Research

Selected research

At most 8 papers: of those listed on Hugging Face Daily Papers in the last 14 days, the most upvoted, then those whose code has the most stars, then the most discussed. Each says where its explanation comes from, and every result is the authors' own unless it says it was independently reproduced.

  1. Waypoint-OPD: Transferring RL-Induced Policy Shifts via Gated On-Policy Distillation

    Bingxi Hou, Guochao Jiang, Guofeng Quan and others ·

    Results reported by the authors; not independently reproduced

    Main idea
    The paper proposes Waypoint-OPD, a distillation method that compares an RL teacher's early and late policy shifts and uses their agreement to gate what the student learns.Our plain-language reading of the abstract
    Why it matters
    It argues that endpoint-only policy shifts can hide inconsistent local preference changes during RL training. The authors claim that gating by early-late agreement improves student performance in math and code reasoning.Our plain-language reading of the abstract
    Practical relevance
    A practitioner can use the provided code to apply gated on-policy distillation from an RL-trained teacher to a student model. The code is at https://github.com/hhhbxloveai/WaypointOPD.Our plain-language reading of the abstract

    Selected on: 664 upvotes · 4 comments · code with 3 stars · 0 citing models on Hugging Face

  2. MJev-Spatial: System One Spatial Intelligence via Native Attention Matching

    Luozheng Qin, Yuhan Zhu, Changhai Zhou and others ·

    Results reported by the authors; not independently reproduced

    Main idea
    MJev-Spatial turns spatial questions into a compact classification problem, scoring candidate answers with the model's existing attention projections instead of generating text.Our plain-language reading of the abstract
    Why it matters
    The authors claim this unified classifier is faster and more accurate than autoregressive baselines for spatial choice, distance, and point tasks.Our plain-language reading of the abstract
    Practical relevance
    A practitioner could use the code at https://github.com/Fr0zenCrane/MJev-Spatial and the mindlab-research/MJev-Spatial-4B model to run fast spatial classification for choice, distance, or point prediction.Our plain-language reading of the abstract

    Selected on: 351 upvotes · 2 comments · code with 8 stars · 1 citing models on Hugging Face

  3. OpenGCUA: Toward a Fully Open Stack for Generalist Computer Use

    Hanzhang Zhou, Long Li, Quyu Kong and others ·

    Results reported by the authors; not independently reproduced

    Main idea
    OpenGCUA is an open stack for a generalist computer-use agent that lets the model choose visual and programmatic actions, use code execution, and make some task-state decisions.Our plain-language reading of the abstract
    Why it matters
    It aims to make frontier computer-use designs more open and studyable, because such implementations remain largely closed. The authors claim their full configuration improves task scores across all eight evaluated models and reduces reported steps per task relative to batched GUI-only execution.Our plain-language reading of the abstract
    Practical relevance
    A practitioner can use the released implementation, prompts, evaluation trajectories, real-device harness, and cross-device SDK from https://github.com/Tongyi-MAI/OpenGCUA to build or test generalist computer-use agents.Our plain-language reading of the abstract

    Selected on: 315 upvotes · 3 comments · code with 75 stars · 0 citing models on Hugging Face

  4. FDE-Gym: The Last Mile of AI Delivery

    Dongxu Yang, Hongda Yan, Renxin Wang and others ·

    Results reported by the authors; not independently reproduced

    Main idea
    FDE-Gym is a benchmark for AI agents that must deliver real customer solutions by clarifying ambiguous requirements, interacting with stakeholders, and modifying live-like systems. It scores work by the customer's success metric, with penalties for extra interactions.Our plain-language reading of the abstract
    Why it matters
    The authors claim it exposes a large gap between solving well-specified tasks and real delivery, especially when unstated constraints must be discovered. They report that the best tested configuration resolved only 14 of 95 tasks.Our plain-language reading of the abstract
    Practical relevance
    A practitioner can use the provided code to run or practice these simulated delivery tasks and evaluate agents under the same grading. The code link is https://github.com/deeplethe/fde-gym.Our plain-language reading of the abstract

