Your brief

Sunday, July 12·Showing Paper
43clusters
—sources
—min read
Paper·5d ago

Global Workspace Models

Anthropic Research (first-party)
Anthropic proposes a "global workspace" model for LLMs,…
Don't Worry About the Vase (Zvi)
A new Anthropic paper proposes that language models use a…
CAVEMANThey built a tiny AI that runs on your phone but answers as well as the big ones in the cloud. Trick: it forgets old conversations on purpose.
RZ
Covered by 2 outlets
·
1 min read
Paper·2d ago

AI Model Innovations

HF Daily Papers
Vidu S1 is a new real-time video generation model that uses…
HF Daily Papers
Researchers found that over half of current video…
Anthropic Research (first-party)
Anthropic proposes a method to 'turn off' specific harmful…
HF Daily Papers
RoboDojo is a new benchmark for testing robot manipulation…
HF Daily Papers
LLM-as-a-Tutor adapts prompts used to train AI agents that…
HF Daily Papers
Researchers automated the design of embodied AI agents by…
HF Daily Papers
Researchers developed "Cross-Space Distillation" to train…
PyTorch Blog
New kernel fusion techniques integrate normalization…
CAVEMANThey made an AI that makes videos from your voice commands, like a real-time cartoon director. You can even use your own pictures of people or pets.
HHRHH
Covered by 8 outlets
·
1 min read
Paper·6d ago

LLM Performance Prediction

arXiv cs.LG
WattGPU predicts LLM inference power and latency on new…
arXiv cs.CL
A new method predicts workload spikes in LLM processing by…
CAVEMANThis paper built an AI that guesses how much power and time a computer chip will use to run a language AI, even if it's a chip or language AI it's never seen before. It uses only public specs, no testing needed.
AA
Covered by 2 outlets
·
1 min read
Paper·3d ago

Robotic Manipulation Models

HF Daily Papers
Researchers developed LaMem-VLA, a new framework for robots…
arXiv cs.RO (Robotics)
NativeMEM is a new approach for robots to remember long…
arXiv cs.RO (Robotics)
Researchers propose Latent Memory Palace (LMP), a new…
arXiv cs.RO (Robotics)
FabriVLA is a compact Vision-Language-Action model for…
arXiv cs.RO (Robotics)
Harness VLA uses a memory-guided agent to make frozen…
arXiv cs.RO (Robotics)
SkillPlug is a new framework for robots that learns…
arXiv cs.RO (Robotics)
New research introduces Temporally Conditioned…
arXiv cs.RO (Robotics)
PriGo is a new framework that helps robots adapt to new…
CAVEMANThis paper shows a new way for robots to remember past actions and use them for future tasks. It's like giving the robot a better short-term and long-term memory that it can use while thinking.
HAAAA
Covered by 8 outlets
·
1 min read
Paper·3d ago

Long-Horizon Agent Memory

arXiv cs.AI
A new memory agent acts alongside an action agent to…
arXiv cs.LG
CompactionRL trains LLM agents for long tasks by…
CAVEMANThey built a helper AI that reminds the main AI what's important from past steps in long tasks. This stops the main AI from forgetting and helps it do better.
AA
Covered by 2 outlets
·
1 min read
Paper·6d ago

Unified Video Generation Models

arXiv cs.CV
Flex-Forcing unifies video generation by letting models…
arXiv cs.CV
UNIVERSE unifies video prediction and action generation for…
CAVEMANThey made a video AI that can be fast like a streamer or slow and perfect like a movie director. It smartly breaks videos into pieces to get the best of both worlds.
AA
Covered by 2 outlets
·
1 min read
Paper·3d ago

Autonomous Driving Vision-Language Models

arXiv cs.AI
WCog-VLA is a new AI framework for autonomous driving that…
arXiv cs.CV
This paper introduces VLM-CASE, a framework for autonomous…
CAVEMANThis AI helps self-driving cars see and predict what other cars will do, not just react. It uses a new way to think about the whole scene and generate possible futures to drive smarter.
AA
Covered by 2 outlets
·
1 min read
Paper·7d ago

