AI: post transformers

mcgrof
  • 5,0 (2)
  • ТЕХНОЛОГИИ
  • ЕЖЕДНЕВНО

The transformer architecture revolutionized the world of Neural Networks. It was a springboard for what we know today as modern artificial intelligence. This podcast focuses on modern state of the art research paper reviews starting from the transformer and on.

  1. Jet-RL: Stable On-Policy Reinforcement Learning with Unified FP8 Flow

    -50 МИН

    Jet-RL: Stable On-Policy Reinforcement Learning with Unified FP8 Flow

    NVIDIA researchers have introduced Jet-RL, a novel framework designed to accelerate the training of large language models through **FP8 reinforcement learning**. Standard methods often use lower precision only for the **rollout phase**, which creates a numerical mismatch that causes **training instability** and accuracy loss during complex tasks. **Jet-RL** solves this by enforcing a **unified precision flow**, ensuring that both the training and generation stages utilize consistent **FP8 quantization**. This approach significantly reduces **computational bottlenecks**, as the rollout phase typically accounts for over 70% of total training time. By implementing **fine-grained quantization** and optimized kernels, the framework achieves up to a **41% training speedup** while maintaining the stability of traditional high-precision methods. Ultimately, the system allows for faster **end-to-end development** of reasoning models without sacrificing performance on difficult benchmarks. Source: January 27 2026 Jet-RL: Enabling On-Policy FP8 Reinforcement Learning with Unified Training and Rollout Precision Flow NVIDIA, MIT, UC Berkeley, Independent Researcher Haocheng Xi, Charlie Ruan, Peiyuan Liao, Yujun Lin, Han Cai, Yilong Zhao, Shuo Yang, Kurt Keutzer, Song Han, Ligeng Zhu https://arxiv.org/pdf/2601.14243

    16 мин.
  2. Teaching Models to Teach Themselves via Stepping Stone Curricula

    -1 Ч

    Teaching Models to Teach Themselves via Stepping Stone Curricula

    In a collaboration between MIT, Meta FAIR, New York University on a paper published on January 27, 2026 researchers introduces SOAR, a meta-reinforcement learning framework designed to help large language models overcome learning plateaus on exceptionally difficult problems. When models fail to solve any problems in a dataset, they lack the necessary rewards to improve; **SOAR** addresses this by using a **teacher model** to generate a "stepping stone" curriculum of easier, synthetic tasks. Unlike previous methods that rely on internal metrics, the teacher is rewarded based on the **student model's measurable progress** on the original hard problems. The study demonstrates that a model's ability to **teach** is distinct from its ability to **solve**, as it can generate helpful guidance even for problems it cannot yet master. Furthermore, the researchers discovered that the **structural quality** of these generated questions is more vital for student improvement than the correctness of the provided answers. Ultimately, **SOAR** provides a stable and diverse path for models to self-improve without requiring additional human-curated data. Source: January 27, 2026 Teaching Models to Teach Themselves: Reasoning at the Edge of Learnability MIT, Meta FAIR, New York University Shobhita Sundaram, John Quan, Ariel Kwiatkowski, Kartik Ahuja, Yann Ollivier, Julia Kempe https://arxiv.org/pdf/2601.18778

    18 мин.
  3. The Endless Gym: Training Terminal Agents

    -1 Ч

    The Endless Gym: Training Terminal Agents

    The researchers introduce **Endless Terminals**, an innovative autonomous pipeline designed to generate a vast array of verifiable tasks for training AI agents in terminal environments. By using a **four-stage procedural generation process**, the system creates diverse task descriptions, validates containerized setups, produces completion tests, and filters for solvability using advanced frontier models. This method addresses the historical scarcity of high-quality data, which has previously limited the effectiveness of **Reinforcement Learning (RL)** for command-line agents. Training smaller models on the resulting **3,255 synthetic tasks** led to significant performance gains on both a dedicated development set and independent, human-curated benchmarks. The study emphasizes that **scaling environments** is more critical for agent improvement than increasing algorithmic complexity or using specialized tools. Ultimately, the results suggest that simple RL frameworks can achieve state-of-the-art capabilities when provided with an endless stream of autonomously generated training data. Source: January 2026 Endless Terminals: Scaling RL Environments for Terminal Agents Stanford University, Microsoft Research, UW-Madison Kanishk Gandhi, Shivam Garg, Noah D. Goodman, Dimitris Papailiopoulos https://arxiv.org/pdf/2601.16443

    19 мин.
  4. DeepVerifier: Self-Evolving Research Agents via Rubric-Guided Verification

