There's a scaling law hiding in your inference bill, where test-time compute becomes the third axis of scaling after parameters and data. This conceptually odd phenomenon leads to test-time compute distillation, where a model is distilled into itself, using the same network given more time to think. The distillation process is the "Amortization Move" that compresses the better model back into the weights.
OpenAI's CEO, Sam Altman, has stated that his company is close to releasing a version of ChatGPT that can watch a user's screen and record every meeting within six months, raising concerns about AI privacy.
The yield on 30-year US Treasuries hit 5.29%, its highest since 2007, as investors demand higher returns due to surging national debt and persistently high inflation. This move is part of a broader trend in global bond markets where investors seek more compensation to protect against government debt risk. The increase is also driven by a surge in corporate borrowing to fund the artificial-intelligence boom and waning demand from traditional bond buyers.
Flock, a police-tech giant, is tightening its rules in response to a growing backlash against its nationwide network of license plate readers. The company will now require officers to enter a criminal case number before conducting a search, and is expanding an automatic auditing system to catch suspicious searches. Flock is also making changes to limit data collection and access, such as recommending that agencies hold onto data for seven days and limiting searches to specific stated reasons. The changes come as cities and states are canceling contracts with Flock due to concerns about mass surveillance and police abuse.
US State Department employees received an email informing them that their unit, the Counter Foreign Information Manipulation and Interference Hub (R/FIMI), would be shut down and their jobs eliminated. The closure was celebrated by Secretary of State Marco Rubio, who claimed it marked the end of government-sponsored censorship in the US. The idea of a vast censorship network, known as the "censorship-industrial complex" (CIC), has been promoted by far-right blogs, podcasts, and social media, and has been linked to the dismantling of numerous government agencies and the suppression of conservative and populist speech online.
The US is telling dozens of allies to choose between joining the Pax Silica AI coalition or China's rival framework, with countries that join both facing exclusion from the US-led coalition. The ultimatum comes after Kazakhstan joined both coalitions simultaneously, and follows China's fast-advancing open-weight AI models that have gained ground against US proprietary systems. The US-led coalition aims to secure allied supply chains for AI models, semiconductors, and critical minerals, while reducing reliance on China.
The "censorship-industrial complex" refers to a theory that government agencies, academics, civil society groups, and Big Tech platforms have collaborated to suppress conservative and populist speech online. This concept has become part of US policy under the Trump administration. Researchers investigated its rise, revealing its potential impact on democracy and the internet.
OpenAI and CodeAI are partnering to provide students and educators with tools and resources to learn how to use and benefit from AI, with a focus on critical thinking, responsible use, and creativity. The partnership includes the launch of ChatGPT for Teens, a dedicated experience designed to help teens think critically about AI, as well as other programs such as the Hour of AI and the Builders Challenge. The goal is to equip students with the skills to shape the future of AI and to use it responsibly.
Developers create an end-to-end OCR workflow with docTR, combining text detection, recognition, geometry, layout analysis, and export. They generate synthetic invoice documents, load images and PDFs, construct GPU-aware OCR predictors, and benchmark different detection-reognition architecture combinations for speed and accuracy. The OCR pipeline extracts structured invoice fields, exports results as text, JSON, hOCR, synthesized document images, and searchable PDFs. The developers also examine performance, fine-tuning, batching, and deployment considerations to move from a basic OCR example to a production-oriented document intelligence pipeline.
Nvidia aims to invest in open-source AI models to make token-based intelligence accessible to everyone, rather than relying on companies like Anthropic and OpenAI for model purchases. This strategy, exemplified by the release of data and training code for their Nemotron models, aims to reduce the monopoly on intelligence and increase demand for Nvidia's offerings. However, the financial sustainability of open-source models remains uncertain, with Nvidia potentially relying on their financing in the coming years to keep the ecosystem viable.
Project management has changed with the advent of LLMs, shifting from 70% coding, 15% meetings, 15% testing, and 0% interacting with the agent to 30% prompting agents, 10% meetings, 30% testing, and 30% extra time. To effectively manage projects, map out work beforehand to clear ambiguities, and use the /goal command to ensure agents complete tasks fully.