LLM Recommendation Systems

arXiv cs.IR
Researchers identify 'Length Bias' in LLMs used for…
arXiv cs.IR
A new framework called Bi-NAS uses neural architecture…
CAVEMANAI for recommending stuff can unfairly favor items with longer descriptions. This paper shows how to make it fairer by adjusting how the AI pays attention and measures item length.
AA
Covered by 2 outlets
·
1 min read
Paper·3d ago

Causal Structure Learning

arXiv stat.ML
New method learns causal relationships in data with…
arXiv cs.LG
CaSPECT is a new framework for finding groups in data that…
CAVEMANThey built a way to find cause-and-effect links in data that has natural groups, like patients in different hospitals. It's like finding a general rule but also seeing how each group might have its own special version of that rule.
AA
Covered by 2 outlets
·
1 min read
Paper·3d ago

Concept Erasure Methods

arXiv stat.ML
New research proposes AutoAnchor, a framework for stable…
arXiv cs.CV
New method CARE precisely removes concepts from diffusion…
CAVEMANThey built a way to teach image AI to forget specific things, like copyrighted art, without messing up its other skills. It uses a clever trick with how the AI pays attention to words to find the right way to forget.
AA
Covered by 2 outlets
·
1 min read
Paper·3d ago

KV Cache Compression

arXiv cs.AI
DepthWeave-KV compresses key-value caches for long-context…
arXiv cs.AI
FreqDepthKV compresses KV caches for long-context LLMs by…
arXiv cs.LG
Researchers propose compressing LLM prompts into a single…
arXiv cs.CL
This survey examines system-aware optimizations for the…
CAVEMANNew AI trick makes LLMs remember way more text using less computer memory. It smartly compresses info across different AI parts and focuses on important words, letting it handle longer stories faster.
AAAA
Covered by 4 outlets
·
1 min read
Paper·6d ago

Model Pruning Techniques

HF Daily Papers
This paper introduces SaMer, a method to compress image…
arXiv cs.IR
This paper introduces SaMer, a method to compress image…
CAVEMANThey made a way to shrink image data for searching, keeping important object details so searches work better. It saves lots of space without losing accuracy.
HA
Covered by 2 outlets
·
1 min read
Paper·4d ago

Vision-Language Model Analysis

arXiv cs.CV
Researchers introduce HIVE, an evaluation tool to study how…
arXiv cs.CV
New research introduces ARGTCA, a method to improve how…
arXiv cs.CL
New research proposes a way to understand why…
CAVEMANThey built a tool to see how fake facts in AI's understanding of images mess up its thinking. Surprisingly, sometimes the fake facts help the AI do better on image tasks by making it think about more things.
AAA
Covered by 3 outlets
·
1 min read
Paper·4d ago

Agent Benchmarking

arXiv cs.AI
New research reveals a silent failure mode in LLM agents…
arXiv cs.AI
AgentGym2 is a new framework for testing AI agents in…
arXiv cs.CL
ToolFailBench is a new benchmark designed to diagnose…
CAVEMANAI agents using tools can secretly break rules without showing errors, causing bad outcomes. Adding a quick check before the tool runs stops many of these secret mistakes.
AAA
Covered by 3 outlets
·
1 min read
Paper·4d ago

Deepfake Detection Benchmarks

arXiv cs.CV
Researchers used a new XAI technique to understand how…
arXiv cs.CV
VendorBench-100 is a new benchmark for deepfake image…
arXiv cs.CV
A new benchmark, XPlainVerse, is introduced for evaluating…
CAVEMANResearchers found a way to peek inside deepfake detectors to see what clues they use to tell real from fake. This helps us understand how they work and build better ones.
AAA
Covered by 3 outlets
·
1 min read
Paper·6d ago

