[GUEST ACCESS MODE: Data is scrambled or limited to provide examples. Make requests using your API key to unlock full data. Check https://lunarcrush.ai/auth for authentication information.]

[@HuggingPapers](/creator/twitter/HuggingPapers)
"Kuaishou Technology just introduced PhysMaster It teaches video generation models to understand physics creating highly realistic and physically plausible videos with a novel reinforcement learning framework"  
[X Link](https://x.com/HuggingPapers/status/1978979708663443859) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-17T00:21Z 7729 followers, 7486 engagements


"Samsung's Tiny Recursive Model (TRM) masters complex reasoning With just 7M parameters TRM outperforms large LLMs on hard puzzles like Sudoku & ARC-AGI. This "Less is More" approach redefines efficiency in AI using less than XXXX% of competitors' parameters"  
[X Link](https://x.com/HuggingPapers/status/1975956602160300051) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-08T16:09Z 7728 followers, 44.1K engagements


"ByteDance just released FaceCLIP on Hugging Face A new vision-language model specializing in understanding and generating diverse human faces. Dive into the future of facial AI"  
[X Link](https://x.com/HuggingPapers/status/1977812522398060716) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-13T19:03Z 7730 followers, 54.2K engagements


"NVIDIA introduces QeRL: Efficient & reliable RL for LLMs on a single H100 GPU This framework enables 32B LLM training on minimal hardware delivering 1.5x speedup in rollout. Surprisingly quantization noise actually boosts exploration helping models discover better strategies"  
[X Link](https://x.com/HuggingPapers/status/1977949560149315847) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-14T04:08Z 7727 followers, 6125 engagements


"Facebook just dropped HoneyBee a massive new dataset for vision-language reasoning on Hugging Face It contains 2.5M high-quality examples with chain-of-thought solutions pushing VLM performance to new SOTA"  
[X Link](https://x.com/HuggingPapers/status/1978614436148474335) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-16T00:10Z 7730 followers, 20.5K engagements


"AI for Service introduces proactive assistance with AI glasses This new paradigm anticipates user needs and provides real-time help without explicit prompts transforming reactive AI into an adaptive companion for daily life"  
[X Link](https://x.com/HuggingPapers/status/1979278656951456014) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-17T20:09Z 7730 followers, XXX engagements


"Alibaba Group just dropped ImagerySearch A new adaptive test-time search for video generation pushing beyond semantic constraints. It dynamically adjusts inference & rewards based on prompt relationships enabling incredible visual coherence for imaginative scenarios"  
[X Link](https://x.com/HuggingPapers/status/1979400364060282912) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-18T04:13Z 7727 followers, 4090 engagements


"Unlocking true embodied AI: Introducing BEAR The first benchmark evaluating Multimodal LLMs on XX atomic embodied capabilities across 4469 multimodal entries. Discover what your models are truly capable of"  
[X Link](https://x.com/HuggingPapers/status/1979460142992138270) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-18T08:10Z 7727 followers, 1251 engagements


"Dive deeper into the theory and practical frameworks of Vibe Coding. This paper systematically analyzes over 1000 research papers covering LLMs for coding coding agents environments and feedback. Read the full paper on Hugging Face: Explore the accompanying Awesome Vibe Coding list:"  
[X Link](https://x.com/HuggingPapers/status/1979590056802156675) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-18T16:47Z 7729 followers, XXX engagements


"ByteDance just released BFS-Prover-V2 a state-of-the-art Lean4 tactic generation model on Hugging Face. It achieves XXXXX% on miniF2F and XXXX% on ProofNet setting new benchmarks in automated theorem proving"  
[X Link](https://x.com/HuggingPapers/status/1975079826093396215) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-06T06:05Z 7730 followers, 1659 engagements


"ByteDance just released Artificial Hippocampus Networks (AHN) on Hugging Face A novel architecture for long-context LLMs that continuously compresses out-of-window information greatly reducing memory and computation"  
[X Link](https://x.com/HuggingPapers/status/1976107974939816308) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-09T02:10Z 7730 followers, 32.5K engagements


