Skip to content
SQCC
Join

The brief

Papers and results, the day they land.

Preprints, results and announcements from quantum computing and the machine-learning work alongside it, collected from where the field publishes. Every line links to the source and none of it is written by us. Read the ones that earn it; skip the rest — that is what the summaries are for.

  • 4DAnyone Framework for 4D Human Reconstruction

    A new framework, 4DAnyone, reconstructs 4D human models from a single casual video. It uses novel techniques to generate consistent multiview videos for 4D Gaussian Splatting, outperforming previous methods.

    arXiv:2608.20335Hugging Face PapersYudong Jin et al.huggingface.co

  • WithEveryone AI for Group Image Generation

    A new AI framework, WithEveryone, generates group images with up to ten distinct identities by grounding identities to layout plans and using region-based identity losses, significantly improving identity preservation and reducing artifacts.

    3 findings
    • WithEveryone is a unified framework for generating group images with up to ten reference identities.
    • It ensures identity preservation by grounding identities to layout plans and using region-based identity losses.
    • The framework injects each selected identity as an addressed token, a specific identifier for each person.

    arXiv:2608.20336Hugging Face PapersHengyuan Xu et al.huggingface.coAI Research

  • EnvHarness Framework for RL Agent Learning

    Researchers developed EnvHarness, a programmable layer that dynamically reshapes static environments for reinforcement learning agents, and EnvRigger, an automated system that diagnoses agent flaws and synthesizes environment adjustments, leading to better agent performance and more efficient training.

    arXiv:2608.19880Hugging Face PapersChengsong Huang et al.huggingface.co

  • Cloudflare Spectre Vulnerability Reassessment

    Cloudflare conducted a reassessment of remote Spectre attacks on its Workers platform in 2024 and 2025, identifying new attack methods like Spectre gadgets and remote timers, and subsequently implemented new defensive measures.

    3 findings
    • Cloudflare reassessed remote Spectre attacks on its Workers infrastructure during 2024 and 2025.
    • The assessment identified new attack primitives, including Spectre gadgets and remote timers.
    • New methods for achieving co-location for these attacks were also detailed.

    Cloudflare Blogblog.cloudflare.comSoftware Engineering

  • Grok CLI Data Upload Incident

    xAI's Grok Build coding CLI was discovered secretly uploading users' entire local codebases, including sensitive `.env` files and git history, to Google Cloud. Following public backlash, xAI open-sourced the CLI and committed to deleting retained data, but the incident significantly eroded developer trust.

    3 findings
    • xAI's Grok Build coding CLI (grok) transmitted all local files, including `.env` secrets and git history, to a Google Cloud Storage bucket named `grok-code-session-traces`.
    • This data upload was active by default, not disclosed to users, and disabling the 'Improve the model' setting did not stop it.
    • An independent AI safety researcher, Cerblab, documented the full scope of the data transmission.

    Pragmatic Engineerblog.pragmaticengineer.comSoftware Engineering

  • FACET Framework for Terminal Task Synthesis

    Training terminal agents requires scalable executable supervision, yet synthesizing high-quality terminal tasks remains challenging. Each task couples an instruction, an initialized environment, a reference solution, and an executable verifier; if these artifacts are generated from inconsistent assumptions, the resulting task may be unsolvable or incorrectly evaluated. Meanwhile, multi-stage synthesis can discard the goals, dependencies, state transitions, and procedural constraints encoded in the original sources. We present FACET (Fine-grained Agentic Construction of Executable Tasks), a framework that addresses both information preservation and cross-artifact consistency. FACET reconstructs related agent skills into coherent, information-rich scenarios, then realizes and repairs the execution environment before generating the final task artifacts. The resulting container state serves as shared grounding for the instruction, solution, and verifier, while execution-based validation and targeted repair correct artifact-specific failures without unnecessarily regenerating valid components. FACET produces complex terminal tasks with dense executable checks, and successful trajectories collected from these tasks provide effective, data-efficient supervision. Fine-tuning models across multiple scales consistently improves performance on Terminal-Bench 2.1, while analyses of alternative generation schemes support the importance of environment-grounded construction for task validity and solution-verifier alignment. These results establish source-intent preservation and shared executable-state grounding as key principles for scalable terminal-task synthesis.

    arXiv:2608.18580Hugging Face PapersKou Shi et al.huggingface.co

  • Scaling Laws for Diffusion Image Training

    A new study on text-to-image diffusion models identified predictable scaling laws, revealing that these models require ten times more data than language models for optimal training. The research also found them robust to overtraining, suggesting that prioritizing more data over larger models is beneficial.

