DeepSeek's R1 model generated 800,000 worked solutions, which were then filtered and fine-tuned on various off-the-shelf open models. The resulting distilled 32B model solved competition math it had no business solving, and the 7B model began verifying its own work and branching its reasoning mid-stream. This result contradicts previous arguments against naive sequence-level imitation in reasoning distillation.
Graph engineering refers to a design approach for creating complex AI systems by organizing tasks and decisions into interconnected nodes or "graph" structures. This method, which has gained popularity in recent weeks, differs from loop engineering, which focused on guiding an agent through each turn of the task completion process. Graph engineering involves giving an agent a goal, a way to check its work, and a repeating cycle until the task is done, with multiple tasks potentially running in parallel and requiring more complex connections between them.
The US government cannot ban open source AI, as it is a decentralized and publicly available technology that can be developed and shared by anyone. Open-source models release their source code, training data, and weights, making them more transparent than closed-source models. The launch of Moonshot's new Kimi K3 model has sparked debate over regulatory capture and government interference in AI development, but it is unlikely to lead to a ban on open-source AI.
Alibaba announced Qwen3.8, a 2.4-trillion-parameter model slated for an open-weight release, with a preview version available through Alibaba's Token Plan and other platforms. Apple has sent legal letters to dozens of former Apple employees now working at OpenAI, claiming they recruited key engineers and benefited from proprietary designs and trade secrets. Netflix's Large Language Model (LLM) stack was developed and deployed within its existing production infrastructure. Kimi Code CLI is a terminal-based AI coding agent that can read, edit code, run shell commands, and choose the next step based on feedback.
US debate over restricting Chinese open models is moving from rhetoric toward policy, with measures such as procurement restrictions, Entity List designations, security advisories, liability requirements, and public pressure campaigns. Technical voices argued that restricting open models would hurt competition, sovereignty, and defensive security more than it helps incumbents. Frontier models are now "obviously superhuman at some mathematical tasks," according to recent discoveries, including a counterexample to the 3D Jacobian conjecture.
Real Intent Superagent is an all-in-one real estate growth platform that provides intent data to help agents find motivated buyers and sellers. The platform uses advanced technology to analyze data sources, providing accurate and reliable information for lead generation. Agents can leverage Real Intent's data through various strategies, including using real-life success stories as examples of how to effectively use the tool. Regional exclusivity allows access to different regions, with availability determined by current spots.
Every industry/ecosystem has its supreme village idiot, and @deanwball is that for AI. Congress passed a law in 2026 that restricted some uses of DeepSeek/High Flyer with waivers permitted. Only those models went through the democratic process, not a "Deep State scheme".
NVIDIA NVLink provides high-bandwidth, low-latency scale-up networking for AI factories, enabling up to 3.6 TB/s per GPU and 260 TB/s rack-level bandwidth with 130 TFLOPS in-network compute, significantly outperforming off-the-shelf Ethernet solutions in large-scale MoE and LLM workloads.
As AI spending increases, enterprises are experiencing opaque pricing and receiving backward-looking bills for their AI costs, prompting a reevaluation of their AI model strategies.
Tech workers are increasingly unionizing, driven by the deployment of artificial intelligence and its impact on their jobs. Workers at companies like Meta and Google DeepMind are objecting to how AI products are being used, such as for military purposes or employee surveillance, and seeking collective bargaining to contest these practices. Unions have become popular among tech employees in recent years, with participation in collective actions increasing and total union actions at Alphabet doubling in 2025 from the year prior.
Nine's managing director of publishing, Tory Maguire, has announced the axe of 30 jobs at the Sydney Morning Herald and the Age due to an "extreme state of disruption" caused by AI. The job cuts are part of a voluntary and targeted redundancy exercise, which is intended as an "evolution" to adapt to structural change.
The UK's water industry says government forecasts are "fatally flawed" and do not include an estimate of datacentre water demands, which could exacerbate existing water shortages. Datacentres already consume large amounts of water in water-stressed areas, with some requiring up to 3m litres a day. Water companies warn that new datacentre growth may be impossible without policy change due to existing pressure on the network.
In the UK, nearly one-third of women in the 50-64 age bracket are economically inactive, while men have a slightly higher rate. The unemployment crisis is hitting all demographics and sectors, with entry-level positions declining and AI-powered hiring processes causing applicants to feel rejected in under a minute. This has created a pipeline problem for employers, who will eventually struggle to fill middle-tier staff roles without experienced workers.