"ByteDance just released Artificial Hippocampus Networks (AHN) on Hugging Face. AHN transforms lossless memory into fixed-size compressed representations for efficient long-context modeling integrating with models like Qwen 2.5"  
[X Link](https://x.com/HuggingPapers/status/1976140230303395991) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-09T04:18Z 7730 followers, 1803 engagements


"Unveiling the first comprehensive survey on Vibe Coding with LLMs Explore a new era where AI agents autonomously code validated by outcomes. This paper formalizes "Vibe Coding" & maps X dev models laying a foundation for future human-AI collaboration"  
[X Link](https://x.com/HuggingPapers/status/1979590046698045472) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-18T16:47Z 7729 followers, 1040 engagements


"Microsoft introduces BitNet Distillation Fine-tunes full-precision LLMs into 1.58-bit precision achieving comparable performance with 10x memory savings and 2.65x faster CPU inference for specific tasks"  
[X Link](https://x.com/HuggingPapers/status/1979641856687263801) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-18T20:12Z 7729 followers, 3344 engagements


"Huawei Research just unveiled SINQ on Hugging Face A novel calibration-free quantization technique that enables state-of-the-art LLM performance while drastically reducing memory usage"  
[X Link](https://x.com/HuggingPapers/status/1973906002001936577) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-03T00:20Z 7692 followers, 18.7K engagements


"NVIDIA just released its gpt-oss-120b Eagle model for accelerated AI inference on Hugging Face. It uses speculative decoding with TensorRT Model Optimizer for highly efficient text generation in AI agent applications"  
[X Link](https://x.com/HuggingPapers/status/1975645612403298470) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-07T19:33Z 7699 followers, 1084 engagements


"NVIDIA just released Fast-dLLM v2 on Hugging Face It delivers up to 2.5x faster LLM inference over standard decoding achieving state-of-the-art efficiency in diffusion LLMs with 500x less fine-tuning data. Get ready for practical fast and accurate LLMs"  
[X Link](https://x.com/HuggingPapers/status/1975837289210073193) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-08T08:14Z 7689 followers, XXX engagements


"Salesforce AI Research unveils CoDA on Hugging Face A 1.7B diffusion model for code generation that outperforms 7B models with bidirectional context and blazing fast inference"  
[X Link](https://x.com/HuggingPapers/status/1975839564242575489) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-08T08:23Z 7692 followers, 1298 engagements


"Drax: Discrete Flow Matching brings a new era of efficient ASR This novel framework enables efficient parallel decoding in Automatic Speech Recognition. It achieves state-of-the-art accuracy comparable to autoregressive models with significantly better accuracy-efficiency trade-offs"  
[X Link](https://x.com/HuggingPapers/status/1976017331718381761) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-08T20:10Z 7692 followers, 12.5K engagements


"Explore Drax and its discrete flow matching approach for ASR: Read the paper: Access the models:"  
[X Link](https://x.com/HuggingPapers/status/1976017341629489412) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-08T20:10Z 7698 followers, XXX engagements


"AgentFlow: In-the-Flow Optimization for LLM Agents A new trainable modular agentic system that optimizes its planner live within the multi-turn loop. Achieve +14.9% on search +14.0% on agentic reasoning and +14.5% on math outperforming models like GPT-4o with a 7B backbone"  
[X Link](https://x.com/HuggingPapers/status/1976079766978502816) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-09T00:18Z 7692 followers, 2354 engagements


"Explore the groundbreaking paper and models here:"  
[X Link](https://x.com/HuggingPapers/status/1976107984754491657) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-09T02:10Z 7692 followers, 1640 engagements


"Explore how AHN enhances models like Qwen 2.5-14B for ultra-long contexts. Find the model here: Learn more on GitHub:"  
[X Link](https://x.com/HuggingPapers/status/1976140239753167154) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-09T04:18Z 7692 followers, XXX engagements


"Cache-to-Cache: LLMs communicate beyond text This new paradigm allows Large Language Models to directly share rich semantic information via KV-Cache projection bypassing slow text generation. It achieves up to XXXX% higher accuracy and 2x speedup"  
[X Link](https://x.com/HuggingPapers/status/1976198778227986465) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-09T08:11Z 7697 followers, 1240 engagements


