AI & ML
A self-evolving agent takes biology protocols from intent to validated bench execution
ProtoPilot is a multi-agent system that turns biological goals into standard operating procedures, SDK-compliant instrument code, and feedback-revised workflows, benchmarked on 294 synthetic- and molecular-biology tasks derived from 98 gold-standard protocols with device-level validity gates. It reports a Top@3 expert-preference rate of 90.2%, an overall protocol-to-code gate pass rate of 89.5%, and an 88.24% Opentrons pass rate versus 32.35% for OpenTrons-AI, with wet-lab runs yielding Sanger-confirmed products and feedback-corrected PCA-assembled DNA. The result pushes autonomous experimentation past plausible-sounding text toward a verifiable route from protocol design to physical execution.
Quantum Computing
A hybrid quantum surrogate offloads the nonlinear collision step in fluid simulation
The scheme uses parameterized quantum circuits with data re-uploading to implement partial Fourier series that model the non-unitary Bhatnagar-Gross-Krook collision dynamics across the full physically admissible relaxation range without retraining. Validated against the classical BGK operator on the Taylor-Green vortex and double-shear-layer benchmarks, it reproduces energy dissipation and shear-driven instabilities, suggesting a practical hybrid path for quantum computing in fluid engineering.
npj QI proposes partial error correction tailored to intermediate-scale machines
Published in npj Quantum Information, the paper sets out a framework for partial error correction aimed at intermediate-scale quantum computers, targeting the regime between raw NISQ operation and full fault tolerance. As a peer-reviewed primary result it addresses how limited error-correcting resources can be deployed on today's devices.
Two central non-local quantum computation tasks are proven equivalent up to constant overhead
Sidestepping known complexity-theoretic obstructions, this Quantum paper studies the relative hardness of non-local quantum computation tasks via resource-efficient reductions between them. It proves f-measure and f-route are equivalent under O(1)-overhead reductions, simplifying existing proofs, transferring properties to f-measure, and giving sub-exponential upper bounds for all functions plus efficient protocols for functions in Mod_k L.
Thermodynamically consistent learned master equations characterize open quantum systems
This Quantum paper introduces a data-driven open-quantum-system model with learnable, thermodynamically consistent nonlinear terms, keeping the result interpretable by directly estimating the system Hamiltonian and linear environment-coupling components. It is validated on synthetic two- and three-level data and on experimental two-level data collected from a quantum device at Lawrence Livermore National Laboratory.
Post-Quantum Crypto
First circuit-level construction of a quantum One-Shot Signature scheme is detailed
Researchers from QPerfect and BTQ Technologies report the first concrete, circuit-level implementation of a One-Shot Signature (OSS) scheme, moving the primitive from abstract cryptographic theory to executable quantum circuits. The joint arXiv preprint provides explicit logical resource estimates and structural parameter matrices for the architecture; as a single trade-reported vendor collaboration it is noted here pending independent corroboration.
AI & ML
Learning from failed trajectories lifts a computer-use agent by 6.6 points without training
Rather than fine-tuning only on successful trajectories, this work uses an LLM to diagnose failure modes, propose inference-time fixes, and emit lightly human-verified code patches that upgrade the agent. Applied to the state-of-the-art OpenCUA-72B on the OSWorld benchmark, it improves success from 42.3% to 48.9% at no training cost and only modest inference overhead, positioning failure-driven improvement as a complement to success-based pipelines.
A 111B Korean-English agent model is adapted from a post-trained base for single-GPU serving
Developed by Cohere and LG CNS, LuckyStar 111B trains from the fully post-trained Command A model and uses preamble conditioning to switch between concise and tool-oriented reasoning. Combining multilingual SFT, RL with verifiable rewards for multi-step tool use, Korean language-consistency rewards, and 4-bit quantization for single-GPU serving, it improves math reasoning, function calling, and agentic NL2SQL while preserving general instruction-following.
Co-evolving reasoning principles replace static rewards for medical multimodal reasoning
Evo-PI generates, evaluates, and iteratively evolves explicit reasoning principles that guide model reasoning while model behavior refines the principles that supervise it, adapting supervision to the model's deficiencies. Instantiated on medical visual question answering as a high-stakes testbed, it reports consistent gains up to 24.6% across eight benchmarks and multiple model backbones, with code released.
Robotics
A humanoid-manipulation benchmark for the chemistry lab exposes a precision gap
Labimus reconstructs over 30 functionally faithful assets from real organic-chemistry workstations, defining six atomic operations and a seven-step solid-weighing workflow with closed-loop instrument readouts. Benchmarking three representative policies under procedural layout and environmental perturbations, it finds that policies which nominally complete tasks can still miss the quantitative tolerances experimental protocols demand, separating task completion from experimental validity.
Reinforcement learning over prompts, not actions, extends generalist robot policies
SARL treats an expressive generalist policy as a controllable skill prior and learns to modulate its language inputs through online interaction, composing already-present skills to solve tasks outside the pretraining distribution. The authors report that this prompt-space optimization yields structured, semantically meaningful exploration and unlocks complex long-horizon behaviors, significantly outperforming action-space RL adaptation across real-world and simulated benchmarks.
Scene-graph memory plus belief planning cuts myopic backtracking in semantic navigation
The method incrementally maintains a hierarchical 3D scene graph over objects, zones, and regions as a compact global state, then fuses semantic priors with exploration evidence through finite-horizon rollouts on an HSG-based simulator. Across high-fidelity simulation tasks it reports average gains of 9.4% success rate and 5.0% SPL over prior state of the art, with the largest improvements in long-distance navigation.
A VLA model adds temporal memory and parallel decoding for faster, less myopic control
MIRTH targets the temporal myopia, reasoning gaps, and autoregressive-decoding overhead of single-frame VLA models with three additions: memory hubs compressing long- and short-term scene evolution, mutual-information-optimized latent reasoning tokens, and vector-wise parallel action decoding. It reports state-of-the-art performance with emergent error-recovery behavior on the LIBERO simulation benchmark and a real-world LeRobot platform, with code and datasets released.
A meta-policy allocates test-time compute for robot control, trimming latency by a third
ELASTIC formulates the split between sequential denoising refinement and parallel action sampling as a meta-MDP, using RL to learn adaptive per-state compute schedules for a frozen generative control policy. It Pareto-dominates fixed and single-axis scaling on simulated manipulation and, on the real-world pi-0.5 vision-language-action model, matches best-of-10 success while reducing wall-clock latency by 34%.
A 6k-trajectory visual-tactile dataset targets dual-arm dexterous humanoid manipulation
Built on the Unitree G1 humanoid, RoboTacDex provides 6,000 multi-modal trajectories with multi-view RGB and depth, tactile feedback, and detailed semantic annotations, collected via a millisecond-synchronized multi-camera system. It emphasizes challenging tasks requiring dual arms and dexterous hands; benchmarking three imitation-learning models shows successful trials and moderate generalization across the task suite.