A 500-million-parameter model scored 41.32 on ROUGE-1, compared with GPT-4.1's 40.04, in a counseling-dialogue benchmark. That's a useful result for compact models. It measures overlap with reference replies, rather than how much a conversation helps a person.

The numbers come from Wei, Zhou and Wang's MoPHES preprint, version 1, posted in October 2025. Looking at the other measures changes the story: GPT-4.1 led the automated response-quality ratings, and PsyChat also scored above MoPHES.

How the researchers built MoPHES

MoPHES stands for Mobile Psychological Health Evaluation and Support. It uses two fine-tuned MiniCPM4-0.5B models: one estimates mental-state labels, and the other generates conversation. The authors also demonstrate deployment on mobile devices.

For training data, they combined Chinese-language material from PsyQA and EmoLLM. Filtering and deduplication reduced 113,552 question-answer pairs to 34,827 single-turn pairs. These were then expanded into five-turn dialogues using GPT-4o-mini and filtered again, leaving 34,381 conversations. The resulting corpus contains generated conversations derived from source questions and answers, rather than recordings of real therapy sessions.

The dialogue benchmark contained 100 samples across five counseling topic categories. The researchers tested two ways of supplying conversation history. Under the label strategy, earlier replies came from the reference dialogue. Under the output strategy, the model received its own previous replies. The second setup more closely resembles an ongoing conversation, where an imperfect earlier answer can affect the next one.

The ranking depends on the measure

Table VII reports ROUGE-1 of 41.32 for MoPHES and 40.04 for GPT-4.1 under the label strategy. ROUGE-1 counts shared individual words with a reference answer. A higher score means greater lexical overlap; it doesn't establish better counseling or a better outcome for a user.

Table VIII asks a different question. GPT-4.1 automatically rated responses against criteria for understanding, empathy, professional expertise, helpfulness and safety. Although the paper calls these "manual" dimensions, the evaluator was a language model, rather than a panel of human experts.

Model	Aggregate response rating, label strategy, out of 10
GPT-4.1	8.685
PsyChat	7.425
MoPHES	7.204

MoPHES ranked second among the tested noncommercial models on this aggregate, behind PsyChat. It did well for its size, but the benchmark doesn't support the claim that it beat every larger model on overall response quality.

DeepSeek-R1-7B had particularly low BLEU and ROUGE scores. The authors suggest that excessive reasoning produced overly long replies. That's their interpretation of this test; it doesn't establish that reasoning models generally fail at psychological support.

Why the compact design is useful

The practical result is a demonstration that a specialized 0.5B model can run locally and compete on selected dialogue measures. Training used a single A100 GPU. Those details make the system relevant to developers exploring smaller models and on-device inference.

Local inference can keep message processing on a device. An application's actual data handling still depends on its storage, telemetry and other services, so a model's size alone doesn't settle privacy.

The next research question is whether these benchmark gains carry over to people using the system over time. The paper reports Chinese-language data and automated evaluations, with no clinical outcome trial of MoPHES. For a different study design, see our Therabot trial analysis.

Nearby's current purpose and features are described in the product facts. It's an emotional-support and self-help service; MoPHES results don't establish Nearby's architecture or effectiveness.

Frequently asked questions

Did MoPHES outperform GPT-4.1?

On ROUGE-1 under the label-history strategy, yes: 41.32 versus 40.04. On the automated aggregate response rating in the same strategy, GPT-4.1 scored higher: 8.685 versus 7.204.

Were the replies rated by human experts?

The reported ratings in Table VIII were generated by GPT-4.1 using the authors' rubric. They aren't human expert ratings or patient outcomes.

Was MoPHES trained on real therapy dialogues?

The source material was Chinese question-answer data. After filtering, 34,827 single-turn pairs were expanded using GPT-4o-mini; a further filter left 34,381 five-turn conversations. Calling those conversations real therapy sessions would misdescribe the corpus.

Is this computational psychiatry?

The fields can overlap, but a language model trained to generate counseling replies isn't automatically a model of a mental disorder. Our computational psychiatry overview explains the distinction; our ChatGPT article focuses on conversational support.