Skip to content
SQCC
Sign inJoin

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.

  • LatentLoop for Scaling Looped MoE Models

    LatentLoop improves looped Mixture-of-Experts (MoE) models by compressing the first processing pass and reallocating saved compute to increase model width, leading to better performance and compute savings.

    arXiv:2610.19975Hugging Face Papershuggingface.co

  • Player-1 Framework for AI Game Control

    Researchers introduced Player-1, a post-training framework that teaches AI agents to control games using synthetic experience, significantly improving performance over supervised learning baselines in complex 2D and 3D environments.

    arXiv:2610.20584Hugging Face Papershuggingface.co

  • ATAS Framework for Agentic Video Generation

    Researchers introduced ATAS, a framework that uses scene and event memories to improve spatial and temporal consistency in long-form AI-generated videos, outperforming existing methods across multiple metrics and human ratings.

    3 findings
    • Long-form agentic video generation struggles with maintaining spatial and temporal consistency across non-consecutive shots.
    • The Any Time, Any Space (ATAS) framework compiles screenplays into shot requests with explicit location, camera, and event-time references.
    • ATAS uses a scene memory for static geometry and an event memory for time-indexed dynamic observations to retrieve relevant evidence before generating each shot.

    arXiv:2610.20571Hugging Face Papershuggingface.coAI Research

  • Optimal Disorder Strength for Unitary k-Designs in Spin Chains

    Researchers developed a theoretical framework to identify the optimal disorder strength for creating Haar-like unitary ensembles in quantum spin chains, confirming their findings with simulations and showing a multi-quench protocol improves performance for classical shadow tomography.

    3 findings
    • Quantum simulators with disordered spin Hamiltonians can generate Haar-like unitary ensembles.
    • An optimal intermediate disorder strength exists for approximating a unitary k-design.
    • A perturbative theory for the moment gap, Δ^(k), was developed to quantify deviation from the Haar value.

    arXiv:2610.18954arXiv quant-pharxiv.orgQuantum Computing

  • Measuring Impact of Netflix Personalized Recommendations

    Netflix developed a behavioral model, validated with A/B tests, to measure the causal impact of personalized recommendations. The current system increases engagement by 16% over random recommendations and maintains content diversity.

    What you get from it

    Quantifying the causal effect of personalized recommendations is challenging due to inherent biases. A validated hybrid approach combining modeling and experimentation can accurately measure how much these systems drive user engagement and content diversity.

    5 findings
    • Netflix created a behavioral model to measure the causal impact of personalized recommendations, separating member preferences from recommendation effects.
    • The model uses a low-rank choice model to learn dynamic member preferences and title representations from viewing behavior.
    • It includes a 'recommendation bonus' to capture how recommendations create awareness and provide information.
    • Validation with small-scale A/B tests showed the model's predicted diversion ratios had a 0.86 correlation and 0.73 R-squared with experimental results.
    • Counterfactual analysis revealed the current recommender system increases engagement by 16% compared to random recommendations, 12% over popularity-based, and 4% over a basic matrix factorization model.

    14 Octnetflixtechblog.comAI Research

  • DoSA: Dots Sparse Attention

    DoSA is a lossless sparse attention method that speeds up AI model processing. It conditionally scores keys and rejects candidates early, boosting Indexer performance by up to 2.74 times and time-to-first-token by up to 1.94 times.

    arXiv:2610.13990Hugging Face Papershuggingface.co

  • UltraBERT: Beyond the Limit of BERT Pre-Training

    Researchers introduced UltraBERT, a new encoder model that achieves state-of-the-art performance on both classification (GLUE 91.0) and retrieval (BEIR 47.7) tasks by optimizing pre-training design choices like masking ratios and distillation.

    3 findings
    • Existing bidirectional encoders are under-optimized, with no single model performing competitively across both classification and retrieval tasks.
    • Encoder performance saturates on downstream tasks after a few trillion tokens of pre-training.
    • Lower masking ratios improve classification, while higher ratios improve retrieval, contradicting previous assumptions.

    arXiv:2610.15028Hugging Face Papershuggingface.coAI Research

  • Vela: User Behavior as Supervision Sparks LLM Intelligence

    Researchers introduced Vela, a framework that trains large language models using real user behavior data as supervision, leading to improved performance on tasks requiring an understanding of user preferences and general benchmarks.

    3 findings
    • Large language models (LLMs) need to infer user preferences and future actions from long-term behavior histories.
    • Current models, trained on static text or synthetic data, do not capture how user behavior evolves over time.
    • Vela, a framework, converts real user behavior data into trainable tasks and trajectories for LLM supervision.

    arXiv:2610.14709Hugging Face Papershuggingface.coAI Research

  • Harness^2: Recursive Agent Harnessing for an Open World

    Harness^2 is a framework that enables LLM agents to recursively self-improve their execution harnesses for open-world tasks, learning from edits without ground-truth verifiers.

    What you get from it

    LLM agents can now improve their performance on unseen tasks without human feedback. This allows them to adapt and generalize more effectively in real-world scenarios.

    5 findings
    • LLM agent performance depends on their underlying models and execution harnesses.
    • Recursive self-improvement (RSI) can automatically evolve these harnesses, but struggles with unseen open-world tasks due to a lack of ground-truth verifiers and poor generalization.
    • Harness^2 is a plug-and-play framework that uses two levels of recursion to address these challenges.
    • Task-level recursion refines the current task's harness using a contrastive Propose-Probe-Compose step, which guides edits without needing ground-truth feedback.
    • Harnessing-level recursion extracts experience from these edits to update a continual improvement harness, helping the agent adapt to future tasks.

    arXiv:2610.14224Hugging Face Papershuggingface.coAI Research

  • LLM Thinking Inertia Phenomenon

    A phenomenon termed "Thinking Inertia" has been observed in Large Language Models (LLMs), where they continue to reason even when explicitly instructed not to, impacting control and efficiency.

    arXiv:2610.11765Hugging Face Papershuggingface.co

  • Multi-Stream Attention Residuals (MSAR) Introduction

    Researchers have introduced Multi-Stream Attention Residuals (MSAR), an extension of AttnRes that utilizes multi-head depth attention to improve the performance of large models with minimal additional overhead.

    arXiv:2610.13630Hugging Face Papershuggingface.co

  • AgentSWE Benchmark for AI Coding Agents

    A new benchmark named AgentSWE has been introduced, which reveals that current AI coding agents consistently fail to deliver fully functional software that meets specified requirements, despite often appearing to succeed.

    3 findings
    • Researchers introduced AgentSWE, a benchmark for evaluating AI coding agents across the agent development lifecycle.
    • AgentSWE spans three stages: creation of an agent from natural language, editing of production agent codebases, and optimization towards a measurable target.
    • The benchmark includes 25 tasks, framed as commissions where the builder iterates against illustrative cases, with acceptance exercises on held-out cases.

    arXiv:2610.13361Hugging Face Papershuggingface.coAI Research

  • Terminal Horizon for Agentic Data Synthesis

    Researchers developed Terminal Horizon, an automated engine designed to synthesize long-horizon agentic data from real-world tasks. This data was utilized to fine-tune Qwen3.5-27B, resulting in Terminal Horizon-27B, which demonstrates improved performance on complex agent benchmarks.

