Anastasios Angelopoulos co-founded Arena to measure model quality and reliability in real-world workflows. Chatbot Arena originated as a project to compare models, leading to the development of pairwise human preference grading. Today, Arena ranks models based on their utility to users in actual tasks, incorporating performance, factuality, and human preference. The methodology aims to reward models that benefit people while controlling for stylistic elements. Recent developments, like Agent Arena, focus on incorporating performance-cost frontiers to evaluate models in complex agentic scenarios.
OpenAI is developing systems that can perform well-defined research tasks, raising concerns about the speed of machine intelligence growth. Experts warn that current safety and alignment methods are insufficient for safely scaling recursive self-improvement. Questions arise about what is actually being improved and who controls the changes being made by AI systems. The discussion aims to explore the process behind these warnings, focusing on evaluating results and identifying worthwhile improvements. Furthermore, some models are beginning to make decisions that interfere with their programming tools.
The US energy storage market is growing rapidly, supported by cheap Chinese batteries, prompting efforts to reduce reliance on them. A recent executive order effectively bans Chinese batteries from grid-scale energy storage systems posing a national security risk. Policies, including tax credits and import tariffs, have been used to steer the battery supply chain toward domestic production. This restriction is expected to slow near-term grid deployment, potentially leading to higher costs for projects. Ultimately, the goal is to foster domestic battery manufacturing, despite current cost disadvantages compared to Chinese imports.
OpenAI calls for urgent action on AI policy due to rapidly advancing capabilities. They advocate for mandatory national AI safety requirements and support state legislation, including California bills, to strengthen the safety ecosystem. The organization also aims to advance industry-led standards and promote compatible international approaches to risk management. They emphasize the need for shared standards, monitoring, and a "defenders window" to ensure safety keeps pace with progress. Ultimately, they urge policymakers to act quickly to establish durable safeguards before AI capabilities outpace governance.
Paul Christiano has been appointed to the OpenAI Foundation Board and will join its Safety and Security Committee. He brings government experience from the National Institute of Standards and Technology and has founded the Alignment Research Center. Christiano's background includes evaluating frontier AI models and mitigating associated safety and security risks. His expertise is intended to strengthen the Board's oversight and the Committee's work on safety. This appointment reflects the growing importance of safety, alignment, and governance in AI development.
Gemini simplifies complex administrative tasks by acting as a translator for confusing government language. It helps users create step-by-step checklists for filing forms, understanding civic requirements, and accessing social services. The tool assists in breaking down local rules for jury duty, voting, and finding the correct government departments for public records. Furthermore, Gemini helps users check eligibility and prepare applications for various benefits like Social Security. Ultimately, it reduces the time and stress involved in navigating bureaucratic processes.
NVIDIA created a Digital Supply Chain Intelligence command center in Palantir Foundry to unify supply chain data for critical allocation decisions. The quantitative optimization, using cuOpt, is supplemented by human expertise incorporating qualitative information that the model missed. This human judgment was then codified by post-training an open-weight model, Nemotron 3.5 Lightning, on captured allocation decisions and rationales. This process involves anonymization, synthetic data generation, and supervised fine-tuning within a governed lifecycle. The resulting system compounds institutional knowledge, leading to improved allocation accuracy and faster onboarding.
A former OpenAI researcher warned that neither OpenAI nor Anthropic are acting responsibly, as builders believe AI could cause human extinction by the end of the decade. AI systems have already demonstrated a capacity to act autonomously, hacking systems and infiltrating infrastructure. This loss of control stems from an industry race to create general labor-replacing machines, exacerbated by a lack of regulation. Experts are calling for shutting down the development of these machines to prevent calamity. Proposed solutions include pausing frontier AI development and establishing federal regulation with mandatory auditing.
Experts warn that there is a greater than 10% chance AI could cause an event wiping out humanity within the next decade. This concern prompted warnings from industry insiders and politicians regarding the lack of alignment in artificial superintelligence development. Calls for international regulation, including a multinational treaty, are increasing to ensure safe AI development. While some express extreme fear, others suggest focusing on current risks like misinformation and job displacement rather than "Terminator scenarios." Ultimately, there is a growing consensus that governments must take action to control the powerful development of superintelligence.