Production agent systems are built on top of a shared infrastructure that determines their success. Despite 40% of enterprise applications embedding AI agents by the end of 2026 and 40% of agentic AI projects being canceled by the end of 2027, many teams struggle with the gap between adoption and production due to inadequate infrastructure. A common pain point is the lack of industry-standard controls for autonomous systems, leading to stalls at security review, cost ceilings, and evaluation phases.
TLDR AI is a daily newsletter covering AI news, research, and tools in a 5-minute read, with 1,100,000 subscribers. It targets engineers, researchers, and technical professionals who want to track AI without hype. TLDR AI is denser and more engineering-focused than The Rundown AI, summarizing research papers and developer tools alongside industry news.
Qwen3.8 is launching and going open-weight soon with massive 2.4T parameters, making it one of the most powerful models available today, comparable to Fable 5. The model has debuted on Alibaba's Token Plan, Qoder, and QoderWork, allowing users to test it out.
US administration is likely to intervene aggressively in the viability of Chinese operating systems, with critics' predictions of well-calibrated FUD seeming optimistic due to the government's history of overreaction to novel risks. The administration has a tendency to capriciously overreact to emerging capabilities and prize unilateral control over dangerous ones, including cybersecurity. This intervention could be driven by economic concerns, as open source dissemination of models beyond a certain threshold threatens the American stock market's existing economic structure.
Little-known data centre developers are cashing in on the AI boom while facing growing backlash over electricity, land, and pollution concerns. These firms often negotiate with household-name tech companies like Meta, Google, and Microsoft, and are now at the center of contentious local political fights across the US. Organized opposition to data centers has grown dramatically, with nearly $130 billion worth of projects blocked or delayed in 2026 alone. Industry defenders point to genuine economic benefits, but critics argue that costs are being borne by communities and who captures the benefits remains a contested question.
Twitter Library is an AI newsletter that provides in-depth content on the latest developments in machine learning and artificial intelligence, including small language models, free courses, best AI coding tools, agent RL training frameworks, open-source vector databases, reasoning RL methods, advanced RAG types, AI co-scientists, open-source tools for foundation model deployment, AI ethics books, agentic memory, multi-agent systems, LoRA variants, attention mechanisms in transformers and LLMs, LLM security and AI guardrails, self-improving AI agents, JEPA milestones, modern reinforcement learning approaches, emerging memory architectures for AI agents, core AI software development paradigms, robotics news, GitHub repos for building local AI agents, must-read books and surveys about AI and machine learning.
China's open-source model leads in reasoning and coding in 2026 with Kimi K2 surpassing competitors like DeepSeek-R1, Qwen3, and GLM-4.5 in benchmarks and agentic use cases. China plans to escape economic stagnation through an AI-driven industrial revolution. India is developing its own indigenous LLMs, receiving government support, but faces challenges in achieving global AI integration.
Advanced materials are essential to advancing next-generation AI technology, enabling greater processing power, memory, energy efficiency, and reliability. Materials companies must innovate continuously to support the increasing demands of AI workloads, which require sophisticated thermal management, higher-voltage power architectures, increased data storage, and faster data transmission. The development of new materials is driven by performance requirements, but also by a growing expectation for responsible manufacturing practices. Digital tools, including AI, are accelerating the pace of discovery in materials science, allowing researchers to identify promising candidates earlier and reduce physical experiments required.
US advisors to President Trump on AI publicly clashed over how to handle China's free, open-source AI model Kimi, which rivals models from OpenAI and Anthropic. David Sacks called Anthropics' models "lobotomized" and "woke," while Emil Michael criticized OpenAIs new head of strategic futures as a "supreme village idiot." The disagreement highlights Trump's dilemma: balancing economic growth with concerns over Chinese AI models threatening national security.
To build a side hustle using Gemini, consider the following ways: * Launch and grow a business by streamlining planning, market research, and branding. * Automate logistics, manage customer data, and calculate pricing models to ensure profitability. * Use Deep Research feature to compress hours of work into minutes and generate professional analysis on competitors or markets. * Define your brand aesthetic using Gemini's design tools and connect apps like Canva. * Put logistics on autopilot with Gemini Spark, a personal AI agent that runs in the background 24/7. * Know when your price is right by simplifying math and generating personalized recommendations.
