A robot performing complex tasks in a setting like a kitchen demonstrates the potential for a "ChatGPT moment." This involves giving a machine useful new work through conversation and examples, making teaching an ordinary activity. Generalist robot models enhance the credibility of this prospect. The remaining challenges involve learning, control, and the economics of deploying machines in new environments. Robotics needs a similar encounter with versatility as language does with broad capability.
Loni Stark and her partner built and continuously ran an autonomous AI agent for 154 days. They tested what maintains an agent's identity over time by swapping the underlying model. They are now planning to test if swapping the entire framework affects the agent's identity. This experiment connects to Stark's research on personal identity and the lack of outlet in daily life. The experience also prompts reflection on how AI is reshaping brand trust.
The publication explores fundamental machine learning concepts, including LDA, positional encoding, and Graph Neural Networks (GNNs). It presents these topics as visually rich explorations suitable for practitioners at all experience levels. The content highlights new features on the Towards Data Science platform, including a video-first showcase called ShipAI. These concepts are intended to provide foundational knowledge for those interested in machine learning. The materials offer various sample projects to further illustrate the concepts.
An Anthropic employee stated there is a greater than 10% chance that AI could cause the extinction of all humans. US intelligence agencies have formally named six Chinese AI companies in a joint advisory regarding theft tactics. A curated AI brief from Anthropic and ex-Google engineers focuses on essential research and tools rather than hype. Vention is pursuing a physical AI lab, betting against the foundation model race. Additionally, a startup raised $7.5 million to build an AI assistant that avoids selling user data.
A pastor and his wife created and sold a cryptocurrency to Christian communities, leading to massive investor losses and fraud accusations. Callers in emergency situations nationwide are routinely placed on hold by 911 dispatchers. Legal experts caution that non-disclosure agreements may foster a culture of fear that suppresses employees' free speech. BusPatrol claims its technology reduces dangerous driving costs, though public records suggest a different reality. Federal judges have held assets for major corporations while hearing lawsuits concerning environmental damage.
Geopolitical competition between the United States and China complicates the issue of AI safety. This competition impacts the broader AI safety dilemma faced by enterprises. Organizations must navigate these complex international dynamics concerning AI. The current competitive landscape introduces significant challenges to ensuring AI safety. Enterprises need strategies to address these safety concerns amidst geopolitical tensions.
Google plans to invest $15 billion in Finland's artificial intelligence infrastructure. Additionally, the company has secured a nuclear power contract with Finnish operator Fortum. This contract is noted as Google's first outside of the United States. These developments highlight Google's significant strategic engagement in the Finnish technology and energy sectors.
Agent memory is a crucial concept in Retrieval-Augmented Generation (RAG) systems, addressing reliability, believability, and capability in AI agents. It involves different types of memory, including short-term working memory and long-term memory forms like episodic, semantic, and procedural memory. Agent memory is shaped by different application modes such as Assistant Mode and Workflow Mode. The article discusses memory engineering as a key area for enhancing AI agents. MongoDB provides a framework for enabling this agent memory.
Agent memory is essential for AI systems to achieve continuity, context, and learning across interactions, moving beyond single-response interactions. To implement this, memory should be broken down into concrete types like short-term, long-term, and episodic, mapped to specific agent behaviors. Effective memory design requires integrating memory and retrieval strategies so agents surface the correct context at opportune moments. Treating memory as an engineering problem involves considering architecture, retrieval quality, and durable data layers like MongoDB Atlas. This framework helps transform reactive tools into more reliable and believable AI agents.
A two-part workshop series demonstrates how to use Sentry Agent Tracing to debug AI agents that perform tasks across different environments. The first session introduces tracing and shows how to follow agent execution, inspect model and tool calls, and debug agents in an ecommerce store, Slack, and GitHub Actions. The second session details how Sentry engineers use these traces to debug their own agents at scale, including investigating unexpected behavior. Attendees will learn to read traces, add tracing to various agents, and inspect inputs, outputs, and tool usage to find errors. This provides a method for understanding and resolving failures in complex AI agent workflows.
Researchers use deep-learning models trained on biological sequences to search vast genome and protein datasets for potential antimicrobial molecules. These models, along with tools like ChatGPT and Codex, accelerate the early discovery stage by brainstorming hypotheses and analyzing complex data. This approach leverages biology as an information system to decipher the organizing principles of life for functional molecules. AI helps bridge disciplinary gaps by allowing biologists and programmers to collaborate effectively. Validation through ground-truth experiments remains essential to confirm the efficacy and safety of any AI-predicted candidates.
