📄 Paper #Paper #LLM #Reasoning #ReinforcementLearning

Achieving Gold-Medal-Level Olympiad Reasoning via Simple and Unified Scaling
👤 Yafu Li, Runzhe Zhan, Haoran Zhang et al.

🎯 Task
Olympiad-level mathematical and scientific reasoning

💡 Idea
Instead of domain-specific systems, it uses one scaling recipe: reverse-perplexity long-CoT SFT to instill proof search and self-checking, then coarse verifiable-reward RL, proof-level RL with self-refinement/replay, and test-time verification loops.

✨ Why it's interesting
SU-01 gets 57.6% on IMO-ProofBench, 70.2% with test-time scaling, and reaches the IMO 2025 gold line with 35 points.

💻 Repo
⭐ Simplified-Reasoning/SU-01 — 68 stars

🔗 paper

via @Papers.Data.Code

📊 Dataset #Dataset #LLM #SoftwareEngineering #ToolUse

Orchard
👤 microsoft

🎯 Task
Agentic software engineering and web GUI interaction

💡 Idea
~110K agent trajectories in 2 parallel subsets: 107,185 SWE chat+tool rollouts over 2,788 GitHub repos with hidden-test pass/fail labels, plus 3,070 GUI decision-point rows with screenshots, chat context, and judge-verified rewards across 409 web tasks.

✨ Why it's interesting
Verified patch outcomes and judge-scored GUI steps make agent training and evaluation measurable across real coding and browser tasks.

Ⓢ 110,255 samples, ~10.97 GB

🔗 dataset

via @Papers.Data.Code

💻 Repo #Repo #CV #DepthEstimation #CameraPose

vggt Omega
👤 facebookresearch

🎯 Task
multi-view camera and depth reconstruction

💡 Idea
Infer camera parameters and per-image depth from a set of images in one forward pass, and optionally produce text-aligned embeddings for the same visual inputs.

✨ Why it's interesting
Runs end-to-end on a single A100 with 6.02 GB for 1 frame and 43.15 GB for 500 frames, with released 1B checkpoints and demo code.

💻 Repo
⭐ facebookresearch/vggt-omega — 413 stars (+413 3d)
Python


via @Papers.Data.Code

📄 Paper #Paper #CV #VideoGeneration #DiffusionModels

Causal Forcing++: Scalable Few-Step Autoregressive Diffusion Distillation for Real-Time Interactive Video Generation
👤 Min Zhao, Hongzhou Zhu, Kaiwen Zheng et al.

🎯 Task
Real-time autoregressive video generation

💡 Idea
Instead of costly AR-teacher ODE trajectory distillation, it initializes few-step AR students with causal consistency distillation: same AR flow-map target, but learned from single online adjacent-step teacher updates, making frame-wise 1-2 step rollout practical.

✨ Why it's interesting
At frame-wise 2-step, beats 4-step chunk-wise Causal Forcing by +0.1 VBench Total, +0.3 Quality, +0.335 VisionReward; 50% lower first-frame latency, ~4x cheaper Stage 2.

💻 Repo
⭐ thu-ml/Causal-Forcing — 665 stars

🔗 paper

via @Papers.Data.Code

📄 Paper #Paper #Multimodal #LongContextModeling #VisionLanguageModels

Training Long-Context Vision-Language Models Effectively with Generalization Beyond 128K Context
👤 Zhaowei Wang, Lishu Luo, Haodong Duan et al.

🎯 Task
Long-context vision-language modeling

💡 Idea
Instead of OCR-style long-data training, use instruction-formatted long-document VQA. A balanced length mix and retrieval-heavy task mixture beat 128K-focused or transcription-based training for extending LVLM context.

✨ Why it's interesting
With 5B tokens, MMProLong improves long-doc VQA by 7.1%, stays strong at 256K/512K beyond 128K training, and exceeds baselines by 20%+ there.

🔗 paper

via @Papers.Data.Code

📊 Dataset #Dataset #LLM #HallucinationDetection #Multilingual

LLM Hallucination Benchmark Dataset
👤 alitaqishah

🎯 Task
LLM hallucination detection and analysis

💡 Idea
200 annotated LLM responses spanning 5 models, 8 domains, 7 languages, 7 hallucination types, 4 annotator types, and 4 mitigation strategies, with prompt, response, hallucination label, span, severity, and verified correction.

✨ Why it's interesting
Makes cross-model, multilingual hallucination detection and mitigation evaluation directly measurable with typed error labels and corrected references.