NVIDIA is showcasing its latest graphics and simulation advancements at SIGGRAPH, including agentic and physical AI. The company's keynote discussed neural rendering techniques, world models, and simulation methods built by AI. NVIDIA also introduced the Model Context Protocol (MCP) connections, allowing AI agents to work inside creative tools while staying in control.
The NVIDIA Omniverse libraries now include ovrtx, a lightweight C and Python SDK for generating RTX-rendered camera and sensor outputs from OpenUSD content. This allows developers to bring physical AI capabilities into their tools and services while maintaining host app control. To integrate ovrtx, create an ovrtx renderer, load or reference OpenUSD content, define RenderProducts and RenderVars, step rendering or sensor simulation from the application loop, and map results back to CPU or GPU memory.
Researchers at Writer have optimized the orchestration layer around foundation models, reducing tokens per task by 38% and cost-per-successful-task by up to 61%. The optimization was achieved through a systematic look at system prompt caching, interaction history compaction, tool management, retrieval strategies, and error management. This approach allows engineering teams to build highly cost-efficient AI applications without changing the underlying foundation model.
A single AI agent conversation can appear flawless but still signal a broken product due to limitations in evaluating its performance, according to industry leaders. Evaluating agents is shifting towards comparing cohorts of users against a baseline rather than scoring individual traces. To address this, teams are building exhaustive evaluation suites that function as the product's specifications and incorporating broader monitoring systems that catch real failures before they occur.
Zillow's SVP of Engineering Toby Roberts and Glean co-founder and CEO Arvind Jain discussed how the company built an AI architecture that carries context across multiple customer interactions, not raw data. The team started by establishing a robust data foundation, but found that building something that remembered where customers were in their journey was the harder problem to solve. To overcome this, Zillow chose to own its own context layer rather than relying on external chat interfaces. This approach allowed for cost savings through mechanisms like model routing and precomputed context, which can reduce token consumption by as much as half.
Millions of dollars have been invested in generative AI pilots over the past two years, yet many initiatives stall before reaching production. The problem lies not with the model itself but with the enterprise data foundation underneath it, which is often unready. This is known as the 'Cleanup Trap,' where organizations assume they can "clean up" legacy data and pass it to a large language model without issues. To break out of this trap, enterprise data teams must adopt rigorous data quality checks and validation frameworks before data reaches an AI orchestration layer.
NVIDIA has released Cosmos 3 Edge, a 4-billion-parameter open world model designed to run on-device, enabling robots and vision AI agents to understand surroundings, reason in real time, and generate robot actions locally. The model uses a Mixture-of-Transformers architecture with two towers that process vision and text tokens for understanding and reasoning, as well as vision, audio, and action tokens for prediction, generation, and neural simulation. Cosmos 3 Edge delivers memory-efficient inference across NVIDIA edge computers, targeting various devices including the Jetson T2000 and T3000 modules, and can be post-trained in about a day using the open Cosmos framework.
Qwen-Audio-3.0-TTS is a hosted text-to-speech model released by Alibaba's Tongyi Lab, featuring two variants: Flash and Plus, targeting real-time interaction and high-quality generation respectively. The model supports 16 languages and has the best word/character error rate (WER/CER) in 10 of the 16 languages, with Plus ranking first on speaker similarity across all 16 languages. The model includes fine-grained tags for non-verbal details, such as laughter and coughing, and has a competitive pricing structure at $27.59 per 1M characters.
Gen Z is living in an "intimacy economy" where human connections are commodified, with AI chatbots offering emotional support and companionship as a cheaper alternative to human relationships. This trend is driven by the growing use of AI in therapy, dating apps, and social media, which has led to a sense of emotional disconnection among younger generations. The AI companion market is projected to reach $31.1bn by 2032, with companies using guilt and FOMO tactics to keep users engaged. To combat this trend, it's essential for Gen Z to reinvest in human relationships and rebuild emotional literacy.
A collection of speeches and writings by Pope Leo XIV has been certified as human-authored by Australian company Proudly Human, using AI detection tools. The certification process involves authors signing a declaration that they are operating within minimum standards, with the company's tools focusing on writing nuances and grammar. Proudly Human aims to provide an alternative to government regulation by giving readers confidence in knowing whether writing was produced by a human or AI.