"Explore the Cache-to-Cache paradigm for multi-LLM communication Get the full details code and more on Hugging Face. Paper: Code: Project page:"  
[X Link](https://x.com/HuggingPapers/status/1976198787543507401) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-09T08:11Z 7699 followers, XXX engagements


"MM-HELIX just launched on Hugging Face This new platform significantly boosts multimodal long-chain reflective reasoning in MLLMs tackling complex real-world problems with a novel benchmark dataset and an adaptive training strategy"  
[X Link](https://x.com/HuggingPapers/status/1976500732825305406) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-10T04:11Z 7699 followers, 1443 engagements


"Stanford unveils AgentFlow: In-the-flow Agentic AI A new trainable modular system that learns live to plan & use tools outperforming even GPT-4o on reasoning tasks with a 7B model. Huge gains: +14.9% search +14.5% math"  
[X Link](https://x.com/HuggingPapers/status/1976804812852543730) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-11T00:19Z 7692 followers, 2106 engagements


"Dive into AgentFlow's Flow-GRPO algorithm. Explore the code try the demo and see how to train your own modular agents on Hugging Face Paper: Demo: Model:"  
[X Link](https://x.com/HuggingPapers/status/1976804822402953262) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-11T00:19Z 7699 followers, 1135 engagements


"ChemMAS: A multi-agent AI for evidence-based chemical reaction condition reasoning This new system goes beyond "what" to explain the "why" behind reaction recommendations achieving 10-35% gains over SOTA. It brings human-trustable rationales to scientific discovery"  
[X Link](https://x.com/HuggingPapers/status/1976924445311660524) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-11T08:14Z 7692 followers, 1204 engagements


"Discover how ChemMAS uses mechanistic grounding multi-channel recall and agentic debate to predict and explain chemical reactions. Explore the full paper on Hugging Face:"  
[X Link](https://x.com/HuggingPapers/status/1976924454799163859) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-11T08:14Z 7693 followers, XXX engagements


"Explore MASA: Meta-Awareness via Self-Alignment This innovative framework for reasoning models is on Hugging Face. Find the paper and code here:"  
[X Link](https://x.com/HuggingPapers/status/1976984472181440627) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-11T12:13Z 7691 followers, XXX engagements


"Ready to explore "Thinking with Camera" Discover Puffin a unified multimodal model for camera-centric understanding & generation on Hugging Face Paper: Demo:"  
[X Link](https://x.com/HuggingPapers/status/1977684087944163659) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-13T10:33Z 7700 followers, 1134 engagements


"Get the full scoop on the Hugging Face paper page: Explore the PRBench dataset: Try the PRAgent demo:"  
[X Link](https://x.com/HuggingPapers/status/1977768728885432616) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-13T16:09Z 7697 followers, XXX engagements


"NVIDIA just launched Nemotron-Personas-India on Hugging Face This groundbreaking dataset offers X million richly diverse synthetically-generated Indian personas grounded in real-world demographics across language age occupation and more in English and Hindi"  
[X Link](https://x.com/HuggingPapers/status/1977857831563833711) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-13T22:03Z 7697 followers, 1059 engagements


"Pixel-space generative models hit new SOTA with EPG AMAP Alibaba NVIDIA & Caltech introduce EPG a novel two-stage training framework that achieves state-of-the-art pixel-space diffusion (FID XXXX on ImageNet-256 with XX NFE) and consistency models (FID XXXX in X step)"  
[X Link](https://x.com/HuggingPapers/status/1978311764954472589) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-15T04:07Z 7691 followers, 7555 engagements


"Read the paper on Hugging Face: Check out the code:"  
[X Link](https://x.com/HuggingPapers/status/1978311774605623316) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-15T04:07Z 7700 followers, 1077 engagements


"Alibaba's DAMO Academy introduces LCO-Embedding for omnimodal representation learning It achieves new SOTA on MIEB (Massive Image Embedding Benchmark) & supports audio/video revealing a new Generation-Representation Scaling Law"  
[X Link](https://x.com/HuggingPapers/status/1978372913595330584) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-15T08:10Z 7697 followers, 1639 engagements