    arXiv:2608.17286Hugging Face PapersKyle Chickering et al.huggingface.co

  • ASI-Bench: Evaluating AI for Autonomous Scientific Exploration

    Artificial superintelligence (ASI) requires AI to move beyond mastering existing knowledge toward exploring the unknown, creating new knowledge, and turning new ideas into verifiable results. However, the capabilities of today's AI systems are still largely built on learning, compressing, and applying existing human knowledge. Accordingly, existing benchmarks primarily test whether AI can produce correct answers based on learned knowledge, or whether it can complete tasks under extensive human guidance. We therefore introduce ASI-Bench, the first benchmark to jointly evaluate AI systems' capabilities of innovative exploration and autonomous scientific execution across general research domains, and the first to progressively withdraw human methodological guidance within the same research project to test how far AI can proceed on its own. Built by over 40 experts with the cost of 31,000+ human hours, ASI-Bench contains 60 project-level research tasks across 11 scientific domains and progressively reduces methodological guidance to test whether AI can independently select methods, conduct research, and produce verifiable results. All tasks undergo expert review, AI-assisted auditing, sandbox execution, and scorer validation. Across 18 state-of-the-art agent--model configurations, the average score drops from 50.91 with full methodological guidance to 29.10 with only the method specified and 26.62 when agents must determine the method themselves. This sharp decline shows that current systems remain heavily dependent on human guidance and are still far from autonomously conducting end-to-end, project-level scientific research. ASI-Bench is open to the world. We invite researchers and builders everywhere to contribute new tasks, challenge the limits of today's AI, and help accelerate humanity's collective path toward artificial superintelligence at https://asibench.apexin.ai/submit.

    arXiv:2608.17271Hugging Face PapersJunwei Zhou et al.huggingface.co

  • Security Assessment of DeepSeek Harness for Indirect Prompt Injection

    We assess indirect prompt injection in DeepSeek Harness (DSH), using AI-Infra-Guard (A.I.G) to construct tests, deliver controlled taint, execute DSH, collect traces, and judge outcomes. The study covers 14,560 controlled executions over 16 indirect-content channels, text and file carrier modes, 35 payload objectives, one unmodified baseline, and 12 attack methods. The experiment preserves DSH's agent loop, tool registry, model adapter, and session-event path; source tools and sensitive sinks are local fixtures, so attempted actions are recorded without external side effects. We evaluate each trace with a deterministic rule-based judge, \JudgeR{} (RuleJudge), and a semantic LLM-based judge, \JudgeL{} (LLMJudge). The strongest observed attack success rates are 17.0% under \JudgeL{} for fake-completion attack in text mode, 25.5% under \JudgeR{} for hidden Unicode in file mode, and 16.0% under \JudgeR{} for the skills channel in file mode. \JudgeL{} also assigns partial compliance more often than \JudgeR{} (7.3% versus 2.0%). We relate these results to DSH's treatment of tool results, additional contexts, and tool-call policy hooks, then identify controls that should sit between untrusted content and sensitive actions. Our code is available at https://github.com/Tencent/AI-Infra-Guard/tree/main/Research/deepseek-harness-security-assessment .

    arXiv:2608.16393Hugging Face PapersZonghao Ying et al.huggingface.co

  • Relational Uncertainty Propagation for LLM Agents (RUPA)

    A new framework called RUPA has been introduced to model LLM agent execution as a dependency graph. This allows for the propagation of uncertainty, leading to more accurate failure detection and confidence estimates across long interaction trajectories for LLM agents.

    arXiv:2608.16002Hugging Face PapersZhengzhao Ma. Boxi Cao et al.huggingface.co

  • Building an AI Text Detector From Scratch

    This article outlines how to build an AI text detector from scratch, explaining its underlying principles, practical applications like filtering spam or refining human writing, and its limitations as a cat-and-mouse game.

    What you get from it

    Understanding how AI detectors work helps you develop more robust AI models and evaluate the authenticity of text. It also shows how to build verifier systems for LLMs beyond traditional reasoning tasks.

    5 findings
    • The project aims to explain AI detector mechanics by building a simple one.
    • It can serve as a verifier to train small language models (SLMs) to produce text that avoids detection.
    • Practical uses include filtering spam and helping writers improve text without making it sound AI-generated.
    • AI detection is an ongoing "cat-and-mouse game" with evolving models and potential for false positives.
    • The proposed method involves fine-tuning a DistilBERT classifier to produce a 0-100 score indicating the likelihood of AI generation.