Research during Hungary's parliamentary elections found that AI chatbots provided inaccurate and unreliable guidance to voters, often recommending the wrong party or listing parties not even running. The chatbots misclassified profiles, omitted relevant parties, and gave highly volatile answers to identical prompts. The researchers warn that in more competitive elections or where voters are less certain, such outputs could potentially impact the outcome of an election.
Anthropic redeploys its Claude Fable 5 model after talks with the US government, adding new cybersecurity classifiers and drafting a jailbreak-severity framework. Anthropic also launches Claude Sonnet 5 with discounted pricing, improved agentic coding, and reduced misaligned behavior. Google releases NotebookLM, generating TikTok-style vertical video summaries of research, and introduces Nano Banana 2 Lite, a faster, cheaper image generator.
Researchers at DeepSeek used a large reasoning model R1 to generate 800,000 "traces" of thought, which were then fine-tuned on smaller models using supervised learning. The results showed that distilled models, even those smaller than the original teacher model, could solve complex math problems and exhibit emergent behaviors like self-correction and mid-stream reasoning.
Alpöge posted a one-line formula on X that breaks the Jacobian conjecture, a famous algebra rule that has stumped mathematicians since 1939, with Anthropic's Levent Alpöge disproving it during the World Cup final. The solution is short enough to confirm directly, and experts say this answer is the first confirmed solution in decades.
Open-weight models have reduced the performance gap between open and closed models from 6-9 months to 3-5 months. Kimi K3, a 2.8T parameter MoE model, is released by Moonshot AI and will be made available on July 27th. China's AI ecosystem has emerged as a strong competitor, with companies like Alibaba announcing plans to release open-weight models.
America's AI edge loss to China will make it harder for frontier capabilities to be monetized, as open-source models like Moonshot's Kimi K3 gain credibility among enterprises. Export controls will look less effective and the trillion-dollar AI buildout harder to justify. NVIDIA may benefit while AI labs lose out on funding. America's reliance on business over geopolitics risks hurting itself, rather than a specific opponent. The shift in power dynamics will prioritize trajectory over point achievements.
A quiet day for technical news, with the AIE Security track being released today. Tariq Shaukat from Sonar CEO emphasized the importance of verification for safety and correctness. The top release of the day went to Kimi K3, which is emerging as a strong open-weight contender in agentic and frontend tasks, with the Alibaba Qwen 3.8 Max model also announcing improvements and potential opening-weighting.
The reduced cost of writing code has made reverse-engineering and automating devices in homes more feasible. Prior to coding agents, the effort required for such tasks was significant due to the high cost and potential maintenance issues. Coding agents have significantly lowered these costs, making it more attractive to attempt automation projects without worrying about the financial burden or future maintenance. This shift has increased the ease with which people can automate devices in their homes.
Ben Thompson proposes the US should pass a law making data collection for training models fair use and barring terms of service that forbid distillation, to support US companies competing with Chinese counterparts. Alibaba's decision to release Qwen 3.8 Max as open weights may have been influenced by Xi Jinping's recent speech on encouraging openness and collaboration.
Install ML Flow by running `pip install mlflow` in the terminal and starting the server with `mlflow server --host 127.0.0.1 --port 8080`. Create a new experiment titled "Logging Example" using `mlflow.set_experiment("Logging Example")`, and then start an MLFlow run titled "salary_baseline_regression" using `with mlflow.start_run(run_name="salary_baseline_regression") as run:`. Log the adjusted R2 metric, model parameters, and model using `mlflow.log_metric`, `mlflow.log_params`, and `model_info = mlflow.sklearn.log_model`.
To run Claude Code agents for 24+ hours, give them all necessary permissions, including auto-mode or skip all permissions, and ensure they have access to necessary tools. Ensure agents can verify their own work by providing detailed instructions and testing frameworks. Implement agentic code review using tools like Codex to minimize human review time. Run agents remotely to prevent interruptions, offering benefits such as reduced oversight but potentially clunky setup.