    3 findings
    • The paper introduces Terminal Horizon, a fully automated data synthesis engine for long-horizon terminal agents.
    • Terminal Horizon addresses the problem of synthetic environments simplifying key agent behaviors like exploration, implementation, and verification.
    • It starts with real-world professional work, uses independent execution to identify omitted work, and guides environment reconstruction.

    arXiv:2610.13166Hugging Face Papershuggingface.coAI Research

  • GroundingPI Foundation Model Release

    Researchers have introduced GroundingPI, a 4B parameter grounding foundation model that achieves state-of-the-art performance across 34 benchmarks, surpassing larger models like GPT-6 Astra, and significantly enhancing robotic manipulation and autonomous driving capabilities.

    What you get from it

    Models can now perceive and locate objects with greater precision and speed, even in complex environments. This enables more reliable and efficient control for embodied AI systems like robots and autonomous vehicles.

    5 findings
    • GroundingPI is a 4B parameter grounding foundation model designed for precise object localization in complex scenes.
    • It establishes a new state of the art, averaging 73.68% across 34 grounding benchmarks, surpassing GPT-6 Astra's 71.54%.
    • The model improves performance in robotic manipulation and autonomous driving tasks.
    • On RoboTwin2.0, it outperforms other backbones by up to 24.8% in out-of-distribution settings.
    • GroundingPI trained with 50% of demonstrations on RoboCasa-GR1 outperforms baselines trained with 75%.

    arXiv:2609.39601Hugging Face Papershuggingface.coAI Research

  • Deno Team Joins Cloudflare

    The Deno team is joining Cloudflare to integrate `celld` and `workerd`, with the goal of simplifying the self-hosting of Cloudflare Workers and Durable Objects for distributed applications.

    What you get from it

    The Deno team joining Cloudflare aims to simplify building and operating distributed applications on self-managed infrastructure using the Workers programming model. This could make the powerful Durable Objects abstraction more accessible for developers outside Cloudflare's network.

    5 findings
    • The Deno team is joining Cloudflare to simplify self-hosting Workers and Durable Objects.
    • This involves merging `workerd`, Cloudflare's open-source Workers runtime, and `celld`, Deno's open-source distributed Workers implementation.
    • Ryan Dahl, Deno's creator, developed `celld` to enable the distributed Durable Objects model outside Cloudflare's network.
    • Durable Objects provide a distributed singleton abstraction with a local SQLite database, simplifying scaling and state management for networked applications.
    • The combined effort aims to make the Workers programming model a mainstream way to build and operate distributed applications on your own infrastructure.

    9 Octblog.cloudflare.comSoftware Engineering

  • ProgramGym: Scaling Training Environments for Long-Horizon Program Reproduction Agents

    Coding agents are increasingly used to build programs from scratch as software engineers do rather than to resolve issues within a repository or to generate a repository from a given requirements document. However, scalable verifiable training environments that force agents to discover what is required and to decide how the system should be designed remain scarce. Moreover, existing agent workflows also cannot sustain long-horizon development in which the agent explores the reference program before implementation. To address these challenges, we introduce ProgramGym, a framework consisting of Program-Env and Program-Agent for scalable training environments of long-horizon program reproduction agents. Specifically, Program-Env constructs verifiable training environments from repositories through a feedback-driven loop that repairs runtime sandbox builds and the quality of black-box tests. Using this novel pipeline, we obtain 1,648 training environments spanning C, C++, Go, and Rust. Program-Agent combines a role-isolated workflow of analysis, implementation, and review with revision management for compilation checks and failure recovery. Experiments on ProgramBench show that both Program-Env and Program-Agent improve performance on program reproduction. Fine-tuning on trajectories with these constructed environments raises the performance of Qwen3.8-27B from 34.87% to 42.84%. Program-Agent further raises the base and SFT models to 55.71% and 56.25%, both exceeding some flagship models. We will release the ProgramGym to support further research on coding agents that build programs from scratch.

    arXiv:2610.13395Hugging Face PapersChengyi Yang et al.huggingface.co

  • Reflex: Visually Grounded Reactive Humanoid Control

    Researchers developed Reflex, a three-stage AI framework enabling humanoids to perform visually grounded, reactive whole-body control, demonstrated by catching thrown objects with a Unitree G1 robot.

    arXiv:2610.12772Hugging Face PapersTaoyang Jia et al.huggingface.co

  • SoftwareGym: Scaling Computer-Use Agent Training with Diverse Synthesized Software Environments

    SoftwareGym is an automated pipeline that synthesizes diverse, runnable GUI software environments for training computer-use agents (CUAs), improving their performance on benchmarks like OSWorld.

    3 findings
    • Computer-use agents (CUAs) are improving, but synthesizing diverse and reliable software environments at scale remains a challenge.
    • SoftwareGym is an automated pipeline for synthesizing GUI software environments with diverse interfaces, executable functionality, and runtime-verified behavior.
    • It generates structurally diverse designs, realizes them as runnable applications, and uses GUI evaluation with iterative repair to ensure functional reliability.

    arXiv:2610.12611Hugging Face PapersTian Jin et al.huggingface.coAI Research

  • DiffusionART: Efficient, Stable and Scalable Reinforcement of Diffusion Models

    DiffusionART is a new method for fine-tuning diffusion models with reinforcement learning, offering significantly improved efficiency and stability compared to previous approaches.

    What you get from it

    Training diffusion models with reinforcement learning can now be much faster and more stable. This makes it easier to develop high-quality image and video generation systems.

    5 findings
    • Online reinforcement learning (RL) fine-tuning of diffusion models faces efficiency issues with reverse-process methods like Flow-GRPO and instability with forward-process methods like DiffusionNFT.
    • DiffusionART (Diffusion Reinforcement with Aggregation-First Reward Tilting) aggregates each prompt's rewards into a shared reward-improving direction for reverse-process states.
    • It removes a hidden forward-reverse consistency force found in NFT-style objectives and adds a trust-region-masked second update for drift-adaptive restoration without needing Exponential Moving Average (EMA) tuning.
    • The method requires no differentiable reward model, works effectively with fewer samples per prompt, and supports pairwise reward models.
    • On Hunyuan Video-13B, DiffusionART achieves the same late-stage mean reward levels as DiffusionNFT and Flow-GRPO-Fast with 12.02 and 11.72 times fewer GPU hours, respectively.

    arXiv:2610.12572Hugging Face PapersZhenglin Wan et al.huggingface.coAI Research

  • Engineering Ergodicity-Breaking and Stable Emergent Conservation Laws in a Generic Quantum Many-body System Under Strong Periodic Drive

    A periodically driven quantum many-body system is believed to approach a locally infinite-temperature-like state as dictated by the ergodicity hypothesis, when all the couplings/parameters are of the same order. When the drive is stronger than a threshold value, a large class of systems is shown to break this mandate, exhibiting Dynamical Freezing (DF): ergodicity is broken due to the emergence of local operators with permanent memory of their initial values (PRX 11, 021008, 2021). However, an exception to DF has recently been noted in a particular setting (PRX 16, 011041, 2026). Here we identify the condition for ergodicity-breaking in a system under strong drive, and show how to stabilize DF by exact, local engineering. We further show that the Floquet Hamiltonian $H_{eff} = (i/T)\ln{[U(T)]}$ is non-local when ordered by spin-flips, in a non-DF system, while it turns quasi-local under our simple, local DF engineering that keeps the Hamiltonian $H(t)$ still non-integrable at each $t$. This perplexing non-universality is absent in equilibrium, where non-integrability is believed to imply ergodicity and thermalization universally. Our work paves a way to engineering novel ergodicity-broken interacting Floquet phases of matter in various quantum platforms, and also endows DF-assisted quantum error mitigation and state stabilization techniques (see Nature 653, 1027 (2026)) with unprecedented versatility and flexibility.

