PaperlayerView exact revision on arXiv

Source-linked AI summary

OmniSimulator: Aligning Small Language Models for Authentic Heterogeneous Behavior Modeling

Jiawei Chen, Ruoxi Xu, Boxi Cao, Zhuoqun Li, Ruotong Pan, Yunfei Zhang, Zhirui Yang, Jingyan Chen, Tingting Gao, Han Li, Yaojie Lu, Xianpei Han, Le Sun, Xiangyu Wu, Hongyu Lin

arXiv:2610.29472v1cs.CLcs.AIcs.LG

TL;DR

User simulation on short-video platforms remains difficult because native chain-of-thought often fails to model the latent factors underlying real decisions. OmniSimulator trains small LLMs with a five-dimension behavior-oriented reasoning schema and teacher-distilled traces. It consistently outperforms zero-shot and free-CoT baselines across behavior types, with Qwen3-8B’s average score improving by 34.33%.

  • Problem

    Native chain-of-thought often falls back on shallow pattern matching rather than reconstructing latent user states, while real short-video behavior is heterogeneous and cross-scenario.

    Source: §1 · p. 2

  • Method

    OmniSimulator trains small LLMs with a five-dimension behavior-oriented reasoning schema and teacher-generated traces filtered for ground-truth consistency.

    Source: §1 · p. 3

  • Results

    34.33% relative improvement: Qwen3-8B’s average score rises from 30.82 to 41.40, while OmniSimulator consistently outperforms zero-shot and free-CoT baselines across binary, continuous, and textual behaviors.

    Source: §3.2 · pp. 7–8, §5 · p. 10

  • Takeaways & Limitations

    The results suggest that user-simulation performance is constrained not only by model scale but also by the reasoning supervision paradigm.

    Source: §5 · p. 10

    Evaluation covers held-out users from one short-video platform within a temporally overlapping collection window, so temporal or cross-platform generalization is not established.

    Source: §5 · p. 10

Abstract

from arXiv · show

Large language models (LLMs) have shown great potential as general-purpose user simulators for interactive systems. Prior studies have demonstrated that training on high-quality data can substantially improve model reasoning and downstream task performance. However, the upper bound of user simulation is still constrained by the model's reasoning paradigm: native chain-of-thought often degenerates into shallow pattern matching and fails to adequately model the behavioral factors underlying real user decisions. In this paper, we propose OmniSimulator, a novel training framework designed to improve the predictive ability of small LLMs for realistic user simulation on short-video platform environments. Specifically, OmniSimulator organizes behavior-oriented reasoning through a structured schema with five dimensions, carefully selects data from real user-behavior scenarios, and introduces a teacher model to generate high-quality reasoning traces under the proposed schema, helping small models learn behavior-relevant reasoning patterns rather than relying on superficial behavioral correlations. Based on this process, we construct 9,881 high-quality training instances from three months of historical interaction trajectories of 200 real users, averaging about 50 distilled examples per user. Experimental results show that OmniSimulator improves the average score of Qwen3-8B by 34.33%, achieves performance competitive with Claude-Sonnet-4.5, and supports smaller-scale, self-hosted deployment. These results suggest that learning structured behavior-oriented reasoning patterns is a promising direction towards scalable and high-fidelity LLM-based user simulation.

1 Introduction

Short-video user simulation must account for heterogeneous, long-horizon behavior, but native chain-of-thought often fails to reconstruct the latent states behind users’ decisions. OmniSimulator trains small models with behavior-oriented structured reasoning and teacher-generated traces, with reported gains over existing baselines.

1 Introduction

  • Real short-video behavior spans long-horizon, cross-scenario trajectories, while current simulators show limited benchmark performance, overestimate positive actions, and collapse distinct users into average personas.The paper also notes that frontier-model costs become prohibitive when conditioning on long interaction histories.Source: §1 · pp. 1–2
  • Native chain-of-thought often yields shallow pattern matching or generic justifications instead of reconstructing latent user states, and thinking mode brings little or sometimes negative benefit.The proposed bottleneck is whether reasoning follows a behavior-oriented structure, not simply whether the model generates more reasoning tokens.Source: §1 · p. 2
  • OmniSimulator teaches small LLMs behavior-oriented reasoning through a five-dimension schema and teacher-generated traces filtered for ground-truth consistency.The schema covers user profile, historical behavior, content relevance, decision mode, and final decision.Source: Figure 1 · p. 2, §1 · p. 3
  • 9,881 distilled training instances were constructed from three months of interaction histories of 200 real users; the framework improves small models and makes Qwen3-8B competitive with Claude-Sonnet-4.5.The authors also report alleviating overestimation bias and supporting smaller-scale deployment.Source: §1 · p. 3

2 Methodology

OmniSimulator treats user behavior as a latent-state process and trains small models to reason through a fixed behavior-oriented schema. Its framework combines structured rationale supervision with task-aware teacher distillation.