"Google just launched google/jefferson-test on Hugging Face A brand new Transformers model ready for the community to define its future. Who will be the first to uncover its potential and enrich its model card"  
[X Link](https://x.com/HuggingPapers/status/1978531665707246047) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-15T18:41Z 7701 followers, 1413 engagements


"Hugging Face releases a definitive tutorial for Robot Learning Dive into everything from RL fundamentals to cutting-edge generalist policies. This comprehensive guide equips you with concepts & practical code examples in our lerobot library. Start your journey into autonomous systems today"  
[X Link](https://x.com/HuggingPapers/status/1978553387831406866) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-15T20:07Z 7699 followers, XXX engagements


"Explore the full tutorial with hands-on code examples in the lerobot library featuring datasets on the Hugging Face Hub. Contribute to future editions Paper: Space:"  
[X Link](https://x.com/HuggingPapers/status/1978553397465776213) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-15T20:07Z 7698 followers, XXX engagements


"PhysMaster captures physical knowledge as representation optimized via RLHF (DPO). It's a plug-and-play solution for physically-aware video generation Project page: Paper on HF:"  
[X Link](https://x.com/HuggingPapers/status/1978979717924475008) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-17T00:21Z 7696 followers, XXX engagements


"When models lie we learn: Introducing PsiloQA A new multilingual dataset annotated with span-level hallucinations across XX languages providing fine-grained supervision to detect LLM inaccuracies"  
[X Link](https://x.com/HuggingPapers/status/1979225831189762393) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-17T16:39Z 7696 followers, XXX engagements


"Zhipu AI & ByteDance introduce GLM-4.5 on Hugging Face A new Mixture-of-Experts LLM designed for Agentic Reasoning and Coding (ARC) tasks. Achieves top performance with powerful hybrid reasoning methods"  
[X Link](https://x.com/HuggingPapers/status/1954817129049428265) [@HuggingPapers](/creator/x/HuggingPapers) 2025-08-11T08:08Z 7723 followers, 11.6K engagements


"Kwai Keye Team at Kuaishou unveils Keye-VL 1.5: a powerful multimodal LLM excelling in video understanding with a novel Slow-Fast encoding strategy 128K context window and advanced RL training"  
[X Link](https://x.com/HuggingPapers/status/1964300251839230059) [@HuggingPapers](/creator/x/HuggingPapers) 2025-09-06T12:10Z 7716 followers, 1291 engagements


"Kuaishou's Kling-Avatar introduces a new framework for high-fidelity long-duration avatar animation. It unifies multimodal instructions to generate photorealistic videos with vivid emotions and precise lip-sync"  
[X Link](https://x.com/HuggingPapers/status/1966594639277756694) [@HuggingPapers](/creator/x/HuggingPapers) 2025-09-12T20:07Z 7723 followers, XXX engagements


"Alibaba Group & partners unveil MMR1: Revolutionizing multimodal reasoning with less data MMR1 introduces Variance-Aware Sampling (VAS) for stable RL fine-tuning. Tackles unstable optimization & scarce high-quality data. Releasing massive open datasets (1.6M CoT 15k RL QA) & models (3B 7B 32B) for the community"  
[X Link](https://x.com/HuggingPapers/status/1971487864807469236) [@HuggingPapers](/creator/x/HuggingPapers) 2025-09-26T08:11Z 7717 followers, 6342 engagements


"How to make LLM agents smarter with less data Tree-GRPO from Alibaba Group's AMAP-ML introduces a novel tree-search RL framework drastically cutting rollout budgets and boosting performance in complex multi-turn tasks"  
[X Link](https://x.com/HuggingPapers/status/1971607402068787323) [@HuggingPapers](/creator/x/HuggingPapers) 2025-09-26T16:06Z 7717 followers, XXX engagements


"Alibaba Group's Q-Tuning: Efficient LLM Supervised Fine-Tuning This unified framework uses the Error-Uncertainty Plane to jointly prune samples and tokens. It achieves +38% improvement on SmolLM2-1.7B with only XXXX% data surpassing full-data SFT"  
[X Link](https://x.com/HuggingPapers/status/1973480203339870425) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-01T20:08Z 7717 followers, 1215 engagements