    Selected on: 282 upvotes · 2 comments · code with 102 stars · 0 citing models on Hugging Face

  5. RoTA-Bench: Evaluating 3D Spatial Reasoning through Interactive Projection Alignment

    Qingchuan Ma, Yuexiao Ma, Xiang Meng and others ·

    Results reported by the authors; not independently reproduced

    Main idea
    RoTA-Bench evaluates whether vision-language models can infer 3D object geometry from 2D views by choosing rotations to match a target silhouette.Our plain-language reading of the abstract
    Why it matters
    The authors report that models fall far behind humans and can improve image overlap while worsening pose alignment, showing a key gap in precise 3D control.Our plain-language reading of the abstract
    Practical relevance
    A practitioner can use the code at https://github.com/MAC-AutoML/RoTA-Bench to run the benchmark and test a model's interactive 3D rotation alignment.Our plain-language reading of the abstract

    Selected on: 201 upvotes · 2 comments · code with 4 stars · 0 citing models on Hugging Face

  6. Self-Evolving Agent Harness

    Chuyang Wei, Haoxiang Guan, Minghang Zhu and others ·

    Results reported by the authors; not independently reproduced

    Main idea
    The paper proposes a self-evolving agent harness that improves an AI agent's prompts, memory, workflows, and code while keeping the model weights fixed. It frames accepted changes as persistent, feedback-driven revisions with separate rules for proposing and admitting them.Our plain-language reading of the abstract
    Why it matters
    The authors claim this can make agent systems more reliable by validating persistent changes and separating learning, runtime privileges, and consequential actions. They also discuss implications for scientific reliability, physical safety, and institutional accountability.Our plain-language reading of the abstract
    Practical relevance
    A practitioner could use the provided code to build or test a harness that updates agent scaffolding and validates changes before deployment. The code link is https://github.com/zgcagi/Harness-RSI.Our plain-language reading of the abstract

    Selected on: 197 upvotes · 6 comments · code with 165 stars · 0 citing models on Hugging Face

  7. MAI-Gym: An Open All-in-One Platform for GUI Agent Evaluation and RL Training across Mobile and Computer Environments

    Panrong Tong, Quyu Kong, Pengxiang Li and others ·

    Results reported by the authors; not independently reproduced

    Main idea
    MAI-Gym is an open platform for evaluating, collecting data, and training GUI agents with online RL across mobile, computer, and web environments. It uses a shared rollout pipeline and can scale to 20,000 parallel environment instances.Our plain-language reading of the abstract
    Why it matters
    Not written: no explanation has been checked for this paper yet.
    Practical relevance
    A practitioner can use the released code, mocked apps, verified tasks, and model checkpoint to build or train GUI agents. The code is at https://github.com/Tongyi-MAI/MAI-Gym.Our plain-language reading of the abstract

    Selected on: 153 upvotes · 1 comments · code with 24 stars · 1 citing models on Hugging Face

  8. Behavioral Compositionality of Sparse Autoencoder Representations

    Haotian Ma, Ziye Mei, Frederic Sala ·

    Results reported by the authors; not independently reproduced

    Main idea
    The paper asks whether sparse autoencoder features combine additively in model behavior. It introduces SAE Additivity Residuals to measure unexplained behavioral variation beyond a chosen interaction order.Our plain-language reading of the abstract
    Why it matters
    The authors claim learned SAE coordinates are more compositional than random rotations of the same subspace, but still need pairwise interactions for faithful explanations and interventions. This helps practitioners know when individual SAE features can be trusted for joint behavioral effects.Our plain-language reading of the abstract
    Practical relevance
    A practitioner can use the provided code at https://github.com/OpenSAE/SAE-Behavioral-Compositionality to compute SAR and compare learned SAE coordinates with random rotations. They can also test whether adding pairwise interaction terms improves steering or coalition-ordering predictions.Our plain-language reading of the abstract

    Selected on: 105 upvotes · 2 comments · code with 5 stars · 0 citing models on Hugging Face