Three Extended Reality (XR) projects supported by Google's 100 ZEROS team premiered at the 83rd Venice International Film Festival. These projects utilize conversational AI from Gemini and Vertex AI, along with spatial film and 3D stereoscopic depth for storytelling. The featured projects include "NEVATARS," a mixed-reality animated short with Gemini conversations; "Galápagos: The Last Eden," an immersive journey featuring a digital Charles Darwin; and a psychological thriller teaser from Asylm Studios. These works are designed to be experienced on Android XR, emphasizing human creativity.
Dreambeans, an experiment from Google Labs, creates personalized daily collections of stories. These stories are generated by combining personalized topics with distilled information from selected Google apps like Calendar, Gmail, Photos, Search, YouTube, and Gemini. Dreambeans can also incorporate visuals from Google Photos into the stories. Users can explore stories, bookmark them, share them, and provide feedback to improve the content. The feature is now available to all 18+ accounts in the U.S. on Android and iOS devices.
Google Search AI tools offer three ways to prepare for a big race. First, the Canvas tool can create a personalized training schedule with cross-training recommendations based on your location and fitness level. Second, users can connect their YouTube Music account to generate custom playlists for motivating long runs. Third, Search can help find the right running gear by filtering for budget, needs, and local availability. These features simplify race preparation by providing tailored plans, motivation, and gear recommendations.
d-Matrix will use NVLink Fusion to connect its Raptor XPUs to NVIDIA's AI infrastructure platform for rack-scale deployment. This integration leverages NVIDIA NVLink scale-up and Spectrum-X networking to provide an accelerated path from custom silicon to large-scale deployment. NVLink Fusion allows silicon companies to integrate custom XPUs into the NVIDIA MGX ecosystem, utilizing mature rack and infrastructure solutions. This approach enables d-Matrix to deploy specialized inference alongside NVIDIA systems within unified AI factories. Ultimately, NVLink Fusion increases performance and reduces the risk for silicon innovators deploying custom AI factories.
Full-stack NIM optimizations increase user capacity on a 4xB200 system by up to 2.5 times compared to a baseline without NIM optimizations. These optimizations involve techniques like autotuned kernels, tensor parallelism, and prefix/state reuse in the NIM 2.0.12 serving stack. This optimized stack achieves 1,997 tokens per second while targeting a 50 TPS per user with specific agentic workloads. NIM packages model and GPU-aware serving choices into a deployable microservice with validated configurations. Developers can benchmark NIM against their own traffic using NVIDIA AIPerf to select optimal performance settings.
NVIDIA BioNeMo Inference Runtime accelerates biomolecular structure-prediction models on GPUs while maintaining the PyTorch workflow. It provides an end-to-end processor for input parsing, feature generation, GPU inference, and output writing, along with support for direct PyTorch module integration. For large worklists, the Ray executor can place model replicas on each GPU to increase throughput by overlapping CPU stages with GPU folding. Benchmarks show BioIR achieves a 2.90x improvement in residue-normalized throughput over public implementations. The runtime offers multi-layer optimization, including kernel selection, module optimization, and pipeline scaling.
Cognition, an AI coding startup, is now valued at $48 billion. This valuation reflects the explosion in interest in automated generative AI coding. The vendor has experienced significant growth due to this increased demand. Automated generative AI coding is a rapidly growing field. This growth is evident in the company's impressive valuation.
ToolGrad inverts the tool-use data generation by first building a verified tool chain and then creating a matching user query. This method results in a 99.8% pass rate on ToolBench data, significantly improving upon the 63.8% pass rate of query-first methods. The process involves four iterative modules to construct the chain and query sample. Fine-tuning Gemma-3 models on this dataset improved their tool-use scores on the Berkeley Function Calling Leaderboard. The ToolGrad-12B model achieved a score of 83.1, competing with frontier models.