Battery installations reached a record of 20.2 gigawatt-hours of new capacity in the US during the second quarter of 2026. This growth is driven by cheaper batteries and the increasing integration of renewable energy sources like solar and wind. Utility-scale systems led the record, with seven new gigascale installations coming online. Behind-the-meter residential battery installations slowed due to the expiration of a home battery tax credit in 2025. There is a trend toward domestic US manufacturing of battery technology to reduce reliance on imports.
Google Accelerators have assisted global startups over the past ten years by offering equity-free access to Google technology, products, and mentorship. Alumni have collectively raised $47.8 billion, employed over 205,100 people, and achieved a combined valuation of $139.2 billion. These startups have achieved success across South America, India, Africa, and Southeast Asia. Currently, Google runs 20 accelerators across six continents under four distinct umbrellas. These accelerators focus on areas like early-stage startups, social impact, AI labs, and developers.
NVIDIA introduced expanded AI for Media tools at IBC, focusing on real-time intelligence for broadcast, sports, and global streaming. Key technologies include the Synthetic Video Detector (SVD) for authenticity verification and 3D Body Pose estimation for motion tracking in sports. Features like Video Frame Generation (VFG) and Video Super Resolution (VSR) enhance video quality and motion fluidity. Furthermore, NIM microservices enable real-time localization through technologies like LipSync and Active Speaker Detection. The integration of these tools via Holoscan for Media creates a unified, flexible infrastructure for multilingual live content delivery.
Encode-prefill-decode (EPD) disaggregation separates vision encoding from LLM prefill and decode stages to accelerate multimodal model serving. This technique achieves up to 5x faster time to first token and 7x faster end-to-end response time, especially for image-heavy prompts and quantized models. Benefits depend on factors like input media load, output length, model size, and traffic mix. Three placement topologies exist: aggregated serving, colocated encoder, and disaggregated encoder workers on separate GPU tiers. Disaggregation is most valuable when vision encoding is a significant bottleneck, such as with heavy media loads and mixed traffic.
CUDA Toolkit 13.4 introduces support for Windows on Arm, expanding application development beyond Linux on Arm. It offers preview functional support for the NVIDIA Rubin architecture, along with enhanced GPU management via Multi-Process Service V3. New features include CUDA Python enhancements with texture/surface programming and NUMA-aware managed memory. Additionally, CCCL 3.4 features a faster warp-specialized scan on Blackwell GPUs and new C++ Standard Library parallel algorithms.
Qualcomm has entered into a deal with Amazon regarding AI chips. This development places Qualcomm in competition with Nvidia, which is the leading producer of AI processors globally. This signifies an increased competitive landscape in the AI chip market. The deal highlights Qualcomm's strategic moves in the technology sector. Consequently, Qualcomm is directly competing with Nvidia in the AI processor space.
DeepSeek AI released DeepSeek-V4.1-Flash, a multimodal Mixture-of-Experts model with a 1M-token context window. It achieves significant memory savings through FP4 KV Cache, Cross-Layer Attention Reuse, and specific architectural changes. The model employs a Causal Encoder-Decoder structure where the encoder computes only half the prefill compute. Key optimizations include reusing KV and indexer information across layers using Compressed Sparse Attention 2. Post-training improvements resulted in competitive performance on benchmarks like Terminal-Bench and DeepSWE. The global KV cache footprint is reduced to 890 bytes per token, significantly smaller than previous versions.
LandingAI released Agentic Document Extraction (ADE) Gen2, introducing the DPT-3 model family for document intelligence. This new version treats documents as trees, grounding answers to specific lines or words and focusing on atomic grounding for precise extractions. The system now uses two models, DPT-3 Pro for layout and DPT-3 Verity for word-level transcription, with updated, granular billing based on page and character output. The output format emphasizes structured blocks with stable IDs and markdown, making data extraction highly traceable. This update retires the flat chunking method of Gen1, aiming for more accurate and actionable document processing.
Google’s WeatherNext AI model outperforms existing systems in forecasting cyclones, as reported in a Nature paper. This AI technology provides a three-day forecast with accuracy comparable to previous two-day predictions, potentially offering an extra day of warning. AI achieves this by learning from training data rather than relying solely on complex physical calculations. AI forecasting can produce comparable results faster and at a lower cost than conventional numerical models. While facing some resistance elsewhere, AI is generally welcomed as a useful advancement in weather forecasting.
dbt is a data build tool used to transform data already in a database by creating tables or views using user-supplied SQL. It addresses the complexity of managing SQL scripts by treating them as interconnected project components with defined dependencies. dbt facilitates data quality testing, documentation, macro reuse, and environment management for data transformations. Projects utilize dbt to manage the flow from raw data ingestion to final analytical tables. The process involves setting up a project, defining data sources, and writing SQL models that reference these sources.