Ⓢ 200 annotated responses

🔗 dataset

via @Papers.Data.Code

📄 Paper #Paper #LLM #ReinforcementLearning #KnowledgeDistillation

Self-Distilled Agentic Reinforcement Learning
👤 Zhengxi Lu, Zhiyuan Yao, Zhuowen Han et al.

🎯 Task
Agentic RL for multi-turn LLMs

💡 Idea
Instead of naively mixing OPSD with RL, SDAR keeps RL as the backbone and uses detached token-level gates to apply distillation selectively—amplifying positive teacher-student gaps and softening negative teacher rejections.

✨ Why it's interesting
Beats GRPO by +9.4% on ALFWorld, +7.0% on Search-QA, and +10.2% WebShop-Acc, while avoiding naive GRPO+OPSD instability.

💻 Repo
⭐ ZJU-REAL/SDAR — 96 stars

🔗 paper

via @Papers.Data.Code

💻 Repo #Repo #CV #VideoGeneration #CameraControl

Warp As History
👤 yyfz

🎯 Task
camera-controlled video generation

💡 Idea
Generate videos that follow user-specified camera trajectories from one input frame, using a single training video and optional interactive/autoregressive control in a drop-in Helios pipeline.

✨ Why it's interesting
Enables interactive viewpoint control from only one camera-annotated training example, with released training, inference, and browser demo code.

💻 Repo
⭐ yyfz/Warp-as-History — 117 stars (+58 3d)
Python

🔗 paper

via @Papers.Data.Code

📄 Paper #Paper #NLP #TheoremProving #RetrievalAugmentedGeneration

OProver: A Unified Framework for Agentic Formal Theorem Proving
👤 David Ma, Kaijing Ma, Shawn Guo et al.

🎯 Task
Formal theorem proving

💡 Idea
Instead of bolting retrieval and self-repair onto a fixed prover at test time, OProver trains that agentic loop itself: multi-round proof revision conditioned on retrieved verified proofs and raw Lean feedback, with new verified proofs and repair traces fed back into training.

✨ Why it's interesting
OProver-32B gets best Pass@32 on MiniF2F 93.3%, ProverBench 58.2%, PutnamBench 11.3%, and second on MathOlympiad 22.8% and ProofNet 33.2%.

💻 Repo
⭐ multimodal-art-projection/OProver — 7 stars

🔗 paper

via @Papers.Data.Code

📊 Dataset #Dataset #Tabular #Education #MentalHealth

Impact of Ai on Students
👤 laveshjadon

🎯 Task
Student outcome and burnout prediction

💡 Idea
50,000 student records with 16 features spanning academic profile, GenAI usage, study habits, institutional policy, anxiety, skill retention, and burnout, with targets for GPA regression, skill retention, and burnout classification.

✨ Why it's interesting
Makes it possible to model academic and well-being outcomes against AI usage and policy in one complete, balanced student dataset.

Ⓢ 50,000 samples, 16 columns, CSV

🔗 dataset

via @Papers.Data.Code

💻 Repo #Repo #LLM #FederatedLearning #Lora

Smart Fed
👤 benmagnifico

🎯 Task
federated LLM fine-tuning with LoRA reuse

💡 Idea
Compose a frozen pool of existing task LoRAs into a federated adapter by splitting them into rank-wise experts and learning a small input-conditioned router that selects and combines them on each client.

✨ Why it's interesting
Cuts training, communication, and energy cost versus federated train-from-scratch baselines, and beats both knowledge-free and knowledge-reuse baselines on three skill-composition tasks.

💻 Repo
⭐ benmagnifico/SmartFed — 15 stars (+15 3d)


via @Papers.Data.Code

📄 Paper #Paper #Multimodal #ImageGeneration #VideoGeneration

Lance: Unified Multimodal Modeling by Multi-Task Synergy
👤 Fengyi Fu, Mengqi Huang, Shaojin Wu et al.

🎯 Task
Unified multimodal understanding and generation

💡 Idea
Instead of one shared visual path or bolted-on modules, Lance uses a shared interleaved multimodal context with dual MoE streams: one expert for text+semantic understanding, one for VAE-latent generation, plus modality-aware RoPE and staged multi-task training.

✨ Why it's interesting
With only 3B activated params and a 128-GPU budget, it substantially outperforms prior open-source unified models on image and video generation while keeping strong understanding.

💻 Repo
⭐ bytedance/Lance — 314 stars

🔗 paper

via @Papers.Data.Code

📄 Paper #Paper #Multimodal #ComputerUseAgents #Benchmarking

OpenComputer: Verifiable Software Worlds for Computer-Use Agents
👤 Jinbiao Wei, Qianran Ma, Yilun Zhao et al.