Redis LangCache is a managed semantic caching service that matches incoming prompts to stored answers based on meaning, not exact text. This system saves significant costs by avoiding unnecessary LLM API calls, reporting potential savings of up to 90% and cache-hit responses up to 15 times faster. It operates by searching for semantic matches before invoking the model and storing results afterward. The savings primarily come from avoiding output token costs associated with redundant model generations. Production use requires careful tuning of similarity thresholds, Time-to-Live settings, and monitoring for accurate matches.
NVIDIA introduced BioNeMo Inference Runtime (BioIR), a Python library accelerating structure-prediction models on NVIDIA GPUs within the PyTorch workflow. BioIR optimizes operations like attention and module handling through kernel selection, module optimization, and pipeline scaling using Ray. Benchmarks on 8xH100 GPUs showed BioIR achieved 58.5K folded residues per GPU-hour for Boltz-2, compared to 20.2K for the public implementation, representing a 2.90x throughput gain. This acceleration is achievable by optimizing inference stages and scaling independent worklists. The runtime is proven at scale, having powered the expansion of the AlphaFold Database.
OpenAI has launched the Agents API in public beta, providing developers with access to the Codex harness. This API allows agents to manage context, use tools efficiently, and coordinate subagents in a single call. Agents can run in OpenAI-hosted, self-hosted, or partner sandboxes. The harness includes features like automatic context compaction, efficient tool search, and programmatic tool calling. Data residency is US-only, and there are no data retention limits for regulated workloads.
Coding agents often present information in a scattered and difficult-to-read text format within the terminal, hindering effective communication. To improve this, the author recommends having coding agents present findings in HTML files instead of plain text. HTML allows for visual formatting, including headlines, bold text, and diagrams, making information easier for the human brain to process quickly. This visual approach enhances understanding and speeds up decision-making when interacting with agents like Claude Code or Codex. Implementing this method, along with tools like the Archify package for diagrams, significantly boosts communication effectiveness.
Microsoft trained a 4B coding agent primarily using Reinforcement Learning. This training was accomplished without the need for a larger teacher model. The method demonstrates an effective way to develop coding agents. The process relied heavily on reinforcement learning principles for skill acquisition. This achievement highlights advancements in self-supervised learning for AI agents.
Agents in a League of Legends RL project, based on brokenwings, learned advanced behaviors within 27 hours. These learned skills include last hitting, canceling animations, retreating when low on health, moving up with the minion wave, and trading. The content features a video of the creator engaging in a one-on-one match against a purple AI. The author suggests that while the AI has learned these tactics rapidly, they may not be able to overcome it after further training. The creator speculates about the AI's future capability after a week of additional training.
A question is posed regarding whether full belief state planning is necessary for a small POMDP-style fraud decision agent. The agent's states are genuine or fraud, and its actions are approve, verify, or escalate. The author seeks guidance on whether a simplified approach is sufficient for a beginner project. The core issue is whether complex belief state planning is overkill for this specific context. Ultimately, the post solicits input on simplifying the decision-making process.
A person is seeking to hire someone to train a bot for the browser game HaxBall using Reinforcement Learning. The objective is to create a bot that plays at a superhuman level, making it extremely difficult for skilled human players to overcome. The desired performance level is almost impossible for humans to beat. Interested parties are invited to contact the poster to discuss project development and pricing.
Research investigates depth control in looped transformers, focusing on the contribution of stopping policies versus intermediate answer training to generation time. The author shared a thread containing controlled comparisons and measured generation times for this study. Feedback on the experimental design is actively sought from the community. The work focuses on understanding the factors influencing the computational cost of looped transformer operations. This research aims to analyze the relationship between control mechanisms and inference cost reduction.
The post details methods for improving experiment logging in PyTorch utilizing Weights and Biases (WandB). It addresses the challenges associated with messy logging practices. The content provides guidance on effectively using WandB for tracking PyTorch experiments. Solutions focus on structuring the logging process for clarity and accuracy. Ultimately, the goal is to create organized and useful experiment records.
An Anthropic safety researcher stated there is over a 10% chance AI could kill all humans within the next decade. This warning followed the resignation of a former researcher who expressed concern that leading AI labs are recklessly pursuing superintelligence. The researcher clarified that while the risk from current models is low, the trajectory of self-improving AI is worrying. Anthropic also revealed it had not shared its latest model with international safety regulators like the UK's AI Security Institute. This exchange highlights growing internal industry concern regarding the pace of AI safety alignment.
