Rushes: A Human Preference Dataset for Pluralistic Alignment

2026-07-22 · Source: Computation and Language · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics, Software Development & Engineering · Depth: Expert, quick

Summary

Rushes is a newly introduced human preference dataset and benchmark designed to study revealed human engagement preferences within interactive AI-generated narrative environments. Collected via a game interface, Rushes logs 44,226 decision events from 8,167 unique users across six games, capturing sequential, personalized choices and narrative context. The dataset reveals structured, non-random user choice patterns, quantified by low choice entropy. Positioned as a diagnostic benchmark for pluralistic alignment, Rushes exposes a significant "Engagement Gap": leading LLMs, including GPT-5, achieve only 34.23% accuracy in event-level choice prediction, failing to match a simple Popularity Baseline (36.4%) or classical Matrix Factorization (SVD) at 37.7%. This suggests that current population-level objectives, such as those in modern RLHF, are inadequate for capturing diverse, context-dependent engagement signals, causing models to default to majority preferences rather than individual trajectories. The dataset and platform code are publicly available.

Key takeaway

For AI Scientists and Machine Learning Engineers developing LLMs for personalized interactive narratives, this research highlights a critical "Engagement Gap." Your current RLHF strategies, often based on population-level objectives, are likely insufficient to capture diverse, context-dependent user engagement. You should prioritize exploring methods for pluralistic alignment and individual trajectory adaptation, moving beyond majority preferences. Utilize the Rushes dataset as a diagnostic benchmark to rigorously evaluate your models' ability to adapt to unique user preferences.

Key insights

Leading LLMs struggle with personalized engagement in interactive narratives, revealing limitations of current population-level alignment methods.

Principles

• Human engagement preferences are structured and non-random.
• Population-level objectives are insufficient for heterogeneous engagement signals.
• Models default to majority preferences without individual adaptation.

Method

Collect human preference data from interactive branching narratives by logging user choices, candidate sets, and narrative context with persistent user IDs.

In practice

• Benchmark LLMs against Rushes for pluralistic alignment.
• Evaluate models on sequential, personalized engagement behavior.

Topics

• Human Preference Data
• Pluralistic Alignment
• Large Language Models
• RLHF
• Interactive Narratives
• Engagement Gap
• Sequential Decision-Making

Code references

• microsoft/rushes

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• Rushes: A Human Preference Dataset for Pluralistic Alignment — Current LLMs struggle to predict personalized engagement in interactive narratives, underperforming simpler baselines.
• What Do People Actually Want From AI? Mapping Preference Plurality — Human preferences for AI are pluralistic and contextual, challenging current singular alignment methods.
• Large Language Models Should Learn Personalized Rather Than Aggregated Human Preferences — LLMs should learn personalized human preferences, not aggregated ones, to avoid optimizing for a hypothetical "average user" and better respect individual autonomy.
• PolyAlign: Conditional Human-Distribution Alignment — PolyAlign aligns language models to context-specific human response distributions, overcoming limitations of global alignment for improved naturalness and faithfulness.
• Response Time Enhances Alignment with Heterogeneous Preferences — Response time data can restore identifiability of population-average preferences in LLM alignment, even with heterogeneous, anonymous labelers.
• Align AI to Dynamic Human-AI Workflows — AI alignment must evolve from static preference emulation to dynamic, interactive human-AI co-evolution.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computation and Language.