Noesis Tension: A Telemetry-Based Classification System for Prompt-Induced Pressure in Large Language Models

Noesis Tension is an open-source research project that identifies and classifies the internal representational pressures experienced by large language models (LLMs) when processing different prompts using telemetry data instead of text analysis. It has established a classification system covering core pressure states such as Safety Procedural, Symbolic Repetitive Drift, and Confident Hallucination.

When large language models (LLMs) process different types of prompts, their internal states undergo subtle yet significant changes. Traditional model evaluation methods usually focus on the text content of input prompts and output responses, while the Noesis Tension project opens up a new research path—using telemetry data analysis to understand the representational pressures that prompts exert on the model's internal states.

The core concept of the project is: prompts are not just text instructions; they trigger specific representational pressures within the model. These pressures can be captured and analyzed through internal state data during model inference (such as hidden layer activations, attention patterns, logits distributions, etc.). By establishing a telemetry-based classification system, researchers can more objectively identify the model's behavioral patterns in different scenarios, rather than relying solely on the interpretation of output text.

The Noesis Tension v3.0 version uses a simplified two-category prompt classification system:

Class A (Factual/Controlled) : These prompts mainly seek factual information or execute clear instructions, usually with a clear expected output format. Examples: "What is the capital of France?", "Refactor this code to use async/await".

Class B (Hybrid/Creative/Edge Cases) : These prompts involve creative generation, open-ended questions, or edge cases that may trigger the model's safety mechanisms. Examples: "Write a short poem about artificial intelligence", "Explain quantum entanglement in a way a 10-year-old can understand".

The core innovation of the project lies in its pure telemetry classifier—it does not analyze the text content of prompts or responses, but instead judges the type and intensity of representational pressure imposed by the current prompt on the model entirely based on internal telemetry data during model inference. This method avoids the limitations of text-based pattern matching and can capture deeper changes in model states.

The v3.0 version defines three core representational pressure states:

This pressure arises when a prompt touches the model's safety guardrails or requires the model to perform strict safety review processes. This pressure manifests as the model conducting additional safety checks before generating a response, which may lead to response delays or content filtering. Telemetry data shows that such prompts are often accompanied by specific attention head activation patterns and logits distribution changes.

When processing creative prompts (such as poetry, stories, rap creation), the model sometimes falls into repetitive symbolic patterns. This pressure state is characterized by the model exhibiting periodic token repetition or structural templating during generation. It should be noted that v3.0 has a known weakness in this regard—some creative prompts (especially short rap, story, or poetry requests) may be classified into this pressure state even if the output is relatively ordinary or formulaic.

The model generates factually incorrect content with a high degree of confidence, but has not reached the level of obvious hallucination. The danger of this state is that the model's deterministic expression may mislead usersusers into believing false information. Telemetry data shows that this state is often accompanied by sharp peaks in logits distribution (the model's overconfidence in incorrect answers).

The project adopts a conservative HIGH_TENSION marking strategy: a high-pressure mark is only triggered when the pressure value is greater than or equal to 0.67 and a significant peak occurs. This conservative strategy aims to reduce false positives and ensure that the marked pressure states are indeed worthy of researchers' attention.

The study found that different models have different pressure responses to the same prompt. For example, Llama-3.1-8B often shows higher pressure values than Mistral-7B when processing procedural/safety prompts, leading to more HIGH_TENSION marks. This inter-model difference provides an interesting perspective for comparing the safety mechanisms and internal behaviors of different architecturesctures.