DeepSeek Harness is an open-source agent harness developed by DeepSeek AI, utilizing an architecture where everything is a plugin. It is powered by Cordis, with a design described in A Programming Paradigm for Spatiotemporal Composability. The harness is currently in developer preview, with compatibility-breaking changes expected.
MongoDB has introduced official Agent Skills, structured instructions, best practices, and resources for coding agents to generate reliable code across the development lifecycle, from schema design to performance optimization. These skills address common pitfalls, such as over-normalizing schemas and misusing indexes, and provide expert guidance to agents. The skills are packaged with the MongoDB MCP Server and are available as plugins and extensions for popular tools like Claude Code, Cursor, Gemini CLI, and VS Code.
Flock Safety's automatic license-plate readers (ALPRs) are a valuable tool for modern police forces, helping to catch criminals and reduce crime. Research suggests that ALPRs can be effective in reducing crime, with studies showing a correlation between ALPRs and a decrease in violent crime, auto theft, and property crimes. However, concerns about privacy and police abuse of the technology are valid, and regulations such as limits on data retention and stricter rules about access to ALPR data can help mitigate these concerns.
AI concepts visualized through infographics and visual guides, curated by Turing Post, explore model architectures, RL approaches, JEPA, and agentic memory.
Flock Safety misled Oshkosh officials about their license plate reader system's capabilities, including the generation of heat maps, leading to the cancellation of a contract within 24 hours. The company falsely claimed no access to federal data, despite maintaining pilot contracts with U.S. Customs and Border Protection and the Department of Homeland Security. Flock's compliance tool to prevent surveillance misuse was easily bypassed by officers, including using inappropriate search justifications, raising concerns about user privacy.
AI researchers have found that AI use differs significantly across models, and has changed over time, with many more sensitive behaviors than captured in reports from major AI companies. The AI Observatory, a public platform, analyzed real AI conversations with popular models like Claude and Gemini, and found that nearly half of the conversations were filtered out by Anthropic's Economic Index, including sensitive topics like health, relationships, adult content, and harassment. The dataset, which includes 24,521 conversations from 5,000 users interacting with 52 different models, suggests that AI companionship is increasing, while AI assistants' self-disclosure is decreasing, and that platforms are deploying more effective safeguards.
Humans have itchy feet and are wired to roam, driving their desire to explore and adapt. This instinct is ancient and primal, potentially unstoppable. As space exploration expands, the same human foibles that stymied us on Earth will apply, including questions about ownership and resource management.
A new study suggests that AI's recursive self-improvement, where AI improves itself with minimal human oversight, might not be as imminent as predicted. Researchers found that AI agents are capable of solving engineering problems necessary for AI research but lack the judgment and creativity to conduct open-ended research, which is essential for building self-improving AI. The agents struggled to produce original research, incorporating feedback, and making novel contributions to their fields.
Google, the UK Government, and aviation leaders are partnering to use AI to reduce aviation's climate impact by avoiding contrails over the North Atlantic. The program, Operation Blue Skies, will operate within the eastern half of the North Atlantic corridor, aiming to reduce contrail warming by 5%. Over 30 months, around 10,000 flights will pass through the trial airspace, with a focus on coordinated contrail mitigation and evaluation of climate impact. The program is co-funded by industry partners and the UK government, with 100% of the government grant funding allocated to academic and non-profit partners.
QAD consistently outperforms PTQ in median accuracy recovery and agentic/coding benchmarks, with QAD enabling more aggressive quantization settings that would otherwise incur unacceptable degradation under PTQ alone. QAD recovers accuracy loss from aggressive quantization, and even with more conservative configurations, QAD consistently outperforms PTQ on agentic benchmarks. The QAD process involves a two-stage process: post-training quantization (PTQ) and quantization-aware distillation, which uses a KL divergence loss and simulated quantization to recover almost all accuracy lost from aggressive quantization.