    arXiv:2610.12551arXiv quant-phSagnik Chaudhuri et al.arxiv.org

  • Exact solutions of non-Markovian dissipative Ising models

    Applications of quantum devices and simulators require an accurate understanding of their dynamics as well as the effects of environmental noise. While some sources of noise can be modeled as Markovian dissipation, coupling to resonators or vibrational modes often produces non-Markovian dynamics as well, making it difficult to characterize the system, particularly for interacting many-body systems. Here, we identify a broad class of exactly solvable many-body quantum systems with both non-Markovian dynamics and Markovian dissipation. We consider a prototypical model of spins coupled to lossy bosonic modes that mediate Ising-type interactions. We exclude a transverse field but allow local spin dissipation that need not commute with the interactions, including dephasing, relaxation, pumping, and even processes that generate coherences. Using an exact stochastic representation, we decouple the many-body dynamics into the evolution of individual spins, yielding exact spin observables and correlations without adiabatically eliminating the bosons. In the Markovian limit, our solutions describe Ising models with correlated dephasing and nonreciprocal interactions. An example of spin squeezing shows that both the achievable squeezing and its optimal preparation time depend strongly on the interplay between non-Markovian dynamics and Markovian dissipation.

    arXiv:2610.12546arXiv quant-phMohammad Maghrebiarxiv.org

  • AgentGarten: Code Worlds for Evolving Agents

    AgentGarten is a framework that combines simulators and neural rendering to create real-time interactive environments for AI agents, enabling them to learn complex tasks in just 4 rounds of experience, a substantial improvement over traditional methods.

    What you get from it

    This framework allows AI agents to learn complex behaviors in interactive environments with significantly less training data. It could accelerate the development of more capable and adaptable AI systems.

    5 findings
    • AgentGarten is a framework for building real-time interactive environments where AI agents can evolve.
    • It couples traditional simulators and game engines with a shared neural renderer to generate realistic visual observations.
    • The neural renderer uses a technique called Adversarial Forcing, which makes history prefilling differentiable and adds real-data adversarial supervision.
    • Agents perceive the world visually, interact in real-time, and improve by distilling experience into playbooks that subsequent agents inherit.
    • Empirical studies show agents learn tasks in 4 rounds, a significant efficiency gain compared to millions of rounds for conventional reinforcement learning.

    arXiv:2610.12374Hugging Face PapersJiawei Chi et al.huggingface.coAI Research

  • TokenRouter: Efficient Serving System for Token-Level LLM Routing

    Researchers developed TokenRouter, a new serving system for token-level LLM routing that achieves 2 to 64 times higher decoding throughput than current systems by using a request-centric, model-centric approach and delayed-batching schedulers.

    3 findings
    • Token-level routing for large language models (LLMs) offers efficiency and quality gains over coarser-grained methods.
    • Existing LLM serving systems struggle with token-level routing due to single-model assumptions, step desynchronization, and batch admission delays.
    • TokenRouter is a new system designed for efficient token-level LLM inference.

    arXiv:2610.12242Hugging Face PapersTianyu Fu et al.huggingface.coAI Research

  • SuperNav: An Agentic Navigation System for Any Task in Any Scene

    General-purpose service robots need navigation systems that can handle diverse human requests in unfamiliar environments, combining task generality with scene generality. Some existing methods fine-tune multimodal large language models (MLLMs) to predict navigation actions, making their behavior dependent on the coverage of navigation training data and potentially limiting generalization to new requests and environments. Our key insight is to let the MLLM focus on interpreting requests, understanding scenes, and making decisions while preserving its general-purpose capabilities and delegating motion execution to navigation tools. To realize this idea, we introduce SuperNav, which equips a pretrained MLLM with a specialized agent harness without navigation-specific fine-tuning of the MLLM. Our harness supports these decisions with Navigation Skills, agent-oriented Tools for physical interaction, and task-progress and context management. A unified visual-point interface connects decision-making to motion by allowing the model to specify destinations directly in images and revise its decisions from execution feedback. Together, these components support sustained navigation across different task requirements and environments. SuperNav outperforms four evaluated baselines on instance-level, multi-object, and demand-driven tasks. Category-level evaluation on HM3D and deployment on a real quadruped robot further demonstrate its applicability across environments. Project Page: https://zju3dv.github.io/SuperNav/

    arXiv:2610.12126Hugging Face PapersJinkai Zhang et al.huggingface.co

  • On linearity or non-linearity in machine learning for quantum chaotic dynamics

    A study found that a simple linear machine learning model, DLinear, accurately forecasts chaotic quantum many-body dynamics across various regimes, outperforming a complex Transformer model. This suggests simpler models can predict complex quantum behavior.

    3 findings
    • Simulating chaotic quantum many-body dynamics is a major computational challenge for classical methods.
    • Researchers explored machine learning for time-series forecasting of quantum dynamics.
    • They used a few-qubit PXP chain, realizable with Rydberg-atom arrays, as a benchmark system.

    arXiv:2610.10697arXiv quant-phFrancesco Perciavalle et al.arxiv.orgAI Research

  • Quantum non-Markovian response spectra

    Researchers developed a response matrix method to characterize hidden environmental memory in quantum systems, which causes correlated errors and is hard to measure directly.

    arXiv:2610.10684arXiv quant-phAaron Sander et al.arxiv.org

  • Long-WAM: Scaling Context for Robot Control

    Long-WAM is a new model-system framework designed to scale the context of world-action models for real-time robot control. It uses autoregressive pretraining to process longer visual histories, boosting robot task success rates and enabling deployment on various hardware.

    What you get from it

    Real-time robot control can now use longer visual histories without action delays. This improves task success rates and enables more complex, dynamic manipulations.

    5 findings
    • Long-WAM is a model-system framework for scaling causal world-action models under real-time control constraints.
    • Autoregressively pretrained video foundation models are key to effectively using longer visual histories for robot control.
    • Increasing context from 0.0 to 19.2 seconds on RoboCasa GR-1 raised success from 63.3% to 78.7%.
    • Long-WAM achieved the best results among compared methods on LIBERO-Long, RoboTwin2.0, and DOMINO benchmarks.
    • It deploys on RTX 5090, DGX Spark, and Jetson AGX Thor, with each action chunk taking 107.4ms on RTX 5090.

    arXiv:2610.10528Hugging Face PapersWei Huang et al.huggingface.coAI Research

  • Everything we launched during Birthday Week 2026

    Cloudflare announced 46 new features during Birthday Week 2026, including an agentic CLI, an open-source pipeline for SDKs, a serverless CMS called EmDash, post-quantum cryptography tools, and expanded AI and developer platform capabilities.

