Research map / Large language models / Language, safety and interpretability
Large Content And Behavior Models To Understand, Simulate, And Optimize Content And Behavior
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces Large Content and Behavior Models (LCBMs) by adding receiver behavior tokens (likes, views, replay values, CTR) to a Vicuna-13B vision-language model via behavior instruction fine-tuning, and releases the Content Behavior Corpus (CBC).
Problem
LLMs are trained on content without receiver behavior, so they cannot simulate, optimize, or explain how content affects receivers (Shannon's effectiveness level).
Approach
Verbalize receiver behavior (replay values, likes/views ratio, views, likes) as text tokens alongside visual and textual content, and instruction fine-tune a VLM end-to-end to predict behavior from content and content from behavior in one text-to-text space.
Key results
- On content simulation (25-option speech selection), LCBM-13B reaches 48.68% accuracy vs 34.98% for GPT-3.5 and 19.30% for Vicuna-13B.
- On behavior understanding, LCBM gets 72.73% sentiment accuracy and 4.00 human reasoning score vs 2.23 for Vicuna and 1.67 for GPT-3.5.
- On LVU like-ratio zero-shot prediction, LCBM achieves MSE 0.14 on the test set, beating the trained baselines (0.23 to 0.444), though GPT-3.5 scores 0.03.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train Vicuna-13B with behavior instruction fine-tuning | Vicuna-13B · YouTube content simulation | Accuracy: 19.30% -> 48.68% | Helped |
| Train Vicuna-13B with behavior instruction fine-tuning | Vicuna-13B · YouTube behavior understanding | Sentiment accuracy: 65.66% -> 72.73% | Helped |
| Train Vicuna-13B with behavior instruction fine-tuning | Vicuna-13B · YouTube behavior understanding | Reasoning score: 2.23 -> 4.00 | Helped |
| Use behavior instruction fine-tuning | Vicuna-13B · YouTube topic classification | Accuracy: 11.75 -> 42.17 | Helped |
| Use behavior instruction fine-tuning | Vicuna-13B · YouTube persuasion classification | Accuracy: 26.59 -> 32.83 | Helped |
| Train on YouTube behaviors and a small set of email examples rather than email-only data | LCBM · Email marketing, different segment | RMSE: 25.28 -> 14.47 | Helped |
| Train on YouTube behaviors and a small set of email examples rather than email-only data | LCBM · Email marketing, different segments and emails | RMSE: 29.28 -> 27.28 | Helped |
| Add YouTube behavior data to Twitter BFT | LCBM trained on Twitter · Twitter content simulation, brand separated | BLEU-1: 62.29 -> 64.28 | Helped |
| Use GPT-3.5 instead of LCBM | LCBM · LVU test set | MSE: 0.14 -> 0.03 | Helped |
| Use 5-BFT rather than 3-BFT (ablation) | LCBM · YouTube replay values, Past | RMSE: 8.12 -> 11.53 | Hurt |
| Use 5-BFT rather than 3-BFT (ablation) | LCBM · YouTube replay values, Random window 5 | Accuracy: 61.41 -> 64.83 | Helped |
| Use 7-BFT rather than 5-BFT (ablation) | LCBM · YouTube replay values, Random window 5 | Accuracy: 64.83 -> 59.22 | Hurt |
Surprising findings
- GPT-3.5 outperforms LCBM on zero-shot LVU like-ratio prediction (MSE 0.03 vs 0.14).
- GPT-3.5 and GPT-4 perform comparably to or worse than random baselines on several behavior simulation tasks.
Limitations
- Content understanding performance lags GPT-3.5, the largest model.
- GPT-3.5 beats LCBM on the zero-shot LVU like-ratio task.
- Behavior understanding is evaluated by a small number of annotators on subjective 0-5 ratings.
Experimental setup
Datasets: WebVid COCO caption Visual Genome CC3M CC12M Content Behavior Corpus (CBC) LVU benchmark In-house Email Marketing data Twitter corporate account data advertisements dataset (Hussain et al.) persuasion strategy videos
Benchmarks: LVU benchmark Content Behavior Corpus test split advertisements dataset topic/emotion/action-reason persuasion strategy dataset In-house Email Marketing
Metrics: RMSE R^2 accuracy BLEU ROUGE-L MSE human rating 0-5
Baselines: GPT-3.5 GPT-4 Vicuna-13B VideoChat Random
Builds on: BLIP-2 Vicuna EVA-CLIP Uniformer VideoChat visual instruction tuning
Closest papers by meaning
- Teaching Human Behavior Improves Content Understanding Abilities Of VLMs (ICLR 2025 · similarity 0.87)
- CogBench: a large language model walks into a psychology lab (ICML 2024 · similarity 0.85)
- When LLMs Play the Telephone Game: Cultural Attractors as Conceptual Tools to Evaluate LLMs in Multi-turn Settings (ICLR 2025 · similarity 0.85)
- Turning large language models into cognitive models (ICLR 2024 · similarity 0.85)
- Are Large Language Models Sensitive to the Motives Behind Communication? (NeurIPS 2025 · similarity 0.85)
- WildBench: Benchmarking LLMs with Challenging Tasks from Real Users in the Wild (ICLR 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.