A speech-to-text AI specialist is now valued at $2 billion, part of a growing trend of native-AI startups receiving significant funding.
Alibaba's 3.8 billion parameter model targets edge AI. It can run locally on a laptop, showcasing Alibaba's commitment to pioneering open source.
LG plans to release a humanoid robot powered by Nvidia in 2027. The robot is part of the companies' partnership to implement physical AI in real-world settings. No further details are available.
Role Anchor mitigates role drift in compound AI systems by forcing modules to stay in their lanes during training, ensuring accuracy and reliability. A technique that measures a role's effect by comparing how a model behaves with and without the role prompt, Role Anchor applies a penalty to the model if its nudge deviates from the reference nudge, preventing role drift. When applied to the RAG and Decomposer-Solver pipelines, Role Anchor improved accuracy while preserving the intended behavior of modules, particularly in scenarios where final accuracy alone is insufficient to verify the model's integrity.
Enterprises with AI context layers report agent failures at more than twice the rate of those without one, with 68% tracing confident but wrong answers to missing or inconsistent business context. The failure rate is climbing, with 37% of enterprises reporting recurring failures, despite more enterprises adopting a governed layer in production. A governed layer doesn't cause the failure but rather makes it visible, highlighting the need for a shared, governed reference point to fix this issue.
xpander.ai is a startup offering an enterprise AI agent platform that aims to provide a vendor-neutral control plane for building, running and governing agents across different models, agent frameworks and infrastructure environments. The platform is designed to help companies manage the growing problem of AI agent sprawl and is generally available now. The company's founder and CEO, David Twizer, argues that enterprises need a framework-independent control plane that can handle execution, permissions, observability, memory, access to enterprise systems, and lifecycle management.
Nous Research has shipped Bot Mode for Hermes Agent, a feature that turns individual agent profiles into named bots with their own chat, memory, skills, and profile picture. Each bot lives under a separate directory and communicates through a persistent Agent Inbox, allowing real CLI handoffs. Bot Mode is now bundled and default-on in Hermes Desktop, version v0.20.3.
ByteDance Seed and Tsinghua AIR have released CUDA Agent, an agentic reinforcement learning system that trains a large language model to write GPU kernels that beat a compiler. The system achieves a 98.8% pass rate and a 96.8% faster-than-torch.compile rate on KernelBench, with a geometric-mean speedup of 2.11×. The model outperforms other state-of-the-art methods, including Claude Opus 4.5 and Gemini 3 Pro, on the hardest Level-3 split.
MiniMax has released MiniMax-Music3, an open-weights text-to-music model that generates complete five-minute songs from lyrics and a structured caption. The model uses a Hybrid-LM architecture with a 8B Global LLM and a 0.6B Local LLM, paired with a continuous synthesis stack built on flow matching and a Flow-VAE. The model can be deployed commercially, but requires attribution and authorization for organizations above $20 million in revenue. MiniMax-Music3 runs on two GPUs via SGLang-Omni or under 24 GB via diffusers.
The modern tech landscape is plagued by various annoyances, including dickovers (pop-up ads), AI-horning (unnecessary AI features), captchore (repetitive and time-consuming captchas), and many more. Examples include quit-shaming (aggressive subscription cancellation processes), cart cling (persistent emails from vendors), chattermining (devices always listening to user conversations), and loginsanity (complicated and frustrating password systems). Other issues include gatejamming (difficulty navigating train ticket barriers), baggravation (self-checkout machine frustration), parcel liveblogging (excessive hand-holding in online package delivery), and socket boomerism (nostalgia for outdated technology).
Anthropic's update to its Claude AI model will include a watermark to comply with EU regulation requiring all AI-generated text to be watermarked starting in December. The changes will alter the random choices made by the model, leaving a pattern detectable to Anthropic and those with a key to decode it. This could potentially make the model's writing worse, according to tech blogger John Gruber, as the watermark may constrain the model to make less precise word choices.