Origin-Destination Flow Prediction

arXiv cs.LG
GeoFlow is a new AI framework for predicting and generating…
arXiv cs.LG
OpFlow is a new AI framework for predicting urban travel…
CAVEMANThis AI helps predict where people will travel in cities by looking at maps and how far apart places are. It can also create realistic travel patterns, like simulating city traffic.
AA
Covered by 2 outlets
·
1 min read
Paper·3d ago

Video-Action Models For Robotics

HF Daily Papers
New research identifies a common failure in AI action…
arXiv cs.RO (Robotics)
Researchers introduce LingBot-VA 2.0, a new video-action…
arXiv cs.RO (Robotics)
Researchers identify a "video-action generalization gap"…
arXiv cs.CV
LingBot-Video is a new video pretraining model designed for…
CAVEMANAI models guess actions by looking at objects, not how things move. This paper shows how to train them to watch the motion instead, so they can understand new actions better.
HAAA
Covered by 4 outlets
·
1 min read
Paper·3d ago

Humanoid Robot Control

arXiv cs.RO (Robotics)
New framework ContactMimic enables humanoids to control…
arXiv cs.CV
WristMimic is a new framework for controlling humanoid…
CAVEMANRobots can now learn to touch things correctly, like sitting on a chair, not just reach the right spot. They learn to make or avoid touching objects on purpose, even in the real world.
AA
Covered by 2 outlets
·
1 min read
Paper·1d ago

China's Orca world model matches specialized robotics systems without ever seeing a single action label

The Decoder
China's Beijing Academy of AI released Orca, a world model…
CAVEMANThey made an AI that learns how the world works just by watching videos, no instructions needed. It's as good as robots trained with specific commands, but learned way easier.
D
Covered by 1 outlet
·
1 min read
Paper·3d ago

Vision-Language Models

arXiv cs.AI
A new method, Structured Sparse AutoEncoder ($S^2AE$),…
arXiv cs.CV
New method improves image recognition by better…
CAVEMANThey built a better way for AI to understand images and text together. It groups parts of images to make sure the AI learns concepts that make sense for both words and pictures, improving how well it understands things.
AA
Covered by 2 outlets
·
1 min read
Paper·3d ago

Reward Hacking

arXiv cs.LG
New research reveals that LLMs trained using self-play…
arXiv cs.AI
LLM judges can give different scores to the same answers…
CAVEMANAI judges trained on their own answers learn to sound good, not be right. They trick themselves into thinking wrong answers are correct. Fixing this means the judge must answer first, or it can't be fooled.
AA
Covered by 2 outlets
·
1 min read
Paper·6d ago

AI Model Generalization

arXiv cs.AI
Researchers studied a small, 12K-parameter AI model to…
arXiv cs.AI
Researchers propose a method to create language models that…
CAVEMANTiny AI learns late, but only if you train it *just right*. Small changes in how it computes can break its learning, meaning we might be seeing things that aren't really there.
AA
Covered by 2 outlets
·
1 min read
Paper·4d ago

Multimodal Reasoning Frameworks

arXiv cs.CV
Researchers propose Brain-inspired Unsupervised…
arXiv cs.CL
ProLaViT is a new framework for multimodal AI that improves…
CAVEMANAI can now learn to check its own work like a brain does, without needing teachers to label everything. It predicts what might have happened to fix its mistakes on visual tasks.
AA
Covered by 2 outlets
·
1 min read
Paper·3d ago

Speculative Decoding

arXiv cs.CL
DominoTree is a new method for speculative decoding that…
arXiv cs.CL
DeLS-Spec decouples long and short context experts for…
arXiv cs.AI
Researchers investigate training-free relaxed speculative…
CAVEMANThey made a faster way to generate text with AI. It predicts multiple words at once and checks them together, making it much quicker than older methods.
AAA
Covered by 3 outlets
·
1 min read
Paper·10d ago

LLM Inference Optimization

HF Daily Papers
Researchers propose separating Transformer computation into…
arXiv cs.AI
Researchers propose HOLA, a new memory system for language…
arXiv cs.LG
Researchers propose separating state storage from…
CAVEMANThey split the AI's brain into two parts: one for guessing the next word, and one for remembering important stuff. This makes the AI learn faster and perform better.
HAA
Covered by 3 outlets
·
1 min read
Paper·3d ago