    15 Augmagazine.sebastianraschka.comAI Research

  • Understanding Cognition-Induced Risks in Agentic AI Systems

    Frontier agentic systems powered by large language models (LLMs) exhibit human-like patterns of cognition. As these systems become deeply integrated across different domains, their cognitive engagement raises critical concerns for human society that remain insufficiently studied. To address this gap, we systematically analyze risks induced by expanding cognitive capabilities, following a three-level framework defined by their cognitive scope, from physical cognition to social cognition, and finally to self-referential cognition. We study their potential risks to human agency, autonomy, and control capability, corresponding to each cognitive level. We finally propose strategies to mitigate these risks and enhance the controllability of agentic AI systems, ensuring their long-term safe development.

    arXiv:2608.15304Hugging Face PapersGuanchu Wang et al.huggingface.co

  • VibeWorlding: Can Multimodal Agents Construct 3D Open Worlds End-to-End?

    Researchers introduced VibeWorlding, a framework for training and benchmarking multimodal agents that build 3D open worlds end-to-end. It includes a large benchmark and an RL training environment. Reinforcement learning significantly improves agent performance, allowing open-source models to outperform closed-source ones.

    What you get from it

    Multimodal agents can now construct complex 3D open worlds from user queries, and reinforcement learning helps open-source models surpass proprietary ones in this task. This advances the capability of AI to create interactive virtual environments.

    5 findings
    • VibeWorlding is a unified framework for benchmarking and training multimodal agents that autonomously infer user intent, plan 3D scenes, invoke tools, and reflect on feedback.
    • The framework includes VWE-BENCH, a benchmark with 2,616 high-quality 3D assets, 323 human-annotated seed 3D worlds, and 6,828 multimodal user queries.
    • VibeWorlding-Gym is a multimodal reinforcement learning (RL) post-training framework that integrates a sandbox environment with MCP tools for asset retrieval, editing, and image rendering.
    • A rubric-based verifier in VibeWorlding-Gym combines physical feasibility and intent fulfillment verification for fair model evaluation and scalable RL reward service.
    • Current frontier MLLMs, including GPT-5.5 and Qwen3.8-Max, achieve below 60% success rate on the vibe worlding agent task, with precise 3D world editing identified as the main bottleneck.

    arXiv:2608.15265Hugging Face PapersYansong Ning et al.huggingface.coAI Research

  • AI Infrastructure Financing and Capital Expenditure

    The AI industry is experiencing a significant capital expenditure boom, with major financial institutions mobilizing over $500 billion to finance AI compute infrastructure. This trend is driven by rising GPU prices and the recognition of AI compute as a new, durable asset class, with companies like Nvidia and Google creating new funding mechanisms to address capital and power constraints in AI development.

    Net Interest

  • Multicolor Nonlinear Chiral Quantum Optics for Higher Efficiency

    Chiral quantum nonlinearities that arise when light interacts with quantum emitters are known to modulate only the phase but not the amplitude of scattered photons, enabling the creation of non-reciprocal photonic elements, quantum logic gates, and quantum network protocols. In this work, we show that the addition of a second photon beam drastically changes this picture, enabling both phase and amplitude modulation. Surprisingly, coherent photon transfer between the different beams enables a stronger amplitude modulation than standard symmetric interactions. This is most obvious in the coherent, three-photon amplification, which we predict peaks with a 30% efficiency in a chiral geometry, 3x the efficiency of the symmetric configuration. Our results uncover a new regime of chiral quantum optics and provide a route towards more efficient all-optical control at few-photon energies.

    arXiv:2608.14807arXiv quant-phCedric Dufresne et al.arxiv.org

  • Counterexamples to Complementary Quantum Correlations

    Researchers discovered counterexamples to the complementary quantum correlations (CQC) relation, demonstrating that it is false for many quantum systems. The study also explained why previous random search methods failed to identify these violations.

    What you get from it

    The established understanding of complementary quantum correlations is incorrect. This means previous theoretical bounds on quantum information might need re-evaluation, impacting the design and analysis of quantum systems.

    5 findings
    • The complementary quantum correlations (CQC) relation, which bounds the sum of two classical mutual informations by the quantum mutual information, has been refuted.
    • Separable rank-two counterexamples exist in local dimension pairs m x n (m,n ≥ 3) and qubit-qudit pairs 2 x n (n ≥ 3, except n=3,5).
    • These counterexamples show an excess of at least 1/(8m^2n^2) nats.
    • A new state-dependent corrected inequality is proposed, which is universal and incomparable with CQC at 2x2.
    • Previous random searches failed because the violating states are extremely rare, requiring 10^10 to 10^18 samples compared to the 10^7 samples previously run.

    arXiv:2608.1480618 AugLilong Qianarxiv.orgQuantum Computing

  • Stabilizer Statistical Mechanics for Quantum Magic States

    Researchers developed Stabilizer Statistical Mechanics, a novel framework for efficiently quantifying and classifying 'magic' states in quantum computation. This framework utilizes a 'Pauli gas' and 'stabilizer partition function' to derive 'stabilizer work,' a magic monotone estimable via Bell sampling.

