpenAI (Reddit) Dataset

  • OpenAI–Reddit contains human and generated summaries with annotator comparisons, but its annotations are not indexed by stable user identities or reading histories.A curated subset of 1,038 posts yielded 7,713 summaries evaluated by 64 annotators.Source: §A.3 · p. 15

A.4 MPChat Dataset

  • MPChat is a multimodal persona-grounded dialogue dataset whose personas are image–sentence pairs and whose dialogues are grounded in those elements.It contains 15,000 dialogues and 42,531 utterances, split chronologically into train, validation, and test sets.Source: §A.4 · p. 16

B.1 Dataset-Specific Construction

  • Dataset-specific construction maps recorded interactions, derived ratings, or dialogue turns into UIG events while preserving their different provenance.PENS uses click/skip and constructed response transitions; OpenAI–Reddit uses thresholded feedback and randomized trajectories; MPChat infers stance labels from response cues, sentiment, and text similarity.Source: §B.1 · p. 16

B.2 What The Schema Shares

  • Matching schema pairs, commands, and initial states yield identical encoder states, while shared update parameters operate on separate user states.Native action names cease to matter when the mapped role and item representation match; a role match alone does not make items, histories, or users equivalent.Source: §B.2 · pp. 16–17
  • Item-embedding changes yield bounded continuous-state changes under stated assumptions, but uniform long-horizon readout bounds require an additional contraction condition.Continuity of a state does not imply continuity of a discrete recommendation or generated token.Source: §B.2 · pp. 17–18

B.3 Transfer And Pooled Training

This section sets out conditions for transfer and pooled training across compatible domains, along with the augmentation and stability principles underlying PerTIDE’s trace timescales. It characterizes what the geometric traces and their mixtures can represent, and describes decoder contextualization variants and the compute setup.

B.3 Transfer And Pooled Training

  • Transfer risk is bounded by source risk plus conditional-loss discrepancy and L times the Wasserstein distance between input marginals, assuming a shared Lipschitz conditional loss over comparable schema inputs.Comparable input geometry alone is insufficient when labels or command semantics conflict.Source: §B.3 · p. 18
  • Pooling yields a target-risk bound from pooled empirical risk, Rademacher complexity, and transfer discrepancy; under matched distributions and zero discrepancy, its estimation bound is no worse than a domain-only bound using n_t samples.The comparison requires independent examples and compatible transfer assumptions; it is about estimation bounds, not guaranteed lower empirical risk or higher accuracy.Source: §B.3 · p. 19

C UIG Augmentation

  • UIG augmentation expands trajectory diversity by substituting contiguous segments while preserving target prefixes and gaps, then contextually rewrites surrogate responses to improve local coherence.For PENS, these operations create the augmented training pool; augmentation is applied within separate native training and evaluation pools for OpenAI–Reddit.Source: C UIG Augmentation · pp. 19–20

D.1 Discrete-Time SSM Stability

  • A discrete linear trace is asymptotically stable when its transition matrix has spectral radius below 1; for diagonalizable transitions, each modal base must have magnitude below 1.Under exponential discretization, this condition corresponds to negative real parts of continuous-time eigenvalues; Euler-style discretization also requires a step-size constraint.Source: §D.1 · pp. 21–22

D.2 SSM Hyperparameter Choices

  • The SSM hyperparameter section gives sufficient, conservative rules for choosing the short-memory horizon and event-step size, rather than necessary conditions.The spectral formulas alone do not establish that the three branches have ordered retention ranges.Source: §D.2 · p. 22, §D.1 · p. 22

Short-Memory Retention And Lag

  • When the short-trace base aS is small, evidence at lags beyond a short window is geometrically suppressed.Choosing a base and reading the resulting horizon are distinct operations.Source: Short-Memory Retention And Lag · p. 23

Event-Kernel Step Size Selection

D.3 Kernel Interpretation Of The Three PerTIDE Traces

  • Each PerTIDE trace induces a geometric lag kernel, whose normalized expected lag is a/(1 − a); larger modal bases increase effective memory length.Different branch parameterizations can overlap in attainable kernels, so branch names alone do not establish semantic disentanglement.Source: §D.3 · pp. 24–25
  • When 0 < aS < aE < aL < 1, the traces have correspondingly ordered half-lives; that ordering requires this additional basis condition.Increasing a slows geometric decay and increases both expected lag and half-life.Source: §D.3 · p. 25

