TL;DR: The mutation-generated faults and curated benchmarks long used to evaluate quantum software testing tools may be missing 86.3% of the bugs that actually show up in real quantum projects. Title: Are We Testing the Wrong Quantum Bugs? A Study of Fault Representation Gaps in Quantum Software Testing URL: Key points 🔍 Collected 410 verified real-world bugs from 10 repositories (Qiskit, PennyLane, Cirq, and more), classified into 56 bug families 🧬 Mutation-testing tools only cover 20 of those families 📚 Curated benchmarks add just 15 more families 📉 Only 13.7% of real bugs fall into families represented by either source 🧩 The big missing categories: compiler transformation errors (66 bugs), type-handling errors (53), missing input validation (47) ⚙️ Most of these faults live in compilers, intermediate representations, and framework interactions, not in the circuits themselves, so circuit mutation alone can't reproduce them The reminder that a correct circuit says nothing about a correct compilation feels like a lesson that reaches well beyond quantum software testing. #QuantumComputing# #SoftwareTesting#

TL;DR A new benchmark called FDE-Gym recreates the ambiguous requests, hidden constraints, and off-target asks that real client-delivery engineers face. Even the strongest model solves only 15% of 95 tasks, and weaker models end up worse than doing nothing. Title: FDE-Gym: The Last Mile of AI Delivery URL: Points 🎯 95 tasks across 20 scenarios in 18 industries, each scored at 4 instruction levels from fully-specified rules to a possibly off-target original request 📉 The best model, Claude Opus 5.5, scores only 0.56, and solves just 14 of 95 tasks (15%) 🧩 The biggest bottleneck is discovering hidden constraints: models score 0.82 when rules are handed to them, but drop 0.28 once they have to find constraints themselves ⚠️ Weaker models like Claude Haiku 4.5 end up worse than a do-nothing baseline on 79 of 95 tasks 💸 Asking the client questions has a built-in "contact cost," and agents that over-ask lose 0.14-0.22 off their final score 📊 Models that hit 88-90% success on benchmarks like Terminal-Bench still fall to 0.19-0.56 on FDE-Gym The lack of correlation between cost or effort and the final score says a lot: working hard and actually improving the client's numbers are two very different things. #AIAgents# #Benchmarks#

TL;DR Claude generated so much code that Git itself became the bottleneck, so Anthropic rebuilt its source control system from scratch. A team of just 7 people pulled it off in 5 months. Title: Redesigning source control at Anthropic URL: Here are the key points. ・📈 Daily merges jumped more than 10x in 5 months, from about 2,500 to over 30,000 ・⚠️ Naively scaling the old Git setup would have meant 600PB of daily traffic ・🖥️ The new system, Scion, assumes clients are always online and pushes work to the server side ・🗂️ A lightweight Changeset data structure allows instant ingestion with no packfiles needed ・⚡ A 3-tier cache hits 99%+ at every layer, cutting checkouts from minutes to seconds ・💾 Infrastructure CPU and memory footprint dropped 200x, and CI CPU usage fell 25% ・🤖 Claude helps with design and code generation, and even acts as a dedicated "superfan" watching individual components What stands out is both the willingness to question Git's core assumptions and the speed of pulling off the rebuild with Claude in just 5 months. #Anthropic# #SoftwareEngineering#

TL;DR: Adding features to a recommender system is easy, but safely removing them is brutally hard. Pinterest built a production framework that automates that whole cleanup process. Title: FORTE: A Production Framework for Feature Evaluation and Deprecation in Industrial Recommender Systems URL: Key points 🧩 Adding a feature just needs a metric win, but removing one must not regress any metric, a built-in asymmetry that stalls cleanup ⏱️ Daily perturbation experiments automatically track feature importance and rank stability over time 🧪 Grouped perturbation evaluates candidate feature sets together, catching interactions before costly retraining 🚀 Pooled execution resolves all experiments for a model in a single distributed inference pass, cutting runtime 54x for 300 experiments 🔗 Automated lineage tracking finds every downstream consumer model, making deprecation safe 📉 On the Pin Ranker model, removing 8 features cut parameters by 29.9%, boosted throughput 19.8-25.5%, and saved hundreds of thousands of dollars a year Feature removal infrastructure deserves as much engineering investment as feature addition, and this looks like a solid blueprint for it. #RecSys# #MLOps#

TL;DR A new framework safely repairs the "hallucinations" that happen when LLMs translate natural-language optimization problems into math models and solver code. It lifts accuracy by up to 20 points while keeping regressions below 1%. Title: Fix What's Wrong, Keep What's Right: Hallucination Repair for LLM-Based Optimization Modeling URL: Points 🔍 Cross-checks the problem text, math model, code, and data in five directions to pinpoint exactly where and why something is wrong ✂️ Picks the smallest necessary fix scope, from code-only to a partial model edit to a full structural rebuild, based on evidence 📜 24 reusable "RepairCards" spell out, as contracts, what can be changed, what must be protected, and what counts as acceptance ✅ A fix is only kept if it passes both an execution test and an audit; otherwise the original artifact is preserved 📊 Tested across 6 benchmarks and multiple upstream generators, it improves accuracy by 6.3 to 20.1 points while capping regression at 1.0% 🧩 Removing the diagnosis step alone drops the rescue rate by 24.5 points, confirming each mechanism's contribution Turning "repair" into a systematic post-processing step, rather than just better generation, feels genuinely practical and could extend well beyond optimization into other structured-output generation tasks. #LLM# #OptimizationModeling#