📊 Dataset #Dataset #RubricBasedEvaluation #RubricBasedEvaluation #PhysicianWritten

HealthBench Professional
👤 openai

🎯 Task
Clinical response evaluation

💡 Idea
Structured medical eval examples with conversations, physician responses, and scored rubric items, labeled by use case, red-teaming vs good-faith, difficulty, and specialty.

✨ Why it's interesting
Physician answers plus rubrics enable consistent scoring of model performance across clinical use cases.

Downloads: 5.7k | Likes: 43

🔗 dataset 🔗 repo

via @Papers.Data.Code

💻 Repo #Repo #Gbnf #Gbnf #LlamaCpp

Structured Cot
👤 andthattoo

🎯 Task
Reasoning token compression for code generation

💡 Idea
Constrain a model's thinking into short structured fields like GOAL/APPROACH/EDGE or GOAL/STATE/ALGO/EDGE/VERIFY at inference time, then let it generate code normally to reduce verbose CoT and compare free vs constrained runs.

✨ Why it's interesting
No training; 22.4× fewer think tokens on HumanEval+ and +14 pp pass@1 on LiveCodeBench.

💻 Repo
⭐ andthattoo/structured-cot — 196 stars (+154 3d)
Python

🔗 paper

via @Papers.Data.Code

📄 Paper #Paper #LLM #MultiAgentSystems #AgentOrchestration

From Skills to Talent: Organising Heterogeneous Agents as a Real-World Company
👤 Zhengxu Yu, Yu Fu, Zhiyuan He et al.

🎯 Task
Multi-agent organization and coordination

💡 Idea
Talent-Container architecture separates agent identity from runtime, while a Talent Market recruits verified agents on demand and E2R tree search plans, executes, and reviews tasks with formal guarantees.

✨ Why it's interesting
84.67% success on PRDBench, beating prior SOTA by 15.48 points.

💻 Repo
⭐ 1mancompany/OneManCompany — 119 stars

🔗 paper

via @Papers.Data.Code

📋 Weekly Digest | Apr 25 – May 02
#WeeklyDigest

📄 Papers

World-R1: Reinforcing 3D Constraints for Text-to-Video Generation
#TextToVideo #ReinforcementLearning #3DConsistency
RL text-to-video fine-tuning ⟶ boosts 3D consistency PSNR
→ Learn more...

GLM-5V-Turbo: Toward a Native Foundation Model for Multimodal Agents
#MultimodalAgents #ToolUse #ReinforcementLearning
Multimodal agent model ⟶ tool use and GUI interaction
→ Learn more...

Recursive Multi-Agent Systems
#MultiAgentSystems #Reasoning #LatentSpace
RecursiveMAS latent recursion ⟶ +8.3% accuracy, fewer tokens
→ Learn more...

Tuna-2: Pixel Embeddings Beat Vision Encoders for Multimodal Understanding and Generation
#MultimodalLearning #ImageGeneration #VisionLanguageModels
Direct patch embeddings ⟶ unified multimodal understanding generation
→ Learn more...

Step-Audio-R1.5 Technical Report
#AudioReasoning #Rlhf #SpokenDialogue
Rubric-guided audio RLHF ⟶ improves long-turn spoken dialogue
→ Learn more...

From Skills to Talent: Organising Heterogeneous Agents as a Real-World Company
#MultiAgentSystems #AgentOrchestration #TaskPlanning
OMC multi-agent orchestration ⟶ 84.67% on PRDBench
→ Learn more...

💻 Repos

andthattoo/structured-cot ⭐
#Gbnf #LlamaCpp #CodeBenchmarks
Grammar-constrained CoT ⟶ 22.4× fewer think tokens
→ Learn more...

deepseek-ai/TileKernels ⭐
#MixtureOfExperts #Quantization #GpuKernels
TileLang GPU kernels ⟶ LLM ops near hardware limits
→ Learn more...

antirez/llama.cpp-deepseek-v4-flash ⭐
#Cpp #Gguf #Quantization
llama.cpp DeepSeek v4 Flash ⟶ local MacBook inference
→ Learn more...

📊 Datasets

H³D: High-quality Holistic 3D Editing Dataset
#3DEditing #InstructionFollowing #PartLevel
3D editing dataset ⟶ trains part-level 3D editors
→ Learn more...

MathNet v0 — Olympiad Math Reasoning & Retrieval
#CompetitionMath #Multimodal #Retrieval
Multilingual Olympiad math dataset ⟶ reasoning and retrieval benchmark
→ Learn more...

