Hugging Face Daily Papers
AI news source. Articles are auto-selected and adapted by Hamidun News editors.
Latest publications
Mycelium: Researchers Break Down Hallucination Benchmarks Along Three Axes
Julia Belikova and Elisei Rykov analyzed LLM hallucination detection benchmarks along three axes and proposed a new metric, Span Coverage F1, for evaluation at the text-span level.
AV-FABLE: Researchers Assemble 998 Hours of Video to Hunt for Pinpoint Deepfakes
The AV-FABLE benchmark, built from 16,650 original-fake pairs, shows that AV-LLMs don't notice audio and video edits — but their hidden features do. The NFSR method capitalizes on that.
Raspberry Pi 5 off the grid: a signed poisoned knowledge package sailed through every check
The authors ran offline RAG on Qwen2.5 1.5B/3B/7B on a Raspberry Pi 5: signature checks missed 6 of 9 update attacks, while layered verification caught all 9 in 4.5ms.
RoboJev-27B judges robot success from 10 frames: 0.993 AUROC on SO-101
An open 27-billion-parameter VLM watches 10 frames and tells you whether a robot nailed its task: 95.2% balanced accuracy, 0.26 seconds per episode, and about $0.19 per thousand evaluations on an H100.
C3L Breaks the Explanation Drift Loop in Continual Learning: 5 Benchmarks, 20+ Baselines
Tel Aviv researchers show that regularizing against a model's own saliency maps locks in spurious correlations. Their fix, C3L, anchors regularization to class-level semantic concepts.
ReGE Teaches Agents From Their Own Attempts: +9.9 pp Without Fine-Tuning
TIGER-Lab's ReGE has an agent write its own plan from logs of past attempts. GPT-5.6 Luna catches up to Sol (67.4% vs 67.0%) at 28% of the cost.
ACEdit Teaches Image Editors to Train on Their Own Mistakes Without Degrading
ACEdit applies Self-Forcing to image editing: the model learns from its own outputs, while a teacher sees the original image and a composite instruction.
Mine-Sample-Render: A Three-Stage Method for Diverse Generation
Modern generative models are high-quality but monotonous. MSR (Mine-Sample-Render) is a Microsoft Research approach for diverse generation at test time. An explicit corpus-derived sampling procedure replaces built-in ran
Sparse Autoencoders: Pairwise Interactions Lift Rank Fidelity to 0.7790
The new SAR metric shows SAE features add up better than a random basis in all 18 settings, but accurate steering still requires pairwise interactions.
PathFM: generating complete time series trajectories through flow matching
Autogluon researchers have introduced PathFM, a model that generates complete time series trajectories instead of simple point forecasts. On a dedicated benchmark called fps-bench, the model beat TimesFM-3 by 10.6% and g
Weight editing without retraining cut sexist responses from 57.9% to 24.3%
Researchers at AI Interpretability @ Illinois developed a weight-editing method for models. Without retraining, it cut sexism by 33.6 points and boosted truthfulness by over 6 points when models were pressured to lie.
Dusting the Phantom: smart Gaussian allocation for dynamic scenes
A team of scientists from Korean universities has solved a key problem in 3D Gaussian Splatting: fast-moving objects vanishing during video synthesis. Two new strategies — searching for under-reconstructed zones and dyna
DuET cuts piano-to-score transcription error from 6.56% to 3.64%
Researchers from the National University of Singapore have unveiled DuET, a transformer for automatic piano-to-score transcription. On the Gangqinpu dataset, the model cut the error rate from 6.56% to 3.64%.
LexFrame reads LLM hidden states: 0.97 coherence against a lens ceiling of 0.30
A new method reads hidden states through the model's own remaining layers. On SmolLM3-3B and OLMo3-7B it reaches 0.97 coherence, while fixed lenses top out at 0.30–0.35.
