# Generalization İlgili Makaleler

HTX Haber Merkezi, kripto endüstrisindeki piyasa trendleri, proje güncellemeleri, teknoloji gelişmeleri ve düzenleyici politikaları kapsayan "Generalization" hakkında en son makaleleri ve derinlemesine analizleri sunmaktadır.

LeCun Continuously Endorses, New Work VISReg Tackles the Core Challenge of 'Representation Collapse' in JEPA World Models

"VISReg: A New Self-Supervised Learning Method Tackles Representation Collapse in JEPA World Models" Yann LeCun has highlighted the new self-supervised learning (SSL) work VISReg (Variance-Invariance-Sketching Regularization), which addresses the core challenge of "representation collapse" in JEPA-based world models. SSL often collapses, mapping different inputs to similar vectors, losing discriminative power. While methods like VICReg and SIGReg attempted to regularize the representation distribution, they suffered from issues like vanishing gradients during collapse or coupled scale and shape optimization. VISReg overcomes these by decoupling the regularization into independent "scale" and "shape" objectives. It uses a variance term to prevent amplitude collapse (providing constant gradient even during collapse) and a sliced Wasserstein distance (SWD) based sketching target with stop-gradient to align the distribution shape to an isotropic Gaussian, without interfering with scale. This approach, requiring only ~15 lines of PyTorch core code, offers linear computational complexity and scales efficiently across multiple GPUs. Evaluated across 15 datasets (in-domain, out-of-distribution/OOD, dense prediction), VISReg outperforms 7 mainstream SSL methods (MoCoV3, DINO, iBOT, I-JEPA, MAE, data2vec) without relying on heuristic tricks like EMA or stop-gradient. Key results include: superior average OOD accuracy; matching DINOv2's OOD performance using only ~1/10 of the data (ImageNet-22K vs. LVD-142M); better transfer learning fine-tuning results across multiple datasets; and competitive performance in semantic segmentation and generative guidance. VISReg also demonstrates robustness on low-quality data like long-tailed or low-rank datasets where other methods fail. The work provides a more stable, efficient, and generalizable regularization solution for SSL and JEPA models.

marsbit07/28 10:12

LeCun Continuously Endorses, New Work VISReg Tackles the Core Challenge of 'Representation Collapse' in JEPA World Models

marsbit07/28 10:12

Embodied Intelligence 'Gaokao' is Insanely Hard, Humans Score 100, Best Model Only 12.8

Embodied AI Faces a Daunting "Everest": New Benchmark Reveals Huge Gap Between Models and Humans A comprehensive new benchmark for robotic manipulation, RoboDojo, has been released, painting a stark picture of the current state of embodied AI. It serves as a unified evaluation platform covering both simulation and real-world robot tasks. The benchmark assesses five core capabilities: Generalization (adapting to new scenes/objects), Memory, Precision manipulation, Long-Horizon multi-step tasks, and Open semantic understanding. It includes 42 simulation tasks and 18 standardized real-world tasks across three dual-arm robot platforms. The results are sobering. In simulation, the best-performing generalist robot policy achieved an average success rate of only 8.80%. Performance in the real world was slightly higher but still low, with the top model succeeding 12.8% of the time on average. In stark contrast, human experts scored 76.03% in simulation and 100% in real-world tests. The benchmark highlights significant, uneven gaps in current models' abilities. While some excel in specific areas like visual recognition or simple actions, they struggle with reliability, especially in long-horizon tasks where errors accumulate and in open-ended semantic instructions. The low scores, particularly in real-world deployment with physical uncertainties like camera noise and contact dynamics, underscore that today's models are far from being robust, general-purpose operational robots. RoboDojo is more than just a ranking; it's an infrastructure designed for fair, reproducible comparison. Its companion system, XPolicyLab, standardizes the interface for different models to be evaluated. Maintained by an academic consortium without commercial ties, it aims to provide a community-wide "altitude meter" to track genuine progress toward reliable and generalizable robot manipulation.

marsbit07/08 11:49

Embodied Intelligence 'Gaokao' is Insanely Hard, Humans Score 100, Best Model Only 12.8

marsbit07/08 11:49

For the First Time, Pure Human Video Pretrained VLA for Dexterous Manipulation: Deployable with Minimal Fine-Tuning Data

