Milestone 1: A Coherent Emotional State
Hypothesis
A persistent emotional-state module will help Cershe carry relevant emotional context forward, reduce unexplained emotional shifts, and respond appropriately when new events change the situation.
What is being built
An emotional state that persists across turns and sessions. It updates before Cershe responds, using her previous state, new information, relevant memories, and elapsed time. Reactions may settle, strengthen, or change as circumstances develop.
How it will be tested
Cershe will be evaluated with and without the additional module, keeping the language model, base personality, and available event history consistent.
- Continuity: does an upsetting exchange remain relevant a few turns later when the situation is still unresolved?
- Responsiveness: when reassuring information arrives, does her state change appropriately?
- Time: after a pause, does her state evolve according to the defined update rules?
- Same prompt, different state: with the question and context held constant, does changing the recorded state meaningfully affect her response?
- State and words agree: do her descriptions and replies accurately reflect the state recorded by the module?
Independent reviewers will evaluate unseen scenarios using a defined rubric. Coherence does not require one exact emotional reaction or simply mirroring the user.
What success looks like
Emotional transitions that are understandable in context, with better continuity and adaptation than the comparison system. Her state reports should match the logs. Text will be evaluated first, followed by separate evaluations of voice and facial expression.
Cershe Architecture
The three central pieces have different jobs:
| Piece | What it represents |
|---|---|
| GRU | The trained model that computes how her state changes. |
| Emotional state | The changing internal values carried from one interaction to the next. These are stored separately from the model’s weights. |
| Base language model | Generates her response using the conversation, personality, and interpreted emotional state. |
Cershe's trainable parts would be:
- A state-update network, potentially using a GRU, that learns how experiences change the internal state.
- A readout network that learns how to interpret that state for evaluation and communication with the base language model.
If you would like to take part as a tester, please write to anyta.ali@cershe.ai.