# Stability Articoli collegati

Il Centro Notizie HTX fornisce gli articoli più recenti e le analisi più approfondite su "Stability", coprendo tendenze di mercato, aggiornamenti sui progetti, sviluppi tecnologici e politiche normative nel settore crypto.

New Breakthrough in Sign Language Translation by Large Models: CAPO-SLT Achieves Top Scores in Three Chinese Benchmarks

New Progress in Large Model Sign Language Translation: CAPO-SLT Achieves Top Three Scores for Chinese A key challenge in automatic sign language translation (SLT) is that models, even with good visual-text alignment, can be misled during word-by-word generation by locally plausible words lacking sufficient visual evidence. These errors accumulate, causing semantic drift in the full sentence. To address this, researchers from vivo AI Lab identified an issue in reinforcement learning's update rule and proposed CAPO-SLT (Confidence-Aware Policy Optimization for SLT). Without changing the visual encoder or reward function, CAPO-SLT stabilizes generation during RL fine-tuning for more faithful translations. The method dynamically adjusts the update limit (clipping boundary) for each word based on the old policy's confidence. For words with positive advantage, a higher old policy confidence results in a more conservative update limit, preventing over-amplification of potentially erroneous but fluent choices. Lower confidence allows for larger corrective updates. This approach prevents all words from sharing the same update "speed limiter." Trained using pose keypoints (face, body, left/right hand) modeled by ST-GCN, the model undergoes supervised pre-training followed by RL fine-tuning. On the CSL-Daily Chinese dataset, CAPO-SLT using only pose input achieved the highest scores in BLEU-1, BLEU-4, and ROUGE-L metrics, outperforming previous pose-only methods. Gains were also shown on the English How2Sign dataset and the WLASL2000 isolated sign recognition task. Qualitative analysis shows CAPO-SLT corrects critical semantic errors (e.g., changing sentence subjects or predicates) produced by baseline models, leading to translations faithful to the visual evidence. Training curves also indicate greater stability with CAPO-SLT compared to methods like GRPO. The value of CAPO-SLT lies in shifting focus from "visual-text alignment" to "stable generation." It offers a low-invasive path to stabilize existing SLT pipelines, as it modifies only the policy optimization clipping rule without altering the core visual backbone or reward design. Future work will explore learned preference models or SLT-specific semantic evaluators to improve reward mechanisms.

marsbit09/09 11:36

New Breakthrough in Sign Language Translation by Large Models: CAPO-SLT Achieves Top Scores in Three Chinese Benchmarks

marsbit09/09 11:36

Tokenized deposits could raise US credit costs: Dallas Fed economists

The Dallas Fed economists have released a paper analyzing that tokenized deposits, if widely adopted, may increase credit costs for U.S. households and businesses by making bank funding more volatile and less stable. The core concern is that instant settlement, combined with programmable tokens and AI, could enable depositors to move funds between banks rapidly in search of higher yields, significantly increasing deposits' sensitivity to interest rates. The economists modeled that a 10% increase in this sensitivity could reduce banks' capacity to hold long-term assets by an estimated $700 billion over a 10-year equivalent, while deposits staying 10% less time at banks could reduce it by about $580 billion. These figures represent capacity constraints, not direct lending cuts. The analysis comes as U.S. banks actively develop shared blockchain networks for tokenized deposits, such as the new BankChain Alliance and a network by major banks including JPMorgan. While aimed at keeping funds within the regulated system, the economists warn banks might respond to more volatile funding by holding more liquid assets (like Treasuries) or relying more on wholesale term debt—both of which could raise the cost of credit. They referenced Brazil's Pix system as a comparative case where increased instant payment usage led banks to hold more liquidity and reduce credit intermediation.

cointelegraph08/27 06:50

Tokenized deposits could raise US credit costs: Dallas Fed economists

cointelegraph08/27 06:50

Following US Ban on Fable 5, Zhipu AI's Stock Soars 47%

On June 15th, shares of Zhipu AI surged dramatically on the Hong Kong stock market, peaking at a 47.6% gain before closing 32.82% higher. This sharp increase was directly triggered by two recent industry events. On June 12th, Anthropic announced it was suspending global access to its latest flagship models, Claude Fable 5 and Claude Mythos 5, to comply with a U.S. government export control order. The next day, Zhipu AI announced it would open access to its latest open-source flagship model, GLM-5.2, under the permissive MIT license. The Anthropic incident highlighted a critical issue beyond raw model capability: the risk of sudden, unpredictable loss of access to advanced AI models, especially for developers and enterprises deeply integrated with them. This has shifted industry and market focus toward factors like stability, sustainable access, and controllability. Zhipu's move, promoting "frontier intelligence for all," positions its openly available model as a reliable and accessible alternative. The GLM-5.2 model emphasizes "Long Horizon Task" capabilities with a 1M context window, targeting complex, multi-step coding and engineering workflows where maintaining context is crucial. Analysts note this event exposes the risk of dependency on closed-source models subject to single jurisdictional controls, potentially accelerating a shift toward domestic base models and localized deployments. The market's reaction signals a new valuation dimension in AI: providers who can offer stable, long-term, and sustainably accessible AI capabilities are gaining strategic importance.

marsbit06/15 12:01

Following US Ban on Fable 5, Zhipu AI's Stock Soars 47%

marsbit06/15 12:01

The Truth About Global Payments, Revealed by Airwallex

The article discusses Airwallex's approach to global payments, highlighting the key challenges and different strategic paths in the industry. It begins by addressing common user questions about platform reliability, cryptocurrency payments, and the necessity of Airwallex's "heavy" infrastructure model. The core argument is that while many payment platforms appear similar on the surface—offering features like global acquiring and multi-currency accounts—their underlying capabilities differ drastically. The piece identifies three primary paths for global payment providers: 1. **Bypassing Traditional Infrastructure (Web3/Crypto):** This path promises efficiency through stablecoins and on-chain settlements but faces significant regulatory hurdles and offers little advantage over established players for mainstream use, often serving only niche or non-compliant markets. 2. **Aggregating/Packaging Existing Infrastructure:** The most common route, where companies layer a better user experience over legacy banking and partner networks. While fast to market, this approach does not solve fundamental issues like dependency on intermediaries, correspondent banking risks, and compliance fragility. 3. **Building Proprietary Global Infrastructure:** The path chosen by Airwallex and similar firms. This involves obtaining local licenses, building direct regulatory relationships, establishing local teams, and controlling the compliance and technology stack. This is the most difficult and capital-intensive route but aims to internalize complexity. Airwallex's strategy of "heavy" investment in its own infrastructure is framed not as inefficiency, but as a long-term bet to provide clients with greater stability, cost savings beyond fees, and certainty. The platform's "heaviness" absorbs risk and operational complexity, aiming to deliver a "lighter" experience for business customers. The article concludes that in global payments, while shortcuts enable faster growth, mastering the most difficult aspects—the underlying infrastructure—is what creates durable value for clients and sustainable competitive advantage.

链捕手05/28 16:02

The Truth About Global Payments, Revealed by Airwallex

链捕手05/28 16:02

活动图片