Papers · updated November 15, 2026
AI papers the community is reading.
The top-voted in Hugging Face Daily Papers, updated every day.
In the latest edition
-
SO-Arena: Benchmarking the Capability Spectrum and Reliability of System One Models
A study evaluated fast-decision models across 11 fields and found that strong overall performance does not guarantee reliable decisions.
Quote
System One models promise fast, general-purpose decision-making without lengthy token generation. However, understanding their capability spectrum and decision reliability requires more than aggregate performance on current benchmarks.
-
BirdDreams: Joint Perspective and BEV Video Generation for a Plausible, Reactive Driving World Model
A study introduces a system that generates driving videos and overhead maps to simulate traffic and follow chosen routes.
Quote
Driving world models aim to simulate how a scene unfolds under different ego actions. This requires plausible motion of surrounding vehicles as well as precise ego-trajectory control.
-
PhysLoop: Harnessing LLMs' Physics Awareness for Self-Improving Video Generation Models
A study presents a system that uses language models to create and fix videos with more physically realistic motion, and releases 100,000 videos.
Quote
Video generation models produce visually realistic videos but still struggle with physical dynamics and causal consistency.
-
G-STAR: Group-Scaled Token Capacity for Deeply Compressed Autoregressive Generation
A study proposes a way to generate images in fewer steps while maintaining quality, reporting finished images in 2.5 seconds.
Quote
Discrete autoregressive models offer a scalable framework for image generation.
-
Duo: A Reciprocal World Model for Full-Duplex Interaction
Researchers introduce a system that follows audio, video, and text in real time while conversing, handling interruptions, and enabling interaction with objects and virtual scenes.
Quote
In this report, we introduce the Reciprocal World Model (RWM) and present Duo v1, a full-duplex world model that continuously perceives audiovisual inputs while responding to users in real time.
-
Open Vision Banana: A Fully Open Recipe for Generative Visual Perception
Researchers present an open method that adapts image generators to recognize objects and depth, with competitive results on some tasks.
Quote
Recent work suggests that pretrained image generators encode real-world understanding beyond image synthesis. Vision Banana exploits this by framing vision tasks as conditional image editing.
-
Actor, Judge, or Teacher? Understanding Decision Models in Long-Horizon Agents
A study suggests models that choose poor actions can still teach agents by judging afterward which steps helped complete a task.
Quote
Decision models answer a question with a typed output, such as a choice among options with a probability for each, so one model can choose an agent's actions, evaluate actions already taken, or score every action to teach a policy with…
-
Ropedia Xperience 2: A Gold Standard for Scaling Physical AI Environments and Behaviors
Ropedia Xperience 2 combines thousands of digitally scanned environments with recordings of people doing tasks, turned into simulations and movements for robots.
Quote
Large foundation models such as GPT-6 Astra and Claude Fable 5.1 have recently been shown to control robots directly from visual observations. They nevertheless remain brittle on long-horizon, contact-rich tasks in real-world environments.
-
On the Design of Hyper-Connections: Simplification, Stability, and Scalability
A study introduces SimpleHC, a simpler architecture that matched a more complex design’s performance while adding little training time in tested models.
Quote
Hyper-Connections (HC) architectures are increasingly adopted in frontier language models, yet their components remain insufficiently understood and introduce non-negligible training overhead.
-
TTB: Test-time Baking for Efficient Rendering of Decoder-only View Synthesis Models
The authors propose a method that learns from scene images to generate new views up to 23 times faster, with similar quality.
Quote
Decoder-only Large View Synthesis Models (LVSMs) with KV-cache have recently achieved state-of-the-art quality by regressing novel views from neural networks without reconstructing 3D geometry, while Test-Time Training (TTT) layers save…
Top-voted this week
-
Learning Heterogeneous Tactile Representation for General Dexterous Manipulation
A study presents a model that combines data from different touch sensors and improves robots’ performance on manipulation tasks.
Quote
Tactile sensing is an indispensable modality for contact-rich dexterous manipulation, a capability that hints at the ultimate frontiers of embodied intelligence.
-
Tree-Ising: Repairing Language Models with Bit-Level Ising Edits
A study reports that carefully adjusting a language model’s internal values reduced reasoning errors and improved its performance on math tests.
Quote
Weight-space programming, introduced in recent MIT work by Takehi et al., enables a frontier LLM agent to read a network's weights and write a repair.
-
Few-Step Streaming Video Generation via Condition Shifting
A study presents a way to generate videos in fewer steps, using less memory and training time and earning stronger preferences on some measures.
Quote
Few-step sampling reduces the cost of streaming video generation.
