Aura Trajectory Streaming: Auto-Conversion from Execution Topology to Self-Evolving Datasets

Aura Trajectory Streaming Viz

In the AI field, the best data is not crawled from the internet but high-quality execution trajectories produced by an Agent in a real production environment. Aura’s Trajectory Streaming mechanism aims to refine these fragmented execution records into the power for continuous system evolution.

1. Trajectory as Thinking: The Breadth of Data Capture

Whenever a long-range task is successfully terminated in Aura, the system initiates a “thought review.”

1.1 Full Dimensional Topology Recording

What we capture is not just a dialogue, but a complete Execution Topology:

Prompt input and Context injection.
ACO path selection probability of Meta.
WASM execution logs and Product artifacts of Matrix.
Final user satisfaction scores.

2. Trajectory Cleaning and Distillation

Not all execution records are worth learning. The system filters data through a strict algorithm (Distiller):

CoT (Chain of Thought) Integrity Check: Exclude trajectories with logic leaps that are too large or that contain abnormal compensations.
Information Content Score: Discard tasks that are too simple (repetitive) based on information entropy.
Contrastive Learning Annotation: Automatically generate “positive example paths” and “negative example paths” contrast pairs, which is crucial for reinforcement learning (RLHF/DPO).

3. Automated SFT Data Factory

Filtered data is automatically converted into standard ShareGPT or Alpaca formats. This allows Aura to achieve “working by day, evolving by night”:

When tasks are executed, the system acts as an executor to generate data.
During idle time, the system acts as a teacher to fine-tune local models using this data.

Academic & Design Insights

Design Philosophy: Trajectory streaming embodies the “learning through action” Agent design philosophy. We abandoned static datasets in favor of feedback-driven self-evolution.
Technical Breakthrough: By introducing CoT integrity checks and information entropy scoring, we solved the noise pollution problem in fine-tuning, enabling the Agent to transition from an executor to an educator.
Inspiration: Understanding how to transform unstructured logs into high-quality SFT data is the core lifeline for building vertical-domain expert Agents.

4. Conclusion: Breaking the “Capability Ceiling”

Trajectory streaming allows Aura’s capability to no longer be limited by the pre-training level of the base model. By continuously digesting its own successful experience, Aura can spontaneously grow vertical expertise that exceeds the original model’s capability for specific user business scenarios.

Produced by Dark Lattice Architecture Lab.