arXiv:2607.07779v1 Announce Type: new Abstract: Recent developments in AI for Mathematics (AI4Math), especially Large Language Model (LLM)-driven theorem provers, has achieved remarkable success in formal proof generation for well-defined mathematical problems through Interactive Theorem Proving (ITP) languages. However, current systems remain fundamentally limited in tackling frontier research mathematics, such as discovering new theorems or resolving open conjectures, which are often open-ended, under-specified, and involve multiple layers of abstraction. We argue that the next leap in AI4Math systems requires a decisive shift from predefined problem-solvers to research agents that can address frontier mathematical challenges with rigorous formal mathematical reasoning. In this position paper, we provide a systematic review of the field, covering datasets, auto-formalization, and proof synthesis. More importantly, we identify core limitations of existing systems in serving as mathematical research agents, examining issues across datasets, relational structure, mathematical exploration, tool ecosystem, and human-AI collaboration, outlining a strategic road-map for the future of AI4Math.

arXiv:2607.07820v1 Announce Type: new Abstract: Training tool-use agents to improve from their own experience remains challenging, as supervised fine-tuning relies on fixed teacher-distilled trajectories, while sparse-reward reinforcement learning provides weak supervision for long-horizon interactions. We present DeepSearch-Evolve, a self-distillation framework for web agents built on DeepSearch-World, a deterministic and verifiable environment with reproducible search and page-reading tools. DeepSearch-World contains 420K multi-hop QA tasks constructed from entity-level random walks and supports key agentic cognitive behaviors useful for self-evolving, including progress verification, grounded reflection, and failure recovery. DeepSearch-Evolve iteratively performs trajectory generation, filtering, data mixing, and fine-tuning to train stronger agents. Without distillation from more capable models, DeepSearch-World-9B achieves competitive performance compared with open-source agents, reaching 31.2% on BrowseComp, 61.5% on GAIA, and 93.4% on HotpotQA, showing that verifiable environments enable scalable self-evolution for long-horizon web agents. We will release the environment, 420K training pool, validation set, model, and code to facilitate future research on self-improving deep search agents.

arXiv:2607.07976v1 Announce Type: new Abstract: Reinforcement learning (RL) has achieved remarkable success in enhancing the reasoning capabilities of large language models (LLMs). However, widely used critic-free RL methods rely on uniform credit assignment, broadcasting the same advantage to all tokens regardless of their differences. We identify a critical failure mode of this design, which we refer to as Positive-Credit Contamination: low-probability tail tokens that are contextually erroneous receive identical positive credit to plausible ones within the same trajectory, resulting in the indiscriminate reinforcement of flawed reasoning behavior. To mitigate this issue, we propose Tail-Aware Credit calibratiOn (TACO), a method that calibrates uniform credit assignment to suppress undesirable positive updates. TACO first computes a tail-risk score that incorporates the local generation context to assess each token's risk of falling into the unreliable tail, distinguishing unexpected rarity from uncertainty-driven exploration. TACO then uses this score to tune positive credit for risky tokens without removing their gradients entirely, so that recurring useful rare patterns can accumulate reinforcement while incidental noise is progressively dampened. Experimental results across three LLMs and eight benchmarks show that TACO consistently outperforms GRPO-style baselines. Notably, TACO improves training stability, supporting sustained performance gains in long-horizon RL. The source code is available at: https://github.com/xiuyilou/TACO.