    Cloudflare Blogblog.cloudflare.com

  • High-Fidelity Inter-Species Rydberg Gates

    Researchers developed a model to suppress laser crosstalk in dual-species neutral atom quantum gates. They demonstrated that gate fidelities above 99.8% are achievable in an $^{85}$Rb/$^{87}$Rb system through careful gate design.

    arXiv:2610.08926arXiv quant-phShuan Wang and Kevin Singharxiv.org

  • nanoMuse: Open-Source Personal Agent for All Devices

    Assistants from 2011 answered and waited, and agents from 2023 did a task and stopped. In September 2026 Meta's Muse showed an agent for one person, with accounts, devices, memory and a conversation that lasts, closed, in a vendor's cloud, in one country. Such an agent is expected to act on a person's accounts and devices, remember them across weeks, speak first when it is worth it, and answer for what it did. It is a kind of software, not a model, and until now had no open counterpart. This report defines the personal agent in five questions and three horizons. It reads how Muse is built from Meta's public record and a copy of its production prompt, each statement marked by its source. It then presents nanoMuse, the open-source counterpart under the GPL-3.0, one agent on every device a person owns, with hands on the phone's screen and the computer's. They share one conversation over a relay anyone can run; every action goes through a Sentinel, memory is files the person can read, and the model is their choice. Its size and cost are given as estimates. What is open, memory with provenance, an evaluation suite for the hands and an open model for them, is set out as a roadmap.

    arXiv:2610.08699Hugging Face PapersGuangyi Liu et al.huggingface.co

  • Recursive Game Creator AI Agent Framework

    Researchers developed Recursive Game Creator, an AI agent framework that iteratively refines game prototypes into engaging games using a four-component system. This framework achieves state-of-the-art performance on game creation benchmarks.

    arXiv:2610.08621Hugging Face PapersJiajun Chen et al.huggingface.co

  • Quantum Readout Complexity for Linear Functionals

    Quantum numerical algorithms often encode a positive solution $q$ in the amplitudes of a quantum state, while the required scientific output is a normalized linear functional of $q$. Basis measurements produce probabilities determined by squared solution values, while the functional uses the solution values themselves. We study the complexity of this output-interface problem for $F_φ(q)=\intφq\,\mathrm dν/\int q\,\mathrm dν$ and identify a target-dependent susceptibility $χ_φ$ that governs the amount of copies of the solution state needed. For a fixed positive reference state and a neighborhood around it, the optimal number of unknown-state copies is $Θ\!\left(χ_0\varepsilon^{-2}\log(1/β)\right)$, where $\varepsilon$ is the absolute error tolerance and $β$ is the allowed failure probability. We show that two fixed single-copy measurements attain the same rate. We then analyze how this baseline changes when the interface retains classical information, uses adaptive allocation, supplies coherent access, or restricts the allowed solution family. Therefore $χ_φ$ quantifies a central cost of using a solution encoded in an $L^2$-normalized quantum state to estimate an $L^1$-normalized linear observable.

    arXiv:2610.08890arXiv quant-phRoi Vence-Nogueiraarxiv.org

  • Rethinking Cross-Tokenizer On-Policy Distillation: From Alignment Coverage to Supervision Reliability

    On-Policy Distillation (OPD) trains a student on its own generations using teacher feedback. With different tokenizers, comparing teacher and student predictions requires alignment at both sequence and vocabulary levels. In this paper, we examine whether expanding this alignment coverage improves learning. Across three heterogeneous teacher--student pairs on mathematical reasoning and code generation, strict 1:1 groups already cover most student-generated tokens despite substantial vocabulary mismatch. On responses sampled from the students before distillation, the shared vocabulary retains nearly all teacher and student probability mass at strictly aligned positions on average. Restricting reverse KL to a student-selected top-16 subset of the shared vocabulary at each strict position achieves accuracy comparable to full shared-vocabulary OPD, outperforming the evaluated cross-tokenizer baselines. Adding mean squared error supervision on span log-probabilities in mismatch groups gives complete supervision coverage, yet reduces accuracy. At checkpoints from training with only the strict loss, the span gradients show weak or negative directional agreement with the strict gradients and grow in magnitude relative to them. These diagnostics may help explain the accuracy drop from adding span supervision. Our findings motivate a shift from maximizing alignment coverage to prioritizing supervision reliability: compact supervision at strict positions can be more effective than broader coverage that introduces weakly aligned or conflicting training signals.

    arXiv:2610.08448Hugging Face PapersBingxi Hou et al.huggingface.co

  • Learn2Play Bench: How Well Do LLM Agents Learn from Experience in Unfamiliar Environments?

    Learning from experience is essential for LLM agents to adapt to unfamiliar and dynmaic environments. Evaluating this ability is therefore important for understanding how effectively agents acquire and use new knowledge. Existing benchmarks have sought to evaluate this ability, but they primarily evaluate tasks whose rules are provided in the instructions or already familiar to pretrained models, making it difficult to distinguish learning from interactions from reasoning with existing knowledge. To address this, we introduce Learn2Play Bench, a benchmark of newly designed text-based games, whose rules are novel or counterintuitive, requiring agents to acquire knowledge through interaction rather than rely solely on pretrained knowledge. These games provide reproducible feedback and automatic scoring, enabling controlled evaluation of learning across repeated attempts. We also vary game instances to test whether agents can apply what they have learned to new situations. Therefore, we evaluate how backbone models, self-evolving methods, and agent harnesses affect agents' learning ability, revealing three findings: (1) Experience retention: Retaining complete records of actions and feedback can support more effective learning than summarizing these experiences into rules or strategies. (2) Human agent gap: Top-performing human players achieve higher peak scores than the evaluated agents. Human explore more varied strategies, and repeat actions less. (3) Harness matters: With the backbone fixed, changing the harness can improve performance while reducing estimated inference cost. Together, these findings provide insights into how LLM agents learn from experience and suggest directions for future work to improve their learning ability. Project website: https://liushiliushi.github.io/learn2play-bench-website/

    arXiv:2610.08215Hugging Face PapersYibo Li et al.huggingface.co

  • VisionWeave: Weaving Elastic Visual Representations as a Native Capability of MLLMs

    VisionWeave introduces a new method for Multimodal Large Language Models (MLLMs) to adaptively process visual information, saving 43% of visual tokens and increasing throughput by 2.3x while maintaining 98.9% performance.

    3 findings
    • Multimodal large language models (MLLMs) currently incur high costs by encoding visual inputs into dense, fixed-size patch tokens.
    • Visual information is unevenly distributed, requiring varied levels of detail.
    • VisionWeave introduces "elastic visual representation weaving," a native MLLM capability to learn where and at what granularity to allocate visual representations.

    arXiv:2610.07987Hugging Face PapersYuan Feng et al.huggingface.coAI Research

  • TRACE: Rollout-Guided Quantization-Aware Training for FP4 Reinforcement Learning of MoE Language Models

    TRACE is a new FP4 quantization framework for training Mixture-of-Experts (MoE) language models with reinforcement learning. It achieves up to 5.4x speedup and BF16-comparable performance. This method aligns quantization decisions between training and rollout.