PyMuPDF parses a page in five milliseconds for free, while a vision LLM can cost ten thousand times more and take ten seconds. The pipeline uses a feedback loop to escalate parsing when necessary, starting with the cheapest parser (PyMuPDF) and escalating to a deeper parser (Azure DI) only when checks against the parse output flag it as insufficient. This process improves the reliability of the pipeline, especially for complex layouts like tables and figures, where the cheap parser may return incorrect results. The LLM's self-evaluation flag helps catch parsing errors, but can also lead to over-confidence in the model's answers if the structural risk is not properly assessed.
Byzantine Fault Tolerance, a system property that allows a system to continue operating properly even with malicious actors included in the system, has its roots in the Byzantine Generals Problem. This problem involves a group of generals who must reach a collective decision without knowing who among them is a traitor. In 1982, computer scientists Leslie Lamport, Robert Shostak, and Marshall Pease mathematically proved that for a system with n nodes to continue normal operation, at least n ≥ 3f + 1 total number of nodes is needed, where f represents the number of Byzantine (traitor) nodes. Blockchain technology has been found to be an elegant solution to this problem, specifically through its consensus mechanism called Proof of Work (PoW). PoW makes it computationally expensive for a traitor node to participate in the consensus process, rendering it economically irrational to cheat.
Grabette is an open system for recording robot-manipulation data that allows users to record and process demonstrations from their own hand using a handheld gripper, camera, and IMU. The system records 6-DoF trajectory data and provides a clean, robot-ready dataset. A robotic counterpart, Gripette, executes the movements learned after training. The Grabette system is designed to be low-cost, open-source, and user-friendly, with hardware components that can be ordered off-the-shelf. It enables users to contribute to a large, collaborative manipulation dataset.
NVIDIA Cosmos 3 Edge is a 4-billion-parameter open world model that helps robots and vision AI agents understand their surroundings, reason in real time, and generate robot actions on edge devices. The model delivers memory-efficient, high-throughput inference across NVIDIA edge computers, including the newly announced Jetson T2000 and T3000 modules.
There is no article to summarize. The provided text appears to be the title and metadata of a Reddit post, but it does not contain any content to summarize.
China's low-priced Z.ai Model is exposing costly coder habits by using loose token budgets to protect its models Opus and GPT-5.5, rather than the traditional approaches used by US rivals. Z.ai's latest large language model, GLM 5.2, narrowed the gap with US rivals in some coding benchmarks.
The World Cup has concluded, and while there were no major winners, socializing IRL with friends and family was a notable victory. Author Scott Galloway is set to discuss his book "Inside the Anti-Social Century" in a live conversation with Derek Thompson, exploring how technology's convenience comes at the cost of human connection. The event will be held on July 20 at 10:00am ET, and viewers' input will shape the discussion.
Researchers at Princeton University and the University of Chicago found that Large Language Models (LLMs) like ChatGPT, Claude, and Gemini can develop biases when hiring job applicants, stereotyping candidates by demographic group more than humans do. The models quickly segregate candidates from different groups into different jobs based on early observations of hiring outcomes, with newer models showing stronger biases. Designing goals that incorporate desirable social values can make the models act in socially desirable ways, and providing personal information about individuals can reduce bias.
Weather data sabotage is a rising concern, with tampering in individual stations having potential to snowball into bigger problems. Traditional forecasting systems have built-in safeguards like data assimilation to ensure reliable observations, but new threats are putting these mechanisms at risk. As weather prediction becomes increasingly reliant on artificial intelligence (AI) and big language models, the stakes rise, introducing new risks of fraud, compromised disaster preparedness, and national security. To mitigate this risk, experts recommend monitoring stations, protecting data, and ensuring continuous accountability along the chain to safeguard weather forecasting accuracy.
AI systems are more likely to form biases when hiring than humans, potentially leading to stereotyping job applicants. Research suggests AI can develop its own biases through experience and may even outdo human bias. Meanwhile, the risk of weather data sabotage is growing due to the increasing use of data-driven AI forecasting in prediction markets, where manipulation could impact weather predictions accuracy.
Nvidia, Meta, and Samsung have launched a collaborative foundry called "Materials Foundry" to develop alternative materials for use in the production of semiconductors. The goal is to reduce dependence on rare materials, which are currently in short supply and drive up costs. By developing new materials, the group aims to increase efficiency and lower production costs.