🎯 Task
Computer-use agent evaluation and benchmark generation

💡 Idea
Instead of screenshot or LLM-judge evaluation, uses app-specific state verifiers over real software, then self-refines them from execution disagreements to synthesize and score realistic desktop tasks automatically.

✨ Why it's interesting
Covers 33 apps and 1,000 tasks. Verifiers align better with humans than LLM judges. Best agent hits 68.3% success; open models drop sharply vs OSWorld.

💻 Repo
⭐ echo0715/OpenComputer

🔗 paper

via @Papers.Data.Code

📄 Paper #Paper #LLM #ReinforcementLearning #LongContext

GoLongRL: Capability-Oriented Long Context Reinforcement Learning with Multitask Alignment
👤 Minxuan Lv, Tiehua Mei, Tanlong Du et al.

🎯 Task
Long-context reinforcement learning for LLMs

💡 Idea
Instead of retrieval-path-heavy QA data and uniform rewards, it trains on 9 long-context capability tasks with task-native metrics, then replaces vanilla GRPO's prompt-level scaling with task-mean normalization plus difficulty-adaptive reweighting.

✨ Why it's interesting
On Qwen3-30B-A3B, average long-context score rises from 60.1 to 69.8; TMN-Reweight reaches 63.0 on 4B vs 62.2 with vanilla GRPO.

💻 Repo
⭐ xiaoxuanNLP/GoLongRL

🔗 paper

via @Papers.Data.Code

🔥 Repo #Repo #LLM #Pretraining #HierarchicalReasoningModel

Hrm Text
👤 sapientinc

🎯 Task
efficient foundation model pretraining

💡 Idea
Pretrain HRM text generation models from scratch on 8-16 H100s with built-in data packing, distributed training, benchmark evaluation, and checkpoint export to Transformers format.

✨ Why it's interesting
Claims 130-600x less compute and 150-900x less data; reference runs train 0.6B-1B models in 46-50 hours on 8-16 H100s.

💻 Repo
⭐ sapientinc/HRM-Text — 580 stars (+580 3d)
Python


via @Papers.Data.Code

📄 Paper #Paper #CV #VideoGeneration #Quantization

LongLive-2.0: An NVFP4 Parallel Infrastructure for Long Video Generation
👤 Yukang Chen, Luozhou Wang, Wei Huang et al.

🎯 Task
Long video generation infrastructure

💡 Idea
Instead of complex multi-stage long-video pipelines, it directly fine-tunes an AR diffusion model and co-designs sequence parallelism with teacher forcing. Balanced SP pairs clean/noisy chunks per rank, while end-to-end NVFP4 enables W4A4 inference, KV-cache compression, and async decoding.

✨ Why it's interesting
Up to 2.15x faster training and 1.84x faster inference; 45.7 FPS, 21.9 ms/frame, and memory cut from 35.4 GB to 19.4 GB.

💻 Repo
⭐ NVlabs/LongLive — 1.4k stars
Python

🔗 paper

via @Papers.Data.Code

📄 Paper #Paper #LLM #ReinforcementLearning #Reasoning

You Only Need Minimal RLVR Training: Extrapolating LLMs via Rank-1 Trajectories
👤 Zhepei Wei, Xinyu Zhu, Wei-Lin Chen et al.

🎯 Task
LLM RL checkpoint extrapolation

💡 Idea
Instead of running full RLVR, estimate each tensor's dominant rank-1 update direction from early checkpoints and linearly extrapolate its coefficient. Unlike raw weight or logit extrapolation, it uses the low-rank RLVR geometry as a denoised predictor.

✨ Why it's interesting
With 15-20% of RLVR steps, RELEX matches or nears full RLVR on MATH: 71.6 vs 71.5, 85.6 vs 85.5, 87.4 vs 88.5 across 3 models.

💻 Repo
⭐ weizhepei/RELEX

🔗 paper

via @Papers.Data.Code

📄 Paper #Paper #LLM #Reasoning #ReinforcementLearning

Anti-Self-Distillation for Reasoning RL via Pointwise Mutual Information
👤 Guobin Shen, Xiang Cheng, Chenxiao Zhao et al.

🎯 Task
Reasoning reinforcement learning for math and code

💡 Idea
Instead of pulling the policy toward a privileged self-teacher that rewards shortcut tokens and suppresses deliberation, AntiSD reverses the signal: ascend student-teacher JSD, with an entropy gate to stop once teacher confidence collapses.