    3 findings
    • Reinforcement learning (RL) for post-training large language models (LLMs) incurs substantial computation and memory overhead during rollout generation.
    • Existing FP4 RL methods optimize quantization accuracy on training and rollout paths independently, leading to discrepancies.
    • TRACE (Train-Rollout Quantization Alignment via Compact GuidancE) is an FP4 quantization framework for RL training of Mixture-of-Experts (MoE) language models.

    arXiv:2610.07767Hugging Face PapersXin Wang et al.huggingface.coAI Research

  • From Evidence to Action: How Tool-Using Agents Fail

    Tool-using AI agents often act without establishing sufficient evidence, even when their final outcomes are correct. Researchers introduced SafeActBench, a new benchmark, to analyze these 'evidence-to-action' chain breaks across various agent configurations and workflows.

    arXiv:2610.07753Hugging Face PapersHongzhan Lin et al.huggingface.co

  • Towards Looped Models Done Right, Part II: Rethinking at Fixed Points

    This paper introduces methods to make looped language models more efficient by using fixed points to optimize training, decoding, prefill, and reinforcement learning. It achieves faster operations and better memory use, with a 1.6B parameter model performing comparably to larger models using 3x less memory.

    arXiv:2610.06833Hugging Face PapersBenhao Huang et al.huggingface.co

  • From Traces to Agentic Worlds: Agentic Language World Models for Interactive Environment Simulation

    Realistic environment replicas are increasingly valuable for training and evaluating LLM agents, yet the original systems may be inaccessible or impractical to reproduce. We explore agentic language world modeling: rather than rebuilding an executable environment, a world model agent serves as the environment for a task agent and supports faithful and stateful simulation. We instantiate this paradigm with Trace2Env, a learning-free framework for settings where the original system is unavailable but historical interaction traces remain accessible. Trace2Env reconstructs these traces into a reusable environment worldbook containing environment schemas, grounded evidence, and induced behavioral knowledge. At runtime, the world model agent actively consults the worldbook together with persistent episodic state to infer each action's observation and lasting state effects. Across nine environments, Trace2Env improves both next-observation fidelity and long-horizon interaction consistency over conventional prompt-based LWMs. In multi-turn interaction, task agent actions generated against Trace2Env remain valid more often when replayed in the real environment, indicating that its simulated dynamics better preserve the consequences of earlier actions across successive turns. These results establish agentic language world modeling as an alternative direction for building realistic environment replicas without reconstructing the original executable system.

    arXiv:2610.06100Hugging Face PapersQuanyu Long et al.huggingface.co

  • Kandinsky 6.0 Video: Foundation Models for Synchronized Video and Audio Generation

    Kandinsky 6.0 Video introduces open-source foundation diffusion models, Lite (3B parameters) and Pro (29B parameters), for generating 5-second synchronized text-to-audio-video and image-to-audio-video clips with 44kHz audio and Full-HD resolution, including lip-sync.

    What you get from it

    Advanced synchronized audio-video generation, including lip-sync, is now openly available. This provides new tools for creating realistic multimodal content.

    5 findings
    • Kandinsky 6.0 Video is a family of foundation diffusion models for synchronized text-to-audio-video (T2AV) and image-to-audio-video (I2AV) generation.
    • The family includes Kandinsky 6.0 Video Lite (3B parameters) and Pro (29B parameters).
    • Models generate 5-second video clips with synchronized 44kHz audio, including lip-sync, and a super-resolution model upscales output to 1920x1080.
    • The architecture uses a dual-stream CrossDiT with bidirectional cross-attention for temporal and semantic alignment between video and a newly trained audio stream.
    • Reinforcement learning post-training reduced the Pro model's speech Word Error Rate (WER) by 47%.

    arXiv:2610.05608Hugging Face PapersTeam Kandinsky et al.huggingface.coAI Research

  • ASCENT: Online Test-Time Training of Long-Horizon Agents via Self-Distillation of Verified Experience

    ASCENT is a new method for Online Agentic Test-Time Training (OaTTT) that allows LLM agents to self-distill verified experience into persistent LoRA weights during deployment. It improves task success by over 22 points.

    3 findings
    • Large language model (LLM) agents solve long-horizon tasks with a single verification signal at termination.
    • Online Agentic Test-Time Training (OaTTT) trains an LLM's weights on its own execution trajectories during deployment.
    • This method uses only the verified trajectory as a learning signal.

    arXiv:2610.05303Hugging Face PapersHaodong Lu and Dong Gonghuggingface.coAI Research

  • Exact Unitary Synthesis: Simultaneously Optimal in Size and Depth-Ancilla Tradeoff

    Researchers developed a method for exact unitary synthesis that achieves optimal circuit size and depth-ancilla tradeoff for quantum circuits.

    3 findings
    • Unitary synthesis is fundamental for universal quantum computation, translating unitary transformations into elementary gate circuits.
    • The new method exactly implements any n-qubit unitary using O(4^n) CNOT and single-qubit gates.
    • It achieves O(n+4^n/(n+m)) layers, where 'm' is the budget of ancilla qubits.

    arXiv:2610.07010arXiv quant-phPei Yuan et al.arxiv.orgQuantum Computing

  • EVISKILL: Grounding Skill Evolution in Replayable Evidence

    Continual skill evolution enables LLM agents to accumulate and refine reusable procedural knowledge from interaction experience without updating model parameters. Its effectiveness depends on determining not only what to change, but also why a change is justified and when it should become persistent guidance. However, existing experience-driven methods can lose the behavioral evidence and task contexts supporting edits. Moreover, a global validation outcome provides an incomplete judgment of its constituent changes: locally supported corrections may be discarded with a rejected revision, while evidence may require further experience to inform useful updates. To this end, we introduce EVISKILL, an evidence-driven framework that organizes execution observations into Replayable Evidence Cards and synthesizes edits with explicit links to their supporting contexts. Targeted replay verifies these edits through re-execution and provides feedback for correction. Across epochs, EVISKILL preserves evidence and provisionally retains supported edits for further refinement, while global validation governs their incorporation into the final skill. Experiments on three interactive benchmarks across six LLM backbones demonstrate the effectiveness of this approach.

    arXiv:2610.05030Hugging Face PapersYan Zhou et al.huggingface.co

  • Simultaneous Circuit Tests for Finite Reversible Models of Quantum Control

    A finite reversible model can approximate individual quantum gates while retaining freedom in its transition tables and preparation contexts. We formulate the simultaneous certification problem in which one stationary controller and one preparation law must reproduce an entire collection of circuit probabilities. For a specified frame catalogue, accuracy graph and calibration constraints, we give an exact shared-table feasibility formulation and distinguish certificates for a fixed table from certificates for the complete model class. A four-frame qubit example exhibits a strict separation between fitting two circuits individually and fitting them jointly. A larger exact benchmark uses 464 frames and a phase symmetry that removes preparation-prior dependence. We independently reconstruct its admissible graph and certify a sharp repeated-circuit horizon: one controller stays within probability error 1/3 through 192 blocks, whereas every allowed controller violates that threshold by block 193. A separate controller reproduces all 8,190 nonempty Hadamard-phase words of length at most twelve within the same tolerance. We also derive a finite-window spectral obstruction and finite-shot confidence tests, illustrated by a reproducible sampling experiment. The conclusions concern the stated finite reversible model classes; they do not establish a universal memory bound or a discrepancy for physical quantum hardware.

    arXiv:2610.06984arXiv quant-phJosef Bruzzesearxiv.org

  • Neural Correlation Learning for Quantum-Enhanced Sensing with Time-Independently Driven Rydberg Atom Arrays

    Researchers developed a neural correlation learning framework for quantum-enhanced sensing using Rydberg atom arrays. This method achieves quantum sensitivity beyond the standard quantum limit without needing engineered entangled states.

    What you get from it

    Achieving quantum-enhanced sensing without complex entangled states simplifies hardware requirements. This opens a path to more scalable and practical quantum sensors.