➡️ Tomorrow — Computer Vision Monthly

via @Papers.Data.Code

📈 Monthly: Computer Vision | Apr 03 – May 03
#MonthlyDigest #CV

📄 Papers

World-R1: Reinforcing 3D Constraints for Text-to-Video Generation
#TextToVideo #ReinforcementLearning #3DConsistency
RL text-to-video tuning ⟶ boosts 3D consistency PSNR
→ Learn more...

LLaDA2.0-Uni: Unifying Multimodal Understanding and Generation with Diffusion Large Language Model
#MultimodalLearning #ImageGeneration #DiffusionModels
Discrete diffusion LLM ⟶ unifies multimodal understanding and generation
→ Learn more...

Tuna-2: Pixel Embeddings Beat Vision Encoders for Multimodal Understanding and Generation
#MultimodalLearning #ImageGeneration #VisionLanguageModels
Direct patch embeddings ⟶ unified multimodal understanding generation
→ Learn more...

Vista4D: Video Reshooting with 4D Point Clouds
#NovelViewSynthesis #VideoDiffusion #3DReconstruction
4D point cloud reshooting ⟶ best camera and 3D consistency
→ Learn more...

CoInteract: Physically-Consistent Human-Object Interaction Video Synthesis via Spatially-Structured Co-Generation
#VideoGeneration #DiffusionTransformers #HumanObjectInteraction
Diffusion HOI video synthesis ⟶ improves stability and contact realism
→ Learn more...

💻 Repos

facex-engine/facex ⭐
#FaceVerification #Webassembly #CpuInference
Local face embeddings ⟶ browser CPU verification
→ Learn more...

📊 Datasets

Sleep Health & Daily Performance Dataset
#Classification #Regression #HealthConditions
Synthetic sleep health dataset ⟶ benchmarks 3 prediction tasks
→ Learn more...

⚡ Trends

▸ Unified multimodal models increasingly merge visual understanding, generation, and editing end-to-end.
▸ Video generation methods add explicit 3D or geometry grounding for consistency.
▸ Specialized architectural priors target realism in controllable human and camera-centric video synthesis.

🧭 TL;DR

📄 World-R1: Reinforcing 3D Constraints for Text-to-Video Generation
RL adds strong 3D-consistent video generation without architecture or inference changes

⭐ facex-engine/facex ⭐
Tiny local face verification runs fast on browser and CPU

💡 Vision models are converging toward unified, geometry-grounded, practically deployable generation systems.

via @Papers.Data.Code

📄 Paper #Paper #Multimodal #AgentSystems #FoundationModels

Heterogeneous Scientific Foundation Model Collaboration
👤 Zihao Li, Jiaru Zou, Feihao Fang et al.

🎯 Task
Heterogeneous scientific agent systems

💡 Idea
LLM-to-FM interfaces wrap specialized foundation models as agents: a query compiler creates structured calls, a response adapter feeds outputs back to reasoning, and a planner can orchestrate mixed LLM and FM agents.

✨ Why it's interesting
~7% higher utility, ~30% fewer tokens, and ~10% faster than single-LLM agents.

💻 Repo
⭐ Violet24K/Eywa — 18 stars

🔗 paper

via @Papers.Data.Code

📊 Dataset #Dataset #LLM #CodingAgent #AgentTraces

SWE-chat
👤 SALT-NLP

🎯 Task
AI coding session modeling

💡 Idea
205+ repositories of real developer–AI coding sessions with full chat transcripts, tool calls, thinking traces, code changes, and authorship attribution between humans and agents.

✨ Why it's interesting
Combines interaction traces with code edits and authorship labels, enabling study of real human-agent coding workflows.

Size: 205+ repositories

Downloads: 1.5k | Likes: 34

🔗 dataset

via @Papers.Data.Code

💻 Repo #Repo #LLM #DiscreteDiffusion #Distillation

Tide
👤 PKU-YuanGroup

🎯 Task
Diffusion LLM distillation

💡 Idea
Distill large diffusion LLM teachers into a 0.6B student even when teacher and student differ in architecture, attention, and tokenizer, with released training scripts, checkpoints, datasets, and 8-benchmark evaluation.

✨ Why it's interesting
+1.53 avg over BD3LM, +16.48 HumanEval over AR, 22× lower peak memory, and 5.2× faster inference.

💻 Repo
⭐ PKU-YuanGroup/TIDE — 64 stars (+62 3d)
Python


via @Papers.Data.Code

📄 Paper #Paper #Robotics #SLAM #OpenVocabulary

RADIO-ViPE: Online Tightly Coupled Multi-Modal Fusion for Open-Vocabulary Semantic SLAM in Dynamic Environments
👤 Zaid Nasser, Mikhail Iumanov, Tianhao Li et al.