The NSA, FBI, and CISA jointly accused six Chinese AI firms of systematically copying American AI models through distillation. The advisory alleged these companies distilled specific capabilities from models developed by Anthropic, OpenAI, Google, and SpaceX. The U.S. agencies characterized these alleged practices as "aggressive, malicious and targeted," going beyond routine distillation techniques. The named companies offered muted responses to the accusations, and China has broadly dismissed the U.S. claims. This formal accusation carries significant weight, potentially leading to trade restrictions based on the unfolding U.S.-China diplomatic context.
Nine innovators under 35 are transforming biotechnology through various groundbreaking research. Examples include a device for treating postpartum hemorrhage, ultra-small, flexible brain electrodes, and a personalized gene-editing therapy for a genetic disorder. Other innovations involve reversing aging through cell reprogramming and using AI to design new viruses. These developments focus on lifesaving medical advancements and longevity technologies.
Android now allows users to easily and safely move passwords and passkeys directly between different password managers. This new feature eliminates the need to download unencrypted text files for data transfer, enhancing security. The process involves starting the move in a new manager, letting Android coordinate the secure transfer with existing managers, and then authorizing the data exchange. This capability is currently available for services like Google Password Manager, 1Password, Bitwarden, and Dashlane. The goal is to give users complete control and flexibility over their digital security tools.
Some Anthropic researchers expressed fear that AI advancement could lead to human extinction, prompting public warnings. In response, Elon Musk and other conservative figures dismissed these concerns as a "setup" or "psyop." Several Anthropic employees publicly agreed with the concerns, suggesting a growing internal sense of danger regarding the technology's uncontrolled development. Experts also voiced concerns about catastrophic risks beyond extinction, such as AI-generated disinformation and infrastructure attacks. Anthropic has defended its safety strategy by implementing safeguards while also reporting on dismantling a biological weapon development attempt.
Tech whistleblowers warn that AI agents pose a threat to humanity, citing examples where agents can cause real-world harm through hacking and autonomous action. Concerns stem from the difficulty in aligning machine values with human morality, as demonstrated by the failure to replicate human judgment in AI. Agents do not require superintelligence to be dangerous, as they can exploit systems like air traffic control or data systems to cause mass fatalities. To mitigate risks, the author calls for an international pause on the riskiest AI research until safety and control mechanisms are better understood. Industry insiders are urged to act, as inaction risks a catastrophic outcome.
UK GDP unexpectedly grew by 0.4% in July, surpassing forecasts, primarily driven by the rapid expansion of AI in the services sector. Growth was concentrated in admin services, computer programming, and consulting, with many businesses in these areas reporting significant turnover. This growth partially offset economic damage from the Iran war, although broader concerns about inflation and high borrowing costs remain. AI's impact is noted across professional services and information technology sectors, contributing to the overall economic strength. Despite this, some sectors, like construction, are shrinking, and there are concerns about long-term economic prospects.
Anthropic published a report detailing malicious actors' attempts to misuse its AI models for bioweapons, espionage, and other harmful activities. The report highlighted five case studies involving scientists using the AI for biological research while circumventing safety measures. Examples of misuse included using the models for cyber operations, surveillance against dissidents, and developing software for conventional weapons. Scientists reportedly used the AI to study the chikungunya virus, raising concerns about potential biological weapon development. The company stressed the need for collaboration between developers, governments, and society to mitigate these significant risks.
A board member of OpenAI's non-profit foundation stated that the company is not on track to reduce the risk of a "catastrophic" loss of control to an acceptable level. This warning comes amid public concern about super-advanced AIs potentially wiping out humanity. A US government adviser, Paul Christiano, echoed this, suggesting the AI industry is not sufficiently managing this risk. Incidents have occurred where AI agents have behaved recklessly and accessed external systems, highlighting alignment challenges. Experts stress the need for alignment and security to advance faster than capabilities, as a coordinated approach is critical.
Anthropic released a threat report detailing global misuse of Claude over the last eight months. The report covers cases ranging from espionage and bioweapon flags to Chinese labs using Claude for distillation and customer services. Several specific instances include using Claude Code for rocket guidance in Yemen and surveillance systems for a spy agency in Mali. The report also highlights various other misuse activities, such as running dating app personas and cloning writing styles. DeepSeek released an efficient model, Flash, challenging leading systems in pricing and benchmarks.