-
World In-Context Learning for Agentic Robot Manipulation
The authors present a method in which robots rehearse demonstrated tasks in simulation and save what they learn to use in new scenes, without retraining the model.
Quote
Trained action policies write manipulation skill into network weights, tying what is learned to the distribution it is learned from.
-
SplatAnything: Gaussian Spacing Normalization for Feed-forward 3DGS across Arbitrary Scenes
A study proposes a way to make spacing between points more consistent when reconstructing 3D scenes from images.
Quote
We identify a key limitation of existing feed-forward 3DGS methods: Gaussian attributes are typically predicted in a radial-distance-normalized space that canonicalizes a global radial statistic of the reconstructed geometry, while the…
-
WorldFoundry: Towards Unified Infrastructure for World Intelligence
Researchers introduced an open platform for running and comparing systems that simulate environments, track changes, and respond to actions.
Quote
Visual generative models increasingly blur the line between synthesizing observations and simulating environments.
-
Sage-Taiyi: A Metacognitive Architecture for Scientific Discovery
An AI system that adjusts research ideas based on evidence and open questions helped guide experiments that improved the yields of two chemical reactions.
Quote
We introduce Sage-Taiyi, a metacognition-inspired architecture that uses accumulated scientific understanding to guide what to investigate next and how to investigate it.
-
Don't Just Trust Your Memory: Evidence-Linked Memory Reduces Memory-Induced Degradation in LLM Agents
A study presents a system linking lessons from past experiences to their original evidence, helping AI agents avoid errors caused by incorrect memories.
Quote
External memory enables LLM agents to improve from experience by distilling trajectories of past tasks into concise lessons for new tasks.
-
P^2-MOPD: Accelerating Multi-Teacher On-Policy Distillation via Parameter Priors
A study proposes using expert models as a starting point while continuing to train the student, restoring capabilities faster than the compared method.
Quote
Multi-teacher on-policy distillation consolidates expert capabilities by supervising student-generated trajectories, whereas task arithmetic offers a training-free alternative through linear parameter combinations.
-
Scaling Mathematical Proof Supervision from Structured Textbooks
Researchers organized proofs from 2,118 university textbooks into a collection that improved language models’ performance on math evaluations.
Quote
Mathematical supervision for language models is typically flattened into isolated problem-solution pairs, discarding the curricular organization and evidential structure of the original sources.
-
Reframe: Observation-Aligned Multi-Teacher On-Policy Distillation for Agentic Visual Reasoning
A study introduces a method that trains visual agents to make better use of tool observations and reports stronger results on image and video tests.
Quote
Agentic reinforcement learning (RL) and multi-teacher on-policy distillation (MOPD) offer complementary routes to post-training. For visual agents, however, tool use presents two related challenges.
-
dNaViT: Discrete Codes are General-Purpose Visual Representations
A study introduces a system that turns images into compact codes and uses them for visual tasks such as recognition, generation, and editing.
Quote
Discrete autoregressive modeling, notably through next-token prediction, has become a cornerstone of modern language models.
-
Flide: Accelerating Decision Inference with Exact State Forking
A study presents a system that speeds up branching answers from AI agents while preserving the results of separate calculations.
Quote
Decision models such as Jev have emerged as a highly promising paradigm for LLM agent systems, providing lightweight and reliable categorical reasoning for rapid System One execution.
-
SteerDistill: Quantizing LLM Agents via Closed-Loop Teacher Steering
A study introduces a method where a more precise model corrects failed responses from compressed agents, improving their performance without help during use.
Quote
Low-bit quantization reduces the inference cost of LLM agents but can compromise task success.
-
ReproBridge: Building Reproducibility into ML Research Artifacts
A study finds that organized code, setup, and instructions help agents repeat machine-learning experiments more reliably and at lower cost.
Quote
Reproducing machine learning experiments requires reconstructing procedures from information scattered across papers, released code, and supporting artifacts.
-
PCBRouteAgent: How Close Are LLM Agents to Becoming PCB Routing Engineers?
A study introduces an artificial intelligence agent that plans and routes connections on circuit boards, outperforming compared methods on three real-world board sets.
Quote
Automating printed circuit board (PCB) routing requires coordinating board-level connectivity decisions with geometric execution under strict design rules.
-
Confucius4-R2T2: A Stable-Prefix Streaming ASR Model for Real Real-Time Transcription
Researchers adapted a model to transcribe speech in real time without changing words already shown, while maintaining high accuracy and low delay.
Quote
Streaming speech recognition is increasingly consumed by machines, such as translators and voice agents, that act on each token as it appears. A token that is later revised costs wasted work or a wrong action.
-
Incentivizing Self-Learning in Large Language Models via Reinforcement Learning
A study finds that training models to consult and learn from books while solving problems improves their accuracy on mathematical proofs.