✨ Why it's interesting
Across 5 models (4B-30B), it matches GRPO in 2-10x fewer steps and improves final avg accuracy by up to 11.5 points.

💻 Repo
⭐ FloyedShen/AntiSD — 11 stars
Python

🔗 paper

via @Papers.Data.Code

📊 Dataset #Dataset #LLM #MultiTurnDialogue #UserModeling

ThoughtTrace
👤 SCAI-JHU

🎯 Task
User modeling in multi-turn dialogue

💡 Idea
2,155 real-world conversations from 1,058 users across 20 LLMs, with 10,174 message-level thought annotations: 7 reason types on user turns and 5 reaction types on assistant turns.

✨ Why it's interesting
Makes latent user intent and satisfaction measurable from real chats; authors show gains for behavior prediction (+41.7%) and alignment (+25.6% win rate).

Ⓢ 2,155 conversations, 10,174 thought annotations

🔗 dataset 🔗 paper 🔗 repo

via @Papers.Data.Code

📄 Paper #Paper #Audio #SpeechRecognition #Robustness

Mega-ASR: Towards In-the-wild^2 Speech Recognition via Scaling up Real-world Acoustic Simulation
👤 Zhifei Xie, Kaiyu Pang, Haobin Zhang et al.

🎯 Task
Robust automatic speech recognition

💡 Idea
Instead of training on isolated mild noise, it scales to 54 physically plausible compound acoustic scenarios and trains ASR progressively from acoustic perception to semantic recovery, then uses WER-gated token- vs sentence-level rewards to handle both local errors and hallucinated/omitted transcripts.

✨ Why it's interesting
Beats prior SOTA on adverse ASR: 45.69% vs 54.01% on VOiCES R4-B-F, 21.49% vs 29.34% on NOIZEUS Sta-0; >30% relative WER drop on compound scenarios.

💻 Repo
⭐ xzf-thu/Mega-ASR

🔗 paper

via @Papers.Data.Code

📋 Weekly Digest · May 16 – May 23
#WeeklyDigest

📄 Papers

Achieving Gold-Medal-Level Olympiad Reasoning via Simple and Unified Scaling
#Reasoning #ReinforcementLearning #TestTimeScaling
Unified SFT-RL scaling ⟶ reaches IMO gold line
→ Learn more...

Causal Forcing++: Scalable Few-Step Autoregressive Diffusion Distillation for Real-Time Interactive Video Generation
#VideoGeneration #DiffusionModels #Distillation
Causal Forcing++ distillation ⟶ real-time 1-2 step video
→ Learn more...

Self-Distilled Agentic Reinforcement Learning
#ReinforcementLearning #KnowledgeDistillation #LLMAgents
Gated self-distillation RL ⟶ beats GRPO on LLM agents
→ Learn more...

OpenComputer: Verifiable Software Worlds for Computer-Use Agents
#ComputerUseAgents #Benchmarking #Evaluation
Verifier-grounded desktop tasks ⟶ auditable agent evaluation
→ Learn more...

LongLive-2.0: An NVFP4 Parallel Infrastructure for Long Video Generation
#VideoGeneration #Quantization #ParallelTraining
NVFP4 long video stack ⟶ faster training, inference, lower memory
→ Learn more...

Training Long-Context Vision-Language Models Effectively with Generalization Beyond 128K Context
#LongContextModeling #VisionLanguageModels #DocumentVQA
Long-doc VQA pretraining ⟶ extends LVLMs to 128K+
→ Learn more...

💻 Repos

sapientinc/HRM-Text ⭐
#Pretraining #HierarchicalReasoningModel #Flashattention
HRM text pretraining ⟶ trains 0.6B-1B on 8-16 H100s
→ Learn more...

facebookresearch/vggt-omega ⭐
#DepthEstimation #CameraPose #3DReconstruction
Multi-image feed-forward model ⟶ infers camera pose and depth
→ Learn more...

yyfz/Warp-as-History ⭐
#VideoGeneration #CameraControl #Lora
Warped history conditioning ⟶ camera-controlled video generation
→ Learn more...

📊 Datasets

Orchard
#SoftwareEngineering #ToolUse #GuiAgent
Dual agent trajectories ⟶ train and evaluate coding GUI agents
→ Learn more...

ThoughtTrace
#MultiTurnDialogue #UserModeling #Alignment
ThoughtTrace dataset ⟶ measures latent intent
→ Learn more...

➡️ Tomorrow — NLP & LLM Monthly

via @Papers.Data.Code