arXiv:2607.07974v1 Announce Type: new Abstract: Intent detection is a critical task that bridges human intents and system actions in human-machine interaction systems. However, there still exist challenges for detecting out-of-scope (OOS) intents. (i) The traditional methods view the OOS intent detection as a multi-class classification, then the detection accuracy decreases as the class number of the known intents increases; (ii) LLM-embedding methods require large parameters, that makes them difficult to train and practically deploy. Thus, this work proposes a multi-cluster boundary learning method to detect OOS intents via MiniLM embedding (i.e., all-MiniLM-L6-v2) in an one-class classification workflow. The method learns the boundaries of multi-cluster embeddings generated by MiniLM from the training utterances, and then rejects the out-of-domain utterances as OOS intents. Experiments are conducted on public CLINC150, StackOverflow and Banking77 datasets. The results show that the method achieves the state-of-the-art OOS intent detection performance compared the other baselines. Ablation studies are also conducted and the results show that the used MiniLM can better adapt to the workflow and utterance embedding requirements. The code is available at supplementary materials.

arXiv:2607.07937v1 Announce Type: new Abstract: Preprocessing-based methods for stereotype mitigation, such as pre-/post-training on debiased corpora, are widely used in NLP. While these approaches reduce measurable stereotypes for targeted groups, we find they often induce unintended shifts-side effects, where stereotyping or counter-stereotyping can increase relative to neutral baselines for other demographics, including across unrelated demographic categories. We demonstrate these side effects across two model families (encoder-only and decoder-only), multiple preprocessing strategies (removing stereotypical sentences, removing group mentions, and swapping group references), and both pre- and post-training at different data scales on Wikipedia. Standard benchmarks frequently miss these shifts. Using attention-rollout analysis, we observe that such side effects are not accompanied by large changes in attention flow, complicating mechanistic explanations. We discuss implications for evaluation, provide actionable diagnostics, and argue for side-effect-aware, transparent mitigation practices.

arXiv:2607.07895v1 Announce Type: new Abstract: Research on stereotypes in large language models (LLMs) has largely focused on English-speaking contexts, due to the lack of datasets in other languages and the high cost of manual annotation in underrepresented cultures. To address this gap, we introduce a cost-efficient human-LLM collaborative annotation framework and apply it to construct EspanStereo, a Spanish-language stereotype dataset spanning multiple Spanish-speaking countries across Europe and Latin America. EspanStereo captures both well-documented stereotypes from prior literature and culturally specific biases absent from English-centric resources. Using LLMs to generate candidate stereotypes and in-culture annotators to validate them, we demonstrate the framework's effectiveness in identifying nuanced, region-specific biases. Our evaluation of Spanish-supporting LLMs using EspanStereo reveals significant variation in stereotypical behavior across countries, highlighting the need for more culturally grounded assessments. Beyond Spanish, our framework is adaptable to other languages and regions, offering a scalable path toward multilingual stereotype benchmarks. This work broadens the scope of stereotype analysis in LLMs and lays the groundwork for comprehensive cross-cultural bias evaluation.

Maxwell Zeff / Wired: OpenAI's head of safety, Johannes Heidecke, is leaving as OpenAI integrates its research and safety teams; Mia Glaese will become VP of research and safety  —  Johannes Heidecke's departure comes as OpenAI tries to further integrate its research and safety teams.

Johannes Heidecke’s departure comes as OpenAI tries to further integrate its research and safety teams.

Independent cybersecurity journalist Brian Krebs reported in May that a security researcher with cyber firm GitGuardian alerted him to reams of exposed passwords stored in a publicly accessible GitHub repository, which an employee of a CISA contractor had uploaded.

