Life Labo

Overview

SHEN is a dialogue AI designed around a single requirement: light and quick in ordinary moments, deeply supportive when it matters. Rather than a monolithic persona, SHEN composes emotional skill modules and safety paraphrasing modules on top of a base LLM, dynamically blending them by context.

Purpose and UX

In normal situations, it prioritizes tempo and comfort (light banter, natural backchannels). When the user enters a strained state, it shifts to depth, empathy, and caution, delivering a two-stage UX. The switch is controlled by state estimation and component weighting, not prompt tricks, to ensure reproducibility.

Base + Modular Architecture

We fix one base LLM family for the everyday range, then add emotion skills such as empathy, encouragement, boundary setting, and conflict de-escalation as LoRA/Adapter diff modules and blend them according to the inferred state. Merging is restricted to compatible model families and uses stability-focused merge methods to minimize collisions.

Inner Council

Instead of single-shot responses, multiple specialist roles generate short proposals in parallel, and a synthesis role composes one natural response. Roles include empathy, practical next step, boundary control, summarization, and safety paraphrasing. Weights are adjusted automatically by state estimation so the council scales down in normal times and scales up in difficult moments while keeping a consistent persona.

Emotion Estimation and Safe Paraphrasing

We infer user state (anxiety, anger, low mood, calm, and so on) and increase contributions from relevant skill modules. A dedicated safety module intervenes when output risks being sharp, rephrasing rather than refusing. It preserves intent while softening wording and syntax, guaranteeing a minimum level of kindness and emotional consistency.