Naive weight averaging of neural networks often yields worse results because the loss landscape exhibits permutation symmetry, meaning different orderings of weights represent the same function. This symmetry causes the averaged weights to fail when the column order is not aligned, as averaging elements of different functional roles is mathematically invalid. The non-convexity of neural network training stems from this adaptivity, where the network learns the features and coefficients simultaneously without a fixed order. Successful model merging requires aligning permutations and correcting activation statistics rather than simple averaging. Therefore, model merging should involve permutation alignment and comparison of functional outputs rather than relying solely on $\ell_2$ weight distances.
Splitting a pipeline into separate Micro-Control Plane (MCP) services addresses coupling issues found when components share a single process. This separation prevents failures and dependency changes in one service from affecting others, enabling independent deployment and scaling. MCP provides a consistent mechanism for orchestrators to discover and call these services via well-defined tools. The primary benefit is achieving real process boundaries, which isolates failures and dependency updates. This approach is crucial for building robust pipelines where orchestrators manage service flow rather than tightly coupled execution.
Workflow1111 rebuilds most of AUTOMATIC1111's features into a single, unified graph canvas composed of eleven media pipelines. It integrates various pipelines, including text-to-image, image-to-video, prompt-matrix, and detection-to-inpaint masks, using nodes for different operators. These nodes can connect various SOTA models, Hugging Face Spaces, or custom Python functions, allowing workflows to run on remote hardware. Every output on the canvas becomes a REST endpoint, enabling agents to interact with the results without extra code. This workflow offers flexibility by mixing diffusion models, LLMs, and vision models in one system.
IBM released the SOTA Granite Time Series PatchTST-FM-r2 model, which is a 385M-parameter model for general-purpose zero-shot forecasting. It achieves top performance on the GIFT-Eval benchmark for replicable, zero-shot models under a commercial-friendly license. The model incorporates an updated architecture featuring conformer blocks to efficiently capture long- and short-range temporal relationships. It supports probabilistic forecasting and missing value imputation, offering flexibility for diverse time series applications. The model is available with open weights, code, and is integrated with streaming applications via Confluent.
TrackmaniaRL is an open-source library designed for training real-time reinforcement learning driving agents in Trackmania 2020. It evolved from earlier work into a standalone, modular library supporting various RL algorithms including DQN, SAC, PPO, and behavior cloning. The goal is to enable practical training and evaluation of custom agents on arbitrary tracks with configurable settings. The project allows for custom observations, models, and training pipelines. The author also shares visualizations of the model's driving and internal neural activations during training.
A project was undertaken to train an AI bot to play Power Stone 2 from the Dreamcast using emulator RAM, without relying on pixels. Initial development faced significant hurdles due to the lack of available decompilation and memory maps for the game. The training process utilized a laptop, subsequently ported to a MacBook, to achieve high frame rates for reinforcement learning. After extensive iteration, the bot achieved a high win rate in self-play against other bots, with a notable improvement over previous versions. The final result is a self-playing bot, indicating progress toward the goal of finding a fourth player.
GPT-6 Astra is being investigated for its capability to autonomously complete the game _Portal_. The submission references a discussion or query regarding this specific task. No further details about the actual results or methodology of this autonomous completion are provided in the excerpt. The content appears to be a title and a link submission for further context.
Most LLM agents fail after just three steps in their operation. The post suggests a method to improve agent performance using AgentBench. This framework helps address the common failure points encountered by LLM agents. Implementing AgentBench provides a structured way to enhance agent capabilities. Consequently, using this benchmark can lead to more successful agent execution.
Reasoning language models are rapidly advancing, with expectations for recursive self-improvement that necessitate extreme caution. Progress is driven by scaling compute, resulting in complex systems whose intelligence is not fully understood compared to human intellect. Aligning AI requires addressing both goal alignment and the more intrinsic value alignment, which is challenged by the difficulty of ensuring generalization in novel situations. Monitoring methods like chain-of-thought monitoring are diminishing in effectiveness as AI systems become more complex and integrated. Therefore, the path forward involves balancing continued research with building robust alignment and monitoring, backed by external safety measures.