Long-Term Trace: Persistent Evidence Accumulation

  • Repeated positive evidence accumulates into a large persistent long-term state when its modal eigenvalue is close to one.This result concerns repeated evidence in a stable preference direction.Source: Long-Term Trace: Persistent Evidence… · p. 25

Short-Term Trace: Recent Burst Tracking

  • With zero initial state and the sum truncated at lag t −1, the short-term modal state is expressed as a finite sum of past inputs.The supplied passage gives the state representation but does not provide its full formula.Source: Short-Term Trace: Recent Burst Tracking · p. 26

If a(S)

  • When a(S) is small, newest inputs dominate the short-term state while older inputs are geometrically suppressed.The decay is rapid, so recent inputs receive greater relative weight.Source: If a(S) · p. 26

Episodic Trace: Decaying Event Imprint

  • After a finite event window closes, its episodic imprint decays at rate a(E), with τk controlling the modal decay; a shared scalar η cannot match arbitrary target time constants for every fixed λE,k.With exact kernel matching, the post-event imprint is governed by the event time constant.Source: Episodic Trace: Decaying Event Imprint · pp. 26–27

D.4 Capacity Beyond The Branch Names

  • On a finite set of lags, sufficiently rich real geometric mixtures can represent any specified filter, but this gives no infinite-horizon guarantee or assurance that training discovers the coefficients.The representation may require signed coefficients and one mode per constrained lag; a fixed-width encoder cannot invoke it for arbitrary horizons.Source: §D.4 · p. 27

D.5 Decoder Contextualization Variants

  • UFI fuses the predicted query-conditioned representation with the query-document embedding, whereas UGI first uses user-conditioned FiLM modulation to emphasize or suppress representation components (Perez et al., 2018).Both variants produce a user-conditioned control for decoder contextualization.Source: §D.5 · p. 28

E.1 Compute Resources

  • Experiments ran on A100 and RTX 6000-family GPUs, but exact device counts and experiment-to-GPU assignments were not preserved in the run records.Pretrained language models used automatic device placement.Source: §E.1 · p. 29

E.2 Training

This section covers training and paired-bootstrap evaluation protocols, encoder families and personalized baselines, and results on recommendation, generation, cross-domain reuse, retention, and deployment costs. It also examines action labels, history length, summary generation, and sparse-action robustness, while qualifying what the reported metrics and protocols establish.

E.2 Training

  • Training uses one optimizer step per valid row for 45 epochs over 78,274 rows, with deterministic epoch shuffling and a processed prefix that leaves the final two positions as query and target.The main encoder uses pooled MovieLens, MIND, PENS, and OpenAI trajectories, while the source-only transfer encoder uses PENS alone; generation freezes PerTIDE and trains decoder contextualization modules, the final two DistilGPT2 blocks, and final layer normalization using Lgen = LCE + 0.15 Lsem.Source: §E.2 · p. 29

E.3 Statistical Significance Testing

  • Paired bootstrap testing uses 10,000 resamples of identical held-out examples and history prefixes, supporting gains when the one-sided 95% confidence interval’s lower bound exceeds zero.For predictive recommendation, both systems also use identical candidate sets with one ground-truth target and 699 sampled negatives; the procedure measures sampling variability for fixed trained systems, not retraining, split, or negative-pool variability.Source: §E.3 · pp. 29–30

F.1 Encoder Baselines

  • Encoder baselines are grouped by exposed state as profile aggregation, recurrent or recency-bounded state, and long-horizon sequence representations, without treating the families as mutually exclusive assumptions about user dynamics.The comparison includes a single-timescale discretized SSM and a three-branch SSM control; the latter matches PerTIDE’s schema-normalized input and downstream organization but lacks explicit trace-specific retention design.Source: §F.1 · pp. 35–37

F.2 Baseline Personalized Models

  • Personalized headline-generation baselines inject user representations through decoder initialization or attention, while GTP uses a two-stage customization pipeline and Signature Phrase conditions generation on learned history phrases.GTP adds Information Self-Boosting and Masked User Modeling; Signature Phrase retains grounding in the source article.Source: §F.2 · pp. 37–38
  • MovieLens recommendation baselines span a persistent profile (BPR-MF), recurrent or bounded-window states (GRU4Rec and Caser), and long-horizon sequence encoders including S3Rec, Diff4Rec, and SASRec.All candidates are ranked under a one-positive plus sampled-negative protocol; SASRec uses causal self-attention over the interaction prefix.Source: §F.2 · pp. 38–39