Felix Kästle/picture alliance via Getty Images Colleges are trying to adapt to the rise of AI and promote students' independent thinking. The University of Chicago Law School just rolled out an in-class laptop ban for first-year students. As AI becomes more integrated in the legal profession, educators are struggling to strike a balance. AI is starting to make some classrooms look a little more old school. The University of Chicago Law School is requiring first-year students to keep their laptops closed in class this fall, as part of a broader strategy to ensure students learn to think independently as artificial intelligence becomes ubiquitous in the legal profession. The move comes as colleges across the country grapple with how generative AI is reshaping higher education. Business Insider reported earlier this month that Brown University said it recently disciplined dozens of students after uncovering what administrators described as a widespread AI-assisted cheating scandal, underscoring how difficult the technology has made traditional take-home assignments and remote assessments. Rather than trying to ban AI outright, Chicago Law says it is redesigning its curriculum to separate the skills students should develop on their own from those where AI should be embraced. "We need to make sure that the students are learning to think for themselves," Adam Chilton, dean of the University of Chicago Law School, told Business Insider. At the same time, he said, "We can't just naively try to pretend that you can turn off AI or that students won't use it or they don't need to know it." The school's new strategy includes laptop-free first-year classes, in-person proctored exams that prevent students from accessing outside materials, and oral defenses for major research papers to ensure students can explain and defend their work. It is also expanding AI instruction by integrating the technology into legal writing courses, adding more AI-focused classes, and providing students access to legal AI tools, including Harvey and Legora. Chilton said educators have been "a little bit asleep at the wheel" by continuing to assign take-home work that students can complete with AI "without thinking for yourself, without learning for yourself." He said reports of AI cheating at Brown, Harvard, and other schools reinforced concerns that students could advance through their education without developing rigorous reasoning skills. The challenge, he said, is that AI has become indispensable in legal practice. Law firms increasingly expect new hires to use technology efficiently and responsibly, making an outright ban unrealistic. Instead, Chicago's approach is to create what Chilton called "space for both of these modes of learning" — teaching students to think without AI first, then teaching them how to use it ethically once they've built those foundational skills. Read the original article on Business Insider

Phia, the shopping startup founded by Bill Gates’ daughter, Phoebe, and her friend Sophia Kianni is under fire for a practice known as “cookie stuffing,” which helped the product receive commissions and credit for sales it did not actually generate, per a Bloomberg investigation.

Thinking Machines Lab: Thinking Machines says its mission is to build AI that people and organizations can shape and make their own, and that “extends human will and judgment”  —  The mission of Thinking Machines is to build AI that extends human will and judgment.  —  Artificial intelligence …

A close look at how Zhipu built GLM-5.2, from DeepSeek Sparse Attention to critic-based PPO, with a live agentic coding demo traced in W&B Weave.

"Our intent was to provide a useful creative tool and to give people control over whether their public content could be referenced in this way," the company said in a blog post. "We've heard the feedback that this feature missed the mark, so it's no longer available."

Reuters: SK Hynix CEO Kwak Noh-Jung says the memory industry is heading for its worst-ever supply shortage in 2027 and demand will outstrip supply beyond 2030  —  SK Hynix (000660.KS) Chief Executive Kwak Noh-jung said the global memory industry is heading for its worst-ever supply shortage in 2027 …