🎯 Task
Open-vocabulary semantic SLAM

💡 Idea
Tightly coupled bundle adjustment fuses dense RADIO/RADSeg vision-language embeddings with geometry, plus temporally adaptive robust kernels to down-weight moving or displaced objects.

✨ Why it's interesting
Best average ATE on dynamic TUM-RGBD: 1.63 cm; top-3 on Replica semantic mapping.

💻 Repo
⭐ be2rlab/RADIO-ViPE — 74 stars

🔗 paper

via @Papers.Data.Code

📄 Paper #Paper #Multimodal #VideoGeneration #DiffusionModels

UniVidX: A Unified Multimodal Framework for Versatile Video Generation via Diffusion Priors
👤 Houyuan Chen, Hong Li, Xianghao Kong et al.

🎯 Task
Multimodal video generation

💡 Idea
Stochastic condition masking enables omni-directional generation; decoupled gated LoRA adds per-modality adapters only for targets; cross-modal self-attention shares keys/values across modalities for alignment.

✨ Why it's interesting
Competitive with state of the art across tasks; robust in-the-wild with <1k training videos.

💻 Repo
⭐ houyuanchen111/UniVidX — 44 stars

🔗 paper

via @Papers.Data.Code

📊 Dataset #Dataset #TimeSeries #GlobalHealth #CountryLevel

WHO Global Health Indicators for Prediction
👤 patelris

🎯 Task
Global health time series forecasting

💡 Idea
100k+ country-year health records across wide, long, latest-value, and metadata tables: 200+ countries, 2000-2024, 43 indicator definitions, demographics, mortality, spending, immunization, nutrition, and GDP.

✨ Why it's interesting
Wide + long formats and country metadata support cross-country trend analysis, forecasting, and dashboarding.

Size: 100k+ data points; 5,275 rows main table

Downloads: 284 | Likes: 26

🔗 dataset

via @Papers.Data.Code

📄 Paper #Paper #CV #MotionCapture #PoseEstimation

MoCapAnything V2: End-to-End Motion Capture for Arbitrary Skeletons
👤 Kehong Gong, Zhengyu Wen, Dao Thien Phong et al.

🎯 Task
Arbitrary-skeleton motion capture

💡 Idea
Learnable Video-to-Pose-to-Rotation pipeline with GL-GMHA attention. A reference pose-rotation pair plus rest pose anchors each asset's coordinate system, making pose-to-rotation prediction learnable and end-to-end.

✨ Why it's interesting
Cuts rotation error from ~17° to ~10°, reaches 6.54° on unseen skeletons, and runs ~20× faster.

💻 Repo
⭐ animotionlab26/MocapAnything — 166 stars

🔗 paper

via @Papers.Data.Code

💻 Repo #Repo #CV #3DReconstruction #GaussianSplatting

Gen Wild Splat
👤 Vinayak-VG

🎯 Task
Sparse-view 3D reconstruction from unconstrained images

💡 Idea
Reconstructs a 3D Gaussian splat from 2-6 unposed photos, jointly estimating camera poses, depth, and appearance while masking transient objects and optionally refining renderings for multi-view consistency.

✨ Why it's interesting
Produces a 3D Gaussian splat in roughly 3 seconds on a single A6000 GPU.

💻 Repo
⭐ Vinayak-VG/GenWildSplat — 24 stars (+24 3d)
Python


via @Papers.Data.Code

📄 Paper #Paper #Robotics #VisionLanguageAction #EmbodiedReasoning

MolmoAct2: Action Reasoning Models for Real-world Deployment
👤 Haoquan Fang, Jiafei Duan, Donovan Clay et al.

🎯 Task
Vision-language-action robot control

💡 Idea
Embodied-reasoning VLM + flow-matching action expert via per-layer KV-cache conditioning, plus adaptive depth tokens that update only changed regions to cut reasoning latency.

✨ Why it's interesting
Beats strong VLA baselines incl. π0.5 on 7 benchmarks; Molmo2-ER gets 63.8% avg on 13 ER benchmarks.

💻 Repo
⭐ allenai/molmoact2 — 30 stars

🔗 paper

via @Papers.Data.Code

📊 Dataset #Dataset #Tabular #QuantumChemistry #AtomizationEnergy

MSR-ACC/TAE25
👤 microsoft

🎯 Task
Molecular property prediction

💡 Idea
73,040 QCSchema molecular records with CCSD(T)/CBS total atomization energies via W1-F12, covering closed-shell neutral equilibrium molecules with up to 5 non-hydrogen atoms from elements up to argon, plus geometry, graphs, and related energy fields.

