Articles

Browse all AI articles. Use filters to narrow down results.

Showing 24481 to 24500 of 30522 entries.

Most Recent Articles

ArXiv Machine Learning
cs.lgcs.clresearcharxiv machine learningragbenchmarkpolicy

Beyond Mode-Seeking RL: Trajectory-Balance Post-Training for Diffusion Language Models

By Saba Ahmadi, Prasanna Parthasarathi, Yufei Cui

arXiv:2605.13935v1 Announce Type: new Abstract: Diffusion language models are a promising alternative to autoregressive models, yet post-training methods for them largely adapt reward-maximizing objectives. We identify a central failure mode in this setting we call trajectory locking: sampled reward-driven updates over-concentrate probability mass onto a narrow set of denoising paths, reducing coverage of alternative correct solutions under repeated sampling. To address this, we propose TraFL (Trajectory Flow baLancing), a trajectory-balance objective that trains the policy toward a reward-tilted target distribution anchored to a frozen reference model. We make this practical for diffusion language models with a diffusion-compatible sequence-level surrogate and a learned prompt-dependent normalization. Across mathematical reasoning and code generation benchmarks, TraFL is the only evaluated post-training method that improves over the base model in every benchmark-length setting, with gains that persist as the sampling budget increases. The improvements transfer to held-out evaluations: TraFL stays above the base model on Minerva Math and is the strongest method on every LiveCodeBench difficulty split.

ArXiv NLP
cs.clcs.lgresearcharxiv nlpbenchmark

Merging Methods for Multilingual Knowledge Editing for Large Language Models: An Empirical Odyssey

By Kunil Lee, Ki-Young Shin, Jong-Hyeok Lee, Young-Joo Suh

arXiv:2605.13919v1 Announce Type: new Abstract: Multilingual knowledge editing (MKE) remains challenging because language-specific edits interfere with one another, even when locate-then-edit methods work well in monolingual settings. This paper focuses on three issues: the effectiveness of vector merging methods for MKE, the extent to which Task Singular Vectors for Merging (TSVM) can reduce multilingual interference, and the influence of the weight scaling factor and rank compression ratio on performance. We evaluate six merging variants with two popular backbone large language models, two base knowledge editing methods, and 12 languages on the MzsRE benchmark under a large-scale batch-editing setting. Our results show that vector summation with shared covariance is the most reliable overall strategy, whereas simple summation without shared covariance performs poorly. TSVM improves performance in some settings, but its ability to mitigate multilingual interference is limited. We also find that performance is sensitive to both weight scale and rank ratio, with larger-than-default scaling and relatively low rank often yielding better results. These findings clarify the practical strengths and limits of current vector merging methods for MKE and provide guidance for future multilingual knowledge editing research.

ArXiv Machine Learning
cs.lgcs.dccs.niresearcharxiv machine learninggpu

EMA: Efficient Model Adaptation for Learning-based Systems

By Daiyang Yu, Xinyu Chen, Yihan Zhang, Yan Liang, Yaqi Qiao, Fan Lai

arXiv:2605.13942v1 Announce Type: new Abstract: Machine learning (ML) is increasingly applied to optimize system performance in tasks such as resource management and network simulation. Unlike traditional ML tasks (e.g., image classification), networked systems often operate in heterogeneous, long-running, and dynamic environment states, where input conditions (e.g., network loads) and operational objectives can shift over time and across settings. Existing learning-based systems offer little support for adaptation, resulting in costly model training, extensive data collection, degraded system performance, and slow responsiveness. This paper presents EMA, the first model adaptation system supporting learning-based systems to adapt to evolving environments with minimal operational overhead. EMA takes a system-driven, data-centric approach that accommodates diverse system and model designs while addressing two key deployment challenges. First, it reduces expensive model training by introducing state transformers that align the input state of a new environment with previously similar states, allowing models to warm-start adaptation. Second, it addresses the often-overlooked yet costly process of data labeling--collecting ground truth for exploring and training on various system decisions--by prioritizing labeling high-utility data while balancing the tradeoff between training and labeling cost. Evaluations on eight representative learning-based systems show that EMA reduces adaptation costs (e.g., GPU training time) by 14.9-42.4% while improving system performance (e.g., network throughput) by 6.9-31.3%.

ArXiv Machine Learning
cs.lgcs.airesearcharxiv machine learningmetallmagentagentsbenchmark