G.1 Motivation; G.2 PerSEval Formulation; G.3 PSE Metrics

  • PerSEval multiplies responsiveness by an accuracy-sensitive penalty, producing a [0, 1] score where higher values indicate better personalization under the chosen reference-distance convention.Its PSE variants use ROUGE-SU4 distance, Jensen-Shannon divergence, or METEOR-based distance; DEGRESS alone measures responsiveness without penalizing inaccurate summaries.Source: §G.2 · pp. 40–41, §G.1 · p. 40, §G.3 · p. 41

H Prediction Accuracy Metrics; I Reference-Based Content Metrics

J.2.1 MovieLens-1M Recommendation

  • On MovieLens, PerTIDE reaches 17.31/23.78/33.20 MRR/nDCG@10/HR@10, gaining 2.36/3.47/1.68 points over SSD4Rec.BERT4Rec has the strongest baseline HR@10 at 31.71, while SSD4Rec leads baseline MRR and nDCG; removing action conditioning or command readout reduces performance.Source: §J.2.1 · p. 46, Table 13 · p. 46

J.2.2 MIND News Recommendation

  • Against Tri-SSM-B, PerTIDE gains 4.12/5.37/9.15 MRR/nDCG@5/HR@10 points on MIND, while action and readout ablations lower MRR from 46.65 to 40.13 and 42.51.Against MINER, gains are 10.05/9.01/9.18 MRR/nDCG@5/HR@10 points; the controlled comparison holds trajectory representation and three-branch organization fixed.Source: §J.2.2 · pp. 46–47, Table 14 · p. 46

J.2.3 PENS Next-News Prediction

  • On PENS next-news prediction, PerTIDE reaches 9.94 MRR, 9.88 nDCG@5, 11.58 nDCG@10, and 18.32 HR@10, with gains over LSTUR of 0.90 MRR, 1.42 nDCG@10, and 5.39 HR@10 points.With one positive target per candidate set, HR@10 indicates whether the target appears in the top ten, not its position there; the histories are also used for headline generation.Source: §J.2.3 · p. 47, Table 15 · p. 47

J.3 Retention Probes And Trace Complementarity

  • The full L+S+E model leads all six dataset–regime probe combinations and reported MIND, MovieLens, and PENS-generation trace metrics, although isolated specialization overlaps.On MIND, E alone reaches 24.18 MRR, while adding it to L+S raises MRR from 41.26 to 46.65; on PENS, L+S reaches 0.264 PSE-JSD and 0.182 PSE-METEOR versus 0.366 and 0.355 for the full model. PENS S+E underperforms S alone, so the evidence supports complementarity in this setup, not that every added trace helps.Source: §J.3 · pp. 47–48, Table 16 · p. 47, Table 17 · p. 48

J.4.1 PENS Personalized Headlines; J.4.2 Reference Overlap And Human Similarity Ratings

  • On PENS, UGI reaches 0.366/0.341/0.355 PSE-JSD/PSE-SU4/PSE-METEOR, exceeding UFI by 0.034/0.052/0.052.PerTIDE also scores 41.32 RG-SU4 and 27.64 BLEU-4 versus DeepSeek-32B’s 29.42 and 19.31, with human similarity ratings of 3.21 versus 3.12; these measures do not establish factual consistency or user utility. Resource estimates are not matched-hardware benchmarks.Source: §J.1 · p. 43, §J.4.1 · pp. 48–49, Table 18 · p. 48

J.5.1 Frozen-Core Reuse And Schema Distance

  • 15.229 MRR points (15.23 rounded) separate the PENS-trained core with a trained MovieLens head (16.344 MRR) from the random core with a trained head (1.115 MRR).With 3,000 target trajectories, PerTIDE also improves over the strongest baseline by 1.43/2.98/5.26 MRR/nDCG@10/HR@10 points; target alignment improves all three systems.Source: §J.5.1 · p. 49

J.5.2 Role Corruption

  • MRR falls from 12.14 to 2.81 when positive/negative roles are corrupted on fixed PENS trajectories, showing prediction depends on supplied action labels.A 10% corruption rate reduces MRR by 2.01 points, while 50% corruption lowers HR@10 from 32.0 to 15.0.Source: §J.5.2 · p. 50, Table 21 · p. 50

