AI & ML
AI agents rebuild entire programs from behavior alone on the MirrorCode benchmark
MirrorCode is a long-horizon coding benchmark whose 25 targets — spanning Unix utilities, interpreters, static analysis, cryptography, compression and bioinformatics — must be reimplemented without source access, matching the original program's output exactly on held-out end-to-end tests. The strongest model scores 56% overall and can reconstruct gotree, a 16,000-line bioinformatics toolkit the authors estimate would take a human engineer weeks. Probing the frontier is costly: a single attempt on one large task ran roughly $2,600 of inference over 19 days, underscoring how much budget autonomous long-horizon software engineering still demands.
MirrorCode: AI can rebuild entire programs from behavior alone
Quantum Computing
Path-recording oracle gives the simplest pseudorandom-unitary construction yet
Extending Zhandry's compressed-oracle technique from random functions to arbitrary closed subgroups of U(N), the work builds an interpretable path-recording oracle that stores input-output pairs in superposition via the commutant of the group's tensor-power representation. Comparing the simulators for the symmetric group and the unitary group yields a notably simple pseudorandom-unitary construction — a pseudorandom permutation composed with a random Clifford (PC) — improving on the earlier PFC scheme (Metger-Poremba-Sinha-Yuen; Ma-Huang).
Transformer decoder beats BP-OSD on bivariate-bicycle quantum codes
Published in Quantum, a recurrent transformer-based neural decoder for circuit-level noise on Bivariate Bicycle QLDPC codes reaches logical error rates almost five times lower than belief propagation with ordered-statistics decoding (BP-OSD) on the [[72,12,6]] code at a 0.1% physical error rate, while running an order of magnitude faster with far more consistent latency. On the larger [[144,12,12]] code accuracy drops but the speed advantage holds, providing initial evidence that machine-learning decoders can already outperform conventional ones on small QLDPC codes.
Machine Learning Decoding of Circuit-Level Noise for Bivariate Bicycle Codes
Photon-atom blueprint targets fault tolerance with a 2.6% loss threshold
The proposed platform couples single 87Rb atoms to optical cavities to realize near-deterministic photon-atom controlled-phase gates, using photons for long-range connectivity and atoms as reusable resources for photon generation and cluster-state assembly. A hardware-aware noise analysis on the Raussendorf-Harrington-Goyal lattice yields a photon-loss threshold near 2.6% per physical gate (~15% total per trajectory), with the full Clifford set implementable transversally or fold-transversally and two non-Clifford resource-state routes proposed.
Blueprint for a fault-tolerant compound photon-atom quantum architecture
Pasqal files for a $2B Nasdaq listing via SPAC merger
Pasqal, the French neutral-atom quantum hardware developer, has finalized the public filing of its joint Form F-4 registration statement with the U.S. Securities and Exchange Commission for a proposed business combination with the special purpose acquisition company Bleichroeder Acquisition Corp. II (Nasdaq: BBCQ). The cross-border transaction values Pasqal at roughly $2 billion pre-money, moving the company toward a public listing.
Detecting dynamical phase transitions shown to give provable quantum advantage
Researchers prove that deciding a subsystem variant of a dynamical quantum phase transition is as hard as simulating generic quantum circuits, implying an exponential quantum advantage, while precise estimation is intractable even for quantum computers. They also give a quadratically faster algorithm for estimating Hamiltonian-dynamics observables at multiple time points with Heisenberg-limited precision, extending to anomalies in classical systems via encoding.
Low-overhead planar fault-tolerant logical measurements
Published in npj Quantum Information, the work introduces a scheme for fault-tolerant logical measurements in a planar layout that reduces the qubit overhead such operations typically require — a practical ingredient for scalable error-corrected architectures.
Planar fault-tolerant logical measurements with low qubit overhead
Neutral-atom quantum subroutine boosts Monte Carlo Tree Search for the TSP
AtomTreeSearch is a hybrid classical-quantum algorithm that uses a neutral-atom processor to sample a maximal weighted independent set of candidate actions at each Monte Carlo Tree Search expansion. On Traveling Salesman instances up to 60 random-Euclidean and 100 TSPLIB cities it generally matches or outperforms OR-Tools and simulated annealing, with the quantum subroutine producing more diverse, higher-quality branches than classical alternatives.