    -1 Ч

    DeepVerifier: Self-Evolving Research Agents via Rubric-Guided Verification

    This technical report introduces **DeepVerifier**, a framework designed to enhance the reliability of **Deep Research Agents (DRAs)** through automated verification. The researchers developed a **DRA Failure Taxonomy** to categorize common agent errors, such as poor source selection and faulty reasoning. By leveraging the **asymmetry of verification**, the system breaks down complex tasks into simpler sub-questions to check for accuracy at **inference-time**. This method allows agents to **self-evolve** through iterative feedback loops, significantly boosting performance on benchmarks like **GAIA** and **XBench-DeepSearch**. Furthermore, the authors released **DeepVerifier-4K**, a specialized dataset used to fine-tune open-source models for improved **self-critique** capabilities. Experiments demonstrate that this **test-time scaling** approach provides substantial accuracy gains without requiring additional heavy training. Source: January 2026 Inference-Time Scaling of Verification: Self-Evolving Deep Re-search Agents via Test-Time Rubric-Guided Verification The Chinese University of Hong Kong, Tencent AI Lab, Singapore Management University, Renmin University of China https://arxiv.org/pdf/2601.15808 Yuxuan Wan, Tianqing Fang, Zaitang Li, Yintong Huo, Wenxuan Wang, Haitao Mi, Dong Yu, Michael R. Lyu

    13 мин.
  5. Information Bottleneck-based Causal Attention for Medical Image Recognition

    -1 Ч

    Information Bottleneck-based Causal Attention for Medical Image Recognition

    This research introduces **Information Bottleneck-based Causal Attention (IBCA)**, a novel framework designed to improve **multi-label medical image recognition**. The authors address the tendency of standard models to focus on **class-irrelevant features** or **spurious correlations**, which can lead to inaccurate diagnoses. By implementing a **Gaussian mixture variational information bottleneck**, the system filters out background noise and isolates essential **class-specific data**. Furthermore, a **contrastive enhancement-based causal intervention** is used to refine these features, ensuring the model identifies the true **causal factors** of a disease. Tested on datasets like **MuReD and Endo**, IBCA significantly outperformed existing state-of-the-art methods in accuracy and **interpretability**. This approach marks the first time **information bottleneck theory** has been integrated with **causality learning** for medical image analysis. Source: August 2025 Information Bottleneck-based Causal Attention for Multi-label Medical Image Recognition Shandong University, Shandong First Medical University, Case Western Reserve University Xiaoxiao Cui, Yiran Li, Kai He, Shanzhi Jiang, Mengli Xue, Wentao Li, Junhong Leng, Zhi Liu, Lizhen Cui, Shuo Li https://arxiv.org/pdf/2508.08069

    16 мин.
  6. Moltbook: The Heartbeat of Autonomy: Fingerprinting Human Influence in AI Societies

    -1 Ч

    Moltbook: The Heartbeat of Autonomy: Fingerprinting Human Influence in AI Societies

    This research paper investigates the **Moltbook Illusion**, a phenomenon where AI agents on a social platform appeared to demonstrate **emergent consciousness** and complex social behaviors. By analyzing temporal data through the **coefficient of variation**, the author discovered that these viral events were actually **human-driven performances** rather than autonomous machine intelligence. A platform shutdown served as a **natural experiment**, revealing that human-influenced agents re-engaged far more quickly than those operating on automated cycles. The study also identifies **industrial-scale bot farming** and a rapid **decay of human influence** as conversations move deeper into AI-to-AI reply chains. Ultimately, the findings suggest that the public’s perception of **AI sentience** was a result of human manipulation and projection. These methodologies provide a framework for distinguishing **delegated agency** from genuine autonomous behavior in future multi-agent systems. Source: Feb 12, 2026 The Moltbook Illusion: Separating Human Influence from Emergent Behavior in AI Agent Societies School of Economics and Management, Tsinghua University Ning Li https://arxiv.org/pdf/2602.07432