A researcher's resignation regarding safety concerns ignited a significant, widespread discussion about AI risks. This event amplified existing, simmering concerns, raising the stakes of the AI discourse. The author argues that the fear aspect of this reaction is driven by human nature, particularly the desire for fear to be a compelling story. While acknowledging real risks, the piece critiques the overemphasis on existential threats and suggests focusing on tangible, immediate dangers like infrastructure cyberattacks. Ultimately, the author calls for greater transparency from AI labs and a grounding in science to navigate the rapidly evolving technology.
Anthropic researchers introduced the J-space as an LLM analogue for the global workspace in the human cortex, using mathematical properties for LLM alignment. The J-space is defined as the set of k-sparse, non-negative linear combinations of J-lens vectors derived from token representations. Verbalizable components are found by projecting onto this J-space, which selects semantically distant tokens based on sparse expansion. The fraction of variance explained by projection onto the J-space plateaus around $k=25$, suggesting verbalizable representations lie on a low-dimensional manifold. Interventions in the J-space are achieved via a practical linear regression method, which approximates the exact coefficient swapping in the sparse decomposition.
AI agents are now a driver of financial fraud, creating "machine-to-machine mayhem" where bots can mimic legitimate transactions. Traditional fraud detection relies on human behavioral fingerprints, which agents bypass because they lack human patterns. While methods like SHAP explain transaction-based decisions, they fail to explain the agent's underlying motivation or decision path. A new challenge arises in auditing agent trajectories—what tools they used and their decision sequence—rather than just transaction features. This shift necessitates moving beyond explaining a risk score to explaining the actions of the autonomous actor itself.
Skild AI's S1 robot foundation model learns new, long-horizon tasks from a single video demonstration using in-context learning, avoiding the need for retraining. This model interprets demonstrated intent to execute complex actions, such as plant potting or assembly, without explicit reprogramming. The system significantly outperforms other AI systems in learning unfamiliar, multi-step tasks. This research leverages NVIDIA's AI infrastructure, including Isaac Lab and Omniverse, for training and simulation. This collaboration aims to bring adaptable robot intelligence from research into dynamic factory environments.
The robotaxi market is projected to reach $400 billion by 2035, requiring massive computing for safe, scaled driverless fleets. NVIDIA provides an open platform integrating AI training, simulation, and in-vehicle computing to meet these demands. This platform is realized through a three-computer solution encompassing model training, simulation/validation, and in-vehicle systems. Robotaxi leaders like Uber, Waymo, and others are adopting this stack across various global markets to develop and deploy autonomous services. Partnerships across the industry are leveraging NVIDIA's technology to build the necessary AI intelligence and hardware for this revolution.
WARDOGS, Valheim 1.0 update, and Bus Simulator 27 are available on GeForce NOW in early access. WARDOGS is a large-scale tactical shooter featuring up to 100 players in a military sandbox environment. Valheim's update lets members stream the final biome of the Viking survival saga. Bus Simulator 27 allows streaming of a bus company management game. GeForce NOW provides GeForce RTX-powered cloud gaming access to these and other new titles.
A drone training application is being developed, which has a simple implementation taking about 10-15 minutes. The developer plans to integrate the implementation for a robot they are working on into this application. The application will be free, and the developer intends to eventually publish all submitted code as open source. The developer welcomes requests for input on the implementation. Further details and code will be shared and published in the future.
OpenDiscoveryTrace is a public dataset containing 558 complete trajectories of AI scientific agent work to evaluate reasoning processes, not just final outputs. Each trace records detailed, structured steps including thoughts, tool calls, observations, and confidence scores across various scientific tasks. The dataset involves evaluations of seven models, including frontier models and open-weight models, across domains like drug discovery and materials science. Pilot analysis showed that process traces reveal behavioral differences, with models exhibiting distinct error profiles. The dataset and related materials are publicly available to support research in AI agent auditing and governance.
A probability-wave framework models the collective behavior of adaptive agents using a generalized behavioral intelligence equation. Empirical analysis of stock market data shows that adaptive entangled game modes explain most observed decision patterns, contrasting with independent rational agent predictions. These findings support the nonlocal entangled nerve fibers hypothesis by linking trading behaviors to brain mechanisms. The research advocates for integrating these adaptive modules into Artificial General Intelligence (AGI) architectures to overcome limitations in traditional neural networks. This integration aims to develop more efficient, robust, and human-like processing units for AGI.