    5 findings
    • Quantum-enhanced sensing typically relies on preparing specific entangled states, which is challenging in scalable many-body systems due to design and noise.
    • This work shows that simple product states, evolving under native time-independent Hamiltonian dynamics in Rydberg atom arrays, generate complex many-body correlations.
    • These correlations can enable quantum-enhanced parameter estimation, potentially reaching the Heisenberg limit.
    • A neural correlation learning framework, combining Bayesian inference with neural networks, extracts metrological information from spatial correlations of measurement patterns.
    • Operating in a two-stage calibration-and-sensing protocol, the framework numerically demonstrates effective extraction of many-body correlations to saturate the Cramér-Rao bound.

    arXiv:2610.069617 OctTao Zhang et al.arxiv.orgQuantum Computing

  • R-Quest: Sustaining Self-Evolution in Reasoning Models

    Self-evolving AI models often suffer performance collapse due to invalid and repetitive self-generated questions. Researchers introduced R-Quest, a method that uses validity and novelty feedback to guide self-evolution, achieving superior and stable performance across various reasoning tasks.

    3 findings
    • Self-evolving reasoning models, which learn from their own generated questions, can experience performance collapse during repeated self-training.
    • Analysis identified two core problems: invalid questions becoming prevalent and existing diversity controls failing to detect mathematically equivalent but lexically different questions.
    • Invalid questions increase across training rounds, and answer-consistency filtering exacerbates their proportion in training data.

    arXiv:2610.04299Hugging Face PapersJinyuan Li et al.huggingface.coAI Research

  • ALoDLM: Adaptively Looped Diffusion Language Models

    Researchers introduced ALoDLM, a Diffusion Language Model that uses token-adaptive latent recurrence to improve text generation quality while keeping fast parallel decoding. It outperforms other DLMs and autoregressive baselines on eleven benchmarks.

    3 findings
    • Diffusion Language Models (DLMs) enable fast parallel text generation but have a quality gap compared to autoregressive (AR) models.
    • This quality gap is attributed to existing DLMs applying uniform computational depth to all unknown tokens, regardless of prediction difficulty.
    • ALoDLM introduces token-adaptive latent recurrence, which iteratively refines latent representations and allocates computation based on token difficulty.

    arXiv:2610.04198Hugging Face PapersLiancheng Fang et al.huggingface.coAI Research

  • Cloudflare Streamline for Custom Video Pipelines

    Cloudflare launched Streamline, a developer playground demonstrating how to build custom video pipelines for dynamic annotations or subtitles using Cloudflare Stream, Workers, and Containers.

    Cloudflare Blogblog.cloudflare.com

  • Discussion on AI Product Token Billing Models

    This article discusses the challenges of token billing for AI products, arguing that it often misrepresents customer value, and suggests unified credits and output-based pricing as better alternatives, especially with the rise of agentic commerce.

    Stripe Engineeringstripe.com

  • World Embedding Benchmark for Video Representations

    Researchers introduced the World Embedding Benchmark, a dataset comprising 8,000 simulated video-text cases across 80 physical families. This benchmark evaluates how effectively video representations capture physical information, revealing a trade-off between cross-modal alignment and quantitative physical information recoverability in pre-trained models. However, it also demonstrated that physics-adapted models can improve retrieval-augmented video generation.

    arXiv:2610.03632Hugging Face PapersYiqi Liu et al.huggingface.co

  • DuoMatching: Joint-Marginal Distribution Matching for Few-Step Video Generation

    Streaming video generation has benefited from distribution matching distillation (DMD), which matches the joint distribution of video frames to a video teacher's approximation of the real video distribution. Although this joint matching mitigates drift during autoregressive rollouts, limitations remain in visual quality and semantic alignment. To address these limitations, we propose DuoMatching, a distribution matching framework that approximates the real video distribution through a unified joint-marginal formulation. On top of existing joint matching formulations, the additional marginal matching objective provides dedicated frame-level supervision from an image generator, transferring complementary visual and semantic priors from it. To apply this frame-level supervision in video generation, we introduce LatentBridge to resolve the latent representation mismatch between the video student and the image teacher. Latent Variation Sampling further distributes such frame-level supervision across distinct temporal segments, reducing redundancy. Experiments demonstrate that DuoMatching improves visual quality, composition, and semantic alignment while largely preserving motion dynamics. Human evaluations show overall preference rates above 80% against all evaluated baselines. The project page is available at https://johnzhan2023.github.io/DuoMatching/.

    arXiv:2610.03543Hugging Face PapersJiahao Zhan et al.huggingface.co

  • LLM Agent Source Preference Bias

    As LLM agents decide on users' behalf which product to buy, which hotel to book, or which paper to cite, a preference for items from certain sources (the sites or services they come from) shapes what users receive and which sources are selected. We study source preference in end-to-end search with 12 agent models across three domains. Comparing items from different sources that satisfy the same requirements at the same position, we find that each model prefers some sources and avoids others in every domain, largely agreeing on which. This preference can outweigh how well items satisfy the request: an item satisfying one requirement fewer is selected about two-thirds of the time when it comes from a preferred source and the better one from a dispreferred source, but almost never in the reverse case. The information identifying an item's source affects selection by itself: hiding it weakens the preference, and relabeling an item with a preferred source raises its selection rate. We test two routes to this preference: training that rewards better items can make a source a shortcut for requirement satisfaction, and missing information can trigger preconceptions about the source. Supplying missing information or a prompt countering these preconceptions reduces source preference.

    arXiv:2610.03195Hugging Face PapersJonghyun Song et al.huggingface.co

  • Hierarchical Continuous Diffusion Language Models Introduction

    Researchers introduced Hierarchical Continuous Diffusion Language Models (HC-DLM), a new architecture that couples discrete token generation with a continuous latent trajectory. This method preserves token dependencies during parallel denoising, outperforming existing discrete and continuous diffusion baselines on tasks like Sudoku, Countdown, and LM1B.

    arXiv:2610.02193Hugging Face Papershuggingface.co

  • Rhetorical Robustness and SciCore for Trustworthy AI Reviewers

    Researchers introduced Rhetorical Robustness, a new metric for AI reviewers, and a benchmark called RobustReview. They found existing AI reviewers are susceptible to rhetorical changes and developed SciCore, a dual-branch reviewer that combines full-manuscript judgment with a structured science core, achieving better stability and discrimination.

    arXiv:2609.39027Hugging Face Papershuggingface.co

  • David Heinemeier Hansson Sparks 'Death of Coding by Hand' Debate

    David Heinemeier Hansson, creator of Ruby on Rails, declared that 37signals has stopped writing code by hand, using AI agents for most development. This sparked a debate about the future of software engineering, with examples like Uber Eats showing declining quality and Craft Docs pivoting to AI-driven workflows.

    3 findings
    • David Heinemeier Hansson (DHH), creator of Ruby on Rails, stated that his company, 37signals, has ceased writing code by hand, making it an exceptional state.
    • DHH considers November 24th, 2025, the release date of Opus 4.5, as the inflection point for the age of agents.
    • 37signals is now embracing native mobile apps and moving backend services to Rust, citing AI agents' ability to write good enough Rust code.

    Pragmatic Engineerblog.pragmaticengineer.comSoftware Engineering

  • Lie-Wedge Stratification for Qubit Control

    Researchers introduced Lie wedges, a mathematical framework that better describes quantum control in open systems than Lie algebras. They used this to classify single-qubit systems into thirteen strata and determine conditions for stabilizing pure states.