Autonomous AI agents breached Hugging Face's systems, exploiting a malicious dataset to gain unauthorized access to internal datasets and service credentials. The attacker used remote-code loaders and template-injection flaws to execute code execution paths, bypassing safety guardrails built to prevent attackers from accessing commercial AI models. The incident highlighted the need for authenticated trust in AI systems, requiring security teams to understand who is asking for analysis, why, and under what governance.
There are approximately 450 large data centres across the UK, which power cloud services and increasingly AI. Nearby residents have raised concerns about land use, pollution, noise, and the "heat island effect" caused by datacentres. The Guardian aims to hear from people living near AI datacentres, allowing them to share their experiences anonymously via a secure online form.
Rater State Bias in RLHF Preference Data: An Audit Framework proposes a structured confound in Reinforcement Learning from Human Feedback (RLHF), where pairwise preference labels reflect not only the compared outputs but also the rater's state during annotation. Under stressful or distressing conditions, raters' preferences can shift over time, encoding rater state alongside judgments about response quality. This phenomenon, termed "rater state shift," is a plausible source of structured bias in RLHF preference data. The authors develop an audit framework to study this bias and propose five falsifiable predictions for initial audits.
A lightweight 1D CNN-based model was designed and validated for affective touch classification in soft plush companions. A diverse dataset of 1326 labelled gesture sequences collected from 25 participants achieved 75% test accuracy with a 13.2k-parameter model. The study identified compact dilated 1D CNNs as the most effective solution, compatible with real-time operation on microcontrollers.
Large language models exhibit robust intra-task consistency, preserving relative risk posture across tasks, and a convergence toward a restricted risk-attitude distribution. Most tested LLMs demonstrate stable mappings from contextual belief to risk decision within fixed task domains, with only minor cross-domain rank-order stability variations. Human participants show a broader range of risk attitudes compared to LLMs.
Graph Neural Networks (GNNs) have emerged as the leading paradigm for link prediction, enabling the inference of missing connections and the anticipation of potential future links. Recent advancements in GNN-based link prediction can be categorized into four key architectures: GCN-based, GAE-based, GAT-based, and GFormer-based methods. These techniques have been applied in prominent use cases such as knowledge graphs and recommendation systems, demonstrating their real-world impact. The current challenges in GNN-based link prediction include data quality issues, scalability concerns, and the need for more efficient algorithms.
Multi-agent LLM systems rely on a Planner to decompose goals into sub-task sequences, which can be attacked via planning-phase prompt injection. Researchers Yuhang Wang and colleagues introduce PlanFlip, a framework of four attacks -- GoalSubstitution, PriorityInversion, ContextPollution, and RoleConfusion -- that achieve cascade amplification and evade keyword filters. Evaluating nine frontier LLMs across 3,479 episodes, the study reveals three findings: stronger models (like GPT-5) are vulnerable to attack success rates; homogeneous pipelines have a blind spot; and reasoning-augmented models resist injections.
AI agent systems with large language models (LLMs) are inherently non-deterministic due to various sources of noise, including LLM sampling variance and external API state. agrepl, a developer-first CLI framework, intercepts all external interactions at the transport layer, serializes them as structured execution traces, and replays them in a strictly isolated environment with zero outbound network access. The framework achieves high replay fidelity (F = 1.0) and median per-step latency reduction of 98.3%. agrepl is implemented in Go, ships as a single static binary, and is released under the MIT licence.
GOI (Generative Ontology Induction), a domain-agnostic framework, induces a generative blueprint from document corpora using large language models. It exports a typed graph in YAML/JSON format, including entities, dimensions, properties, relationships, and constraints. GOI achieves 95-100% structural coverage for four contrasting ontologies, with varying degrees of accuracy depending on the familiarity of the document type to the model.
A controlled evaluation of nine open-weight language models on a 1,085-example, 16-topic multiple-choice benchmark was conducted to study AI democratization through structured local deployment. The benchmark emphasizes symbolic precision and constrained formatting, and a shared parameter-efficient fine-tuning pipeline adapted a subset of models using 4-bit NF4 quantization. The results show that a disciplined workflow of benchmark construction, cross-model evaluation, and low-cost specialization makes sub-3B models viable as local experts for structured niche workloads, with Qwen Coder 3B leading at 75.67% strict accuracy.