✨ Why it's interesting
Large, accurate, chemically diverse labels enable broad benchmarking and training beyond typical organic-only sets.

Size: 73,040 molecules

Downloads: 360 | Likes: 4

🔗 dataset

via @Papers.Data.Code

💻 Repo #Repo #Multimodal #VisionEncoder #AutoregressivePretraining

GenLIP
👤 YanFangCS

🎯 Task
Generative vision-language pretraining

💡 Idea
Pretrain a ViT-based vision encoder for MLLMs using a single Transformer and a single autoregressive language modeling objective, without contrastive loss, dual towers, or an extra text decoder.

✨ Why it's interesting
Simplifies MLLM vision pretraining and reports particularly strong gains on Doc and OCR tasks.

💻 Repo
⭐ YanFangCS/GenLIP — 49 stars (+49 3d)
Python

🔗 paper

via @Papers.Data.Code

📄 Paper #Paper #Tabular #DiffusionModels #DecisionTrees

Trees to Flows and Back: Unifying Decision Trees and Diffusion Models
👤 Sai Niranjan Ramachandran, Suvrit Sra

🎯 Task
Tabular generation and tree-to-network distillation

💡 Idea
Tree↔flow correspondence maps refined tree partitions to PF-ODEs and diffusion dynamics to hierarchies; GTSM unifies boosting and score matching. TREEFLOW conditions flows on tree paths, and DSM-TREE distills full tree decisions.

✨ Why it's interesting
TREEFLOW is 2× faster; best TSTR on 3/5 and best Wasserstein on 4/5 benchmarks.

🔗 paper

via @Papers.Data.Code

📄 Paper #Paper #LLM #SearchAgents #SupervisedFineTuning

OpenSeeker-v2: Pushing the Limits of Search Agents with Informative and High-Difficulty Trajectories
👤 Yuwen Du, Rui Ye, Shuo Tang et al.

🎯 Task
LLM search agent training

💡 Idea
High-difficulty trajectory synthesis for SFT: enlarge source graphs, expand the tool set, and filter out short trajectories to force longer multi-hop ReAct search without CPT or RL.

✨ Why it's interesting
46.0 BrowseComp, 58.1 BC-ZH, 34.6 HLE, 78.0 xbench; beats Tongyi DeepResearch.

💻 Repo
⭐ PolarSeeker/OpenSeeker — 634 stars

🔗 paper

via @Papers.Data.Code

📄 Paper #Paper #Multimodal #ReinforcementLearning #KnowledgeDistillation

Beyond SFT-to-RL: Pre-alignment via Black-Box On-Policy Distillation for Multimodal RL
👤 Sudong Wang, Weiquan Huang, Xiaomin Yu et al.

🎯 Task
Multimodal reasoning post-training

💡 Idea
Black-box adversarial on-policy distillation with an MoE discriminator separates perception and reasoning feedback, aligning post-SFT outputs to supervision before RL without teacher logits.

✨ Why it's interesting
Boosts average accuracy over SFT→RLVR by +4.4 on 4B and +6.0 on 8B.

💻 Repo
⭐ XIAO4579/PRISM — 53 stars

🔗 paper

via @Papers.Data.Code

💻 Repo #Repo #Multimodal #TextToImage #FlowMatching

Leap Align Code
👤 RockeyCoss

🎯 Task
Preference alignment for text-to-image flow matching models

💡 Idea
Aligns flow-matching image generators with human preference rewards by replacing full-trajectory backpropagation with a two-step leap trajectory, so optimization can target any generation step during sampling.

✨ Why it's interesting
Enables gradient propagation to any generation step while avoiding full-trajectory memory cost.

💻 Repo
⭐ RockeyCoss/LeapAlign_Code — 12 stars (+12 3d)
Python

🔗 paper

via @Papers.Data.Code

📄 Paper #Paper #LLM #ContextLearning #MultiAgentSystems

From Context to Skills: Can Language Models Learn from Context Skillfully?
👤 Shuzheng Si, Haozhe Zhao, Yu Lei et al.

🎯 Task
Context learning for language models

💡 Idea
Multi-agent self-play builds skills instead of updating weights: Challenger makes tasks/rubrics, Reasoner solves with evolving skills, Judge gives binary feedback, and Cross-time Replay picks the most generalizable skill set.

✨ Why it's interesting
Improves CL-bench solving rates, e.g. GPT-4.1 11.1%→16.5% and GPT-5.1 21.2%→25.8%.

💻 Repo
⭐ S1s-Z/Ctx2Skill — 44 stars

🔗 paper

via @Papers.Data.Code