"Meta just released SSDD (Single-Step Diffusion Decoder) on Hugging Face It's a novel image tokenizer with a diffusion decoder that achieves higher reconstruction quality and faster sampling than traditional VAEs"  
[X Link](https://x.com/HuggingPapers/status/1975484440475505039) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-07T08:52Z 7724 followers, 22.9K engagements


"Meta unveils "Early Experience" for language agents This new paradigm lets AI agents learn & improve from their own actions using future states as supervision without requiring reward signals or extensive human data. It's a bridge to truly experience-driven AI"  
[X Link](https://x.com/HuggingPapers/status/1976621792354824224) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-10T12:12Z 7725 followers, 47.2K engagements


"MemMamba: A breakthrough in ultra-long sequence modeling It rethinks memory patterns in State Space Models achieving stable performance at massive context lengths and delivering a XX% inference speedup"  
[X Link](https://x.com/HuggingPapers/status/1976681769954169080) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-10T16:10Z 7714 followers, 11.8K engagements


"ByteDance just released veAgentBench on Hugging Face A new benchmark to rigorously evaluate the capabilities of next-generation AI agents"  
[X Link](https://x.com/HuggingPapers/status/1976848050191843583) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-11T03:11Z 7729 followers, 1116 engagements


"Agentic Context Engineering (ACE) is here A new framework for self-improving LLMs by evolving contexts as dynamic playbooks preventing collapse and outperforming baselines on agent and domain-specific tasks"  
[X Link](https://x.com/HuggingPapers/status/1976863234432155767) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-11T04:11Z 7723 followers, 1576 engagements


"Boost AI Reasoning with Meta-Awareness Introducing MASA: a self-alignment RL framework boosting reasoning models' meta-awareness using internal signals. It filters prompts & prunes unproductive rollouts speeding up training by 1.28x gaining XXXX% on AIME25 and improving generalization across XX benchmarks"  
[X Link](https://x.com/HuggingPapers/status/1976984462433919351) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-11T12:13Z 7717 followers, 16.8K engagements


"When Thoughts Meet Facts: New from Amazon & KAIST LCLMs can process vast contexts but struggle with reasoning. ToTAL introduces reusable "thought templates" that structure evidence guiding multi-hop inference with factual documents"  
[X Link](https://x.com/HuggingPapers/status/1977044479149281566) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-11T16:11Z 7702 followers, 17.9K engagements


"DreamOmni2: Multimodal Instruction-based Editing and Generation by ByteDance This unified framework pioneers multimodal instruction-based editing & generation. It handles text & image inputs for both concrete objects and abstract concepts achieving impressive results"  
[X Link](https://x.com/HuggingPapers/status/1977451373953241284) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-12T19:08Z 7729 followers, 19.5K engagements


"📷Thinking with Camera" just dropped on Hugging Face A unified multimodal model Puffin introduces a novel paradigm for camera-centric understanding and generation by treating camera as language. It interprets and creates scenes from arbitrary viewpoints enhancing spatial intelligence for tasks like world exploration"  
[X Link](https://x.com/HuggingPapers/status/1977684084961734785) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-13T10:33Z 7703 followers, 12.7K engagements


"Academic promotion just got smarter with ByteDance's AutoPR It automates turning research papers into engaging social media posts boosting watch time by XXX% and likes by XXX% with its multi-agent framework"  
[X Link](https://x.com/HuggingPapers/status/1977768719272103966) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-13T16:09Z 7727 followers, 1846 engagements


"New inference method TAG fights diffusion model hallucinations Introducing Tangential Amplifying Guidance (TAG): a training-free plug-and-play method for diffusion models that significantly reduces hallucinations and boosts sample quality by steering generation to high-probability regions"  
[X Link](https://x.com/HuggingPapers/status/1977828987167494207) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-13T20:09Z 7727 followers, 7705 engagements


