Institution profile

Comcast

Industry researchnorthamerica · us
Official website
Research library14linked papers
Opportunities0open roles
Selected work

Representative Papers

When LLM-Inferred User Context Adds Value in Production Streaming Recommendation

Sep 30, 2026

This study addresses the unclear applicability of large language model (LLM)-generated profiles versus aggregated embeddings for streaming media recommendation by systematically evaluating both user representation strategies in a production environment. Methodologically, LLMs are leveraged to generate natural language summaries, while an attention mechanism integrates long- and short-term behavioral features. The findings reveal that representation effectiveness depends significantly on user consumption patterns: habitual users benefit more from aggregated profiles, whereas exploratory users gain greater advantages from the semantic context provided by LLMs. Accordingly, this work proposes a dynamic profile selection strategy conditioned on user consumption states. This approach effectively circumvents the limitations inherent in either method alone, yielding substantial improvements in overall recommendation performance.

0 citationsRead paper

Routing Between Generative and Collaborative User Profiles: A Serving-Time Gate for Controllable Novelty

Sep 30, 2026

This study addresses the high cost and deployment challenges of large language model (LLM) user profiling, alongside the lack of efficient routing mechanisms between generative and collaborative filtering models in recommender systems. We propose a serving-time-only gating network that dynamically routes users to either collaborative sequential recommendation or LLM-based profiling models based on real-time context, enabling a controllable trade-off between novelty and relevance. Our analysis reveals that performance gains stem from precise selective routing rather than LLM generation alone, supported by an adjustable threshold control strategy. Experiments demonstrate that, under a 5% NDCG degradation constraint, the proposed method accurately routes only 12.5% of users to the LLM while improving Novelty@10 by 6.5%, significantly outperforming heuristic and random baselines.

0 citationsRead paper

Mechanisms of Misgeneralization in Physical Sequence Modeling

May 19, 2026

This work identifies and formalizes the phenomenon of “physical misgeneralization,” wherein generative sequential models in physical environments produce global distributions of conserved quantities—such as path length or mechanical energy—that deviate from design intent due to accumulated local modeling errors. The authors introduce controlled synthetic tasks to elucidate the underlying mechanism: local inaccuracies propagate through physical constraints, distorting the global distribution of these quantities. To address this, they propose a data bias kernel that predicts the direction of distributional shift, enabling a structured intervention strategy informed by this prediction. Experiments on maze navigation and double-pendulum dynamics demonstrate that the method effectively anticipates and mitigates distortions in physical quantity distributions, thereby validating both the proposed mechanistic explanation and the efficacy of the intervention approach.

0 citationsRead paper
Recent publications

Latest Papers

When LLM-Inferred User Context Adds Value in Production Streaming Recommendation

Sep 30, 2026

This study addresses the unclear applicability of large language model (LLM)-generated profiles versus aggregated embeddings for streaming media recommendation by systematically evaluating both user representation strategies in a production environment. Methodologically, LLMs are leveraged to generate natural language summaries, while an attention mechanism integrates long- and short-term behavioral features. The findings reveal that representation effectiveness depends significantly on user consumption patterns: habitual users benefit more from aggregated profiles, whereas exploratory users gain greater advantages from the semantic context provided by LLMs. Accordingly, this work proposes a dynamic profile selection strategy conditioned on user consumption states. This approach effectively circumvents the limitations inherent in either method alone, yielding substantial improvements in overall recommendation performance.

0 citationsRead paper

Routing Between Generative and Collaborative User Profiles: A Serving-Time Gate for Controllable Novelty

Sep 30, 2026

This study addresses the high cost and deployment challenges of large language model (LLM) user profiling, alongside the lack of efficient routing mechanisms between generative and collaborative filtering models in recommender systems. We propose a serving-time-only gating network that dynamically routes users to either collaborative sequential recommendation or LLM-based profiling models based on real-time context, enabling a controllable trade-off between novelty and relevance. Our analysis reveals that performance gains stem from precise selective routing rather than LLM generation alone, supported by an adjustable threshold control strategy. Experiments demonstrate that, under a 5% NDCG degradation constraint, the proposed method accurately routes only 12.5% of users to the LLM while improving Novelty@10 by 6.5%, significantly outperforming heuristic and random baselines.

0 citationsRead paper

Mechanisms of Misgeneralization in Physical Sequence Modeling

May 19, 2026

This work identifies and formalizes the phenomenon of “physical misgeneralization,” wherein generative sequential models in physical environments produce global distributions of conserved quantities—such as path length or mechanical energy—that deviate from design intent due to accumulated local modeling errors. The authors introduce controlled synthetic tasks to elucidate the underlying mechanism: local inaccuracies propagate through physical constraints, distorting the global distribution of these quantities. To address this, they propose a data bias kernel that predicts the direction of distributional shift, enabling a structured intervention strategy informed by this prediction. Experiments on maze navigation and double-pendulum dynamics demonstrate that the method effectively anticipates and mitigates distortions in physical quantity distributions, thereby validating both the proposed mechanistic explanation and the efficacy of the intervention approach.

0 citationsRead paper