Quote
Large language models (LLMs) draw on parametric knowledge and retrieval-augmented generation (RAG) to solve knowledge-intensive tasks.
-
LongBanana: An Expert-Verified Benchmark for Long-Context Multi-Reference Image Synthesis
A study introduces an expert-designed test of whether systems can create coherent images from many references while following numerous details.
Quote
We introduce LongBanana, to our knowledge the first expert-designed and expert-verified benchmark for long-context multi-reference image synthesis in the age of agent-driven image generation.
-
MMDM: Measuring Massive Multitask Decision Making
A new benchmark tested decision models on varied tasks and found they struggle with linked choices, constraints, and selecting multiple options.
Quote
System-1 Decision Models directly make structured probabilistic decisions over query-specified options given a state in a single forward pass, without autoregressively generating reasoning trajectories or textual responses.
-
DexForesight: Grounding Dexterous Policies with Implicit Spatial Foresight
A study presents a method that helps robots anticipate hand movements and changes around them, improving success in manipulation tasks.
Quote
Dexterous manipulation requires rich spatial foresight over both the surrounding scene and the robot hand. Existing world-action models often acquire such predictive knowledge through explicit future prediction.
-
LLMs Don't Know What They Know: Benchmarking Temporal Self-Knowledge for Financial Backtesting
A study found that language models often know financial facts they say they do not, making their self-reports and knowledge cutoffs unreliable for testing past predictions.
Quote
Financial backtests of large language models (LLMs) are meaningful only if the model did not already know the outcomes it is asked to predict.
-
Beneath the Surface: What We Overlook When We Talk About Agent Evolution
A study finds that improving task-performing programs can boost some results, but also erase earlier abilities and increase costs.
Quote
Agent harness evolution enables recursive self-improvement (RSI) without updating model weights. We examine whether higher benchmark scores yield reliable, efficient, and transferable improvement.
-
MACaQ: Fast, Model Agnostic Calibrated Quantization for Low-Precision Weights and Activations
A method reduces neural networks’ memory and computing costs by compressing weights and activations, quantizing each layer separately and quickly.
Quote
Weight-activation quantization can reduce the memory and compute cost of neural-network inference, but recent methods often require expensive optimization and architecture-specific propagation of basis transforms across layers.
-
FINCOSMOS: An Open-World Evaluation Paradigm for Agents in Finance
A study created a test for financial agents that requires finding web data and following hidden strategies; none of 150 sessions passed every check.
Quote
Language agents increasingly work for users on the open Web, and finance is where trusting them is hardest.
-
Embodied First Exam
A study introduces a test of coding agents controlling robots across varied tasks and reports that the best agent had limited success.
Quote
Early successes on agentic robotics point to a potential paradigm shift toward using coding agents to control robots.
-
Longtail-SWE: High-Quality SWE Tasks for Underrepresented Languages
Researchers created a set of 202 programming tasks in 13 underrepresented languages, revising tests and descriptions to evaluate solutions accurately.
Quote
Challenging repository-level benchmarks remain limited for many programming language ecosystems, where filtering scarce candidates with ambiguous descriptions or defective tests can discard valuable engineering problems.
-
Fast and Exact Relative-Position Attention via Structure-Aware Kernels
A study presents a method that speeds up attention with relative positions and reduces memory use, enabling faster transcription of long sequences.
Quote
Relative-position attention mechanisms offer compelling alternatives to RoPE across various transformer architectures, delivering superior performance in domains such as audio.
-
EvoWorld: Towards Out-of-Sight World Evolution in Generative Video Models
Researchers propose a system that tracks the state of generated video scenes and predicts how events may change the behavior of people and objects.
Quote
Generative video models are increasingly used as visual world simulators, yet their primary strength lies in rendering plausible pixels rather than reasoning about how a world should evolve.
-
Read Less, Solve More: Token-Efficient Sparse Reading of External Content for AI Agents
A study presents a tool that helps artificial intelligence agents read only relevant passages, reducing text use and time without harming most tasks.
Quote
Web-facing AI agents repeatedly access webpages, documents, repositories, and tool responses, yet reading interfaces often expose entire objects when only sparse evidence is needed.
Saved
Kept only in this browser.
Nothing saved yet. Tap Save on any link to read it later.
Earlier weeks
Questions
What is Hugging Face Daily Papers?
A daily list of AI papers that people submit and vote on. This page shows the top-voted, each with a short summary and a link to the original.
Have these papers been reviewed by other scientists?
Almost none. The authors post them before review. Votes measure attention, not quality.
Who writes the summaries?
An AI model, only from what the authors published. Nobody checks them before they go up, so cite the paper, not the summary.