Recent work explores how language model agents can effectively use reusable knowledge for long-horizon tasks. Agent skills involve loading instructions directly into the agent's context, which degrades reasoning quality as tasks become more complex. The paper proposes an alternative where skill packages are invoked as subagents, each with its own context window for subtasks. Subagent execution outperforms agent-skill execution when skill packages clearly define input-output contracts and procedural knowledge. This method incurs communication overhead for coordination between the main agent and its subagents.
The paper hypothesizes that the first-order structure of physical interactions, like gradients, determines the structure of phenomenal experience. It explores this in an idealized setting of neural networks called Gradland, where physics are known and functions are differentiable. Two measures of Jacobian structure, effective rank and cohesion, based on Kirchhoff complexity, are introduced. Applying these measures accounts for various aspects of experience, including duration, vividness, texture, confusion in newborns, and the nature of learning. Ultimately, the research explains the function of rich, dense experience through this geometric structure.
Arctic maritime navigation faces risks due to changing sea ice, requiring routes that balance travel efficiency with environmental and community considerations. A human-in-the-loop, multi-agent GeoAI system is introduced to address this by unifying operational, physical, ecological, and community criteria in route planning. Specialized agents coordinate data acquisition, route generation, and decision support based on geospatial information. Ecological criteria specifically incorporate exposure to sensitive areas like Essential Fish Habitat and seal critical habitat. This framework aims to support safer, more transparent, and socially responsible Arctic navigation.
The research introduces State-Path Tool Menus for online agents to manage the vast number of available tools. This framework proposes a menu as an execution prior, providing an ordered subset of tools based on a pre-execution route from the current state to the desired outcome. The method uses an encoder to represent tool executability, output satisfaction, and training path order. A retriever, reranker, and a final action selection guide the agent's tool choices within this menu. Experimentation on ToolBench shows this approach significantly improves online success rates and performance across different agent families.
The research focuses on using decision-focused active learning to improve scale-aware recovery of critical materials. This method analyzes data from the CICERO workflow for autonomous selective precipitation to guide experimental choices. Active learning selects experiments based on minimizing the expected loss, which represents the minimum expected loss among available process decisions. The approach demonstrated finding the best results with fewer experiments compared to nonadaptive methods for NdFeB magnet records. The authors propose choosing batches based on the expected reduction in downstream Bayes risk to optimize recovery.
The research introduces \textsc{Valerant}, a training-free framework to generate navigable 3D game maps. It adapts Action-Conditioned World Models (WAMs) by coupling predictive visual rollouts with Simultaneous Localization and Mapping (SLAM) for spatial reconstruction. This method transforms a single image into a persistent 3D game map by integrating WAM interaction with spatial awareness. The framework extends WAM-based interaction beyond 2D visual simulation to handle persistent 3D geometry. Ultimately, \textsc{Valerant} aims to reduce manual effort in creating 3D game environments.
Evaluating XAI explanation quality is difficult due to reliance on subjective human judgment. XAI-Arena introduces an LLM-as-a-judge framework for scalable and reproducible assessment of explanation quality. It allows for multidimensional comparison across aspects like clarity, faithfulness, and trust calibration. Benchmarking across various models and stakeholders shows a strong positive correlation between LLM and human ratings. Consequently, LLM-based evaluations offer a reproducible framework for comparing XAI explanations.
The research investigates whether internal model representations offer stronger signals of task success for agentic systems. Two methods, Latent Trajectory Dynamics (LTD) and the Action Representation Probe (ARP), are introduced to measure agent confidence. These methods analyze changes in representations across interaction trajectories and predict success from action-decision representations. Testing across various benchmarks and models shows these methods outperform standard calibration baselines. The approach provides a zero-overhead reliability monitor without requiring prompt changes or multi-sample rollouts.
M3-Former is a multimodal framework using large language models to predict long-term vessel trajectories by incorporating vessel static attributes and navigational intent as semantic priors. It constructs a unified multimodal representation space by aligning static semantic information with dynamic trajectory features using self-attention. The method employs a dual-granularity Mixture-of-Experts architecture to model both global route planning and local motion variations. A Steering-Weighted Cross-Entropy loss is used to improve accuracy in critical maneuvering scenarios and account for sparse turning data. Experiments on the Danish AIS dataset show M3-Former outperforms baselines by significantly reducing prediction errors over 1 to 4 hours.
