Research map / Large language models / Alignment and preferences

PRIORITY2REWARD: Incorporating Healthworker Preferences for Resource Allocation Planning

AAAI 2025 · Large language models › Alignment and preferences · approach: Reward model design and preference models · 1 citations (OpenAlex)

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.

Presents PRIORITY2REWARD, a demo tool where health workers enter natural-language preferences and an LLM (Gemini Pro) proposes reward functions for restless multi-armed bandit resource allocation in an Indian maternal mHealth program.

Problem

Health workers need to steer automated resource allocation (service calls) in public health programs toward community priorities, but reward design for restless multi-armed bandits is hard to do by hand and the existing Decision Language Model is too slow for real-time use.

Approach

Adapts the Decision Language Model pipeline (LLM generates reward functions, simulates RMAB policy, LLM selects best) for real-time interactive use by swapping the PreFeRMAB policy solver for Whittle index heuristics and caching Whittle indices across reward functions that yield identical per-arm rewards, with a dashboard showing resource and reward distributions across demographic groups.

Key results

• Caching Whittle indices yields roughly 5x speedup for reward function proposal in the tool.
• The tool generates 4 reward functions per preference prompt and runs 10 RMAB simulations with different seeds.
• Dashboard shows how LLM-generated rewards shift resource allocation and engagement across socio-demographic groups versus a default engagement-maximizing reward.

Measured effects reported in the paper

Limitations

• PreFeRMAB is too slow to scale beyond a few hundred arms, motivating the Whittle index substitution.
• Demonstration is a simulated application with anonymized data; no real-world deployment evaluation is reported.

Experimental setup

Datasets: anonymized ARMMAN mHealth listenership data (January 2022 quality improvement study)

Benchmarks: simulated RMAB rollouts on 2100 beneficiaries, budget 210, 12-week horizon

Metrics: engagement resource distribution across demographic groups reward distribution speedup of reward proposal

Baselines: PreFeRMAB default engagement-maximizing reward

Builds on: Decision Language Model PreFeRMAB Whittle index policy

Closest papers by meaning

• A Decision-Language Model (DLM) for Dynamic Restless Multi-Armed Bandit Tasks in Public Health (NeurIPS 2024 · similarity 0.93)
• Context in Public Health for Underserved Communities: A Bayesian Approach to Online Restless Bandits (AAAI 2025 · similarity 0.86)
• Value-Incentivized Preference Optimization: A Unified Approach to Online and Offline RLHF (ICLR 2025 · similarity 0.85)
• Tool-Augmented Reward Modeling (ICLR 2024 · similarity 0.85)
• Explicit Preference Optimization: No Need for an Implicit Reward Model (ICML 2025 · similarity 0.84)
• Moral Alignment for LLM Agents (ICLR 2025 · similarity 0.84)