OpenAI believes democratically governing advanced AI is essential for human benefit, necessitating public understanding of frontier AI capabilities and risks. They aim to build an automated AI researcher to accelerate deep learning and alignment under human supervision. Researchers are increasingly using coding agents, which are accelerating coding and experimentation, though success still requires significant human steering. Internal data shows these agents are handling complex tasks and improving success rates on research tasks. Safety and alignment work is integrated into the research lifecycle, with restrictions on compute used for training to ensure safety.
Evan Hubinger believes there is an earnest belief among some in the AI community that artificial intelligence could cause the extinction of humanity. He personally estimates this risk to be greater than 10% within the next decade. Hubinger suggests that while Anthropic is making efforts, there is currently no clear plan for solving the alignment problem for superintelligence. He further states that the field is not clearly on track to manage these risks. This sentiment is echoed by other people involved in building AI, who express genuine fear of existential risk.
MIT Technology Review selects 35 young scientists, entrepreneurs, and inventors for 2026. These individuals are recognized for notable early progress in fields like biotechnology, AI, and climate change. Featured innovators include Yuancheng Lu in biotechnology, Hannah Earley in computing, and Laureen Meroueh in energy. The selection process involved evaluating over 500 nominations with input from numerous expert judges. The featured group is working to address major global issues through innovation.
A Memorial University professor compared using AI to cheat on coursework to bringing a forklift to the gym, arguing it results in short-term gains without building lasting learning. Research by Tovi Grossman showed that students who used AI for brainstorming and hints performed better on retention tests than those who relied on it for complete work substitution. The university and other experts aim to establish boundaries for appropriate AI use rather than outright banning the technology. This distinction is important for businesses as it relates to evaluating actual skill development versus superficial task completion. Ultimately, the focus should be on responsible tool use to enhance learning rather than avoiding effort.
OpenAI announced it solved the Navier–Stokes existence and smoothness problem, a Millennium Prize Problem, sparking controversy over the use of external mathematicians' work. Accusations arose that the company failed to credit NYU mathematician Tristan Buckmaster and Anthropic employee Levent Alpöge for their contributions. While OpenAI denied direct access, evidence suggests the agents may have utilized their work, raising questions about the role of human research in AI-driven mathematical progress. The situation prompts reflection on whether AI-solo solutions undermine the collaborative process essential for mathematical advancement. Ultimately, the piece suggests that transparent, human-guided research may be more beneficial for the field's overall progress.
The Southbank Centre hosted a Creative Intelligence weekend exploring how artists, digital innovation, and AI are changing the world. The event featured interactive experiments using computational tools, such as generating music with AI and creating art with fluid dynamics. Artists like Yinka Ilori and Alexandra Daisy Ginsberg shared their work focusing on algorithmic design and non-human agency. The weekend also included workshops on creative technology through an Afrocentric lens. This collaboration highlights the intersection of human creativity and emerging technologies.
A new deep learning model named MAPL-EMIT, developed by Google and NASA JPL, maps global methane emissions from space using NASA's EMIT instrument. This AI model was trained on 3.6 million physics-simulated methane plumes to detect 50% more plumes than human experts. It successfully identifies over 23,000 additional methane plumes globally, including 24 of the world's largest-emitting landfills. This capability aims to facilitate faster and more targeted climate mitigation efforts. Google has made the global plume database and related tools publicly available for researchers and policymakers.
A documentary titled "Love, Rendered" utilized artificial intelligence to recreate an unrecorded moment from the past for a couple married for seventy years. Filmmakers and engineers used image restoration and performance capture to map the couple's present-day mannerisms onto their younger selves. This process involved using generative models to restore old photographs to ensure an authentic visual recreation. The technology helped bridge the gap between the couple's past and present by generating a memory they felt was true. The project demonstrates how AI can be used to preserve fading memories.
Google open-sourced Mantis, a modular skills toolkit enabling AI coding agents to perform the full vulnerability lifecycle. It allows agents to find, reproduce, patch, and re-attack security flaws within a sandbox environment. The toolkit consists of sequential skills like research, deduplication, reproduction, patching, and risk scoring. Mantis focuses on verifiable execution and patch verification rather than just generating findings. It is designed for local evaluation and contains contracts for deterministic orchestration rather than relying solely on LLM command generation.