2.1 Problem Formulation

  • User simulation predicts responses from long-horizon, cross-scenario histories and current context, with an unobserved latent state represented by a structured reasoning trace.The latent state is not supervised as an explicit label; the model generates the trace before its behavior prediction.Source: §2.1 · pp. 3–4

2.2 Preliminary Analysis: Revealing Native Thinking Effects

  • Native thinking yields little or no consistent OmniBehavior gain over instruction-following inference, and some variants perform worse.The analyzed traces often use generic justifications or shallow guesses instead of reconstructing latent user states.Source: §2.2 · p. 4

2.3 Overview of OmniSimulator

  • A teacher generates multiple schema-guided traces with task-specific prompting, and rejection sampling retains traces whose predictions match observed outcomes.For accepted candidates, the pipeline selects the trace with the highest teacher mean token log-probability; examples without a match are excluded.Source: §2.3 · pp. 4–6
  • The five-dimension schema guides reasoning through user profile, historical state, content alignment, decision mode, and final decision declaration.The final declaration must agree with the output; decision mode distinguishes rational from random behavior.Source: §2.3 · pp. 4–5
  • Students train to generate both the structured trace and final prediction, learning rationale supervision rather than only answer-level behavior patterns.The model uses the user profile, interaction history, and current context without manually constructed external annotation or additional structured memory.Source: §2.3 · p. 6

3 Experiments

Experiments evaluate OmniSimulator on heterogeneous, long-horizon user behaviors and compare it with zero-shot, free-CoT, and training-control baselines. The results favor the full method, with gains spanning behavior types, reduced overestimation bias, and competitive Qwen3-8B performance.

3.1 Experimental Setup

  • The benchmark covers binary, continuous, and textual behaviors across long-horizon, cross-scenario interaction traces from 200 short-video-platform users.Training data uses another 200 users with no user, trajectory, or target-behavior overlap, though its collection window overlaps temporally with the benchmark.Source: §3.1 · pp. 6–7, Table 1 · p. 7

3.2 Main Results

  • 34.33%: OmniSimulator raises Qwen3-8B’s average score from 30.82 to 41.40, competitive with Claude-Sonnet-4.5 at 40.49 and nearly matching GLM-4.7 at 41.46.It substantially outperforms the zero-shot Qwen3-235B baseline (32.11); baseline scores are from OmniBehavior [9].Source: §3.2 · pp. 7–8

3.3 Overestimation Bias Analysis

  • OmniSimulator alleviates overestimation bias, with its largest gains on binary tasks where generic positive predictions are especially harmful.It better identifies content misaligned with recent user state, weak historical evidence, and likely negative decisions; bias reduction is tightly coupled with binary accuracy.Source: §3.3 · p. 8, Figure 3 · p. 8

3.4 Ablation Study

  • The full schema performs best across behavior types; removing Historical Behavioral State Inference most harms binary behavior, while removing User Profile Grounding most harms textual simulation.The ablation is on Qwen3-1.7B and supports complementary roles for dynamic state reconstruction and persona grounding.Source: §3.4 · pp. 8–9, Table 2 · p. 9

3.5 Training Strategy Analysis

  • Answer-only SFT yields marginal gains, synthetic rationales remain below OmniSimulator, and removing rejection sampling degrades performance.The findings support combining behavior-oriented reasoning structure with quality-controlled distillation rather than merely increasing supervision.Source: §3.5 · p. 9, Table 3 · p. 9

4 Related Work

Prior work uses LLMs to simulate heterogeneous human behavior and explores chain-of-thought and structured reasoning as supervision. OmniSimulator focuses these ideas on reconstructing latent states from realistic, long-horizon user behavior.

4.1 LLMs for Human Behavior Simulation

  • LLM-based simulators offer a unified interface for heterogeneous actions and natural-language responses, while prior work has paid less attention to realistic long-horizon logs shaped by evolving latent states.OmniSimulator targets small models and explicit latent-state reconstruction.Source: §4.1 · p. 10

4.2 Chain-of-Thought and Structured Reasoning

  • Unlike free-form CoT for generic reasoning, OmniSimulator tailors a structured schema to user simulation’s challenge of recovering the latent mental state behind the next behavior.Prior work indicates that how intermediate reasoning steps are organized can matter [8].Source: §4.2 · p. 10

5 Discussion

The paper concludes that structured latent-state reasoning and task-aware rationale distillation improve small-model user simulation. It also identifies limits on generalization and on what the distilled traces establish.

5 Discussion

  • Across binary, continuous, and textual behaviors, OmniSimulator outperforms zero-shot and free-CoT baselines, reduces overestimation bias, and makes an 8B model competitive with Claude-Sonnet-4.5.The authors identify reasoning supervision for behavior-relevant user states, not model scale alone, as a key challenge.Source: §5 · p. 10
  • Evaluation covers held-out users from one short-video platform in a temporally overlapping collection window, so temporal and cross-platform generalization are not established.The traces are behavior-oriented supervision, not direct evidence of users’ cognitive processes; teacher bias and post-hoc rationalization risks remain.Source: §5 · p. 10

LLM Usage Statement

The authors used LLMs only to polish language after drafting, with human review and approval of the final paragraphs.