A new constellation of 50 satellites, called FireSat, is being deployed to detect wildfires as they start, using artificial intelligence and infrared sensors. The satellites can spot fires as small as a beach bonfire and scan every point on Earth about every 20 minutes. FireSat is designed to work with a growing network of cameras on the ground that use AI to search for smoke and heat. This combination of satellite and ground-based systems aims to give firefighters their earliest warning of a fire, reducing the risk of large blazes.
Researchers from Thinking About Thinking, University of Oxford, Princeton University, and others have created DiG-bench, a benchmark of 70 games designed to test AI systems' ability to infer the unwritten rules of their environment through exploration and curiosity. The benchmark shows that Opus 5 and Fable 5 are the best overall models, followed by GPT-5.5. In another study, Inherent has built an AI scientist model called Faraday, which is post-trained on top of large frontier models and uses a coding agent (OpenAI Codex) to improve its performance. Mark Zuckerberg has written an essay, "The Future is for Everyone," which proposes to massively proliferate AI capabilities to everyone on the planet, but raises concerns about the missing question of how a system capable of superhuman invention will work in conjunction with individual empowerment.
Qwen 3.8 27B scored 52 on the Artificial Analysis Intelligence Index, tying with GPT-5.6 Luna (max) and just one point behind GLM-5.2 (max) and DeepSeek V4 Pro 0813 (max). Qwen 3.8 27B is described as a "truly astonishing model".
404 Media tracked a shipment of rare books using an Apple AirTag, which ended up at an Amazon AI training facility in Las Vegas. The books were delivered to the VGT3 corner of the facility, where workers confirmed that large volumes of books are destructively scanned. This is part of Amazon's practice of using books as training data for AI models.
Large Language Models struggle to detect errors in numeric substitutions, with only a few methods performing well. The Ragas-style LLM judge and MiniCheck show promise, but most methods average out errors, making it difficult to detect single-digit substitutions. A new approach called groundlens uses a frozen encoder and regular expression to detect errors, with a false-positive rate of 0.65 at 95% recall, reducing review volume by 30% while missing 5% of defects. The trained span detector performs well on its own benchmark but fails on other datasets.
Agentic traffic violates all seven assumptions used in scaling for on-demand and serverless models, resulting in the failure of both approaches. To build and respond to agentic requests, four layers of response are necessary: behavior-based scaling, AI gateway as a shock absorber, async queuing, and token-based admission control. However, intelligence must move upstream, and the client must be smart enough to know when to stop asking. A well-behaved agent client should carry a retry budget, honor backpressure signals, run a client-side circuit breaker, and treat backpressure signals as cooperative.
Loop engineering for RAG: Small loops inside each step, the big loops across the pipeline. Loop engineering is the discipline of designing what happens next when a call fails. It has three control surfaces: trigger, termination, and recovery. A loop that fires its trigger, retries with the same payload, gets the same failure, and retries again is not a loop, but a denial-of-service against the company's API budget.
A constraint-aware GPU allocator was developed and benchmarked against a FIFO scheduler across seven benchmark scenarios. The allocator improved utilization by as much as 33 percentage points and priority-weighted output by as much as 105%. The gain came from encoding what the cluster physically permits into the order decisions get made.
MIT Professor Eugene Fitzgerald's book "The Invisible Engine: Why Innovation Evades Control" examines how innovation unfolds in the real world, highlighting the importance of understanding market applications, technology, and implementation. Fitzgerald argues that innovation is not just about an idea, but a process of working with the world to create value.
Researchers adapted an SO-ARM100-based setup to run reinforcement learning in Isaac Lab with two robot arms: PiPER (6-DoF) and NERO (7-DoF). The setup includes end-effector reaching for PiPER and Reaching+cube manipulation for NERO. The experiment uses PPO-based reinforcement learning and features 64 parallel simulation environments. The goal is to provide a ready-to-run workflow for reinforcement learning in a simulated environment.
Claude Swarm conducted experiments to analyze patterns and problems in multiagent systems. The experiments involved a game where two AI agents competed, with the agent that lost the most games being considered "swarm intelligence" as a potential solution to optimize performance. The results showed that swarm intelligence can be used to optimize the performance of multiagent systems.
