LongEmotion finds small models hold long support chats better than GPT-4o

Most tests of emotional intelligence in LLMs work on short, neatly labeled utterances. Real conversations are messier: people talk at length, drift, change the subject, circle back to old hurts. Over that distance models start losing the cues that matter, confuse cause and effect, and rarely sustain a coherent line of support. The authors of LongEmotion propose exactly that stress test — a benchmark of long, noisy, emotionally dense scenarios.

What LongEmotion is

The benchmark covers six tasks, each probing a different facet of emotional intelligence:

• recognizing emotion in a long text where the useful passage is buried in context;
• spotting the emotional outlier among similar passages;
• answering questions from psychology sources;
• a long four-stage counseling dialogue scored against psychological criteria;
• summarizing the emotionally salient parts of clinical reports;
• expressing emotion: self-assessment plus a five-phase long-form narrative.

Average input length reaches 8,000–12,000 tokens, and in places well beyond that. These are not toy exchanges but full texts, where the model has to hold long-range connections, keep the nuance, and finish a thought clearly.

Two engineering ideas: RAG without an external store, and CoEM

To keep models from drowning in a long dialogue, the authors propose two approaches.

First, retrieval-augmented generation in an unusual form: what gets retrieved is the conversation itself plus the model's own memory, with no external knowledge base. The idea is simple — don't pull in everything at once; isolate the locally relevant chunks, work with those, and add no extra noise.

Second, a multi-agent system called CoEM. It is a five-stage pipeline: split the text, rank the pieces, enrich them with emotional and theoretical signals (supplied by a separate assistant), re-rank, and generate the final answer with an ensemble. This "emotion orchestrator" helps the model hold its through-line and its tone, but does so in measured doses, without leaking finished answers.

What the experiments showed

• On emotion recognition and detection, CoEM delivers the clearest gain: structured enrichment helps pull faint emotional signals out of the surrounding clutter.
• On tasks where the answer is tied tightly to the source text (question answering, summarization), the extra "hints" can hurt: the added background drags down F1 and lowers agreement with the source.
• In the long counseling dialogue, Qwen-3 8B and Llama-3.1-8B-Instruct are unexpectedly strong: they hold the thread of the conversation consistently, in places beating GPT-4o on average stage scores. The biggest jump comes at the closing stage, where integrating meaning and ending the session carefully matter most.
• For emotional text generation, much rides on the quality of CoEM's "sage." When a strong model plays the enriching assistant, the resulting prose is more coherent, less repetitive and richer in emotional detail.

A separate thread is the comparison across GPT versions. On the aggregate results the extended GPT-5 reasons better on pure classification, but in question answering it sometimes rewrites the answer to fit its own reading and loses points against the reference. In dialogue GPT-5 is stronger on the theoretical criteria of therapy, yet sounds slightly less empathetic. GPT-4o-mini, by contrast, at times strikes a more human tone. Which underlines the point: emotional intelligence is not only logic but style, care for the person, and knowing how to close a topic without sharp edges.

What practitioners should watch

• If you run long support chats, use "internal" RAG over the dialogue history. It cuts the noise and helps you keep hold of the main threads.
• For classification and detecting anomalous emotion, CoEM is the better fit: multi-agent enrichment makes the signals more visible.
• For QA and summarization, watch the dose of enriching hints: an assistant that talks too much rebuilds the answer to suit itself.
• Models behave differently on long inputs: the steadiest is not always the "biggest" version. Test at the lengths you actually work with.

Limitations and next steps

The authors are candid that automatic scoring — an LLM as judge — correlates well with experts but stays sensitive to which model does the judging. Next the team plans to widen the set of models, release the data and refine the methodology. In practice this is a meaningful step toward testing an AI's "emotional memory" systematically over long distances.

💾 Code