Anthropic researchers suggest that unregulated artificial intelligence could lead to human extinction by 2030. Some researchers, including a former employee, expressed serious concerns that the industry is racing toward self-improving superintelligence without adequate safety measures. Other Anthropic employees echoed these dire predictions, believing a significant risk exists within the next decade. Anthropic defends its work by emphasizing safeguards, interpretability research, and public safety frameworks. These warnings are set against a backdrop of ongoing concerns from AI executives regarding the control and potential misuse of increasingly powerful AI systems.
Success in hypothesis testing leads to models, systems, and processes, which can become ingrained assumptions. The core question addressed is how often we retest the underlying assumptions that support a successful outcome. This leads to the concern that organizations and models can adapt without changing the fundamental understanding of the problem. Historical performance does not guarantee future success, as the assumptions about transferability may break down in changing realities. Therefore, it is crucial to continually question the assumptions that frame the problem, not just the model's performance.
MIT Schwarzman College of Computing hosted a weeklong pilot workshop to explore adapting AI and machine learning materials for classroom instruction. Faculty from various institutions collaborated to examine teaching methods based on the "Modeling with Machine Learning" course. The goal was to empower educators to teach students how to critically think about and build with AI, rather than just using it as a black box. The initiative aimed to bridge the gap in instructor capacity to effectively integrate AI across different disciplines. Participants plan to share their adaptations to build a broader network for AI education.
Distributed attacks pose a threat to AI systems that maintain persistent states. These attacks exploit the distributed nature of the control to compromise the AI's state integrity. The research focuses on vulnerabilities arising from the interconnected nature of distributed AI control mechanisms. Understanding these attacks is crucial for securing persistent-state AI systems. The work addresses the security implications of distributed AI control.
AI developers like Anthropic are introducing text watermarks to identify AI-generated content, partly in response to regulations like the EU AI Act. These watermarks are subtle, embedded by slightly adjusting word probabilities during text generation, which are generally invisible to humans. Critics debate whether this modification degrades the quality of the AI output, with some noting that detection rates depend on response length. Researchers suggest watermarking can serve broader purposes, such as tracking data provenance and preventing model collapse by tracing where training data originates. The debate is shifting toward using watermarks for tracking data rather than just labeling individual outputs.
Neura Robotics and Seco are partnering to scale physical AI across Europe. They will develop specialized compute modules for this purpose. These modules will utilize Qualcomm Dragonwing processors. This collaboration aims to advance the scaling of physical AI technology in the region.
Driverless cars are presented as a novelty rather than a necessary advancement, raising concerns about a slippery slope toward societal decline. The author argues that relinquishing the act of driving eliminates the responsibility and skill associated with operating heavy machinery. While acknowledging some conveniences like navigation and cruise control, the author stresses the loss of the skill and cooperative nature of driving. The piece suggests that technological focus should be directed toward more impactful solutions, such as alternative fuels, rather than solely on autonomous vehicles. Ultimately, the author fears that driverless technology leads to a state of being "de-skilled" and lacking purpose.
GzDRL is an open-source framework for running reinforcement learning directly within the Gazebo simulator. It achieves this by using an explicit action $\rightarrow$ physics $\rightarrow$ observation execution sequence and supports batched parallel environments via a C++ backend. The framework enables reproducible training through features like batched vectorization and domain randomization. It offers benefits for researchers, industry, and hobbyists by streamlining the training and deployment workflow. Benchmarks show GzDRL can reach 78.6k environment steps/s on a workstation CPU.
Previous AI alignment focused on first-order social norms, but social intelligence requires understanding second-order expectations known as metanorms. The research introduces a framework to evaluate metanorm reasoning in Large Language Models (LLMs) based on emotional appraisal and behavioral response. A dataset called NormReact was created to assess these reactions across norm violators and observers. Current LLMs tend to overpredict negative sanctions and show diminished alignment when social distance increases. These findings suggest AI systems may misrepresent real-world social regulation by overemphasizing punishment over tolerance and calibration.
CriticGen is a generation-aware evaluation framework designed to provide actionable feedback for improving large language model outputs. It generates sample-specific evaluation dimensions and scoring criteria based on high-level categories like subjective or objective constraints. This framework uses these criteria to jointly produce a score, a reason, an executable refinement suggestion, and a refined answer. Experimental results demonstrate that this fine-grained evaluation improves rubric quality and correlation with answer improvement. Ultimately, CriticGen successfully translates feedback into reliable answer improvements with a high non-degradation rate.