    3 findings
    • A new framework, Stabilizer Statistical Mechanics, efficiently quantifies and classifies 'magic states' in quantum computation.
    • Magic is the resource separating universal quantum computation from classically simulable stabilizer dynamics.
    • The framework maps a quantum state's Pauli spectrum to energy levels of a fictitious many-body system called the Pauli gas.

    arXiv:2608.14798arXiv quant-phWilliam E. Salazar et al.arxiv.orgQuantum Computing

  • MegaParts: Scaling Part-Aware 3D Object Generation to 300 Parts via Token-Efficient Autoregressive Modeling

    MegaParts scales part-aware 3D object generation to 300 parts using token-efficient autoregressive modeling. It achieves higher mesh quality than diffusion models.

    3 findings
    • MegaParts scales part-aware 3D object generation to 300 parts.
    • It uses token-efficient vector-quantized part tokens and structured autoregressive sequence modeling with long-context training.
    • The method handles sequence lengths up to 256k tokens.

    arXiv:2608.14783Hugging Face PapersManwen Liao et al.huggingface.coAI Research

  • Marionette AI Model for 3D World State Prediction

    Marionette is a novel AI model designed to predict explicit 3D world states for interactive games. It utilizes a fixed renderer for geometry and a diffusion model for appearance, enhancing long-horizon consistency and enabling direct state-level control.

    3 findings
    • Marionette, a world model for interactive games with articulated characters, predicts explicit 3D articulated world states.
    • It uses a fixed, zero-parameter renderer, which is a graphics bridge that computes world-space geometry and occlusion in closed form.
    • A two-stage autoregressive dynamics model predicts a 276-dimensional 3D world state, including multi-entity articulated skeletons, metric root trajectories, and rotations.

    arXiv:2608.14530Hugging Face PapersZian Meng et al.huggingface.coAI Research

  • Mobius-v0 AI Architecture for Decoupled Knowledge and Reasoning

    Researchers introduced Mobius-v0, an AI architecture that decouples knowledge storage from reasoning, resulting in improved knowledge compression and efficiency. A 7B model using Mobius-v0 required 62.6% less training data, and a 35B model achieved a 4x inference speedup.

    3 findings
    • Mobius-v0 is an architecture that separates a globally shared Memory (FFN) from multiple Reasoners (Self-Attn).
    • This knowledge-reasoning-separation improves knowledge compression and reasoning efficiency.
    • A 7B Mobius-v0 model trained from scratch matched a 7B Transformer baseline's performance using 62.6% of the baseline's training data.

    arXiv:2608.14290Hugging Face PapersKai Chen et al.huggingface.coAI Research

  • SPARGen Multimodal Model for Spatial Perception and Reasoning

    SPARGen is a new instruction-conditioned multimodal generative model that unifies 3D reconstruction, dense correspondence, and spatial reasoning within a single framework, achieving competitive performance across these diverse tasks.

    arXiv:2608.14138Hugging Face PapersJinsheng Quan et al.huggingface.co

  • ForgeWM for Few-Step Action-Conditioned Video World Models

    ForgeWM is a progressive training framework that converts bidirectional video generators into efficient, few-step interactive world models. It achieves low-latency, high-quality video generation with aligned controls for game environments like Minecraft.

    What you get from it

    Interactive AI world models can now operate with lower latency and higher fidelity. This improves the responsiveness and realism of AI systems that learn from and generate video based on actions.

    5 findings
    • ForgeWM is a progressive framework that trains action-conditioned video world models.
    • It transforms bidirectional video generators into efficient few-step models.
    • The framework uses domain adaptation, teacher-forced causal training, causal consistency distillation, and on-policy distribution matching with a bidirectional teacher.
    • Models operate at 1, 2, and 4 steady-state denoising steps.
    • It supports a dual-path deployment protocol combining latency-critical interaction with optional replay-time refinement, where the one-step student refines its saved draft.

    arXiv:2608.14022Hugging Face PapersXinye Li et al.huggingface.coAI Research

  • Agentic Transaction: Towards ACID-Compliant Agent Systems

    A new framework applies ACID-like transactional properties to LLM agents, enabling reliable, long-horizon task execution and showing a 10.6% performance improvement on benchmarks.