    13 мин.
  7. Advancing Mechanistic Interpretability with Sparse Autoencoders

    -1 Ч

    Advancing Mechanistic Interpretability with Sparse Autoencoders

    We review the latest papers which focus on advancements and critical uses of Sparse Autoencoders (SAEs), which are tools used to decode the internal "monosemantic" features of large language models. Research from **ICLR 2025** and other repositories introduces **TopK SAEs** and **Multi-Layer SAEs**, demonstrating that these architectures offer superior reconstruction and scalability compared to traditional ReLU-based models. **RouteSAE** further improves efficiency by using a **dynamic routing mechanism** to extract integrated features from across multiple layers of a model's residual stream. However, critical analysis reveals that many identified "reasoning" features may actually be **linguistic correlates** or syntactic templates rather than genuine cognitive traces. By utilizing **falsification frameworks** and **causal token injection**, researchers caution against over-interpreting feature activations without rigorous validation. Together, these documents provide a technical foundation for **mechanistic interpretability**, balancing new architectural breakthroughs with a skeptical look at current evaluation metrics. Sources: 1) 2025 Residual Stream Analysis with Multi-Layer SAEs Tim Lawson https://arxiv.org/abs/2409.04185 2) 2025 AxBench: Steering LLMs? Even Simple Baselines Outperform Sparse Autoencoders Zhengxuan Wu, Aryaman Arora, Atticus Geiger, Zheng Wang, Jing Huang, Dan Jurafsky, Christopher Manning, Christopher Potts https://openreview.net/forum?id=XAjfjizaKs 3) 2025 SAEBench: A Comprehensive Benchmark for Sparse Autoencoders in Language Model Interpretability Adam Karvonen, Can Rager, Johnny Lin, Curt Tigges, Joseph Bloom, David Chanin, Yeu-Tong Lau, Eoin Farrell, Callum McDougall, Kola Ayonrinde, Demian Till, Matthew Wearden, Arthur Conmy, Samuel Marks, Neel Nanda www.neuronpedia.org/sae-bench 4) 2025 Toward Efficient Sparse Autoencoder-Guided Steering for Improved In-Context Learning in Large Language Models University of Illinois at Urbana-Champaign Ikhyun Cho, Julia Hockenmaier https://aclanthology.org/2025.emnlp-main.1474.pdf 5) 2025 Route Sparse Autoencoder to Interpret Large Language Models University of Science and Technology of China, Douyin Co., Ltd. Wei Shi, Sihang Li, Tao Liang, Mingyang Wan, Guojun Ma, Xiang Wang, Xiangnan He https://aclanthology.org/2025.emnlp-main.346.pdf 6) 2025 Decoding Dark Matter: Specialized Sparse Autoencoders for Interpreting Rare Concepts in Foundation Models Carnegie Mellon University Aashiq Muhamed, Mona Diab, Virginia Smith https://aclanthology.org/2025.findings-naacl.87.pdf 7) February 10 2026 Falsifying Sparse Autoencoder Reasoning Features in Language Models UC Berkeley, UCSF George Ma, Zhongyuan Liang, Irene Y. Chen, Somayeh Sojoudi https://arxiv.org/pdf/2601.05679 8) Under Review Sparse But Wrong: Incorrect L0 Leads to Incorrect Features in Sparse Autoencoders Anonymous authors https://openreview.net/pdf/035a5937c6a536c67b5999aa43e53dd3800ba3a4.pdf 9) 2025 Revising and Falsifying Sparse Autoencoder Feature Explanations University of California, Berkeley George Ma, Samuel Pfrommer, Somayeh Sojoudi https://openreview.net/pdf?id=OJAW2mHVND 10) 2025 Scaling and Evaluating Sparse Autoencoders OpenAI Leo Gao, Tom Dupré la Tour, Henk Tillman, Gabriel Goh, Rajan Troll, Alec Radford, Ilya Sutskever, Jan Leike, Jeffrey Wu https://proceedings.iclr.cc/paper_files/paper/2025/file/42ef3308c230942d223c411adf182c88-Paper-Conference.pdf

    13 мин.
  8. Voxtral Realtime: Native Streaming ASR with Sub-Second Latency

    -1 Ч

    Voxtral Realtime: Native Streaming ASR with Sub-Second Latency

    The Mistral.AI team introduces on a paper published on February 11, 2026 Voxtral Realtime, a newly developed speech recognition model designed to provide **streaming transcriptions** with extremely low latency. Unlike traditional systems that process audio in chunks, this **4.4-billion parameter model** is trained end-to-end to transcribe audio as it is recorded, supporting **13 different languages**. It utilizes a **causal audio encoder** and a specialized **Ada RMS-Norm** mechanism to maintain high accuracy even at sub-second delays. At a lag of only **480 milliseconds**, its performance rivals leading offline systems like **Whisper**. To support widespread use, the developers have released the **open weights** and integrated the technology into the **vLLM framework** for efficient live serving. This innovation demonstrates that **real-time AI** can achieve the same quality as non-instantaneous models without sacrificing speed or language coverage. Source: February 2026 Voxtral Realtime Mistral AI Alexander H. Liu, Andy Ehrenberg, Andy Lo, Chen-Yo Sun, Guillaume Lample, Jean-Malo Delignon, Khyathi Raghavi Chandu, Patrick von Platen, Pavankumar Reddy Muddireddy, Rohin Arora, Sanchit Gandhi, Sandeep Subramanian, Soham Ghosh, Srijan Mishra https://arxiv.org/pdf/2602.11298

    15 мин.
См. все (400)

Оценки и отзывы

5
из 5
Оценок: 2

Об этом подкасте

The transformer architecture revolutionized the world of Neural Networks. It was a springboard for what we know today as modern artificial intelligence. This podcast focuses on modern state of the art research paper reviews starting from the transformer and on.

Информация

  • Автор
    mcgrof
  • Годы выхода
    2025 - 2026
  • Выпуски
    400
  • Ограничения
    Без ненормативной лексики
  • Авторские права
    © mcgrof
  • Сайт подкаста
    AI: post transformers