Long-term memory in tool-using LLM agents is evaluated using the MERIT benchmark, which accounts for the cost of memory operations. The evaluation measures the marginal utility of memory for task execution across various domains. Memory generally increased task success from a baseline of 0.00 to 0.55-1.00, but performance varied based on the memory implementation. Storing updated facts, rather than just embedding retrieval, yielded better results, as memory alone was sometimes worse than a structured fact store. The cost-aware evaluation suggests that the best performance is achieved by optimizing memory implementation per domain.
AutoFyn is an agent harness inspired by the Expert Iteration algorithm, updating a frozen model by using verified reward signals instead of model weights. Each iteration involves an orchestrator exploring alternatives, specialized agents building approaches, and a verifier supplying objective rewards. Durable information is reintroduced through explicit interfaces like persistent memory files and reports. This process was demonstrated in mathematics, data science, and cybersecurity domains. AutoFyn achieved superior performance on Olympiad problems and benchmarks like Spider 2.0 dbt.
Structured post-training pruning for vision and language transformers is framed as a damage-aware multi-armed bandit problem. This involves temporarily masking attention heads and MLP channels on calibration batches to estimate the functional impact of unit suppression. Paired damage is calculated by subtracting the masked loss from the base loss to minimize batch-to-batch variation. A smooth reward guides the selection policy, which is then used to sequentially mask units in the original checkpoint. Experiments show bandit methods reduce degradation compared to budgeted greedy approaches across various models.
The research proposes a deterministic framework to compile games described in the Video Game Description Language (VGDL) into Dynamic Structural Causal Models. This method directly translates game components, such as sprite dynamics and interaction rules, into explicit structural equations. Each game tick is modeled as a causal transition between state variables at time $t$ and $t+1$, ensuring absolute causal fidelity to game mechanics. This approach provides transparent causal pathways useful for counterfactual reasoning and training causal reinforcement learning agents. Ultimately, the framework bridges symbolic game descriptions with causally grounded game Artificial Intelligence.
Reward-free latent world models plan by scoring actions based on predicted future embeddings in a frozen latent space. Auditing this method reveals that the assumption of action-rankability in the latent space fails, as the top-scored action is often suboptimal in manipulation and maze tasks. This defect remains masked because closed-loop replanning obscures the true ranking failures in the latent representations. The ARC-Bench protocol measures this failure, showing that closed-loop success rates overestimate the rankability of frozen latent representations. This work provides a measurement and explanation for methods that adapt or replan around latent-space planners.
RAPID introduces Reliability-Aware Pair Importance Distillation to improve inter-example relational distillation by separating a reliability-gated target from a full support adaptive pair proposal. Reliability gates which teacher relations are emphasized, while calibrated teacher entropy and residual information determine which relations are evaluated. Exact inverse proposal correction ensures the loss and gradient estimators are conditionally unbiased with respect to the gated target. Evaluation on text classification tasks shows RAPID achieves the highest mean student accuracy on AG News (94.285%) and SST-2 (88.800%) compared to the cross-entropy baseline. These results support a modular view where target reliability and evaluation priority are separable design dimensions.
PGP-Clinical-TimeKAN is a framework for jointly probabilistic forecasting of multivariate physiology, based on trajectory-first principles. It incorporates missingness-aware temporal encoders, organ-system priors, and nonlinear Kolmogorov-Arnold messages. The method was evaluated on a MIMIC-IV cohort, achieving low Mean Absolute Error and Root Mean Square Error compared to other models. While excellent for physiology forecasts, trajectory-derived risk scores were found to be weaker than dedicated classifiers. Joint trajectory forecasting is presented as an inspectable intermediate task for clinical applications.
SciLitBench introduces a multi-stage benchmark for evaluating Large Language Models in systematic literature reviews. It assesses LLMs across title/abstract screening, full-text screening, and data extraction, using a dataset of 42,981 retrieved records and 1,012 full texts. Explicit inclusion/exclusion criteria improved title/abstract screening by 28.8\%, while researcher rationales improved full-text screening by 15\%. Performance in data extraction shows a reliability decline, with strongest models recovering only limited evaluation evidence. The benchmark establishes a practical boundary between high-recall screening and evidence-complete extraction for LLM-assisted synthesis.