LLM Usage Statement

  • LLMs were used only for language polishing; the authors conducted the research and reviewed, edited, and approved the final text.The authors drafted initial paragraphs before LLM-assisted editing.Source: LLM Usage Statement · p. 13

Ethics Statement

The study reports that its behavior data was authorized for research and benchmark development, processed with privacy protections, and used only after sensitive information and unsuitable content were removed or anonymized.

Ethics Statement

  • Only behavior data authorized for research and benchmark development was included, after quality checks, sampling, and privacy-preserving preprocessing.Personally identifiable and sensitive information was removed or anonymized, and unsuitable or harmful content was filtered before experiments.Source: Ethics Statement · p. 13

A Experimental Parameters and Details

The supplementary experiments report preliminary results for native thinking, detail the training configuration, and describe the filtering and selection process used to produce the final distilled training set.

Preliminary Analysis

  • Native thinking showed no consistent advantage: Qwen3-8B changed from 30.82 to 30.77, while Qwen3-235B rose from 33.15 to 33.79.The analysis also reports gains for Qwen3-32B (31.87 → 32.58) and Qwen3-4B (28.29 → 29.09), and a decline for Qwen3-14B (31.72 → 31.59).Source: A Experimental Parameters and Details · p. 13
  • The benchmark comparison table covers binary, continuous, and textual behaviors, with its average defined as the arithmetic mean of six task scores.Higher average scores indicate better performance.Source: Table 4 · p. 13

Training Configuration

  • Models were fully fine-tuned for two epochs in bf16, using a 38,000-token maximum sequence length and a learning rate of 1 × 10−5.Teacher-driven distillation used GPT-OSS-120B with 16 rollouts per instance and decoding temperature 1.0.Source: A Experimental Parameters and Details · p. 13

Distillation Data Statistics

  • The final training set contains 9,881 instances, selected after answer-consistency filtering and teacher-log-probability ranking of candidate traces.The pipeline began with 13,389 source instances and 214,224 candidate traces; 31,642 traces were accepted, covering 9,881 sources, for a 73.8% item-level retention rate.Source: A Experimental Parameters and Details · p. 14

B.1 Textual Judge Validation

Independent LLM judges and blinded human assessment provide validation for the textual evaluation, with the human scores for OmniSimulator and Claude-Sonnet-4.5 reported as comparable.

B.1 Textual Judge Validation

  • Across 800 textual targets, GPT-OSS-120B and Qwen3.5-35B had example-level Spearman correlations of 0.93 and 0.85 with Claude-Sonnet-4.5.Their system-level correlations across five systems were 1.00 and 0.90, respectively.Source: §B.1 · p. 14
  • OmniSimulator scored 50.83 in blinded human evaluation, comparable to Claude-Sonnet-4.5 at 51.41, with a paired 95% confidence interval of [-1.42, 2.55] for the score difference.The assessment covered 50 shared textual targets across four systems and used three annotators; agreement was ICC(2,k) = 0.74 and Krippendorff’s α = 0.68.Source: §B.1 · p. 14
  • The validation table presents independent LLM-judge results for the full textual test set and blinded human-assessment results on 50 shared targets.The caption identifies these as separate panels.Source: Table 5 · p. 14

B.2 Reasoning Trace Validation

The appendix validates sampled reasoning traces, identifies the prompts used for training and evaluation, and lists four case studies with their corresponding prediction-comparison tables.

B.2 Reasoning Trace Validation

  • 92% of 200 sampled reasoning traces were judged valid, with substantial inter-annotator agreement (Fleiss’ κ = 0.72).Three annotators assessed logical consistency, completeness, and relevance to behavioral evidence across binary, continuous, and textual behaviors.Source: §B.2 · p. 14, Table 6 · p. 15

C Experimental Prompts

  • The appendix presents prompts for generating training CoT data and evaluating binary, continuous-value, and next-utterance prediction.Figure 4 covers response-generation prompts across five dimensions; Figure 5 covers heterogeneous-task prediction.Source: C Experimental Prompts · p. 15, Figure 4 · p. 15, Figure 5 · p. 16

D.1 Case 1

  • The appendix includes a video-completion prediction case study with its results in Table 7.The example is described as anonymized and processed according to data privacy and usage policies.Source: D Case Study · p. 17, Table 7 · p. 17

D.2 Case 2

  • The appendix includes an author-follow prediction case study with its results in Table 8.The example is described as anonymized and processed according to data privacy and usage policies.Source: D Case Study · p. 17, Table 8 · p. 18

D.3 Case 3

  • The appendix includes a continuous watch-time prediction case study with its results in Table 9.The example is described as anonymized and processed according to data privacy and usage policies.Source: D Case Study · p. 17, Table 9 · p. 19

D.4 Case 4

  • The appendix includes a next customer-service utterance prediction case study in an e-commerce scenario, with results in Table 10.The example is described as anonymized and processed according to data privacy and usage policies.Source: D Case Study · p. 17, Table 10 · p. 20

Related papers

Loading 2610.29472v1…