OmniXRef: A Unified Framework for Video Editing with Multimodal References
Adobe Research has unveiled OmniXRef, a framework for video editing with multimodal references. It lets creators combine text, images, and video to control motion, camera work, style, and effects. On the OmniXRef-Bench b
Scene-Designer Learns to Design and Critique 3D Scenes from Text Using 50,000 Examples
Scene-Designer learns to design and critique 3D scenes from text. Two datasets of 35,000 and 15,000 examples. Scene-Critic hits 89% accuracy in scene comparisons, outperforming GPT-5.2.
RAP compiles 25,000 paper-poster pairs to teach AI how to lay out posters in HTML/CSS
A new dataset called RAP contains 25,000+ pairs of conference papers and their author-made posters from events held between 2022 and 2026. Researchers at Meituan built a fully automated pipeline that converts posters int
ProD Stops AI Agents From Looping: Tested on 8 Models (0.6B–8B)
Small language models get stuck in bad strategies when tackling complex tasks. Researchers from Penn State and Microsoft CoreAI propose ProD, a distillation method that penalizes looping during training by exploiting the
ECHO: Multi-Step Forecasting Without Add-Ons, Half the Cost of MTP
ECHO embeds multi-token prediction directly into the final layer—no auxiliary modules needed—at half the cost of MTP. Gains show up in reasoning and code, already deployed in production at 200B parameters.
UniAVC: Joint Audio-Video Synthesis via Asymmetric Noise Scheduling
Researchers at HKUST have developed UniAVC, the first model for simultaneous audio-video synthesis. The system trains on each modality's data separately, then aligns them using paired examples and independent noise sched
FineSplat: Fine-Grained Patches and Dual Grouping for Detailed View Synthesis
KAIST researchers have developed FineSplat, a novel view synthesis method for 3D scenes that uses fine-grained patches and a dual grouping attention strategy. It achieves superior results with 4× fewer Gaussians thanks t
CoworkEnv synthesizes training environments for AI coworkers: Qwen3-VL-8B jumps from 46.6 to 64.1
CoworkEnv is a new framework for training compact AI agents that work with tools in multi-step scenarios. It combines task generation, validation, and repair — and the results are striking: Qwen3-VL-8B-Instruct jumped fr
RoboValue Tests Robot Action Evaluators on 35 Tasks and Finds Blind Spots
The RoboValue benchmark covers 15 simulated and 20 real-world manipulation tasks. It shows that value models confuse step order and handle recovery from failures poorly.
Where VLMs' Physical Self-Improvement Comes From: A Scaling Law for LLM Agents
The PhysRSI-eval study measures how Vision-Language models improve on physical manipulation tasks. The authors evaluated various VLMs across 18 tasks using a five-dimensional metric, revealing which agent capabilities bo
HoloBoxing teaches a humanoid to box in 7 minutes on a single GPU
Researchers froze a ready-made humanoid whole-body controller and trained a compact boxing planner on top of it: about seven minutes of additional training on a single GPU.
Alibaba DAMO Releases RynnUnify: A 16B Model That Plans in Video and Drives a Robot
The 16B-parameter RynnUnify combines scene understanding, video generation and robot control: 51.5% success on 20 RoboCasa365 tasks and 46.7% on three real-world tasks.
RSIHub Packs 4 Agent Evolution Methods Into One Framework: Up to +13.5 pp
The open-source RSIHub provides shared infrastructure for comparing methods of AI agent self-improvement. The largest pass@1 gain in testing was 13.5 percentage points.
Easy-3Dfy builds a 3D scene from a single photo and outperforms SAM-3Dfy
Ziang Zhang's team split object generation from scene-layout estimation. Easy-3Dfy beat SAM-3Dfy on geometric metrics on Scene-Maker and 3D-FUTURE200.
LoCA Combines Sentinel-1 and Sentinel-2 and Wins 68.8% of Comparisons Against SatCLIP
The new LoCA coordinate encoder builds embeddings from a (latitude, longitude) pair using two satellite sensors. On 17 tasks it beat SatCLIP, GeoCLIP, GT-Loc, AlphaEarth and Copernicus-FM.
OPST: A Model Fine-Tunes Itself on Its Own Questions, No Humans Required
Researchers from Google show that an LLM can improve its reasoning by training on questions and answers it generated itself. No human prompts, only topic descriptions.