TraceLite Cuts Token Overhead for Software Engineering Agents
A new method reduces the growing interaction traces of software engineering agents to lower token use and inference cost.
Research Digest
Preprints and papers, rewritten in plain English.
Quantum Broker Aims to Route Jobs Across Heterogeneous QPUs
A proposed broker would distribute quantum computations across cloud QPUs based on quality-of-service needs.
Multi-granularity method flags AI-involved code from tokens to files
A new arXiv paper proposes detecting and localising AI-involved code at granularities ranging from individual tokens to entire files.
GraphRAG Framework Assesses Post-Wind-Event Building Recovery
A proposed engineering GraphRAG framework connects observed physical damage to safety, habitability, and recovery inferences from heterogeneous post-disaster evidence.
Object-Oriented Harnesses Proposed for Diverse Agent Tasks
The paper argues that a single global agent harness is ill-suited for heterogeneous tasks and points to an object-oriented approach—though the abstract ends before details.
Requirements Engineering as a Key to Reducing Software Project Risks
An arXiv review highlights how requirements problems drive late delivery, budget overruns, and cancellations, and why improving requirements engineering matters.
Measuring Whether LLM Code Repairs Stay Minimal and Faithful
A new arXiv paper argues that code-editing models should be judged on minimality and fidelity, not just correctness.
Embodied Playground benchmark targets long-horizon robot reasoning
A new robotics benchmark evaluates whether foundation-model agents can sustain coherent reasoning and reliable physical execution across interdependent task stages.
REPOCOD Benchmark Finds LLMs Still Fall Short of Programmers
A new repository-level coding benchmark suggests large language models are not yet ready to replace human programmers.
Host-Side Tokenized Logging Speeds Up RTL Simulation
A new technique offloads printf-style formatting from cycle-accurate RTL simulation to the host, cutting a major debug bottleneck.
ACE-Brain-0.6: Scaling Spatial Grounding for Physical AI
A new arXiv preprint links general multimodal ability to strong spatial perception and introduces ACE-Brain-0.6 for scaling spatial grounding.
Attributing Failures to Visual Corruptions in VLA Models
A new method uses counterfactual reasoning to determine whether a specific visual corruption caused a vision-language-action model's mistake in a given episode.
Hip Exoskeleton XoNav Uses Haptic Cues to Guide Walkers Without Vision
A four-actuator hip exoskeleton aims to make navigation guidance more intuitive by delivering haptic cues directly through the body.
Action Influence Maximization Could Fix Video Models Ignoring Commands
A new training objective aims to stop action-conditioned video models from ignoring commanded actions, improving controllability.
Legged-Aerial Robot Design Uses Separate Control Contraction Metrics
A new arXiv paper addresses the difficulty of combining legged locomotion with aerial flight in one robot.
Quadrotor control adapts its predictive model to changing winds in real time
A disturbance-aware Koopman control framework lets quadrotors keep their predictive models accurate as aerodynamic conditions shift.
Sensitivity-Orthogonal Learning Keeps Sideslip Estimation Adaptive
A new method decouples learned model correction from online cornering-stiffness adaptation so each uses the filter innovation without interfering.
KITE Lets Robot Skills Transfer Across Embodiments Without New Demos
A new method decouples interaction intent from robot-specific kinematics, enabling zero-shot transfer of manipulation skills to structurally different robots.