ScienceTrace

The Instruction Provenance Problem in Autonomous AI Agents

ScienceTrace Research Proposal | Autonomous AI agents that read external content and call tools face a structural blind spot: they cannot reliably tell an instruction from their authorized user apart from one hidden inside content they merely read. We define this instruction provenance problem and propose Provenance-Bound Instruction Graphs (PBIG), a lineage-tracking framework for gating agent actions on instruction origin rather than content alone.

Where Should a Reasoning Model Spend Its Thinking? A Proposed Uncertainty-Adaptive Test-Time Compute System

ScienceTrace Research Proposal | We propose UARCS, an inference-time system that allocates a fixed test-time compute budget across a reasoning trajectory using step-level calibrated uncertainty, rather than spending it uniformly or after the fact.

Beyond Detection and Tracking: A Proposed Deep Learning Framework for Motion, Uncertainty, and Anomaly Understanding in Video

ScienceTrace Research Proposal | We propose DCMU-Net, a deep learning framework that goes beyond conventional object detection and tracking to jointly model motion patterns, trajectory uncertainty, environmental context, and unusual movement in video.

Sublime: A New Sketch Algorithm Rethinks How Big Data Systems Track Endless, Skewed Data Streams

A SIGMOD 2026 paper called Sublime redesigns the sketches that let big-data systems count trillions of events in real time, with memory and error that adapt as the stream keeps growing.

How Machine Learning Could Decode the Growth of Black Holes

ScienceTrace explores how combining astrophysics, machine learning, statistical inference, and physics-based simulations could help decode how supermassive black holes grow and shape the galaxies around them.

ScienceTrace Proposal: A Scientific Validation Framework for Self-Evolving AI Software

ScienceTrace proposes SV-SEAIS, a framework combining AI-driven software evolution with independent scientific validation — testing whether AI can discover better algorithms while producing reliable evidence that its improvements are genuine.