DailyPapers

Paper: huggingface.co/papers/2605.02… Code, model weights, and complete training data released including OpenFAST tokenizer and three new datasets spanning low-to-medium cost platforms.

MolmoAct2 by AllenAI A fully open action reasoning model for real-world robot deployment, featuring the largest open bimanual dataset (720 hours), a specialized spatial reasoning backbone, and adaptive-depth reasoning that cuts latency while outperforming GPT-5 and Gemini Robotics.

Ctx2Skill improves GPT-4.1 solve rates from 11.1% to 16.5% on CL-bench. Paper: huggingface.co/papers/2604.27… Code: github.com/S1s-Z/Ctx2Skill Dataset: huggingface.co/datasets/ssz11…

Ctx2Skill: Self-evolving context skills A framework that autonomously discovers skills from complex contexts through multi-agent self-play. No human labels or external feedback needed. The resulting natural-language skills plug into any LM to enhance context learning.

Paper page: huggingface.co/papers/2604.23… Demo video: cdn-uploads.huggingface.co/production/upl…

Prox-E enables fine-grained 3D shape editing without training It abstracts shapes into geometric primitives, uses VLMs to edit them, and guides 3D generation to preserve identity while making structural changes.

AllenAI just released new OlmPool long context models on Hugging Face Exploring how seemingly minor architectural choices impact long context extension, with full training checkpoints from 8k to 14k context lengths. huggingface.co/allenai/F_pre_…

Paper: huggingface.co/papers/2605.00… Enabling precise control over 3D world layout and scale through user-defined segment maps.

Map2World Generate 3D worlds from any segment map and text. This framework ensures global-scale consistency across expansive environments, while a detail enhancer network preserves fine-grained details without compromising scene coherence.

Paper: huggingface.co/papers/2605.00… Models: huggingface.co/collections/Ya… Code: github.com/YanFangCS/GenL…

ByteDance released GenLIP: Let ViT Speak A minimalist generative pretraining framework that trains Vision Transformers to predict language tokens directly from visual tokens using a single autoregressive objective, outperforming baselines with only 8B training samples.

Paper: huggingface.co/papers/2604.27… Code: github.com/web2bigtable/w…

Web2BigTable A self-evolving bi-level multi-agent framework for internet-scale information extraction. Orchestrator-workers coordinate via shared workspaces to tackle both wide and deep search tasks, achieving 7.5× SOTA on WideSearch and 73% on XBench-DeepSearch.

Paper: huggingface.co/papers/2605.00… Project page: finch.agibot.com/research/lwd LWD uses DIVL and QAM to learn from successes, failures, and human interventions across the fleet, continuously improving a single generalist policy without imitating only demonstrations.

Learning while Deploying: Fleet-Scale RL for Generalist Robot Policies A new framework that turns robot deployment into a continuous training loop, enabling 16 dual-arm robots to improve from real-world experience and achieve 95% success on long-horizon tasks like brewing tea and making cocktails.

Paper: huggingface.co/papers/2605.00… Model: huggingface.co/houyuanchen/Un… Code: github.com/houyuanchen111… Project: houyuanchen111.github.io/UniVidX.github…

UniVidX: A Unified Multimodal Framework for Versatile Video Generation Enables omni-directional generation across RGB, intrinsic maps, and alpha channels using diffusion priors with stochastic condition masking—trained on fewer than 1,000 videos for SIGGRAPH 2026.

Discuss: huggingface.co/papers/2604.15… Hallucinations arise from semantic interference during fine-tuning. Self-distillation mitigates this by regularizing output distributions.

Fine-tuning increases hallucinations New research shows SFT causes factual errors by interfering with pre-trained knowledge. The authors propose self-distillation to learn new facts without forgetting, plus selective parameter freezing to reduce hallucinations while preserving performance.

Paper: huggingface.co/papers/2604.27…

Edit-R1: Reasoning verifier-based RL for image editing Moves beyond simple scorers to chain-of-thought verifiers that break instructions into verifiable principles. Trains editing models via GRPO with fine-grained rewards, outperforming Seed-1.5-VL and scaling up to 7B parameters.