J.5.3 History Length

  • Extending visible history from 10 to 40 events changes MRR from 11.874 to 11.926 and HR@10 from 27.8 to 28.6 on 500 PENS trajectories.The target and candidate set remain fixed; the reported changes are modest under this protocol.Source: §J.5.3 · p. 50, Table 22 · p. 50

J.5.4 Eventized OpenAI–Reddit Summaries

  • On eventized OpenAI–Reddit summaries, PerTIDE scores 0.285/0.261/0.294 PSE-JSD/PSE-SU4/PSE-METEOR, versus DeepSeek-14B’s 0.243/0.095/0.109.Chronology and positive/negative roles are constructed from feedback, with augmentation performed separately within native training and evaluation pools; this is not evidence of unconstrained cross-domain user-history transfer.Source: §J.5.4 · p. 50, Table 23 · p. 50

J.5.5 Sparse Positive-Action Histories

  • Under sparse positive-action histories, PerTIDE’s PSE-JSD drops 42.35% (0.366 to 0.211), while PSE-SU4 and PSE-METEOR drops are 28.45% and 26.2%.DeepSeek-R1-32B’s PSE-JSD changes from 0.263 to 0.152 (42.21% drop), with PSE-SU4/PSE-METEOR drops of 30.4%/47.2%; GTP’s PSE-SU4 changes from 0.017 to 0.009. Trace-specific causal attribution would require the corresponding controlled ablation.Source: §J.5.5 · pp. 50–51, Table 24 · p. 51

J.6 Dialogue As An Additional Application And Boundary

Dialogue histories with inferred stance roles extend PerTIDE beyond explicit news interactions, but the reported results distinguish generation gains from weaker predictive and speaker-identification performance.

J.6 Dialogue As An Additional Application And Boundary

  • PerTIDE exceeds Discretized SSM by 8.48/7.08/9.09/3.52 points in RG-2/RG-L/BLEU-4/BERTScore for dialogue generation.The results test dialogue eventization under imperfect preference logs.Source: §J.6 · p. 52, Table 25 · p. 52
  • For response prediction, PerTIDE reaches 45.64/54.25 R@1/MRR, below text-only SBERT at 51.32/64.76.The cross-input comparisons measure performance differences, not a causal upper bound on missing images or persona.Source: §J.6 · p. 52, Table 26 · p. 52
  • For speaker identification, PerTIDE reaches 41.77/47.19 R@1/MRR, while CLIP with multimodal persona reaches 62.17/73.08.Text persona improves several comparator configurations.Source: §J.6 · p. 52, Table 27 · p. 53
  • The dialogue results support applicability to inferred events, not uniform superiority; compatibility theory does not measure missing-persona or multimodal effects here.Generation improves in the reported comparison, while response prediction and speaker identification remain more difficult.Source: §J.6 · pp. 52–53

K Human Reference-Similarity Evaluation (PENS)

The PENS evaluation uses blinded human ratings to compare personalized headline summaries, while the prompting protocols specify history-conditioned generation and the assets retain their own usage conditions.

K Human Reference-Similarity Evaluation (PENS)

  • PerTIDE receives the highest average human-judgment score, 3.21, above DeepSeek-32B at 3.12 and the listed specialized baselines.Participants rated reference–generation similarity on a 1–6 scale with model identities hidden.Source: K Human Reference-Similarity… · p. 53

L LLM Prompting Protocol

  • LLM runs use temperature 0.4 and nucleus sampling with top_p = 0.9.The prompting overview applies these settings to generative tasks.Source: L LLM Prompting Protocol · p. 55

L.1 LLMs As Summarizers (Personalized Headline Generation)

  • For personalized headline generation, LLMs receive user history and a query document, emphasizing clicked topics, deprioritizing skipped topics, and requiring factual consistency.The output is constrained to one headline field.Source: §L.1 · p. 55

L.2 LLMs As Dialogue Generators

  • For dialogue generation, LLMs receive the entire discourse as context and are instructed to condition the next response on persona information and discourse.Source: §L.2 · p. 55

M Assets And Usage Terms

  • Assets retain their own attribution and access conditions; derived event trajectories grant no redistribution rights, and provider licenses alone do not certify downstream compliance.Source: M Assets And Usage Terms · p. 55

Related papers

Loading 2610.01375v1…