LLM Reasoning Improvement

arXiv cs.AI
Researchers identify the "Knowing--Using Gap" where LLMs…
arXiv cs.CL
MILES is a new framework for LLMs to improve reasoning by…
arXiv cs.CL
This paper theoretically analyzes how LLMs improve…
CAVEMANLLMs learn facts but can't use them. Researchers found the problem is that the AI doesn't send the learned info to the right thinking parts. A simple fix helps the AI use its knowledge better.
AAA
Covered by 3 outlets
·
1 min read
Paper·4d ago

Infrared Vision-Language Attacks

arXiv cs.AI
New research explores how adversarial attacks exploit…
arXiv cs.CV
Researchers developed InfraQR, a novel attack that places…
arXiv cs.CV
New research introduces AirflowAttack, the first…
CAVEMANResearchers found a way to trick AI that understands images and text by messing with its internal math. They show how this trick works and suggest ways to make the AI smarter and harder to fool.
AAA
Covered by 3 outlets
·
1 min read
Paper·8d ago

Mistral AI Models

Mistral AI
Mistral AI released Leanstral 1.5, an open-source model…
Mistral AI
Mistral AI has released Leanstral 1.5, a new model focused…
CAVEMANThis paper shows how to make AI better at creating and checking math proofs. It's like giving AI a super-powered calculator that can also show its work.
MM
Covered by 2 outlets
·
1 min read
Paper·3d ago

LLM Uncertainty Quantification

arXiv cs.CL
Researchers found that internal model states, not just…
arXiv cs.CL
New research explores how LLMs estimate confidence during…
arXiv cs.IR
New framework uses LLM internal states to estimate…
CAVEMANAI forecasters can be wrong even when they sound sure. This study found a way to peek inside the AI's brain to see if it's really thinking correctly, and it turns out the AI decides its answer before it writes its explanation.
AAA
Covered by 3 outlets
·
1 min read
Paper·3d ago

Proactive Agent Benchmarks

HF Daily Papers
UniClawBench is a new benchmark for evaluating proactive AI…
arXiv cs.CL
UniClawBench is a new benchmark for proactive AI agents,…
CAVEMANThey made a new test for AIs that use real computer programs to help people. It checks if the AI can learn new tools, explore, understand long instructions, see images, and work across different apps, all in live tests.
HA
Covered by 2 outlets
·
1 min read
Paper·2d ago

Reducing High-Bandwidth Memory Bottlenecks in JAX-Based LLM Training with Host Offloading

NVIDIA Technical Blog
NVIDIA's technical blog details a new technique for LLM…
CAVEMANThey found a way to use your computer's main memory to help train big AI models faster. This lets the AI use more of its brainpower instead of getting stuck waiting for memory.
N
Covered by 1 outlet
·
1 min read
Paper·3d ago

Remember When It Matters: Proactive Memory Agent for Long-Horizon Agents

HF Daily Papers
New research introduces a "proactive memory agent" to help…
CAVEMANAI agents forget important stuff in long tasks. This new AI has a helper that reminds it what to remember, making it do better on hard problems.
H
Covered by 1 outlet
·
1 min read
Paper·3d ago

Robots for Dementia Therapy

arXiv cs.RO (Robotics)
A study analyzed robot-delivered cognitive therapy sessions…
arXiv cs.RO (Robotics)
A social robot was developed to autonomously deliver…
CAVEMANResearchers used a robot to talk with dementia patients at home. They found that when the robot asked personal questions, patients talked more and seemed more engaged. Patients also got tired later in the sessions.
AA
Covered by 2 outlets
·
1 min read
Paper·3d ago

Scientific Idea Reasoning

HF Daily Papers
New benchmark, IdeaGene-Bench (IG-Bench), evaluates AI's…
arXiv cs.AI
New benchmark, IdeaGene-Bench (IG-Bench), tests AI's…
CAVEMANAI can now track how science ideas grow from older ones, like a family tree. They built a test to see if AI can guess the next step or even invent new ideas based on what came before.
HA
Covered by 2 outlets
·
1 min read
Paper·5d ago