Quantum-enhanced Monte Carlo Tree Search framework for combinatorial optimization problems
Lie-group diffusion model synthesizes hardware-aware quantum circuits
The method pairs a circuit-skeleton selector that chooses an entangling template with a diffusion model that generates single-qubit gates directly on the curved SU(2) (S^3) manifold. On transverse-field Ising and Heisenberg-XXZ three-qubit simulation targets it outperforms comparable baselines and can produce circuits tuned to constraints such as large or small rotation angles, while balancing fidelity against complexity.
Lie Group Diffusion Models for Hardware-Aware Quantum Circuit Synthesis
EU launches Quantum, Generative-AI and Virtual-Worlds skills academies
The European Commission officially launched three digital skills academies — covering Quantum, Generative AI and Virtual Worlds — as part of one of Europe's most ambitious pushes yet to build the workforce its critical technologies will require. The announcement came during the Digital Skills Awards 2026 ceremony.
Europe Launches Quantum Skills Academy to Expand Workforce Development
Post-Quantum Crypto
Benchmarking variational quantum attacks on a toy symmetric cipher
The study builds a unified, modular framework for variational quantum cryptanalysis of Simplified DES, comparing design choices across convergence behavior, success probability and effective time complexity. It reports clear performance hierarchies among configurations and shows carefully optimized designs can significantly outperform naive quantum search, positioning S-DES as a practical testbed for NISQ-era attacks on symmetric ciphers.
A Modular Benchmark of Variational Quantum Attack Algorithms for S-DES
AI & ML
Quantum-simulation benchmark exposes limits of generative molecular design
NMO swaps proxy drug-likeness oracles for quantum simulations and imposes hard structural constraints, acting as both an ML testbed and a nanotechnology discovery engine. Notably, sophisticated molecular-optimization methods underperform simpler baselines on its rugged, physics-based tasks; a new baseline using a constraint-aware representation and domain-agnostic pretraining surpasses prior physical-property results and surfaces previously unknown structural motifs.
Beyond Drug Discovery: The Nanotechnology Molecular Optimization (NMO) Benchmark
Activation steering as procedural memory for LLM agents
NPM represents agent memory implicitly, distilling procedural skills from historical contrastive experiences into activation-space steering vectors rather than textual guidelines, directly activating task-relevant internal representations. Across four agent benchmarks it matches explicit-instruction baselines, and combining implicit steering with explicit workflows yields complementary, more robust task execution.
Neural Procedural Memory: Empowering LLM Agents with Implicit Activation Steering
Capability-testing router hardens multi-agent LLM systems against injection
ANTAP (Automatic Non-Textual Agent Picker) discards self-descriptions and static surrogates, instead probing agents to measure their true competence and routing via a purely non-textual algebraic projection — a 'linguistic firewall' that renders metadata-based attacks inexpressible. It reports near-zero attack-success rate against description-based injection (versus 67.3% or more for the baseline) and roughly 20% lower ASR against adaptive embedding attacks.
Linguistic Firewall: Geometry as Defense in Multi-Agent Systems Routing
Google DeepMind releases Nano Banana 2 Lite and Gemini Omni Flash
Google DeepMind announced two new Gemini-family models, Nano Banana 2 Lite and Gemini Omni Flash, available for developers to start building with. The vendor post carries no benchmarks or concrete performance figures in the candidate, so the models' capabilities remain to be independently corroborated.
Start building with Nano Banana 2 Lite and Gemini Omni Flash
Microsoft's SkillOpt treats agent skills as trainable parameters
Microsoft Research's SkillOpt reframes the manual editing of agent instructions, or 'skills,' as a training process, aiming to make agent behavior more reliable without changing model weights. The vendor post presents the approach without concrete benchmark figures in the candidate.
Robotics
LLM-driven search discovers tactile-perception networks for robots
TacEvo evolves neural architectures for vision-based tactile sensing using LLM-generated code mutations and crossovers inside a MAP-Elites quality-diversity loop guided by architectural-diversity and efficiency descriptors. On ViTacTip force regression and grating classification it produces trainable architectures ~95% of the time, improves best validation fitness by 56-96% over 20 generations, and matches the expert baseline on force prediction while outperforming it on fine-grained grating classification.