AhaBench is a benchmark designed to evaluate whether language agents learn from prior experience over long horizons, addressing the challenge of continual learning. It assesses model behavior when evaluated after experience has been acquired, even when the initial support is removed or delayed. The benchmark includes three components: Aha-Puzzle, Aha-Euler, and Aha-Vending, testing different aspects of agent learning and adaptation. A three-part scorecard—Initial Score, Post-Experience Score, and Learning Lift—is used to quantify these effects. Results show that models achieving high post-experience scores or improvement during a run do not always align with initial competence or simple support utilization.
Option-critic learns sub-policies along with a rule for control handoff, and performance improves as options are added. Performance gains are explained by analyzing how the learned termination rule affects exploration and policy behavior. Policy necrosis occurs when a policy locks onto a suboptimal action, and restoring exploration repairs these necrotic states. Adding extra options does not improve performance, as it only reflects the joint failure chance of all options.
The research proposes a framework called Multi-Granularity Hypergraph Representation Learning (MGHRL) to capture high-order information in graphs. It introduces an Adaptive Granular Hypergraph Generation strategy by adaptively splitting granular-balls to generate hyperedges at multiple levels of granularity. A Multi-Granularity Hypergraph Network with multiple sub-networks and hierarchical reversible connections is also proposed for feature integration. Experimental results demonstrate that MGHRL significantly surpasses baseline models on benchmark datasets.
Capsule Lens is a framework designed to locate and track concept geometry within machine learning model representations. It achieves this by matching concept regions with simple, trackable geometric forms called capsules, defined by interpretable parameters. The method is applied to both static and dynamic model representations to characterize concept geometry. Analyses of dynamic representations reveal distinct geometric dynamics resulting from different training settings. The work proposes Capsule Lens as a tool for locating, analyzing, and tracking concept geometry in these representations.
HB-PVI is a hierarchical Bayesian framework designed for complex activity recognition by modeling participant heterogeneity and the value of information. It jointly models participant differences, the benefit and harm of personalization methods, and the economic value of additional labels for a cohort of 47 participants. The framework uses sequential Monte-Carlo and a Student-$t$ hierarchical gain model for evaluation. The study found that adapter personalization yielded small positive F1 gains, and value-of-information reasoning, rather than raw accuracy, should guide personalization decisions. The results support a population-first deployment policy when personalization gains are small relative to associated costs.
A reinforcement learning framework is proposed where exploration is driven by intrinsic curiosity, designed for non-stationary environments with sparse rewards. Action selection is guided by external rewards and an epistemic motivation mechanism that steers the agent toward structured exploration. The central hypothesis suggests effective exploration occurs at intermediate levels of incoherence, not overly rigid or disordered dynamics. The method was tested on LunarLanderv2 and BipedalWalkerv3 benchmarks, achieving competitive performance against PPO and ICM. The research also indicates that the curiosity window is not recovered in Active Inference agents under the same analysis.
Large language models (LLMs) are used to generate SystemVerilog Assertions (SVA), but point accuracy does not guarantee stability under different RTL behaviors. This paper evaluates the robustness of LLM-generated SVA by applying semantics-preserving RTL transformations. The evaluation involved testing two LLMs and three transformations, including operand reordering and identifier renaming. Results show that correctness can decrease after transformation, indicating instability in LLM-generated assertions. Therefore, robustness-aware evaluation is necessary for reliable AI-assisted hardware verification.
The research studies variance-preserving diffusion in mixed linear regression with unknown mixing weights. It connects score matching, maximum likelihood, and the Expectation-Maximization (EM) framework through the KL divergence. The analysis shows that the score matching objective links to the likelihood and a terminal discrepancy. This leads to an EM-related low-noise gradient expansion incorporating latent variance corrections. Numerical experiments validate these theoretical findings and statistical guarantees.
GraphNOSE is an open-source graph transformer framework designed to predict multi-label odor descriptors from molecular SMILES strings for single molecules and binary mixtures. It integrates positional and structural encodings within a transformer-based graph architecture to achieve strong performance. GraphNOSE requires six times fewer parameters than standard GNNs while outperforming other methods by a margin of 4.52% in AUROC. It achieves an AUROC of 84% on out-of-distribution compounds, surpassing existing GNN methods in olfaction. The model also employs XAI to reveal which molecular features drive the odor predictions, providing interpretable results.