Illustration by Thomas Fuller/SOPA Images/LightRocket via Getty Images Apple has filed suit against OpenAI, alleging the AI giant stole trade secrets. In a filing, Apple says employees who went to OpenAI had access to its systems after leaving. Apple argues that OpenAI's hardware business is "rotten to its core" and built on stolen secrets. Apple's new lawsuit against OpenAI reads less like a routine trade-secret complaint and more like a play-by-play of how the iPhone maker says its rival built its consumer hardware ambitions. Filed Friday, the suit accuses OpenAI of "a coordinated pattern of misconduct at an institutional level," including everything from exploiting a security bug to recruiting Apple engineers for "show and tell" interview sessions involving confidential hardware. In a statement to Business Insider, an OpenAI spokesperson said the company has "no interest in other companies' trade secrets." "We remain focused on building innovative technology that empowers people everywhere," the spokesperson said. Here are the biggest allegations. OpenAI allegedly exploited an Apple security bug According to the complaint, Apple's investigation began after former employee Chang Liu left for OpenAI in January. Apple alleges Liu failed to return a company laptop and later discovered he could still access Apple's internal systems because of what the company describes as an authentication bug. Rather than reporting the issue, Apple said that Liu downloaded dozens of confidential engineering files while working at OpenAI, including documents related to unreleased products, technical specifications, presentations, and manufacturing processes. Apple also alleges Liu encouraged another Apple engineer who was interviewing at OpenAI to study confidential Apple materials ahead of her interviews and advised her how to avoid attracting attention from Apple's security team while copying files. The complaint says that the two moved conversations to a private messaging app to avoid detection. Apple says OpenAI turned interviews into intelligence-gathering sessions Apple also accuses OpenAI of systematically using its job recruitment process to obtain trade secrets. The buzzy AI company, which has filed the initial paperwork in a lead-up to an initial public offering, has become one of the most in-demand employers in tech. Amid the influx of talent, several hardware engineers at Apple have decamped to the AI juggernaut. The lawsuit claims that Chief Hardware Officer Tang Tan, a former Apple vice president who spent 24 years at the company, asked Apple job candidates to bring physical components for "show and tell" sessions during interviews. OpenAI employees would also ask candidates for prototypes and information about vendors, the complaint alleges. One then-Apple employee screenshotted and downloaded files from a "highly confidential Apple project," the lawsuit says, alleging that Tan later asked about that project during the interview. "OpenAI's recruiting practices suggest it hires these individuals at least in part because of the confidential Apple-specific knowledge and expertise they have and could improperly obtain," the complaint says. Apple says the misconduct went all the way to OpenAI leadership Apple says it raised its concerns with OpenAI earlier this year and asked the company to investigate whether Apple's confidential information had entered its business. According to the complaint, OpenAI never responded, prompting Apple to continue its investigation before filing suit on Friday. "This is the tip of the iceberg," the complaint reads. "Apple lacks visibility into what's been happening behind closed doors at OpenAI, where such misconduct is normalized and exemplified by leadership." "This much is clear, however: at every level, from members of its Technical Staff to its Chief Hardware Officer, and in coordination with business partners, OpenAI has been stealing Apple's trade secrets and confidential information," the complaint continued. "As a natural result, OpenAI's nascent hardware business now rests on the shakiest of foundations, rotten to its core by its illegal reliance on misappropriated trade secrets." Apple said it is "left with no choice" but to seek damages in an amount to be determined at trial, and court orders preventing OpenAI and the individual defendants from possessing or using Apple's trade secrets. The suit marks a dramatic escalation in what had been one of Silicon Valley's highest-profile AI partnerships, following Apple's integration of ChatGPT into Apple Intelligence in 2024. The relationship has frayed as OpenAI ramped up its own consumer hardware ambitions by hiring former Apple executives and engineers. Apple's suit isn't the only high-profile claim that OpenAI has engaged in dishonest business practices. The ChatGPT creator is also embroiled in a legal battle with The New York Times over claims that it allowed its chatbot to use copyrighted material to train its systems. Elon Musk also previously filed suit against OpenAI, alleging that CEO Sam Altman and the company's president, Greg Brockman, "stole" the original OpenAI nonprofit and unjustly enriched themselves by changing it to a for-profit model. A jury handed OpenAI a legal victory in May, concluding that Musk had missed the three-year statute of limitations to sue. Tan, Liu, and representatives for Apple did not immediately respond to requests for comment from Business Insider. Read the original article on Business Insider

Corbin Bolies / Variety: Meta says it will discontinue a feature that allowed users to generate images in Meta AI using public Instagram accounts, following days of criticism  —  Meta said it will discontinue an AI feature that allowed users to generate images using public Instagram accounts following days of criticism …

Data centers are driving up the company’s use of electricity—and carbon pollution.

Marcus Mendes / 9to5Mac: In response to Apple's trade secret theft lawsuit, OpenAI says “we have no interest in other companies' trade secrets”  —  OpenAI has issued a formal statement in response to Apple's lawsuit accusing the company of trade secret theft.  Read it below.  —  OpenAI denies Apple's allegations

Reinforcement learning uses error information to adjust control algorithms.