Hugging Face Daily Papers
AI news source. Articles are auto-selected and adapted by Hamidun News editors.
Latest publications
Diffusion LLMs Spend Over Three Times as Many Passes on a Story as on Grammar
An analysis of two diffusion models across 10 tasks found that a story token in DiffusionGemma costs over three times as much as a grammar-fix token, and decoding speed helps pick correct answers.
EgoX Merges Camera, Hands and Objects Into One 3D Scene, With 65.6% Lower Hand Error
From first-person video, EgoX reconstructs camera, scene, hands and objects in shared metric coordinates in a single pass. Camera error is 68.9% lower.
OpenVLHarness Gives Multimodal Models Memory and Tools: +9.7–12.6 Points
OpenVLHarness, a harness with tools and persistent multimodal memory, lifted the average score of eight MLLMs by 9.7–12.6 points on 22 benchmarks, using at most 37% of competitors' tokens.
Stateful Harness Lifts World Model Accuracy on Hidden State from 32% to 92%
Researchers split rule execution from video rendering: Stateful Harness with a frozen MiniMax-H3 gets the correct outcome in 92% of WorldState-Bench pairs, versus 32% without it.
Hybrid LLMs: Recurrent Layers Suppress Attention, and Staged Training Helps
Researchers describe Competing Learning: in hybrid models, recurrent layers learn retrieval faster and suppress attention. They tested staged training on models up to 1B.
Robots Share Experience: 96-Machine Fleet Boosts Packing Success From 79.2% to 98.3%
Researchers have unveiled Collective Experience Learning, a paradigm where deployed robots share task experience to improve a shared policy. The system uses failure contexts and retrieves relevant successful segments fro
Halcyon beats LlamaGen and JiT-T2I with 18–275x less pretraining compute
Researchers trained Halcyon from scratch on 100M image-text pairs with no language pretraining: gFD 48.0 vs 63.3 for LlamaGen, at 18.2x less training compute.
MemT-1.0 builds memory into the model itself: 66x smaller storage at 128K
MemT-1.0 puts memory inside the model's own computation, built on Gated DeltaNet-2. At a 128K-token history, Nano speeds up inference 1.86x and cuts storage 66x.
ActionRoPE Turns Player Actions into Coordinates and Cuts Scene Drift by 66%
ActionRoPE controls game world models through world coordinates with no new parameters, cutting revisit drift by 66% compared with the best baseline.
Heartbeat Thinking Teaches Voice Agents to Work While You're Still Talking: Latency Down 36.4%
A new scheme for voice agents starts reasoning and tool calls mid-speech. On FDBench v3, tool coverage rose by 16 percentage points and response latency fell by 36.4%.
OpenDS-World: The Best AI Agent Solved 39% of Deep Search Tasks on the Live Web
The new OpenDS-World benchmark tested 12 top models on 200 live-web tasks. The best scored 39.0%; humans, across 1000 independent attempts, scored 21.1%.
SEAM teaches TTS new words without audio and cuts data needs by 75%
SEAM initializes embeddings for new TTS tokens from known subtokens, without a single audio recording, and cuts training data requirements by up to four times.
GPT agents lose up to 11.3 points of accuracy when the simulator report sits in a file
When PHREEQC's long output moves from context into a file, GPT-5.1 and GPT-5.2 answer worse by 11.3 and 7.9 points. Claude Opus 4.6 and Sonnet 4.6 show no drop.
VerifierBench: 19 LLMs Wrote Verifiers, and the Best Scored Only 62.1
A new benchmark tests whether agents can write their own code to check solutions. Across 222 tasks, 19 models scored 10.6 to 62.1, and the verifiers most often reject correct answers.
Strong AI agents hand the move to a human in 80% of cases when the human's path is shorter
A two-player maze experiment found that models that can solve the task alone delegate selectively: about 80% of first decisions go to the human when the human's route is shorter.
RouteWeaver teaches an LLM router to pick its own mode, reaching 61.4% accuracy
RouteWeaver unifies single-round, multi-round and agentic LLM routing in one policy. It averages 61.4% accuracy, 11.03% above the strongest baseline.
SpatialScout: Best VLM Agent Builds Only 41.71% of a Room's 3D Map
The new SpatialScout benchmark tests whether VLM agents can explore scenes on their own and map objects. Claude Opus 5.5 scored 41.71%, and only 3.25% of targets reached IoU above 0.9.
Omni-Woodpecker Cuts Hallucinations in Omni-LLMs Without Retraining: +12.9 on CMM
A Nanjing University team proposes OWP, a method that corrects an omni model's hidden states before decoding and adds 12.9 points on CMM for Qwen2.5-Omni-7B.
Researchers Explain How JEPA Selects Features: Relationships Matter, Not Lone Features
Mosh Levy, Aly Lidayan and Yoav Goldberg introduce "predictive closure": JEPA retains not individually predictable features, but sets of features that predict each other.
TILDE Teaches Robots From Foundation Models, Improving Results by Up to 4.1x
Researchers propose TILDE: an Embodied Foundation Model teacher steps into the student's training, while AutoDAgger collects recovery data twice as fast as humans.
Deep Fusion Memory cuts retrieval 32x and beats baselines by 12.7%
Deep Fusion Memory injects external knowledge into a frozen LLM's hidden states with no retrieval at inference: 32x fewer queries and up to +12.71% on QA tasks.
Biometrics Cracked Using Only 'Accepted/Rejected' Answers: Template Reconstructed with 97.6–100% Accuracy
Researchers show that even a binary 'accepted/rejected' answer is enough to reconstruct a protected biometric template, and a diffusion model of faces cuts the number of queries needed by 5x.
IF-MAP Finds Which Agents to Improve: Twice the Accuracy Gain
IF-MAP ranks agents in an LLM system by improvement potential: +3.81 pp from prompt optimization vs 1.85 with uniform selection, and up to 65% token savings.
ID-TAP finds the right persona for an LLM agent in ten pairwise comparisons
ID-TAP teaches LLM agents to pick a persona per query from binary response comparisons, and posts the best aggregate result across seven benchmarks within a ten-round budget.
LayerEditor Teaches an Image Editor to Return Results as RGBA Layers
LayerEditor combines instruction-based image editing with generation of ordered RGBA layers. The team built the LayerEditor-100K dataset and separate benchmarks for layer-level editing.
Design Anything Keeps the Brief Intact in Large Projects: +25% Across 24 Artifacts
Design Anything's graph harness stores the designer's intent outside the model. With GPT-5.6 Sol High it satisfies 92.5% of brief constraints, and at 24 artifacts its lead over baselines reaches 25.1%.
CoRE Teaches AI to Choose Its Own Knowledge-Base Retrieval Rules Without Labeled Data
CoRE jointly trains a rule router and a retriever without query-rule pairs. It was tested on three STaRK-QA domains: scientific discovery, biomedicine and e-commerce.
MTEdit: 500 thousand editing trajectories teach models not to lose quality over 8 steps
Researchers built MTEdit, a dataset of 500 thousand multi-step image edits, and fine-tuned Qwen-Image-Edit to the best result among open models.
Xiaomi-Lens-0, a 4B model, beats Qwen3.5 by 5.6 points on RefCOCO
Xiaomi releases Lens-0: a 4B model built on Qwen3.5 that draws masks from its own visual latents, gains 5.6 points on RefCOCO and leads on cIoU.
OfficeTown Simulates 23 Organizations and Trains Agents for Office Work on 406 Tasks
Researchers built OfficeTown, a generative simulation of offices that produced a benchmark of 406 tasks and 4,836 criteria. Training on its data noticeably improved Qwen3.5-9B on GDPval.