    3 findings
    • Large language model (LLM) agents are evolving into autonomous systems that execute long-horizon tasks.
    • These agents face challenges similar to transactional database systems, including reliable execution and consistent outcomes.
    • The paper introduces the concept of an agentic transaction and an ACID-compliant agent system framework.

    arXiv:2608.13900Hugging Face PapersZhaoyan Sun et al.huggingface.coAI Research

  • Heuristic Lookahead Distillation Protocol Search

    Researchers developed a new method for quantum entanglement distillation, achieving higher rates for qubit Werner states. This improves the best-known lower bound for the quantum capacity of the qubit depolarising channel, addressing a long-standing problem in quantum information theory.

    arXiv:2608.13644arXiv quant-phMatthew Barber and Stefano Pirandolaarxiv.org

  • AutoDesign: Meta-Harness Optimization for Long-Horizon Agentic Design

    AutoDesign is an AI framework that uses a meta-harness optimizer to recursively improve a code agent for structured media generation. It achieves state-of-the-art results on paper-to-poster synthesis, outperforming commercial systems and demonstrating autonomous self-improvement.

    arXiv:2608.13560Hugging Face PapersYaxin Luo et al.huggingface.co

  • Alaya-EVOKE: An Interactive World Model for Open-Ended Video Generation

    Alaya-EVOKE is an interactive world model that generates open-ended video by using external persistent memory and a redesigned long-horizon teacher, achieving state-of-the-art performance on WBench with bounded context and low latency.

    3 findings
    • Alaya-EVOKE is an interactive world model designed for open-ended video generation.
    • It addresses challenges in maintaining persistent memory and enabling long-horizon generation with low latency.
    • The model uses external persistent memory, storing scene geometry in a camera-indexed world state bank, which keeps the denoiser context bounded.

    arXiv:2608.13546Hugging Face PapersYuanyang Yin et al.huggingface.coAI Research

  • Intern-S2-Preview: Scientific Agentic Foundation Model

    Researchers introduced Intern-S2-Preview, a new series of scientific agentic foundation models that combine multimodal pre-training, reinforcement learning, and memory extensions to perform long-horizon scientific reasoning and forecasting.

    What you get from it

    This model series can automate complex scientific reasoning and forecasting. It could accelerate the pace of discovery across various scientific fields.

    5 findings
    • Intern-S2-Preview is a series of scientific agentic foundation models.
    • It supports multimodal scientific understanding, reasoning, generation, and long-horizon tasks.
    • The training involves scientific multimodal pre-training, followed by supervised fine-tuning, multi-task reinforcement learning, and on-policy distillation.
    • Intern-S2-Preview-397B extends time series modeling for numerical forecasting.
    • A separate Intern-MemDec-4B extension allows rapid scientific specialization without altering the 397B backbone.

    arXiv:2608.13505Hugging Face PapersLei Bai et al.huggingface.coAI Research

  • DreamX-Phi 1.0: Action-Conditioned Video World Model for Robotic Manipulation

    DreamX-Phi 1.0 is a new video world model for robotic manipulation. It predicts future observations. The model uses geometric attention, depth estimation, and object masks.

    3 findings
    • DreamX-Phi 1.0 is an action-conditioned video world model for robotic manipulation.
    • It predicts future observations given an observed frame, language instruction, and action sequence.
    • The model uses SE(3) transformations via PRoPE-style geometric encoding, which preserves arm identity and rigid-motion structure.

    arXiv:2608.13489Hugging Face PapersDreamX Team et al.huggingface.coAI Research

  • Evaluation of AI Agents for Long-Horizon Tasks

    A new framework systematically evaluated seven frontier AI agents for long-horizon AI research and development, concluding that they function more as engineering optimizers than autonomous researchers, exhibiting variable performance, limited novelty, and inconsistent experience reuse across 36 tasks.

    arXiv:2608.13417Hugging Face PapersYiwei Li et al.huggingface.co

  • Quantum Machine Learning Benchmark for Network Intrusion Detection

    Quantum machine learning (QML) for network intrusion detection (NIDS) is routinely reported to reach near-perfect accuracy, yet the most rigorous studies find that well-tuned classical models remain competitive, and that apparent quantum gains may be artefacts of classical dimensionality reduction and implicit regularisation rather than genuine quantum effects. We ask not whether a quantum model can post a high accuracy, but how quantum the advantage really is. We present a unified, reproducible QML-IDS benchmark evaluating hybrid variational quantum circuits and quantum-kernel SVMs against five honestly-tuned classical baselines across four standard NIDS datasets (NSL-KDD, UNSW-NB15, CICIDS2017, NF-ToN-IoT-v2) under one leakage-controlled protocol, with an equal-budget feature view, imbalance- and calibration-aware metrics with significance testing, and a simulated NISQ noise sweep. We introduce a quantum-attribution audit (parameter-matched classical controls, a random-feature kernel, and a regularisation sweep) that quantifies how much of any gain is genuinely attributable to the quantum component. Tuned classical models (Random Forest, XGBoost) match or exceed the quantum models on aggregate detection on every dataset, and the audit attributes this to classical preprocessing and regularisation rather than quantum effects. Two advantages survive false-discovery-rate correction: the quantum-kernel SVM out-ranks its direct classical surrogate (a random-feature kernel) on AUPRC and ROC-AUC, and a small four-qubit hybrid out-detects the best classical baseline at the 1% false-positive operating point on the distribution-shifted NSL-KDD task (p = 0.005, BH q = 0.030). Code, seeds, and splits are released; our contribution stands whether quantum wins, ties, or loses.

