Engineering
SWE Refactor Bench: Testing Whether Coding Agents Can Actually Migrate Legacy Code
SWE Refactor Bench is a new benchmark of 20 whole-repository code migrations across four categories of technical debt, designed to…
Engineering
AnTrap: Benchmarking GUI Agent Robustness Against Runtime Anomalies
AnTrap is a benchmark that injects realistic dynamic anomalies — such as unexpected pop-ups and action misuse — into Android…
Engineering
The Handoff Tax: What Happens When You Switch Models Mid-Agent-Run
This study measures the cost-quality trade-off of switching between cheaper and more capable models partway through a long-running coding agent…
Engineering
Gated Recurrent Transformers: Trading Parameters for Depth via Recurrence
Gated Recurrent Transformer is a new architecture that brackets a single shared transformer core, iterated multiple times, between fixed prelude…
Engineering
Prefix Sliding: Capping Memory for Long Test-Time Reasoning
Prefix Sliding is an inference technique that discards reasoning tokens outside of a fixed prefix (containing key instructions and tools)…
Engineering
RetrievalRouter: A Query-Aware Router for Document Retrieval Pipelines
RetrievalRouter is a lightweight router that learns from query text alone which retrieval pipeline — varying by modality and by…
Engineering
Self-OPD: Teacher-Free On-Policy Distillation for Flow Matching Models
Self-OPD is a teacher-free on-policy distillation framework for flow matching models that replaces a separate, task-specific teacher with the student…
Engineering
Modern LLMs: What the FLIP is an Engram!?
A Hugging Face blog post explains “engrams,” a technique used in modern large language models where embeddings hash sequences of…
Engineering
Piloting the world’s first double-blind AI evaluations
Google DeepMind, together with Singapore’s AI Safety Institute, OpenMined, AVERI, and MLCommons, piloted what it describes as the first double-blind…
Engineering
Writing Down the Line Between Luck and Skill
FINAL-Bench launched a live financial forecasting benchmark that scores entrants against a “luck ceiling” – the 95th-percentile return of 20,000…
Engineering
Introducing the EdgeFirst Model Zoo
EdgeFirst released a Model Zoo of YOLO object-detection and segmentation models, spanning YOLOv5 through YOLO26, validated across multiple edge hardware…
Engineering
Extremely Fast and Accurate Transcription with Granite Speech 5.0 Turbo CTC
IBM released granite-speech-5.0-470m-turboctc, a pair of compact, encoder-only speech recognition models built from a stack of 16 Conformer blocks with…