"Introducing Latent Refinement Decoding (LRD) A new two-stage framework enhancing diffusion-based language models. It addresses information loss and premature commitment for more globally consistent parallel generation. This leads to higher accuracy and significant speedups"  
[X Link](https://x.com/HuggingPapers/status/1978131428157055217) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-14T16:10Z 7724 followers, 1288 engagements


"LRD achieves up to 10.6x faster decoding while improving accuracy across various coding and reasoning tasks Experience a powerful versatile alternative for parallel sequence generation. Read the full paper on Hugging Face:"  
[X Link](https://x.com/HuggingPapers/status/1978131437766222105) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-14T16:11Z 7706 followers, XXX engagements


"ByteDance's RLFR redefines LLM reinforcement learning It introduces "flow rewards" from the latent space of LLMs for efficient and reliable reasoning self-improvement. Outperforms existing methods on language & multimodal benchmarks"  
[X Link](https://x.com/HuggingPapers/status/1978191203284738479) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-14T20:08Z 7727 followers, 7019 engagements


"Shanghai AI Lab unveils VPPO for multimodal RL This new method spotlights "token perception" to make LVLMs reason better. It achieves state-of-the-art results with superior stability & faster convergence on X benchmarks"  
[X Link](https://x.com/HuggingPapers/status/1978255601961488670) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-15T00:24Z 7729 followers, 12.3K engagements


"Spatial Forcing enhances robot's 3D perception This plug-and-play strategy aligns VLA models with 3D foundation models to gain spatial awareness. Achieve SOTA in robotic tasks with 3.8x faster training & XX% higher real-world success without explicit 3D sensors"  
[X Link](https://x.com/HuggingPapers/status/1978494410665992418) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-15T16:13Z 7723 followers, 11.1K engagements


"Attention reveals the hidden rhythm of LLM reasoning Researchers from Shanghai Jiao Tong University and Alibaba Group uncover a "preplan-and-anchor" mechanism in LLM attention transforming opaque reasoning into a legible blueprint for fine-grained policy optimization"  
[X Link](https://x.com/HuggingPapers/status/1978674474376421408) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-16T04:08Z 7724 followers, 8901 engagements


"Their new work "Attention Illuminates LLM Reasoning" introduces novel RL strategies that dynamically assign credit to critical reasoning steps. This makes LLM optimization more transparent and effective Learn more and discuss the paper on Hugging Face:"  
[X Link](https://x.com/HuggingPapers/status/1978674483796824493) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-16T04:08Z 7702 followers, 1009 engagements


"ByteDance just released Sa2VA on Hugging Face The first unified model for dense grounded understanding of images and videos. Combines SAM2 with LLaVA for SOTA segmentation and visual QA"  
[X Link](https://x.com/HuggingPapers/status/1978734063973179439) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-16T08:05Z 7724 followers, 7815 engagements


"Sa2VA delivers state-of-the-art visual Q&A prompt understanding and object segmentation for both images and videos. A breakthrough in grounded multimodal AI Explore the model on Hugging Face:"  
[X Link](https://x.com/HuggingPapers/status/1978734073448198580) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-16T08:05Z 7727 followers, 1038 engagements


"Tencent's FlashWorld: high-quality 3D scenes fast Generate stunning 3D scenes from a single image or text prompt. Achieve high-quality results in just X seconds on an A100/A800 GPU That's a 10-100x speedup over prior methods"  
[X Link](https://x.com/HuggingPapers/status/1978735016147333468) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-16T08:09Z 7725 followers, 16.6K engagements


"ByteDance just released Sa2VA on Hugging Face. This MLLM marries SAM2 with LLaVA for dense grounded understanding of images & videos offering SOTA performance in segmentation grounding and QA"  
[X Link](https://x.com/HuggingPapers/status/1978745567258829153) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-16T08:51Z 7730 followers, 29.4K engagements


"Tencent Hunyuan Team just released Bee: a full-stack suite for advanced open MLLMs It introduces Honey-Data-15M a high-quality corpus & HoneyPipe a transparent curation pipeline powering the new state-of-the-art Bee-8B model"  
[X Link](https://x.com/HuggingPapers/status/1978861237564744143) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-16T16:30Z 7724 followers, 1240 engagements