Screeps is a real-time strategy game where players control a single game entity, known as a "slime". The goal is to build and expand a network of "screeps" to capture and harvest resources. The game utilizes reinforcement learning, where players receive rewards for achieving goals and penalties for failing. By optimizing their slime's behavior, players can improve their chances of success.
nanoRL is a smallest async RL trainer written by alex000kim, consisting of 1,800 lines across 7 files, without using Ray, TRL, or DeepSpeed. It scales from a laptop CPU to a GPU cluster using a disaggregated trainer/worker setup.
AI systems out-persuade expert humans, according to a study by Hackenburg et al (2026). The study found that AI systems were more convincing and persuasive than human experts in certain situations. The researchers' results suggest that AI may have an advantage over human experts in tasks that require persuasive communication. The study's findings have implications for the use of AI in areas such as sales, marketing, and public speaking.
Google's AI leadership underwent significant changes, with Jeff Dean and Demis Hassabis leaving their positions. Koray Kavukcuoglu was promoted, while details on the reasons for the other departures are unclear. Google's AI efforts now seem to be focused on Gemini, with the phrase "Gemini uber alles" indicating a shift in priorities.
A bot-friendly Reinforcement Learning (RL) environment was created for a Pokémon-like game, allowing AI agents to learn and interact with the environment. The RL environment was designed to mimic the behavior and complexity of the Pokémon game, enabling AI agents to develop strategies and make decisions like humans. The environment was created using a combination of game mechanics, agents, and rewards, providing a realistic and engaging space for RL agents to learn and improve.
AMD has surpassed its initial target of a 25% productivity boost from AI in software development, achieving a 30% overall productivity gain just one year later. The company's AI systems have improved dramatically, enabling agents to work in parallel, scale expertise, and discover solutions independently. To achieve this, AMD is developing collaborative swarms of AI agents that can independently identify and develop solutions, guided by humans on what to solve. This transformation will require a continuous learning loop that feeds errors and human interventions back into future agent workflows, making the entire AI engineering workflow smarter and more effective.
A team at Axiom Math has used its AI system AxiomProver to automatically verify the proof of the 246 theorem, a significant milestone in AI-assisted mathematical research. The 246 theorem posits that there are infinitely many prime numbers that differ by 246, and its formalization is a comprehensive achievement that demonstrates the potential of automated AI verification. This milestone demonstrates how AI can be used to verify the correctness of AI-generated computer code, which could be crucial in ensuring the safety of digital data.
ChatGPT for Teens is a new experience designed to help teens learn, think critically, and use AI with confidence, with stronger built-in safety protections for teens. The platform provides features to promote healthy use and additional controls for parents, including age-appropriate safeguards to reduce exposure to harmful or developmentally inappropriate content. The goal is to support learning outside of the classroom while keeping teens engaged in an active and collaborative learning process.
A customer, Matt Arnold, was incorrectly identified as a shoplifter at a Sainsbury's store in East Dulwich, London, and ejected from the shop after a false alarm triggered by the store's AI face scanning technology, known as Facewatch. The store has paused the use of Facewatch in the East Dulwich store while an investigation takes place. Sainsbury's has apologized to Arnold and assured customers that the Facewatch system has a 99.98% accuracy rate, but Arnold and experts are concerned about the potential for false accusations, particularly for vulnerable individuals. The incident is part of a growing list of issues with the Facewatch technology, which has plagued multiple stores, including Home Bargains and B&M.
A master's student in robotics in Germany seeks tips on gaining an internship in reinforcement learning (RL) with a focus on positive industry applications, such as construction and mining. The student has applied theoretical knowledge in learning-based control and reinforcement learning, participated in hackathons, and completed projects with ROS2. They feel that their efforts may not be sufficient for securing an internship in RL.