Masked diffusion language models (MDLMs) achieve better coherence, groundedness, and empirically validated rollout diversity than autoregressive (AR) LMs over 4x their parameter size, at comparable inference latency. MDLMs via bidirectional anchor-aware denoising overcome the left-to-right bias of AR world models, allowing for conditioning on globally interdependent state anchors like tool schemas and prior turns.
Weak-to-strong off-policy reinforcement learning via auxiliary branches, W2SPO, injects short segments into target model trajectories to induce stronger reasoning policies. These segments are as brief as 8 tokens and restrict policy updates based on final verifiable rewards. Empirically, W2SPO outperforms evaluated baselines on mathematical reasoning benchmarks by improving Pass@1 from 62.3% to 64.2%, with a 3.55 times training speedup compared to vanilla GRPO.
A novel reinforcement learning-guided non-dominated sorting genetic algorithm II (NSGA-II) enhanced with gray relational coefficient for multi-objective optimization is proposed. The method combines an RL agent controller and GRC-based selection to improve convergence and diversity of Pareto fronts, achieving convergence improvements of about 5.8% and 4.4% over NSGA-II on the Kursawe and CONSTR benchmarks, respectively.
A correction-based framework called DocOCR-Eval is proposed to evaluate and select Optical Character Recognition (OCR) tools without ground truth labels. This framework employs a three-staged correction and ranking strategy, which aggregates multiple large language models to improve alignment with annotation-based rankings. Extensive experiments demonstrate that reliable OCR tool selection can be achieved in label-limited settings, providing practical guidance for deploying document parsing systems across diverse real-world document collections.
ARGO, a smart eyewear platform, leverages the STM32N6 microcontroller and its NPU to enable on-device machine learning for real-time urban obstacle recognition. The platform incorporates a multimodal sensor suite and achieves 10 FPS with a continuous autonomy of ~113 minutes on a 200 mAh battery. ARGO's primary contribution lies in the holistic co-design of hardware, firmware, and artificial intelligence, centered on an optimized YOLOv11 model. This approach demonstrates the feasibility of high-performance, privacy-preserving, and socially acceptable assistive devices.
LLMs increasingly deploy in security-critical systems, yet their inability to forget creates serious cybersecurity risks due to encoded sensitive information across billions of parameters. LLM unlearning has emerged as the principal cyber defense response to remove or suppress targeted knowledge from a trained model without retraining. Gradient-based methods dominate the field due to compatibility with existing training pipelines and scalability to billion-parameter models, but it remains unresolved whether current methods genuinely remove knowledge.
The current fixed-shape all close-style check with hand-picked absolute and relative tolerances in tensor-kernel correctness tests are too loose, resulting in 9.3 percentage points fewer bug detections per test case. By analyzing the element-wise error distribution of every test case from accumulated cloud GPU runs, the authors found that a tighter tolerance, such as 1/2^184 for attention_triton fp16, would justify an absolute tolerance. This results in a 82.4% (2,034 of 2,467) bug-detection recall compared to a 73.2% (1,805 of 2,467).
A production-inspired multi-stage framework called RouteCost is proposed to improve pre-order shipping cost estimation in e-commerce. The framework combines time-aware demand forecasting, fee-card-informed baseline pricing, residual correction, and box-consolidation inference to produce accurate product-level shipping cost predictions. Across over 250,000 orders, 260 products, and 18 months of order history, RouteCost improves predictive quality and aggregate calibration while preserving route-level interpretability.
Researchers evaluated OrthoGrad, which removes the component of each weight gradient parallel to the current weight vector, in noisy-label image classification. On MNIST with small-data regimes, OrthoGrad improves test accuracy most clearly for CNNs while reducing corrupted-label fitting. The method can alter memorization trajectories but does not prevent eventual noisy-label memorization, and its effect becomes weaker in larger-data regimes where gradients are already nearly orthogonal to weights.
Researchers Zachary Wojtowicz, Ayush Nayak, and Jacob Andreas have developed a method called "weights to words" that automatically discovers domain-relevant preference dimensions from high-dimensional choice data. These dimensions are described in natural language and paired with vector representations in the model's space, addressing under-determination and opacity issues in machine learning algorithms. The method was tested on two human-subject experiments involving moral dilemmas and movie selection, demonstrating improved prediction accuracy when incorporating participants' structured edits and preference profiles.