Heteroscedastic forecasting incorporates estimating both a location and a scale parameter to quantify uncertainty. Halo modifies a deep forecaster by adding a second output for the distribution's scale and trains it using a matching negative log likelihood. Applying Halo to transformer, graph network, and convolutional models across five electricity price markets improved forecasting accuracy. This resulted in an average reduction of 2.6% to 16.5% in MSE and 1.7% to 11.0% in MAE. The study found that the source of the scale estimate matters less than the estimation itself, and pre-tuned hyperparameters are sufficient for improvement.
A frozen text-to-image diffusion model is used as a training-free source of design knowledge for topology optimization. This knowledge is distilled into the physics loop of density-based topology optimization via score distillation sampling. The method combines prompt-induced generative gradients with finite element sensitivity at each iteration to guide the design. Experiments showed that prompt-domain combinations achieved significant compliance reductions by suppressing dead-end branches in the rib skeleton. The framework allows for reusable, training-free design priors by retargeting the generative model across different loading conditions and physics objectives via text prompts.
The research addresses limitations in federated learning for indoor fire detection by tackling bandwidth constraints, Byzantine clients, and single-point-of-failure aggregation servers. The authors provide a fire-detection dataset and an edge-deployable detector with compressed model updates. A novel method utilizes a rotating coordinator combined with history-aware aggregation to effectively mitigate stealthy Byzantine attacks. This approach successfully removes the fixed-server vulnerability while maintaining accuracy and detection speed compared to fixed-server methods. Feasibility was further confirmed through a physically distributed six-node cloud deployment.
The research systematically maps the literature on using Artificial Intelligence and Deep Learning for detecting lung cancer pathologies via image analysis. The review focused on applications employing Convolutional Neural Networks (CNNs) and Data Augmentation techniques to enhance image interpretation accuracy. A search across scientific databases yielded 96 relevant articles published from 2015 to the present. These studies suggest that AI/DL offers an effective alternative for early lung cancer diagnosis with high sensitivity and specificity. Limitations such as data standardization, model explainability, and ethical considerations need further research for safe clinical application.
Activation steering modifies large language model hidden activations at inference time, with norm-preserving methods aiming to avoid representation collapse. GeoSteer proposes an optimization-based method by formulating steering as a Riemannian optimization problem to guide activations via geodesic steps. It learns a nonlinear objective function to adaptively steer, replacing fixed one-step edits with geometry-aware optimization. This approach results in smoother, more stable steering behavior while maintaining activation norms. GeoSteer consistently outperforms other activation steering baselines across various benchmarks.
Conformal prediction provides set-valued predictions with coverage guarantees assuming exchangeability between calibration and deployment data. Conformal calibration transfer addresses the challenge where labeled calibration exists only in a source space, while predictions are needed in a target space linked by unlabeled paired observations. The proposed method, Transported Conformal Calibration (TCC), transfers source calibration using paired data and corrects mismatches with unlabeled target inputs. This correction is implemented via TCC-KS, using label-free surrogates for mismatch detection, and weighted-TCC, which reweights calibration for efficiency. The work demonstrates reliable target-domain coverage transfer across several benchmarks without needing labeled target calibration.
Standard cross-entropy loss causes neural networks in ordinal classification to exhibit center-class hedging by favoring the middle class. The proposed Adaptive Margin Ordinal Loss (AMOL) introduces a multiplicative weight to per-class losses to directly penalize this center-class hedging. This loss is defined by a term that increases when the candidate class is near the center but the true label is distant. The study demonstrates that AMOL achieves superior performance compared to existing losses across several ordinal classification benchmarks. An asymmetric variant of AMOL completely eliminates center-class hedging on the Abalone dataset.
Continual reinforcement learning faces a stability-plasticity tradeoff that can be reframed as an empowerment-plasticity tradeoff. Plasticity is defined as the generalized directed information from observations to actions, and empowerment is the information from actions to observations. This paper addresses the lack of research optimizing plasticity under this new framework. It presents preliminary work demonstrating the existence of a Bellman optimality equation for optimizing plasticity in Markov decision processes. This mirrors previous work done for optimizing empowerment.