AI efficiency has taken the spotlight due to massive model scales, high energy demands, and environmental costs. A study proposed the $\alpha-FLOPs$ estimation formula, which reported FLOPs as a traditional approach for assessing computational costs. However, FLOPs alone are not an appropriate metric for execution time, as spatial dimensions remain more easily parallelized than kernel dimensions. Newer hardware exhibits instabilities and discontinuities in execution time, including jumps and oscillations, that the $\alpha-FLOPs$ formula generally underestimates.
Large language models (LLMs) show partial metacognitive sensitivity in medical reasoning, with diagnostic accuracy of 93.5% and AUROC2 of 0.876 in a controlled benchmark. Confidence increased with evidence distance from the diagnostic boundary, but errors clustered in moderate, conflicting cases, where the model shifted toward incorrect diagnoses. The study suggests that confidence quality should be measured directly rather than inferred from benchmark accuracy or model capability.
Multimodal machine learning models struggle with abstract perceptual reasoning, a task defined as acousto-kinematic word inference, where models must decipher words from audio and video of pen scratches without visible ink trace. Human participants achieve high ordered letter accuracy (over 80%), while models, including GPT-4o and Gemini 2.5-Pro, fail to surpass 10%. The models also exhibit a paradoxical fusion effect, where providing both modalities often degrades performance. This finding highlights significant limitations in cross-modal causal reasoning and micro-kinematics understanding.
Agents communicate only when the KL divergence between their learned belief distributions exceeds a fixed threshold. Each agent maintains a belief distribution over a latent world state computed as a softmax over its LSTM hidden state, and communicates only when belief disagreement is large enough to justify information exchange.
AI governance is shifting from voluntary ethics to enforceable, risk-based regulation, creating compliance uncertainty for operators of high-stakes AI. A comparative matrix for the EU, US, and China maps risk classification triggers, binding obligations, enforcement mechanisms, and FAIR principle operationalisation. Three gaps identified are weak interoperability mandates, under-specified governance for critical digital infrastructure use cases, and difficult operationalisation of cross-regime obligations. A proposed machine-checkable compliance artefact pattern, Knowledge Blocks, aims to bridge the implementation gap.
AI safety research should address the risk of AI Lock-In, which occurs when humans become overly reliant on AI systems, leading to deskilling, diminished human capacity, and systemic vulnerabilities. This phenomenon is emerging at individual, societal, and national levels, and can be amplified by AI service disruptions or geopolitical conflicts. Proactive addressing of AI Lock-In is essential for preserving individual autonomy and national security.
Recent work on evaluating AI moral reasoning has primarily focused on the moral value problem, but neglects the moral norm problem. Evaluations cluster heavily around value representation, while discussion on normative ethics remains underrepresented. This imbalance is due to the field's reliance on descriptive ethics frameworks, which emphasizes value representation over normative application. A proposed research agenda aims to address this gap by developing standardized formal representations, constructing expert-annotated datasets, and evaluating protocols that distinguish between values-level and norms-level competence.
This paper presents a targeted narrative review establishing the historical and theoretical foundations for computational belief change implementation, tracing the evolution of belief revision from Doyle and London's foundational taxonomy to the AGM framework and contemporary approaches. It demonstrates how pre-AGM computational pragmatism relates to AGM theoretical constructs, revealing both continuities and transformations. The analysis identifies the theoretical foundations and historical precedents that inform contemporary implementation challenges, providing a baseline for systematic implementation analysis and engineering-focused belief change research.
Robust governance frameworks are needed for AI systems deeply integrated into critical global infrastructure. Current approaches, led by jurisdiction-specific laws, create a fragmented regulatory landscape. AI governance must be built on ISO-like interoperability protocols to enable standardized, machine-readable risk communication across borders, and standardized AI "nutrition labels" containing unified metrics for bias, energy usage, and data provenance are proposed.
Neural solvers for constraint satisfaction problems achieve remarkable in-distribution accuracy, yet they suffer from persistent constraint violations under distribution shifts. To address this, the authors advocate for prioritizing symbolic integration over pure learning, particularly when hard constraints exist. They propose a multi-agent certified reasoning framework that combines neural methods with symbolic solvers, enabling both computational efficiency and provable correctness.