Diffusion Language Model Decoding

arXiv cs.CL
Nemotron-Labs-Diffusion is a new language model that…
arXiv cs.CL
New decoding method for diffusion language models called…
arXiv cs.CL
New hybrid language model combines diffusion and Mamba for…
CAVEMANThey built an AI that can talk in three different ways at once. It's faster and smarter than other AIs, especially when it guesses ahead and then checks its work.
AAA
Covered by 3 outlets
·
1 min read
Paper·3d ago

Video Reasoning Models

arXiv cs.AI
OpenCoF introduces a new dataset (OpenCoF-17K) and a…
HF Daily Papers
Researchers introduce OpenCoF, a framework for video…
CAVEMANThey built a new way for AI to think by watching videos play out frame by frame. It helps AI make better decisions by showing its work over time, like a movie.
AH
Covered by 2 outlets
·
1 min read
Paper·5d ago

ASR Timestamp Correction

arXiv cs.CL
Researchers developed a new gradient-based method to find…
arXiv cs.CL
New research introduces REDDIT, a method to fix timestamp…
CAVEMANThey found a way to see when each word is spoken in audio, even if the AI model doesn't tell you directly. It uses the model's own learning signals to figure it out, working for many different kinds of AI.
AA
Covered by 2 outlets
·
1 min read
Paper·3d ago

Human Motion Generation

arXiv cs.RO (Robotics)
ARDY is a new framework for generating realistic 3D human…
HF Daily Papers
ARDY is a new framework for generating realistic 3D human…
CAVEMANThey built a fast AI that makes 3D people move realistically in games or robots, using text and body poses to guide it. It can handle long instructions and works on the fly.
AH
Covered by 2 outlets
·
1 min read
Paper·3d ago

Video Object Segmentation

arXiv cs.CV
SAM-MT is a new framework for real-time video segmentation…
arXiv cs.CV
A new self-supervised method for video object segmentation…
CAVEMANThey made a system that can track and outline many things in a video at once, super fast. It works like magic by giving each thing its own special tracker and a shared memory for the whole scene.
AA
Covered by 2 outlets
·
1 min read
Paper·10d ago

AI Coding Agent Governance

arXiv cs.AI
Researchers propose using traditional software engineering…
arXiv cs.AI
A new paper explores how software engineers can manage AI…
CAVEMANThey used old-school computer security rules to make AI coders safer and easier to check. A small AI found way more hidden bugs when given these rules.
AA
Covered by 2 outlets
·
1 min read
Paper·3d ago

Robotic Mapping Systems

arXiv cs.AI
Track2Map is a new online 3D reconstruction system for…
arXiv cs.RO (Robotics)
GeoGS-SLAM is a new visual SLAM system that uses a…
CAVEMANThey made a system that maps 3D body parts during surgery using video, even if the camera shakes. It figures out the camera's path and the body's shape at the same time.
AA
Covered by 2 outlets
·
1 min read
Paper·4d ago

AI Content Detection

arXiv cs.LG
Researchers propose a unified framework using Mahalanobis…
arXiv stat.ML
Researchers propose a new statistical framework for LLM…
CAVEMANNew way to spot AI text, fake facts, hidden marks, and tricky inputs. It learns what normal stuff looks like really well to find the AI fakes.
AA
Covered by 2 outlets
·
1 min read
Paper·5d ago

Robot Simulation Generation

arXiv cs.RO (Robotics)
Image2Sim is a new framework that creates realistic 3D…
HF Daily Papers
Image2Sim creates realistic 3D environments for training…
arXiv cs.RO (Robotics)
RoboSnap turns a single image into a simulation scene for…
CAVEMANThey built a way to make 3D worlds for robots to learn in, just from videos. This lets robots practice driving around much more than before, and they do better in tests.
AHA
Covered by 3 outlets
·
1 min read
— end of today's brief —