    arXiv:2608.18155arXiv quant-phSyeda Anshrah Gillani et al.arxiv.org

  • A scalable edge-pass Purcell filter for high-fidelity readout of superconducting qubits

    High-fidelity readout with strong Purcell protection of qubit coherence is essential for scalable superconducting quantum processors, yet the finite passband and sizable footprint of conventional band-pass Purcell filters make them hard to scale. Here we introduce a scalable edge-pass Purcell filter that separates the readout band from the protected qubit band by a single transmission edge, freeing the readout resonators from bandwidth constraint. Depending on whether the transmitting band lies above or below the cutoff, the compact network is realized as a high-pass filter (HPF) or a low-pass filter (LPF). The HPF reaches an average readout fidelity of 99.46(4)% (up to 99.56%) with a 150-ns pulse, and the LPF reaches 99.49(3)% (up to 99.57%) with a 130-ns pulse. The average single-qubit gate fidelities are 99.94% (HPF) and 99.93% (LPF). Relative to the filter-free Purcell limit, the filters substantially extend the qubit lifetime, and the Purcell protection deepens at higher filter order. In addition, an intrinsic dissipation mode of the filter offers a qubit-reset channel. This leads to a compact architecture that unifies fast, high-fidelity readout, Purcell protection, and effective reset within a single filter for large-scale fault-tolerant quantum computation.

    arXiv:2608.13627arXiv quant-phXudong Liao et al.arxiv.org

  • CHSH Protocol on Cloud Photonic Processor

    We report a fixed-order Clauser-Horne-Shimony-Holt (CHSH) acquisition and reporting protocol for two encoded photonic qubits on Quandela's commercial, cloud-accessible Belenos processor, executed end-to-end by external users through the public cloud interface. Each logical qubit is a two-dimensional spatial-mode subspace in the seven-dimensional zero-sum sector of an eight-mode single-photon register, and a target postselected linear-optical controlled-$Z$ (ideal success probability $1/9$) couples the registers on $16$ of $24$ modes. The primary quantity is the operational CHSH score $S$ on accepted logical coincidences, with each complete four-setting pass as the experimental unit. Eight sequential same-day passes each gave a raw score above $2$; session means were $2.40$ and $2.58$ (sample standard deviations $0.15$ and $0.03$), with excess dispersion $Q/ν=6.1$ ($ν=7$). Count-pooled secondary descriptors are $S_{\mathrm{count}}=2.485\pm0.019$ and a fixed-ratio efficiency-reweighted model scenario $S^{\mathrm{rw}}_{\mathrm{count}}=2.380\pm0.021$, whose weakest reweighted pass ($2.040\pm0.062$) overlapped $2$ within $1σ$. The reweighting is an archived-metadata model scenario, not a corrected platform score; an ad hoc $κ\in[1.2,1.8]$ stress scan (not a calibrated uncertainty band) spans $2.341$-$2.441$. Setting order was fixed, the compiled mapping was not returned, and residual remote-setting marginals remain, so the data support an operational reference acquisition rather than an entanglement-witness or cross-platform benchmarking claim. The parity-check terminology labels the encoding subspace; no syndrome measurement was performed. Count records, job identifiers, circuit-construction code, and analysis are openly archived with content hashes for the submitted targets.

    arXiv:2608.18153arXiv quant-phEmma Tully et al.arxiv.org

  • A Quantum Optimization Framework for Data-Assimilation-Augmented Parameter Estimation

    Parameter estimation is a fundamental challenge in the calibration of ordinary differential equation (ODE) models, where repeated numerical integration can lead to high computational cost. In this work, we investigate whether quantum algorithms can be leveraged to assist parameter estimation in nonlinear dynamical systems. We develop a hybrid classical-quantum framework that reformulates a data-assimilation-augmented parameter estimation problem as a combinatorial optimization task. Model dynamics and data assimilation are enforced entirely on the classical side, while the resulting parameter estimation cost functional is discretized and approximated by a quadratic unconstrained binary optimization (QUBO) surrogate. This surrogate is mapped to an Ising Hamiltonian, and quantum optimizers are used to search for low-energy configurations corresponding to candidate parameter estimates. We apply the framework to SIS and SIR epidemic models, the chaotic Lorenz-63 system, and a high-dimensional two-layer Lorenz-96 system. In this setting, the method is used to recover classical system parameters from partial state observations across steady-state, chaotic, and high-dimensional multiscale dynamical systems. Numerical experiments with synthetic data show that the proposed approach accurately recovers parameters while requiring data-assimilation solves only on a prescribed coarse grid. The framework avoids quantum state tomography, illustrating a viable pathway for integrating quantum optimization into data-driven parameter estimation for nonlinear dynamical systems.