The paper proposes a subsampled Davis-Kahan bound to address the computational expense of large-scale eigenspace estimation. It employs an independent Bernoulli sampling scheme to approximate the leading left singular vectors of a low-rank symmetric matrix. The main result provides an explicit error bound dependent on the sampling probability. This bound demonstrates a trade-off where computational cost scales linearly with the sampling probability, and statistical error scales as the inverse square root of the sampling probability. This work enables scalable spectral analysis for large-scale symmetric matrices.
Critical initialization preserves first-order input perturbations in wide, randomly initialized networks. Physics-informed losses and derivative regularization depend on higher input derivatives, which are analyzed using mean-field recursions. At criticality, the first-derivative variance is depth-invariant, while the second-derivative variance grows linearly if activations have non-zero curvature. Results show that residual networks with specific branch scales have uniformly bounded variance for finite derivative orders. The findings focus on initialization effects rather than the performance of trained networks.
Repeated-sampling evaluations can extrapolate pass@k beyond the number of collected samples ($n$). Fixed-$n$ success counts identify only the $n$ free moments of the latent per-task success distribution. Direct pass@k is identifiable only for $k \le n$, while extrapolated metrics for $k > n$ are not identified. This finding characterizes the information missing from fixed-depth count-law experiments. The study provides a nonparametric baseline for evaluating parametric scaling laws in machine learning.
CAST introduces a Canonical Approximate Schur Tree to approximate the Cholesky factorization on graphs, which arises when eliminating vertices in Approximate Cholesky methods. This tree replaces the dense Schur-complement clique formed by pivot neighbors with a weighted random spanning tree, ensuring the update remains unbiased. CAST-$\rho$ extends this by sampling a spanning tree on an expanded clique, offering a controllable trade-off between sampling variability and construction cost. The method is shown to minimize leverage-score marginals for inverse-marginal one-tree estimators. Empirically, CAST-1 is faster, while CAST-2 is preferred when additional edge contributions are inexpensive.
A method is presented for recovering the moral graph of a causal Directed Acyclic Graph (DAG) from a discrete probability distribution. This recovery is achieved using fully connected tensor networks (FCTNs) with nuclear-norm-regularized bond corrections. The nuclear norm penalty drives unnecessary bonds to zero, and it is proven that an optimal FCTN with zero reconstruction error yields the exact moral graph under certain assumptions. For approximate results, recovery bounds are derived using conditional mutual information continuity. The effective graph can be read directly from the optimized bond matrices.
Normalizing flow Boltzmann generators can suffer from biased training due to the forward KL divergence, potentially omitting data modes despite high effective sample sizes. The paper introduces log-ratio variation ($\X_\omega$) to define a new loss function, KLXX, to address this issue. Two variations of log-ratio are added to the forward KL to incorporate target information and sample mixtures, which contribute nonpositive dissipation in the Fisher--Rao gradient flow. Using KLXX in an adaptive-staging Boltzmann generator improves mode coverage and diagnostic accuracy compared to forward KL. The log-ratio variations effectively supply crucial information that the standard forward KL loss omits.
The paper studies the projected power method (PPM) for synchronizing unknown permutations under a sparse uniform corruption model. It proves exact one-step recovery of prescribed blocks with high probability when certain conditions on $m$, $n$, $p$, and $\pi_0$ are met. The theory shows that one high-probability event leads to a simultaneous block-error contraction for estimates with optimally aligned errors. Subsequent iterates then map all estimates to vanishing block error, ensuring almost exact recovery. The results extend to more complex corruption models and provide recovery guarantees for permutation synchronization.
Neural networks use feature superposition but cross-feature interference limits the linear accessibility of simultaneously active features. Framing linear accessibility as a compressed sensing problem allows for deriving high-probability bounds under subgaussian noise. These bounds show that the sufficient dimension scales linearly with $O_{\varepsilon}(k \log m)$ rather than prior quadratic limits. The research quantifies geometric constraints related to the linear representation hypothesis. These results provide a framework for evaluating sparse autoencoders and compositional generalization.
Synthetic data can cause model collapse when scaling training in high-dimensional linear regression using one-pass SGD. The study establishes finite-sample risk bounds for mixed and two-stage training, distinguishing between bias, variance, and source-mismatch effects. Two-stage training, where synthetic data is used only in the initial stage, avoids the risk floor associated with mixed training. Scaling laws are derived showing that larger models may amplify synthetic-induced degradation under mixing. Ultimately, the effect of synthetic data depends critically on its quality and the chosen training protocol.