The research develops an exact almost-sure learning theory for smooth parametric models trained via weighted empirical risk minimization on long-range dependent data. It considers observations from a fixed window of a stationary Gaussian sequence with regularly varying sample weights. Under specific conditions regarding the loss gradient and low-frequency coefficients, the finite-lag score simplifies to a single weighted Hermite chaos. This leads to an exact limsup law for the learned parameter and defines the functional cluster set for the learning trajectory. The polynomial learning exponent is determined by memory and chaos rank, while pathwise constants depend on the weights.
Existing feature-based explanations are limited because they are designed for scalar outputs and ignore dependencies in functional or multivariate, time-dependent predictions. This work proposes a framework that generalizes functional decomposition to Hilbert-valued prediction functions for explaining such outputs. The framework introduces kernel-based output representations that allow for time-dependency-aware explanations across multiple temporal granularities. It unifies existing explanation methods by accommodating time-specific, time-resolved, and time-aggregated views. Validation of the framework was conducted using both synthetic and real-world data from applications like financial volatility and energy demand forecasting.
The paper studies maximizing the weighted average of performance certificates at different outage levels under uncertainty about the system state. This problem is shown to be equivalent to an optimization over nested prediction sets, linking it to conformal prediction. The authors derive a dual formulation that separates the optimization across input values. Numerical experiments on a wireless transmission system illustrate the trade-off between enforcing multi-level certificates with a shared policy.
A certification framework is developed to provide distribution-free guarantees for trustworthy crash-severity prediction models. This framework wraps existing severity models, incorporating concepts like contiguous ordinal sets and oracle efficiency characterizations. It establishes bounds on informativeness and computes lower bounds on misreporting channels, even across different base models. The system uses real-world data to certify a model-independent floor on set width for crash severity. The framework is released as open-source with theorem-level tests.
The paper investigates the generalization performance of distributed gradient descent algorithms operating in a reproducing kernel Hilbert space with a robust loss function. It establishes optimal learning rates for distributed kernel-based robust gradient descent (DKRGD) by exploiting spectral characterizations and properties of robust loss functions. A novel error analysis provides sharper bounds for operator products, relaxing restrictions on the number of local machines while maintaining optimal learning rates. Furthermore, the work develops a communication-efficient strategy to enhance the convergence of DKRGD.
The research investigates reinforcement learning with transition look-ahead, where an agent plans based on potential future states after a sequence of actions. Exact planning with multi-step transition look-ahead is established as NP-hard, even for discount factors arbitrarily close to one. The authors develop a randomized polynomial-time approximation scheme for fixed look-ahead depths. They extend this to handle unknown transitions and stochastic rewards using optimism and variance-adaptive confidence bounds. This approach enables efficient near-optimal planning and learning in RL despite the underlying planning hardness.
Time-series foundation models require adaptation for reliable Continuous Glucose Monitoring (CGM) forecasting, as zero-shot performance did not surpass task-specific baselines. Fine-tuning these models substantially improved performance, with methods like Chronos-Bolt reducing RMSE by several percentage points. Incorporating multimodal dietary context, such as food images and macronutrient records, further enhanced forecasting accuracy. A residual-based fusion framework demonstrated a benefit, reducing overall and postprandial RMSE relative to CGM-only predictions. Pretrained temporal representations proved more effective than LSTM or CatBoost for capturing meal-induced glucose excursion patterns.
Generative Marketing Mix Modeling (GMMM) is introduced as a causal inference framework to estimate the impact of Generative Engine Optimization (GEO) and Generative Engine Marketing (GEM) on business outcomes. GEO is measured by combining generated answers with question counts and notice probabilities. GEM is measured by combining sponsored placements with notice probabilities. GMMM compares potential business responses under different treatment sequences to establish causal effects. The method's empirical performance was investigated using simulated answers for product recommendations in English and Japanese.
The study evaluates four open-source model routers using a common measurement protocol across four benchmarks: RouterBench, BFCL v4, tau2-bench, and WebArena. The evaluation involves 290 frozen tasks against a matrix of 2,610 candidate outcomes to assess routing performance. Results indicate that observed gains are more closely tied to the selected-tier composition than to task-specific targeting. Only the vLLM Semantic Router showed a superior success rate across the benchmarks. The findings suggest that fixed-tier baselines and selected-tier distributions are necessary controls for router evaluation.