For the first time, a purely human-video-pretrained Vision-Language-Action (VLA) model for dexterous manipulation requires only a small amount of data for fine-tuning to achieve successful real-world deployment. Achieving human-level dexterous manipulation remains a core challenge in robotics. While multi-fingered hands offer hardware potential, Visual-Language-Action (VLA) models lag behind due to the high cost of collecting diverse, high-quality robot data. A novel framework, VITRA, developed by Microsoft Research Asia and Tsinghua University, addresses this by automatically transforming massive, unlabeled real-world human activity videos into a structured V-L-A training dataset. Key innovations include precise 3D hand motion annotation from monocular video, atomic action segmentation based on hand-speed minima, and automated instruction generation using VLMs combined with 3D trajectory visualization. This process created a massive dataset of 1 million clips. Pretrained exclusively on this human video data, the VLA model (combining a VLM backbone with a Diffusion Transformer action expert) demonstrates strong zero-shot hand motion prediction in unseen environments. Crucially, it requires minimal fine-tuning (~1.2k demonstrations) on real robot data to achieve high-success-rate dexterous manipulation tasks like grasping, placing, pouring, and sweeping on hardware like the Realman robot with the XHAND1 dexterous hand. The model shows exceptional generalization to novel objects and environments. The research also observes promising scaling behavior, where performance improves with more pretraining data, paving the way for more generalized embodied intelligence.

marsbit06/08 08:54

For the First Time, Pure Human Video Pretrained VLA for Dexterous Manipulation: Deployable with Minimal Fine-Tuning Data

marsbit06/08 08:54

Large Language Models Ace All Exams, Yet Move Farther from AGI: What Does This Paper Reveal?

The article discusses the ongoing challenge of defining and achieving Artificial General Intelligence (AGI). It notes that industry leaders have set vague, often profit- or time-based benchmarks for AGI, while the concept itself lacks a consensus definition—a situation the article compares to a "Rorschach test." It highlights a recent 2025 paper by researcher Michael Timothy Bennett, who proposes a new, measurable definition. Bennett frames AGI not as mimicking human performance on tests, which current large language models (LLMs) have already mastered, but as an "artificial scientist." A true AGI, according to this view, should be able to widely and efficiently adapt to new environments and tasks within real-world constraints (like computational and energy limits), focusing on the *discovery of new knowledge* rather than the replication of existing data. The author contrasts this with the current dominant approach of "scale-maxing"—massively scaling up data, parameters, and compute. While powerful, this method leads to models that fail on out-of-distribution problems and lack core intelligent abilities: they are passive learners, cannot reason causally, and cannot actively experiment or balance exploration with exploitation. The article argues that Bennett's framework offers a crucial shift. It makes AGI a quantifiable engineering problem and proposes new evaluation "adaptation benchmarks" that test an AI's ability to actively learn in novel scenarios. The conclusion is that achieving AGI will require a fundamental reset—a fusion of multiple methodologies beyond simple scaling, moving AI from mimicking patterns to embodying the scientific spirit of inquiry and discovery.

marsbit05/28 00:24

Large Language Models Ace All Exams, Yet Move Farther from AGI: What Does This Paper Reveal?

marsbit05/28 00:24

Anthropic Has Taught Models to Understand Morality and Opened a New Path for Distillation

Anthropic's research "Teaching Claude Why" reveals a new, data-efficient method for AI alignment. Instead of relying on massive reinforcement learning with punishment (RLHF), which only teaches models to mimic safe answers without true ethical understanding, they used a small dataset (3 million tokens) of "difficult advice." This data consisted of detailed moral deliberations, reasoning, and debates, teaching the model the *why* behind decisions. The key was "deliberation-enhanced" Supervised Fine-Tuning (SFT). The model was trained on responses that included a "chain of thought" (CoT) process based on a constitutional framework. This framework included top-level principles, practical heuristics (like the "1000-user test"), and an 8-factor utility calculator (evaluating harm probability, reversibility, consent, etc.) for weighing complex trade-offs. This approach dropped model misalignment rates from 22% to 3% and showed strong generalization to unseen scenarios. The success challenges the old belief that "SFT memorizes, RL generalizes." It shows that SFT can generalize powerfully if the training data has two features: 1) high prompt diversity (many different scenario types) and 2) CoT supervision (showing the reasoning steps, not just the final answer). The model learns the underlying *thinking framework*, not just surface-level behaviors. This method points to a new paradigm for training AI in "non-RLVR" domains—areas like ethics, creative writing, or strategy where there's no single verifiable answer. The formula is: Domain Constitution + Heuristics + Multi-Factor Deliberation Framework + Diverse Deliberative CoT Data = Generalized capability. It represents a new form of "distillation," moving competition from pure compute towards who can best structure expert knowledge into high-quality reasoning datasets.

marsbit05/15 10:55

Anthropic Has Taught Models to Understand Morality and Opened a New Path for Distillation