EvolveMem:Self-Evolving Memory Architecture via AutoResearch for LLM Agents

By Jiaqi Liu, Xinyu Ye, Peng Xia, Zeyu Zheng, Cihang Xie, Mingyu Ding, Huaxiu Yao

arXiv:2605.13941v1 Announce Type: new Abstract: Long-term memory is essential for LLM agents that operate across multiple sessions, yet existing memory systems treat retrieval infrastructure as fixed: stored content evolves while scoring functions, fusion strategies, and answer-generation policies remain frozen at deployment. We argue that truly adaptive memory requires co-evolution at two levels: the stored knowledge and the retrieval mechanism that queries it. We present EvolveMem, a self-evolving memory architecture that exposes its full retrieval configuration as a structured action space optimized by an LLM-powered diagnosis module. In each evolution round, the module reads per-question failure logs, identifies root causes, and proposes targeted configuration adjustments; a guarded meta-analyzer applies them with automatic revert-on-regression and explore-on-stagnation safeguards. This closed-loop self-evolution realizes an AutoResearch process: the system autonomously conducts iterative research cycles on its own architecture, replacing manual configuration tuning. Starting from a minimal baseline, the process converges autonomously, discovering effective retrieval strategies including entirely new configuration dimensions not present in the original action space. On LoCoMo, EvolveMem outperforms the strongest baseline by 25.7% relative and achieves a 78.0% relative improvement over the minimal baseline. On MemBench, EvolveMem exceeds the strongest baseline by 18.9% relative. Evolved configurations transfer across benchmarks with positive rather than catastrophic transfer, indicating that the self-evolution process captures universal retrieval principles rather than benchmark-specific heuristics. Code is available at https://github.com/aiming-lab/SimpleMem.

Techmeme
techtechmeme

Mike Wheatley / SiliconANGLE: UK-based Iceotope, which develops data center liquid cooling tech, raised a $26M Series B co-led by Two Seas Capital and Barclays Climate Ventures  —  Liquid cooling system startup Iceotope Technologies Ltd. said today it has closed on a $26 million Series B round of funding as it strives …

Techmeme
techtechmeme

Atsuko Fukase / Bloomberg: Alphabet sold $3.6B of bonds in its debut yen-denominated bond sale, the largest by a non-Japanese company, as AI-driven capital spending accelerates  —  Alphabet Inc. sold ¥576.5 billion ($3.6 billion) of bonds in the biggest ever yen deal by a non-Japanese company as competition to fund centers and AI infrastructure intensifies.

Datafloq
geoai search visibilitygenerative engine optimization (geo)lovable aisemrush integrationserver-side rendering (ssr)tanstack start:datadatafloqgooglemetallmrag

Lovable Just Made Discoverability a Day-One Feature

By Datafloq Editorial Team

The AI app-builder embeds SEO, AI search optimization, and live Semrush intelligence into its product, addressing … Lovable Just Made Discoverability a Day-One Feature Read more The post Lovable Just Made Discoverability a Day-One Feature  appeared first on .

Techmeme
techtechmemeanthropicpolicy

Anthropic: In a policy paper, Anthropic urges the US and allies to enforce export controls, curb distillation attacks, and export US AI to hold the lead over China by 2028  —  We're releasing a new paper that explains our views on the competition on AI between the US and China.

Product Hunt
startupproduct huntagentagents

Agentspan

By Nick

Discussion | Link

Techmeme
techtechmemeanthropic

Robert McMillan / Wall Street Journal: Security research firm Calif says it used Mythos to help build a macOS kernel memory corruption exploit circumventing Apple's Memory Integrity Enforcement tech  —  During tests in April, researchers found software issues in MacOS, one of the world's toughest targets for hackers

Business Insider
newsbusiness insider

Retail investors testified at the securities fraud trial of Andrew Left, pictured right, in Los Angeles. Bloomberg/Getty Images Andrew Left, a prominent short-seller, is on trial in Los Angeles for alleged securities fraud. A retired firefighter testified he lost most of his 401(k) after investing in stocks Left targeted. Left is accused of manipulating the market, but his defense says he shared his honest opinions. Federal prosecutors called a witness in the Andrew Left securities fraud trial who might resonate with the jury — a retail investor who said he lost his retirement nest egg by investing in stocks that tanked after being criticized by the famous short-seller. Billy Banks, a retired firefighter from Texas, testified at Left's trial in a Los Angeles federal court on Thursday. Prosecutors have accused Left, founder of Citron Research and a familiar face on financial TV news shows, of using his prominence to manipulate the market and deceive everyday stock buyers, earning more than $20 million in the process. Banks said he was working as a firefighter in 2018 when he decided to become more aggressive with his investing. He testified he took about $110,000 that was previously invested in mutual funds in his 401(k) to buy shares in a cannabis company called CV Sciences, with the stock ticker CVSI. Banks said he bought CVSI because he believed the company had great CBD products that "really helped" with stress and pain management. Banks said he was on vacation with his wife and saw his investment had already sprouted to $190,000. He said he remembered thinking "Oh my god" and being "just so excited." "It was like I had been watering this plant for weeks, and here it is," he said. Banks said that shortly after he returned from vacation, Left shared negative comments about the company, and the stock price plummeted. While he still believed in the product, he said he couldn't justify the "beatdown" he was taking as the stock price tanked. "It was like trying to catch a tiger tail. You couldn't catch up with the thing," he said. Banks sold his CVSI shares, losing around $80,000 of his initial investment. He took about $30,000 of the remaining funds and invested in another cannabis stock, hoping to recoup some of his losses. This time, he invested in a company called Namaste. Banks testified he saw Left in a TV appearance in which the short-seller said he believed Namaste's share price would go to zero, and that his reaction was "simply horror." "When he said he was going to short it to zero, I saw my funds going to zero," he said. Banks testified that by the time he sold, he had lost around 80% of his investment in Namaste, calling it "devastating." The defense said Left never recommended anyone buy CVSI or Namaste. They also argued that Left's view on Namaste turned out to be right, and that if Banks had sold all his shares the day Left issued the negative report, he would've lost less money than he did by selling them later. Another retail investor — the kind the prosecution is accusing Left of targeting — also testified on Thursday. Adam Gray was working as a car salesman in South Dakota in 2018 when he said he began trading and learning about stocks. Gray said Left was one of the people he'd follow for their perspective, but that he started to see Left's behavior as potentially fraudulent. Gray, who said he was closely following cannabis stocks, said Left was betting long on one cannabis company while sharing negative information about two of its competitors. "To me, it just seemed off, like that shouldn't be happening," he said. Gray eventually sent a tip to the Securities and Exchange Commission about Left. It's unclear if that tip sparked the investigation into Left. Left's defense said his reports about the cannabis companies he criticized were accurate. The trial is set to continue on Friday. Read the original article on Business Insider

