Can AI Achieve Higher-Quality Empathy? Two Open‑Source Studies Offer New Paths

EMPA: Process‑level Empathy Evaluation

EMPA treats long‑horizon empathy as a task in which a user's latent psychological state evolves over a multi‑turn dialogue. The evaluation pipeline consists of three stages:

Real‑to‑Sim data pipeline – noisy real‑world conversations are distilled into reproducible psychological scenarios.

Non‑scripted multi‑agent sandbox – a user agent, a director agent, a judge agent, and the model under test interact openly.

Empathy Potential Model (EPM) – maps evidence extracted from each turn onto changes in the latent state, enabling trajectory‑level measurement of whether the dialogue produces a sustained positive shift.

Rubric‑Grounded Physics Evaluation

Instead of a single final score, the judge extracts structured evidence for each rubric item at every turn. EPM aggregates this evidence across the dialogue and translates it into a potential‑state signal. This separates evidence generation from scoring, allowing continuous updating of the user’s state rather than a one‑off impression score.

Empirical results in the EMPA paper show that this evaluation is more robust and sensitive than traditional rubric checklists or LLM‑as‑a‑Judge approaches.

MAPO: Mixed Advantage Policy Optimization for Long‑Horizon Dialogue

MAPO provides a reinforcement‑learning algorithm that leverages both immediate and long‑term feedback while remaining critic‑free.

Per‑turn process reward – the EMPA judge supplies an incremental score change between consecutive turns; this delta is used as the immediate reward for the current turn.

Long‑term future return – Monte‑Carlo roll‑outs estimate the cumulative return from the current turn to the end of the dialogue, preserving long‑range strategy information.

Immediate rewards are normalized batch‑wise because their distribution is largely independent of turn index, whereas future returns are normalized turn‑wise due to a strong correlation with dialogue length. The two normalized signals are combined via a convex mixture, stabilizing optimization of long sequences.

Training proceeds by sampling multiple dialogue trajectories from the same initial prompt; each step in each trajectory becomes a training sample.

Experimental findings report that MAPO outperforms GRPO on the EMPA sandbox, achieves performance comparable to Claude‑3.5 with a 32B model, and generalizes to other multi‑turn dialogue benchmarks.

Open‑source resources:

EMPA benchmark code: https://github.com/KAYA-HAI/EMPA-Benchmark-EPMSandbox

EMPA 1000+ datasets: https://huggingface.co/datasets/SalmonTell/EMPA-character_card/tree/main

MAPO code: https://github.com/2200xiaohu/MAPO

EMPA paper (arXiv): https://arxiv.org/abs/2603.00552

MAPO paper (PDF): https://arxiv.org/pdf/2603.06194v1