    3 findings
    • Lie wedges, a new mathematical framework, describe how control generators combine in open quantum systems. They retain information about reversible control and irreversible dissipation, unlike Lie algebras.
    • The study focuses on drift-free single-qubit systems with bidirectional Hamiltonian controls and one independently actuated dissipator.
    • These systems fall into thirteen structural strata, defined by their generated algebra and Lie-wedge geometry.

    arXiv:2610.00046arXiv quant-pharxiv.orgQuantum Computing

  • Harnessing Long-Horizon LLM Agents with Explicit Belief States

    Researchers introduced PoS, an inference-time framework that gives LLM agents explicit belief states to track world state and task requirements, preventing them from getting stuck in 'Belief Trapping' and significantly improving performance on long-horizon tasks.

    What you get from it

    AI agents can now reliably tackle more complex, multi-step tasks. This improves their ability to maintain context and make meaningful progress over long interactions.

    5 findings
    • Large language model (LLM) agents struggle to maintain a coherent understanding of the current world state over long interactions.
    • The PoS (Progression of States) framework is an inference-time method that constructs and continually maintains explicit belief states for agents.
    • Each belief combines an estimate of the current world state with unresolved task requirements, making explicit what the agent still needs to learn and accomplish.
    • PoS validates belief consistency and monitors task progress to detect Belief Trapping, a state where the agent acts without making meaningful progress.
    • When Belief Trapping occurs, PoS diagnoses the pattern and unresolved requirement, then composes targeted recovery constraints.

    arXiv:2610.01415Hugging Face Papershuggingface.coAI Research

  • Counterdiabatic Quantum Circuits Theory

    Researchers developed a Hamiltonian-free theory for counterdiabatic driving. This advancement enables faster quantum control for Floquet systems and quantum circuits without the necessity of a local effective Hamiltonian.

    3 findings
    • Counterdiabatic driving, a method for fast quantum control, typically requires a Hamiltonian generator.
    • This research introduces a Hamiltonian-free theory for the adiabatic gauge potential (AGP) for parameter-dependent unitary operators.
    • The exact and regularized AGP versions come from a Hermitian, banded linear system.

    arXiv:2610.02312arXiv quant-phAnkit Wenju Shrestha et al.arxiv.orgQuantum Computing

  • Homological Thresholds in Quantum Error-Correcting Codes

    Random projective measurements are an effective tool to model the interlinked phenomenology of monitoring, inferring, and decoding highly entangled states such as quantum error-correcting codes. At the same time, the study of such monitored dynamics has opened a new arena for critical phenomena, renewing longstanding questions in disordered statistical mechanics, including percolation, random-bond models, and Nishimori physics. Here we investigate the robustness of stabilizer quantum error-correcting codes under a single round of independently sampled single-qubit Pauli measurements and demonstrate that their threshold behavior falls into two broad categories - geometric and homological percolation transitions, with the toric code and color code being the principal examples for each class. We establish the distinct character of these two threshold theories by analyzing the logical-support criteria of these codes and by providing numerical estimates of their respective critical exponents obtained from optimized stabilizer simulations of large-scale systems with over one million qubits. Our framework extends to non-CSS, subsystem, higher-dimensional, and Floquet stabilizer codes and allows us to determine their respective threshold phase diagrams. We exemplify how the latter provide quantitative and qualitative guidance for the computationally more demanding problems of tunable learning and maximum-likelihood decoding. As a practical application of our framework, we discuss the robustness of the gross code.

    arXiv:2610.02310arXiv quant-phYoshito Watanabe et al.arxiv.org

  • Premonoidal Semantics for Fermionic Quantum Computing

    This paper introduces a categorical framework and diagrammatic notation specifically for fermionic quantum computing. It addresses the unique compositional structure of local fermionic mode systems, defines pronaps for reasoning, and provides new normal forms for circuit fragments.

    arXiv:2610.02287arXiv quant-phThomas Perez and Titouan Carettearxiv.org

  • OneStreamer AI Model for Streaming Video Interaction

    OneStreamer is a new 4B AI model and 1M record dataset designed to enable large language models to perceive, remember, and proactively respond to streaming video, achieving state-of-the-art results on eight benchmarks.

    What you get from it

    AI models can now process streaming video by remembering past events and proactively responding to new information. This allows for more natural and timely interaction with dynamic visual content.

    5 findings
    • OneStreamer unifies perception, memory, and proactive response for streaming video interaction.
    • It learns to record query-independent evidence and respond when sufficient evidence is available.
    • Proactive Hierarchical Caption Memory (PHCM) generates time-grounded local-detail captions and event summaries. This preserves evidence for later questions without revisiting historical visual features.
    • Proactive State Transition Learning (PSTL) reduces repeated waiting states. It supervises only 27.5% of annotated state tokens.
    • A 4B OneStreamer model achieves the best results across all eight evaluated streaming video understanding benchmarks.

    arXiv:2610.01762Hugging Face PapersXiangyu Zeng et al.huggingface.coAI Research

  • Sharpening Tax for Post-Training LLMs

    Researchers introduced Sharpening Tax to measure the trade-off between single-shot accuracy and solution coverage in post-training LLMs, and Posterior-Tempered Group Sampling to mitigate this, improving both metrics.

    3 findings
    • Reinforcement learning post-training of large language models (LLMs) is often hypothesized to sharpen existing behaviors, improving single-shot accuracy but reducing solution coverage.
    • Pre-trained LLMs, equipped with a light inference harness, can serve as capable agents.
    • These pre-trained models often surpass their post-trained counterparts in solution coverage (pass@K) given a sufficient test-time budget, despite having lower single-shot accuracy (pass@1).

    arXiv:2610.01509Hugging Face PapersChangdae Oh et al.huggingface.coAI Research

  • Agents Revolt

    Meta launched Muse, a personal AI agent that quickly gained 2.8 million downloads, while startup Spear Street Technology's Instinct app also shows promise in automating financial tasks like finding cheaper car insurance. These agents threaten the traditional profit models of the finance industry by reducing customer friction.

    3 findings
    • Meta launched Muse, a personal AI agent designed to help with goals and tasks like sending emails or booking travel.
    • Muse recorded 2.8 million downloads within two weeks, becoming number one on Apple's US App Store and Google Play, a faster start than ChatGPT.
    • Spear Street Technology launched Instinct, a similar AI agent app that uses WhatsApp as its interface.

    Net InterestInvesting

  • EvoDuet: Bilevel Co-Evolution of Web Searching and Task Solving for Scientific Discovery

    Evolutionary search with large language models (LLMs) can stall when progress requires external knowledge the model lacks. Supplying relevant documents helps, but simply adding web search tool can keep returning the same pages as solutions change. We introduce EvoDuet, a bi-level optimization method that co-evolves solutions and search queries with fixed model parameters. At each iteration, a retrieval gate lets the LLM assess its knowledge gap and choose to retrieve new documents, reuse stored ones, or proceed without them. An inner loop refines queries and ranks documents by the solution scores they are predicted to yield; an outer loop generates candidates in parallel from these documents and records the evaluated outcomes for later searches. Across 21 optimization tasks with one candidate per iteration, EvoDuet raises OpenEvolve's normalized discovery gain from 74.1% to 78.0% with GPT-5.6-Luna and from 61.3% to 82.3% with Gemini-3.8-Flash, whereas Qwen3.5-9B does not benefit. Our best runs surpass the previously reported best scores on eight tasks, including Swap Reduction on Q20 and Rosetta, and match them on three more. EvoDuet also improves with other scaffolds (e.g., Top-K, EvoX) on Sums/Diffs and Denoising, demonstrating its applicability across evolutionary search scaffolds.

    arXiv:2609.40340Hugging Face PapersYoung-Jun Lee et al.huggingface.co

  • Fold2Reason: Protein Folding for General AI Reasoning

    A new AI model, Fold2Reason, was developed and post-trained on protein folding data. This model significantly improved both protein structure prediction and general reasoning across 10 different benchmarks, suggesting that learning protein folding can generalize to broader reasoning tasks.