Techmeme
techtechmemeanthropicrag

George Hammond / Financial Times: Sources: Anthropic has agreed to the terms of a $30B fundraising at a $900B valuation, with Sequoia, Dragoneer, Greenoaks, and Altimeter co-leading the round  —  AI lab has tapped Dragoneer, Greenoaks, Sequoia Capital and Altimeter Capital to lead new round

Techmeme
techtechmeme

Forbes: Sources: xAI co-founder Igor Babuschkin plans to raise up to $1B at an up to $5B valuation for a new AI research startup, with General Catalyst possibly leading  —  He's the latest superstar AI researcher to take advantage of the wave of investment dollars going into so-called ‘neolabs,’ which have no product or revenue.

Wired AI
businessbusiness / artificial intelligencetechwired aiopenai

The Real Losers of the Musk v. Altman Trial

By Maxwell Zeff, Paresh Dave

A federal jury is now deciding whether Elon Musk will win his lawsuit against OpenAI and Sam Altman—but the trial has made everyone look bad.

Simon Willison Weblog
aisimon willison weblogllm

<p><strong>Release:</strong> <a href="https://github.com/datasette/datasette-llm-limits/releases/tag/0.1a0">datasette-llm-limits 0.1a0</a></p> <p>This plugin works in conjunction with <a href="https://github.com/datasette/datasette-llm">datasette-llm</a> and <a href="https://github.com/datasette/datasette-llm-accountant">datasette-llm-accountant</a> to let you configure a per-user (or global) spending limit for LLM usage inside of Datasette. Configuration looks something like this:</p> <pre><span class="pl-ent">plugins</span>: <span class="pl-ent">datasette-llm-limits</span>: <span class="pl-ent">limits</span>: <span class="pl-ent">per-user-daily</span>: <span class="pl-ent">scope</span>: <span class="pl-s">actor</span> <span class="pl-ent">window</span>: <span class="pl-s">rolling-24h</span> <span class="pl-ent">amount_usd</span>: <span class="pl-c1">1.00</span> </pre> <p>Tags: <a href="https://simonwillison.net/tags/llm">llm</a>, <a href="https://simonwillison.net/tags/datasette">datasette</a></p>

Business Insider
newsbusiness insiderragagent

"Entourage" creator Doug Ellin attended the securities fraud trial of Andrew Left. Allen Berezovsky/Getty Images "Entourage" creator Doug Ellin attended Andrew Left's securities fraud trial in LA court on Thursday. Ellin called Left a friend and said he was there in support of the famous short-seller. Left is accused of market manipulation. His defense says he shared his honestly held opinions. Andrew Left's entourage at his securities fraud trial in Los Angeles on Thursday included a very famous "Entourage" guy — Doug Ellin, the creator of the hit HBO show. Business Insider spotted Ellin in court, watching as prosecutors laid out their case against Left. The famous short-seller is accused of manipulating the market and deceiving retail investors, raking in over $20 million in the process. Ellin told Business Insider he was there to support Left, who he called a "good friend." He said Left was mentioned on "Entourage," which ran on HBO from 2004 to 2011 and told the story of rising movie star Vincent Chase and his crew. A quick search turned up a season three "Entourage" episode that may have contained the reference that Ellin was referring to. "Sorry, Ari" has a scene where Chase plans to meet with a potential new agent who is never seen onscreen but is named "Andy Left." The real Left rose to fame as a short-seller, betting against companies like Valeant Pharmaceuticals and GameStop. Left was charged with securities fraud in 2024. Thursday was the fourth day of his trial, which is expected to last three weeks. Read the original article on Business Insider

OpenAI Blog
productaiopenai blogchatgpt

Preview a new personal finance experience in ChatGPT for Pro users in the U.S. Securely connect your financial accounts and get AI-powered insights and guidance grounded in your financial context, goals, and priorities.

OpenAI Blog
openai academyaiopenai blog

See how business operations teams can use Codex to create initiative briefs, strategy updates, leadership decision packets, progress updates, and more from real work inputs.