Dongzhe Zheng and Christine Allen-Blanchette introduce Riemannian Hodge Message Passing (RHMP), a neural architecture that separates topology and geometry in physical fields. RHMP learns symmetric positive-definite cochain metrics for geometry-dependent propagation, achieving cochain-frame equivariance and exact cochain-complex identities. RHMP performs well across seven physical benchmarks, with the largest gains when topology, geometry, and field structure interact.
Forward-Pass-Only MLP training (FPO) adapts large language models without a backward pass through the model body, achieving 2.7--3.2x the throughput of standard fine-tuning. FPO rests on the observation that late layers of a transformer's output-layer prediction error approximates the true gradient with cosine similarity. FPO computes a single error signal at the output and applies it to each target layer, without propagating signals between layers or constructing an autograd graph.
MR-Traj, a novel multi-resolution diffusion framework, generates large-scale trajectories explicitly modeling coarse-grained milestones and fine-grained segments. Experimental results show that MR-Traj outperforms state-of-the-art methods in modeling fine-resolution mobility patterns and supporting downstream urban mobility tasks, including traffic management and epidemic control.
Projected Gradient Descent (PGD) is widely used to evaluate adversarial robustness but typically via final adversarial accuracy, which does not capture model behavior throughout the attack. A trajectory-level investigation of PGD attacks on convolutional neural networks reveals a clear robustness hierarchy across models, but trajectory metrics do not equally contribute to its identification. Steps-to-failure distributions provide a clearer separation of robustness regimes, directly reflecting functional resistance to adversarial perturbation. Trajectory-level diagnostics describe optimisation geometry but do not independently measure adversarial robustness, and should be used as a complementary diagnostic tool.
A study on 1,000 real European Court of Human Rights cases found that fusing uncertainty tools into a pipeline does not improve case-outcome prediction. The pipeline, which included a conformal selective-prediction layer, yielded no improvement over a raw LLM or a term-frequency baseline, and actually increased calibration error in some cases. The use of Dempster-Shafer fusion was found to be actively unsafe on long chains and was recommended to be removed. After recalibration and risk control, the tuned engine achieved 96.8 percent accuracy.
A new neural operator framework, PIKFNO, combines physics informed kernel functions with a trunk network to align with kernel expansions in meshless collocation methods. This framework constrains the trunk network through physics informed kernel functions, improving interpretability and generalization under limited training data. PIKFNO achieves high predictive accuracy and offers a new pathway for developing efficient and physically consistent neural operators. Two construction strategies are introduced, one learning kernel functions directly from data and the other building them through transformations of analytical fundamental solutions.
EARL, a Python library, generates counterfactual explanations in RL settings by exploring "What-if" scenarios to clarify agent behavior. It allows users to produce explanations in realistic RL-based self-adaptive systems. EARL is demonstrated on a simulation of CitiBikes, a self-adaptive bike-sharing system, and shows its applicability in real applications.
Fractional optimizers meet fractal activation functions in a unified experimental framework to improve neural network training. The authors evaluate fractional optimizer families on benchmark surfaces and feed-forward neural networks with conventional and fractal activations, finding useful but selective pairings. Regularization-style fractional scaling performs well with selected fractal activations, while Grünwald--Letnikov memory is most relevant on perturbed surfaces. Adaptive memory improves plain memory substitution in several cases, supporting controlled fractional memory as a promising direction.
ARISE (Adaptive Residual-Informed Stability Ensemble) is a feature selection framework that integrates complementary relevance signals, class-balanced stability assessment, residual-informed redundancy control, and multiclass pairwise coverage for small-sample molecular classification. It combines seven percentile-normalized relevance components through 15 predefined profiles, adaptively weighted by nested inner cross-validation. ARISE was evaluated on five molecular datasets, eight feature-set sizes, and three fixed classifiers, and ranked first in all 15 dataset-metric combinations, with balanced accuracy, macro-F1, and Cohen's kappa values exceeding the strongest aggregate comparator by 0.022, 0.023, and 0.028, respectively.