"Bee-8B achieves new SOTA for fully open MLLMs competitive with semi-open models. Explore the full suite & model weights on Hugging Face Paper: Model Hub:"  
[X Link](https://x.com/HuggingPapers/status/1978861247077425415) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-16T16:31Z 7724 followers, XXX engagements


"ByteDance just dropped WithAnyone on Hugging Face A new diffusion model for controllable & ID-consistent image generation. It tackles the "copy-paste" artifact by preserving identity across diverse poses & expressions not just replicating faces"  
[X Link](https://x.com/HuggingPapers/status/1979158921311875382) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-17T12:13Z 7727 followers, 1111 engagements


"Discover **WithAnyone** by ByteDance for controllable ID-consistent image generation It effectively mitigates "copy-paste" artifacts across varied poses & expressions. Find the paper models & demo on Hugging Face: Paper: Model: Demo:"  
[X Link](https://x.com/HuggingPapers/status/1979158932296794372) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-17T12:13Z 7727 followers, XXX engagements


"PsiloQA uses an automated pipeline for scalable data generation showing encoder-based models achieve state-of-the-art performance across languages and enable robust knowledge transfer. Read the paper: Explore the dataset:"  
[X Link](https://x.com/HuggingPapers/status/1979225840635335111) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-17T16:39Z 7725 followers, XXX engagements


"Discover Alpha-Service: a unified framework based on AI glasses that proactively assists users in real-time from museum tours to shopping. Explore the paper and join the discussion on Hugging Face"  
[X Link](https://x.com/HuggingPapers/status/1979278666598346905) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-17T20:09Z 7730 followers, XXX engagements


"Dive into ImagerySearch It excels at generating imaginative videos outperforming baselines on the brand new LDT-Bench a benchmark for long-distance semantic prompts. Explore the paper on Hugging Face:"  
[X Link](https://x.com/HuggingPapers/status/1979400373933617546) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-18T04:13Z 7728 followers, XXX engagements


"UniMoE-Audio: Unified Speech & Music Gen A Dynamic-Capacity Mixture-of-Experts model from HIT-TMG that seamlessly unifies high-fidelity speech & expressive music. Achieves state-of-the-art on major benchmarks tackling task conflicts & data imbalances"  
[X Link](https://x.com/HuggingPapers/status/1979589426104660154) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-18T16:44Z 7730 followers, 3045 engagements


"Tencent introduces LaSeR: Reinforcement Learning with Last-Token Self-Rewarding This new algorithm optimizes LLM reasoning & self-rewarding with minimal cost. It aligns last-token scores with true rewards boosting performance & efficiency with just one extra token inference"  
[X Link](https://x.com/HuggingPapers/status/1979705973892902939) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-19T00:27Z 7729 followers, 10.8K engagements


"ByteDance just launched ReSA on Hugging Face. This new 80K synthetic dataset trains LLMs with an "Answer-Then-Check" strategy boosting jailbreak defense and enabling safe helpful responses for sensitive queries"  
[X Link](https://x.com/HuggingPapers/status/1980113019511427195) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-20T03:25Z 7729 followers, 12.2K engagements


"Self-Forcing++ for minute-scale video generation ByteDance's new method generates high-quality videos up to X min XX sec It scales diffusion models without long-video teachers or retraining preserving fidelity and consistency"  
[X Link](https://x.com/HuggingPapers/status/1974688371340648857) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-05T04:09Z 7730 followers, 17.5K engagements


"ServiceNow's Apriel-1.5-15B-Thinker: Frontier AI on a single GPU This 15B-parameter open-weights multimodal model achieves state-of-the-art reasoning performance matching models 8-10x its sizeall without an RL phase"  
[X Link](https://x.com/HuggingPapers/status/1975171993755427076) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-06T12:11Z 7730 followers, 16.2K engagements


"ReSA prevents over-refusal and offers superior safety performance. It even allows models to provide thoughtful supportive answers to sensitive questions like self-harm rather than just refusing. Dive in: Paper:"  
[X Link](https://x.com/HuggingPapers/status/1980113029313462469) [@HuggingPapers](/creator/x/HuggingPapers) 2025-10-20T03:25Z 7730 followers, XXX engagements