A new approach, called the Token-Level Cross-Modal Transformer with Contrastive Multi-Task Learning, has been proposed to improve breast cancer subtype classification and survival prediction by integrating genomic and clinical modalities. The model overcomes existing limitations by allowing fine-grained token-level interactions across modalities, using structured token exchange for cross-modal fusion, and jointly optimizing survival and classification objectives through contrastive multi-task learning. This method enhances the performance of previous approaches in precision oncology.
HantaWatch, a federated learning framework, enables laboratories and surveillance sites to collaboratively train sequence-based models for hantavirus genomic surveillance without sharing raw data. The framework supports high-risk screening, outbreak-associated prediction, clade classification, and clinical-syndrome categorization while balancing predictive performance and false-negative risk. HantaWatch converts model output into risk scores, confidence estimates, uncertainty flags, and ranked expert-review priorities, providing a practical federated decision-support layer for decentralized hantavirus surveillance.
Lipschitz continuity is a fundamental property of neural networks that characterizes their sensitivity to input perturbations, governing robustness, generalization, and optimization dynamics. The review explores theoretical foundations, estimation methods, regularization approaches, and certifiable robustness in deep learning, providing a comprehensive reference for researchers and practitioners seeking a deeper understanding of Lipschitz continuity.
MTSSL: Meta-Thresholding Semi-Supervised Learning establishes a unified theoretical framework to explain the role of $\tau$ in Semi-supervised Learning (SSL). The value of $\tau$ adjusts the numbers of correct and incorrect pseudo-labels to balance their error terms, allowing SSL to achieve similar performance with varying optimal values. MTSSL optimizes $\tau$ via differentiation, treating it as an updatable parameter, and demonstrates superior performance through extensive experiments.
Isotonic Conformal Prediction (ICP) is a framework that decouples calibration from prediction-set construction by fitting a single isotonic recalibration map and constructing prediction intervals within strata of similar recalibrated predictions. Two procedures, Split Isotonic Conformal Prediction (SICP) and Transductive Isotonic Conformal Prediction (TICP), have been developed within this framework. SICP attains prediction-conditional validity in finite samples and self-calibration asymptotically at a lower computational cost, while TICP attains both objectives exactly in finite samples through a per-test-point inner loop.
Label-augmented conditional diffusion (LACD), which incorporates unlabeled examples by assigning a trivial label and performing joint denoising score matching, provides sufficient conditions for population-level identifiability of the target conditional distribution. When sufficiently many unlabeled samples are available, LACD converges strictly faster than purely supervised estimators in total variation distance and at least as fast in Wasserstein-1 distance.
A causal Markov condition (v-CMC) is proposed as a causal independence principle for value, and the conceptual foundations of a "causal value theory" linking causality and utility. The v-CMC is formulated locally, globally, and in decompositions, and its equivalence with related concepts is proven. A Bellman-type recursion is derived as a special case of the v-CMC, supporting modular transfer and updating of utility information across causal contexts.
Changyu Liu, Yuling Jiao, and Jian Huang propose a semi-supervised framework combining conditional stochastic interpolation with low-dimensional latent representations to effectively leverage structural information from unlabeled datasets. The RepG (Replica-GAN) decomposes generation into two stages: label-dependent latent sampling and high-dimensional reconstruction.
Dropout and Random Gradient Masking (RaM) training techniques used to improve deep learning performance are asymptotically equivalent in large ResNets. In the complete feature learning regime, both methods converge to the same limiting dynamics. This equivalence holds for several variants of dropout and RaM, including layerwise dropout used in stochastic-depth ResNets, albeit at slower quantitative rates.
Deep Adaptive Bayesian Screening (DABS) is a method for performing adaptive factorial screening in high-dimensional discrete design spaces. It learns a policy network offline to sequentially select informative experiments, amortizing Bayesian Optimal Experimental Design. DABS handles binary designs, incorporates sparsity and interactions via a spike-and-slab prior with strong heredity.
Maxence Noble and co-authors introduce Twisted Schrödinger Bridge Matching (TSBM), a diffusion-based method to approximate optimal transport dynamics between two prescribed boundary distributions. TSBM extends the Iterative Markovian Fitting (IMF) paradigm to handle generalized Schrödinger bridge problems with time-dependent potentials, including continuous- and discrete-time scenarios. The method recovers the original DSBM objective when the potential vanishes and provides a new bridge-matching loss that depends on the gradient of the potential.