Backed bya16z Speedrun
Tenera
ResearchBlogTeamDocsCareers
← All research
Literature Review·July 1, 2026·Tenera Research

Literature Review

Tenera ResearchWritten by Tenera Research

Tenera's ability to carry user behavior simulation from academic research into industry rests on the contributions of the researchers listed here. Their work on clickstream modeling, session-based interaction prediction, and behavioral simulation is the scientific foundation that makes what we are building possible. We are grateful to each of them.

User Behavior Simulation from Clickstreams

These papers directly address the problem of predicting or replicating user actions from observed behavioral sequences.

Can LLM Agents Simulate Multi-Turn Human Behavior? Evidence from Real Online Customer Behavior Data

First author: Yuxuan Lu · Co-authors: Jing Huang, Yan Han, Bingsheng Yao et al. · ACL 2026 · arXiv:2503.20749

The first rigorous evaluation of whether LLM-based agents can replicate human behavior across multi-turn interactions using real customer data. Demonstrates both where LLMs succeed and where behavioral fidelity breaks down, providing the benchmark framing that defines the problem Tenera solves.

OPeRA: A Dataset of Observation, Persona, Rationale, and Action for Evaluating LLMs on Human Online Shopping Behavior Simulation

First author: Ziyi Wang · Co-authors include Yuxuan Lu, Dakuo Wang et al. · ACL 2026 · arXiv:2506.05606

Introduces the first public dataset capturing user personas, browser observations, fine-grained web actions, and self-reported rationales from real online shopping sessions. Establishes the benchmark for evaluating how accurately LLMs can predict next user actions.

Customer-R1: Personalized Simulation of Human Behaviors via RL-based LLM Agent in Online Shopping

First author: Ziyi Wang · Co-authors: Yuxuan Lu, Yimeng Zhang, Jing Huang, Dakuo Wang · arXiv 2025 · arXiv:2510.07230

Uses reinforcement learning to train an LLM agent conditioned on explicit user personas for step-wise behavior simulation in online shopping. Shows that RL-based persona conditioning materially improves behavioral alignment over population-level baselines.

Shop-R1: Rewarding LLMs to Simulate Human Behavior in Online Shopping via Reinforcement Learning

First author: Yimeng Zhang · Co-authors include Tian Wang, Jiri Gesi, Ziyi Wang, Yuxuan Lu, Jiacheng Lin · ICLR 2026 · arXiv:2507.17842

Trains LLMs via reinforcement learning to simulate how users browse, compare, and select products in online shopping. Demonstrates that RL-tuned behavioral simulation outperforms prompted baselines.

SimPersona: Learning Discrete Buyer Personas from Raw Clickstreams for Grounded E-Commerce Agents

First author: Zahra Zanjani Foumani · Co-authors include Alberto Castelo, Shuang Xie et al. · Shopify · 2026 · arXiv:2605.14205

Extracts discrete behavioral personas from raw clickstream data and uses them to ground LLM-based e-commerce agents. Achieved 78% conversion-rate alignment with real buyers across 42 live storefronts.

Session-Based and Sequential Recommendation

The neural architectures here model how user intent evolves across a click session.

Self-Attentive Sequential Recommendation (SASRec)

First author: Wang-Cheng Kang · Co-author: Julian McAuley · ICDM 2018 · arXiv:1808.09781

Self-attention over item sequences models how user intent evolves across interactions. SASRec is the foundational architecture against which behavioral traces are calibrated, defining the standard for sequential click prediction.

Neural Attentive Session-based Recommendation (NARM)

First author: Jing Li · Co-authors: Pengjie Ren, Zhumin Chen, Zhaochun Ren, Tao Lian, Jun Ma · CIKM 2017 · arXiv:1711.04725

Attention-based encoder for capturing both general user preferences and short-term intent within a session. NARM's dual attention mechanism directly informs how recent clicks are weighted against longer-horizon behavioral history.

RepeatNet: A Repeat Aware Neural Recommendation Machine for Session-Based Recommendation

First author: Pengjie Ren · Co-authors: Zhumin Chen, Jing Li, Zhaochun Ren, Jun Ma, Maarten de Rijke · AAAI 2019 · arXiv:1812.02646

Models the repeat consumption pattern: users often return to items they have interacted with before. Repeat-aware modeling produces authentic behavioral traces that reflect real revisitation patterns in clickstreams.

Self-Supervised Hypergraph Convolutional Networks for Session-based Recommendation (DHCN)

First author: Xin Xia · Co-authors include Hongzhi Yin, Junliang Yu et al. · AAAI 2021 · arXiv:2012.06852

Captures higher-order item co-occurrence in sessions via hypergraph convolution. Extends session modeling beyond pairwise transitions to multi-item click patterns that more accurately reflect real browsing.

Click Models and Clickstream Analysis

These papers model the probabilistic structure of user click behavior, the foundational theory for clickstream-based behavior extraction and position-bias correction.

Click Models for Web Search

Authors: Aleksandr Chuklin, Ilya Markov, Maarten de Rijke · Springer Cham, 2015 · Publisher

The canonical reference on probabilistic models of user click behavior. The cascade model, dependent click model, and user browsing model are the theoretical underpinning of how observed click sequences are interpreted and position bias is corrected.

Cascading Bandits: Learning to Rank in the Cascade Model

First author: Branislav Kveton · Co-authors: Csaba Szepesvari, Zheng Wen, Azin Ashkan · ICML 2015 · arXiv:1502.02763

Formalizes the cascading click model as a bandit problem: learning to rank items by observing which one a user first clicks. Informs the understanding of position bias and click attribution in observed behavioral data.

Improving Web Search Ranking by Incorporating User Behavior Information

First author: Eugene Agichtein · Co-authors: Eric Brill, Susan Dumais · SIGIR 2006 · ACM Digital Library

Demonstrates that user behavior signals (clicks, dwell time, and browsing patterns) are strong features for improving ranking quality. The behavioral signal framework that underpins clickstream feature engineering.

Learning User Interaction Models for Predicting Web Search Result Preferences

First author: Eugene Agichtein · Co-authors: Eric Brill, Susan Dumais, Robert Ragno · SIGIR 2006 · ACM Digital Library

Builds explicit probabilistic models of user interaction from clickthrough logs, an early and influential formalization of user behavior prediction at scale.


Tenera is an applied research effort that stands on the shoulders of this body of work. The reason we are able to carry user behavior simulation into industry is precisely because of the contributions of all these researchers. We are grateful to each of them.

Tenera

© 2026 Tenera. All rights reserved.

PrivacyTermsCareers