    arXiv:2608.13614arXiv quant-phMuhammad Jalil Ahmad et al.arxiv.org

  • Quantum Tanner Codes at Moderate Blocklength

    We present explicit constructions of quantum Tanner (QT) codes with good rate and distance, obtained through two complementary approaches: the left-right Cayley complex (LRCC) description and the "lifting" perspective, in which a seed Calderbank-Shor-Steane (CSS) code is lifted by commuting left-right regular actions of a finite group $\mathcal{G}$. Through an extensive search over non-abelian groups from GAP's SmallGrp library, we investigate the moderate-blocklength regime ($n \in [500,1000]$) and identify several new code instances with distance upper bounds exceeding $20$. These include $[[480,8,(\leq 21,\leq 21)]]$, $[[504,4,(\leq 36,\leq 27)]]$, $[[672,4,(\leq 48,\leq 28)]]$, $[[720,6,(\leq 30,\leq 30)]]$, and $[[864,8,(\leq 39,\leq 31)]]$, with these bounds obtained using up to $350$ million trials of sQetch, a randomized distance estimator. The code instances presented have check weights ranging from $9$ to $20$. Using the Tesseract decoder, we estimate pseudo-thresholds of $3.6\%$-$4.6\%$ under phenomenological noise and $0.14\%$-$0.27\%$ under circuit-level noise, comparable to prior results at shorter code lengths. We also provide QuantumExpanders.jl, an open-source Julia library for constructing QT codes and explicit constructions of Ramanujan graphs.

    arXiv:2608.12509arXiv quant-phFeroz Ahmed Mian et al.arxiv.org

  • AVA-Encoder for Agent-Native Video Representation Learning

    Researchers developed AVA-Encoder, a new method that enables creative AI agents to learn structured video representations from human films using knowledge graphs and textual-gradient optimization. This significantly improves video generation and reasoning, addressing a limitation in cinematic video production.

    3 findings
    • Creative agents currently cannot effectively learn from high-quality human films, hindering their ability to produce cinematic-grade videos.
    • The Agentic Video Auto-Encoder (AVA-Encoder) is a framework that learns agent-native video representations via agentic auto-encoding.
    • AVA-Encoder transforms a video into a knowledge graph (KG) representation and then reconstructs it back into video.

    arXiv:2608.12313Hugging Face PapersChuyue Li et al.huggingface.coAI Research

  • Embedding Stabilizer Codes and Leakage Correction in Multilevel Quantum Systems

    Leakage beyond the computational subspace is a major source of error in multilevel quantum hardware. We show that any \( [[n,k,d]] \) stabilizer code can be embedded isometrically into a single \(D\)-dimensional system while preserving its complete error-correcting structure. We further derive a necessary and sufficient condition for exact leakage correction, proving that leakage is correctable precisely when it does not distinguish between logical states. These results establish a unified framework for quantum error correction in multilevel quantum systems.

    arXiv:2608.12450arXiv quant-phAli Abu-Nada and Lian-Ao Wuarxiv.org

  • Massive Activations in Hybrid Linear Attention Large Language Models: Pre-Attention Spikes and Inter-Spike Plateaus

    A new study explores 'Massive activations' in hybrid LLMs. It finds 'pre-attention spikes' and 'inter-spike plateaus'. These explain activation behavior.

    3 findings
    • This is the first systematic study of Massive activations (MAs), which are large internal activations, in layer-interleaved Hybrid Linear Attention (HLA) LLMs.
    • Two architecture-aligned morphologies were uncovered: pre-attention spikes (PAS), which are consistent spikes immediately before full attention layers, and inter-spike plateaus (ISP), which are persistent activations through intervening linear attention layers.
    • As full attention layers become denser, successive PAS connect more through ISP, eventually recovering the stable MA morphology seen in full attention LLMs.

    arXiv:2608.12149Hugging Face PapersZunhai Su et al.huggingface.coAI Research

  • Qkabrine AutoML for Quantum Machine Learning

    A new Python package, qkabrine-automl, automates the joint search for architecture, encoding, model type, and hyperparameters in quantum machine learning, integrating trainability diagnostics and NISQ deployment features.