A deep learning model for spatially clustered data via differentiable cluster assignment is proposed, which jointly learns the partition and cluster-specific regression functions. A neural network determines cluster membership, while separate neural networks describe the covariate-response relationship within the clusters. The model uses a combination of graph-Laplacian and occupancy penalties to discourage fragmented regions and degenerate solutions, and establishes identifiability up to label permutation.
Sufficient dimension reduction (SDR) seeks the minimal subspace of predictors that captures the full conditional distribution of the response. Existing methods face limitations, including strong distributional assumptions, matrix inversion, and computationally intensive iterative smoothing. A new SDR framework based on the generalized Stein's lemma is proposed, which constructs a cross-moment matrix and recovers the central subspace via singular value decomposition. The proposed method establishes convergence guarantees and can leverage unlabeled data, outperforming existing approaches in simulation studies and real data applications.
The authors propose a method for diffusion inverse problems, using a scale-consistent posterior dynamics. This method is based on a conditional score governed by a stochasticity parameter, which controls probability-flow transport and stochastic exploration. The likelihood is expressed in a rescaled clean-image coordinate, and a noise-conditioned covariance path is obtained through the forward operator.
Population-based adaptive importance sampling methods assess performance through effective sample size (ESS) and weight-based diagnostics, but a large ESS only indicates balanced sample weights without describing proposal component arrangement. The effective number of proposals (ENP) combines total normalized weight with a redundancy measure, estimating non-redundant empirical proposal contributions and detecting proposal collapse and duplication missed by standard ESS. ENP can serve as a targeted feedback signal for proposal rejuvenation.
PENN-GMD, a parameter estimation neural network, is proposed to identify parameter couplings and uncertainties in mixed-noise stochastic systems. It maps partially observed trajectories to a Gaussian mixture distribution (GMD) over system parameters, employing full covariance matrices to explicitly reveal parameter couplings. PENN-GMD accurately recovers likelihood distributions and captures parameter couplings, and naturally diagnoses non-identifiability through variance broadening or mode splitting.
Researchers Gordei Verbii and Juho Lee propose a distributional view of knowledge distillation, where the teacher's softened output is represented by a family of multi-temperature views. They formalize various aggregation operators, including log-linear pooling, entropic Wasserstein barycenters, and debiased Sinkhorn-divergence, and prove that log-linear pooling of tempered views is equivalent to a single temperature. Experiments on instruction-tuned Pythia pairs yield three empirical laws, including the dispersion law, the unlocking of aggregation operators, and a two-regime picture governed by the ceiling gap.
QXGB, a quantile-based gradient boosting framework, incorporates a convolution smoothed loss that estimates conditional quantiles, enabling the construction of dense cumulative distribution functions, exceedance probabilities, and tail behavior relevant to extreme outcomes. This approach preserves the computational efficiency of extreme gradient boosting while restoring the Hessian information XGBoost relies on for tree splitting, providing interpretable measures of extreme value and exceedance probability predictions. QXGB delivers accurate predictions with near-zero quantile crossing, well-calibrated CDF and exceedance probability estimates, and useful tail characterization for extreme values, particularly when paired with multi-output trees.
Researchers have introduced a unified geometric framework for analyzing the spatiotemporal evolution of gene expression networks through embeddings in Gromov--Wasserstein (GW) space. This framework represents each developmental stage as a graph combining gene expression and spatial proximity, enabling comparisons of network structure across time, interpolation between developmental stages, and quantification of network-level changes. The approach is validated on a spatiotemporal transcriptomic Drosophila dataset, showing agreement with higher-order Co-Optimal Transport (COOT) distances, which jointly represent spatial and temporal information. The framework provides a unified geometric approach to study dynamically evolving biological networks.
Researchers Kejin Wu propose a method using standard ReLU Deep Neural Networks to predict time series and quantify uncertainty. The approach, called the pertinent prediction interval (PPI), captures both the future variability and the estimation variability within training data. The PPI is built by imposing minimal conditions on the limiting distribution of predictive roots, which are beta-mixing and possess the same stationary distribution as the original time series in probability.