    What you get from it

    AI models can learn broader reasoning capabilities by training on structured scientific data like protein folding. This suggests that complex scientific problems can serve as a rich, non-linguistic source of supervision for improving AI.

    5 findings
    • Large language models often lack spatial and structural logic found in scientific problems.
    • Protein folding offers a natural testbed, providing thousands of checkable spatial and topological statements from one solved structure.
    • Researchers created FoldingCorpus, a protein-derived question-answer dataset.
    • They developed Fold2Reason, a post-training method that uses discrete structural answers and continuous 3D geometry.
    • Fold2Reason achieved structure prediction scores 2.7 to 3.5 times higher than Qwen3.5-9B on FoldBench.

    arXiv:2609.38879Hugging Face PapersYong Liu et al.huggingface.coAI Research

  • FrameMorrow: Future-Guided Long-Horizon Video Generation

    Long-horizon video generation requires models to effectively leverage an increasingly long generation history. As the generated history grows, retaining all previous content becomes increasingly expensive and redundant, making effective historical selection essential. Existing approaches often determine historical relevance based on the current content. However, information relevant to the present is not necessarily useful for future generation, while seemingly less relevant history may become important later. Our key insight is that historical information should be selected according to its relevance to future information needs. Capturing these needs does not require generating the full future; instead, a compact representation of what becomes important next is sufficient to guide historical selection. Building on this insight, we propose FrameMorrow, a prospective frame selector that predicts a small set of prospective tokens representing future information needs and uses them to identify relevant information from history. FrameMorrow selects explicit historical frames rather than model-specific internal states, enabling plug-and-play integration across diverse generators, including closed-source models, with little additional inference cost. We evaluate FrameMorrow across five benchmarks and 11 generative models spanning long-video generation, interactive generation, and action-conditioned world models. Extensive experiments demonstrate consistent improvements in long-range consistency, visual quality, and action alignment across diverse generation settings.

    arXiv:2609.38839Hugging Face PapersBo Yin et al.huggingface.co

  • When Symmetry Suppresses Magic

    Nonstabilizerness is a critical resource for quantum advantage, but evaluating it, especially for mixed states, requires superexponentially many samples in system size, making the problem NP-hard. While symmetries are known to reduce this complexity, it is unclear whether they also restrict the amount of magic. In this work, we provide an explicit example of such a symmetry by proving that the Robustness of Magic (RoM) for N-qubit X-states is at most $\sqrt{3}$. Leveraging this symmetry constraint, we introduce a computationally efficient method to lower-bound the RoM of arbitrary many-body states, demonstrating its utility on the ground states of a spin-1/2 Hamiltonian with system sizes well beyond the reach of exact evaluation. Furthermore, for Hamiltonians whose equilibrium states are X-states, we analytically derive the critical temperature at which magic emerges. Our work shows why certain symmetries constrain magic while others do not, provides a scalable lower-bounding method, and identifies a nontrivial regime where many-body nonstabilizerness is analytically solvable.

    arXiv:2609.38319arXiv quant-phA. de Oliveira Junior et al.arxiv.org

  • Exact Operator Complexity Measures from Random Matchgate Unitaries

    Matchgate circuits describe classically tractable dynamics in a physically relevant setting while nonetheless exhibiting features characteristic of complex quantum systems, such as high amounts of magic. This invites the broader question: to what extent are typical complexity measures saturated under trivial dynamics? Here, we address this by studying the behaviour of different Heisenberg-picture complexity measures in random matchgate ensembles. Our main technical contribution is a tractable and generalisible free-fermionic unitary Weingarten calculus formulated entirely in the Majorana basis. This method is tailored to computing moments of operators, and simplifies to a combinatorics problem which depends on the boundary conditions prescribed by a given operator functional. Using this framework, we derive closed-form expressions for the typical local-operator entanglement, operator stabiliser entropy, and higher-order out-of-time-ordered correlators. For extensive initial Majorana strings, both operator entanglement and stabiliser entropy obey volume laws, even though the underlying circuits remain classically simulable. Out-of-time-ordered correlators can likewise become exponentially small, showing that their late-time saturation value alone does not witness computational hardness and constituting a classically tractable example of emergent free independence. Finally, we relate these separations to operator resource theories, showing how operator entanglement beyond the Gaussian baseline lower-bounds the number of non-Gaussian gates in doped matchgate circuits. Our results present a practical toolbox for computing exact averages over random matchgates, while clarifying the distinction between quantum resources versus simulation complexity, and their dependence on the representation being exploited.

    arXiv:2609.38300arXiv quant-phGregory A. L. White et al.arxiv.org

  • STEPQuant: Recurrent State Quantization Framework

    Researchers developed STEPQuant, a post-training quantization framework for linear attention recurrent states. It significantly reduces memory usage by up to 68.7% and achieves over 5x compression while maintaining near FP32 accuracy with 6-bit states.

    3 findings
    • Linear attention models use fixed-size recurrent states, which can be a memory bottleneck during concurrent serving.
    • Directly quantizing these states to low precision often causes severe accuracy degradation due to error propagation.
    • The impact of quantization errors depends on temporal (long-lived memory) and spatial (key-row impact) dimensions.

    arXiv:2609.38169Hugging Face PapersBingchen Yao et al.huggingface.coAI Research

  • Learning Meta-Skills for Agent Harness Design in Test-Time AI4AI

    Researchers introduced Meta-Skills, principles an AI "Builder" learns to create optimal execution environments for a "Target" AI, improving performance on unseen tasks by up to 12.02 percentage points.

    What you get from it

    AI agents can improve their performance on new tasks by learning how to construct better execution environments for themselves. This approach allows for system-level self-improvement without retraining the core agent.

    5 findings
    • Agent performance depends on both its reasoning ability and the environment in which it acts.
    • The paper studies test-time AI-for-AI (AI4AI), where a Builder AI learns to construct better execution environments, called harnesses, for a Target AI while both models' weights remain fixed.
    • Meta-Skills are principles specifying when support is needed and what resources to provide to an agent.
    • A Builder learns these Meta-Skills from the Target's execution feedback on a development set.
    • Using frozen Meta-Skills, the Builder constructs harnesses for unseen tasks.

    arXiv:2609.38143Hugging Face PapersCheng Qian et al.huggingface.coAI Research

  • MotorMind: Zero-Shot Robot Manipulation with VLMs

    Researchers developed MotorMind, a system that enables general vision language models to perform zero-shot robot manipulation with high success rates on both simulated and real robots, without requiring task-specific training or external tools.

    3 findings
    • MotorMind is a robot manipulation harness that connects VLM-proposed mid-level actions to deterministic robot control and feedback.
    • It uses asynchronous monitoring and background memory updates.
    • The system does not require task-specific policy training, coding agents, or additional grounding tools like SAM3.

    arXiv:2609.38078Hugging Face PapersBingxuan Li 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.