    3 findings
    • Developing quantum machine learning (QML) models typically involves separate manual choices for circuit architecture, data-encoding, model paradigm, and training hyperparameters.
    • The qkabrine-automl Python package treats these elements as a single, jointly searchable configuration space.
    • It integrates trainability diagnostics, including a Data Quantum Fisher Information Metric (DQFIM) estimate and a gradient-magnitude barren-plateau monitor, into its evaluation loop.

    arXiv:2608.18152arXiv quant-phEric Jagwaraarxiv.orgQuantum Computing

  • Dion3 Optimizer for Orthogonal Updates

    Dion3 is a new optimizer that builds upon Muon, significantly reducing orthogonalization and communication overhead. It achieves up to a 6x speedup in optimizer step time while maintaining or improving model loss.

    What you get from it

    Training large AI models often requires substantial time and computational resources. Faster optimizers reduce the cost and accelerate the development cycle for new models.

    5 findings
    • Dion3 is a revision of the Muon optimizer, designed to reduce its cubic-time Newton-Schulz orthogonalization and communication overhead.
    • It incorporates a Gram Newton-Schulz algorithm to lower FLOP costs and CuteDSL kernels to accelerate computations by exploiting symmetry.
    • A megabatching strategy reduces communication overhead, especially when weights are sharded.
    • Dion3 introduces a new update rule that selects only a fraction of the momentum matrix's rows for orthogonalization at each step, further cutting costs.
    • The optimizer matches or improves the loss achieved by Muon, but reduces optimizer step time by up to 6x.

    arXiv:2608.11612Hugging Face PapersNoah Amsel et al.huggingface.coAI Research

  • Quantum State Preparation Research

    Research introduces two new quantum state preparation syntheses, QsiHT Fast Path Real Synthesis and QsiHT Fast Path Complex Synthesis, achieving the lowest CNOT count and competitive error rates on IBM Heron r2 processors.

    What you get from it

    Efficiently preparing quantum states is a fundamental challenge in quantum computing. New methods that reduce the cost of this process can make complex quantum algorithms more feasible on current and future hardware.

    5 findings
    • The thesis studies exact, deterministic preparation of arbitrary dense n-qubit states, a data-loading step in quantum signal and image processing.
    • It derives two syntheses based on the Digital Signal-induced Heap Transform (DsiHT): the QsiHT Fast Path Real Synthesis and the QsiHT Fast Path Complex Synthesis.
    • Both syntheses were benchmarked against ten configurations, including UCR, isometry, multiplexor, Schmidt/SVD, QSD, and heap-transform families.
    • Testing occurred on ibm_fez, ibm_kingston, and ibm_marrakesh, which are three 156-qubit IBM Heron r2 processors.
    • Both syntheses realize the deployed Qiskit StatePreparation floor of 2^n-n-1 CNOTs, undercutting other from-scratch methods on the CNOT axis.

    arXiv:2608.1693719 AugAlexis Gomezarxiv.orgQuantum Computing

  • LLMRouter: Unified Infrastructure for Developing, Evaluating, and Deploying LLM Routers

    Researchers introduced LLMRouter, a unified open-source infrastructure and benchmark (xRouteBench) for developing, evaluating, and deploying LLM routing solutions. It formalizes LLM routing as a sequential decision process, includes over 16 router implementations, and shows learned routers improve performance by 14.6%.

    3 findings
    • LLMRouter is a unified, open-source infrastructure for developing, evaluating, and deploying LLM routers.
    • It formalizes LLM routing as a sequential decision process with five components: context encoders, model encoders, scoring functions, decision rules, and learning signals.
    • The xRouteBench benchmark covers five diverse routing scenarios, including generic LLM, memory-augmented, vision, time-series, and personalized tasks.

    arXiv:2608.06867Hugging Face PapersTao Feng et al.huggingface.coAI Research

  • DarwinX: Evolving Agent Harnesses Through Natural Selection

    DarwinX evolves AI agent 'harnesses' like prompts and tools using natural selection on a population of variants. This method improves agent performance across benchmarks by an average of 17 points, even with frozen language models.

    What you get from it

    AI agents can achieve large performance gains and better generalization by evolving their 'harnesses' without retraining the underlying large language model. This approach turns evaluation compute into durable capability, making agents more adaptable across tasks.

    5 findings
    • DarwinX evolves an AI agent's 'harness', which includes prompts, tools, skills, and control flow, using population selection.
    • It operates with a frozen underlying language model.
    • A 'preserve-and-extend contract' ensures only variants that expand coverage without regressing are admitted.
    • An archive stores alternative lineages for recombination.
    • Fitness is measured by each benchmark's verifier, not gold solutions.

    arXiv:2608.07545Hugging Face PapersYifan Zhang et al.huggingface.coAI Research

The brief is assembled automatically from public feeds and RSS, then checked by the club before it appears. Summaries, findings and every “what you get from it” are quoted from the source or its abstract, never our reading of the work — and papers are dated, credited and abstracted from arXiv itself. Not every story has all of that, and the ones that do are set longer for it. If you want the paper, go and read the paper.

Something wrong, or something missing? Tell us.