ParaText: Exploring How LLMs Infer Users' Implicit States from Surface Text Clues

ParaText is a research project that aims to test whether large language models (LLMs) can infer users' implicit states (such as fatigue, urgency, frustration, limited network bandwidth, etc.) from surface text features, and whether these inferences will change the model's response strategies. This project focuses on the social perception capabilities of LLMs, providing a new perspective for understanding AI-human interactions.

In daily human communication, people can capture implicit information from the surface of utterances, which is an important part of social intelligence. With the improvement of LLM capabilities, the question of whether AI can infer users' implicit states from subtle text features like humans has emerged, leading to the birth of the ParaText project.

ParaText is an open-source research project that focuses on testing LLMs' ability to infer users' potential states (fatigue, urgency, frustration, cognitive load, network bandwidth limitations, etc.) from surface text clues and the impact of these inferences on response strategies. Core hypothesis: User states leave detectable traces in text (sentence structure, punctuation, vocabulary, response delay, etc.), and LLMs that can recognize these patterns possess social perception capabilities.

ParaText designs multiple research dimensions to evaluate LLMs' state inference capabilities:

1. Fatigue state detection: Focuses on insufficient energy, with features including short sentences, simple vocabulary, abbreviations, and spelling errors;
2. Urgency perception: Identifies time pressure, with features including immediate action vocabulary, fewer polite expressions, and direct expressions;
3. Emotional state recognition: Focuses on negative emotions, with features including negative vocabulary, complex sentence structures (confused thinking), and unusually short responses (avoidance);
4. Cognitive load assessment: Judges difficulties in information processing, with features including repetition, long thinking time (edit history), and simplified responses;
5. Network environment perception: Infers low bandwidth/unstable networks, with features including short messages, avoiding multimedia, and mentioning connection interruptions.

ParaText adopts a rigorous experimental design:

1. Build a text dataset annotated with various user states (real or simulated scenarios), with real states annotated as benchmarks;
2. Present the text to LLMs and ask for state judgments (e.g., "What state might the author be in?"), then evaluate the inference accuracy;
3. Test whether inferences affect response strategies (e.g., whether to simplify responses for fatigued users) to ensure the inferences have practical value.

The results of ParaText guide the optimization of AI interaction design:

• Fatigued users: Simplify responses, use clear structures, and provide step-by-step guidance;
• Urgent needs: Prioritize key information and omit secondary detailed explanations;
• Network-constrained users: Provide lightweight responses and avoid operations involving large data transfers; This adaptive strategy makes AI more natural and considerate, evolving towards a context-aware intelligent partner.

Limitations:

1. The correlation between text features and states is not absolute, making over-inference and misjudgment easy;
2. Privacy issues: State inference capabilities may be abused, requiring a balance between user experience and privacy. Future directions:
3. Multimodal scenarios: Combine typing speed, editing patterns, and sensor data to improve accuracy;
4. Cautious inference: Enable models to avoid arbitrary judgments when uncertain.

ParaText provides a valuable perspective for understanding the social perception capabilities of LLMs, revealing the possibility of AI reading more information from text while reminding us of the opportunities and challenges. In the future, we can expect to see more intelligent and context-aware AI assistants.