AI tokens are rapidly becoming a core component of corporate economics, with companies beginning to report token expenses. This signals the emergence into a new "token economy." The author questions how this unit should be measured and whether it will spur new software solutions, such as an Enterprise Resource Planning (ERP) system for tokens. Currently, the token has become established as the atomic unit within the AI economic framework, overshadowing parameters or GPUs. This shift necessitates a discussion on properly accounting for and managing token usage across businesses.
Midjourney unveiled a prototype for the Midjourney Scanner, described as a full-body ultrasonic CT system capable of capturing detailed internal images without radiation. The device utilizes a large ring setup with over 358,000 ultrasonic elements and processes vast amounts of data, aiming for resolutions down to 0.5 mm. While impressive in hardware specifications—including processing around 40 GB per slice—the current system is early-stage, requiring about 20 minutes per scan due to infrastructure limitations and reportedly not yet incorporating AI for imaging. Separately, Midjourney also announced the physical opening of a "Midjourney Spa" location in San Francisco.
To set up a new podcast subscription, open the provided link on your mobile device. Locate and tap the designated "Add to podcast app" button within the message. This action integrates the content directly into your preferred podcast application. Following these steps ensures you can easily access the show's episodes through your usual listening platform.
Recursive Self-Improvement (RSI) is the concept of AI participating in its own development by proposing ideas, running experiments, and improving components. This shifts human researchers to roles focused on setting goals and validating results rather than minute implementation details. Historically, RSI builds upon ideas from John von Neumann concerning self-reproducing automata and Irving John Good's theory of an "ultraintelligent machine." Currently, the field is in its early stages, with recent advancements shown by Anthropic, Recursive, and Sakana AI focusing on creating better development loops. In this framework, advanced models write code that improves training infrastructure or propose experiments to enhance post-training capabilities.
The article offers five valuable resources for learning system design, a topic frequently required in technical interviews. System design expertise is becoming crucial even for entry-level roles, possibly to guide AI tools. One highlighted resource, ByteByteGo, provides a helpful newsletter and YouTube channel for understanding the subject. Another section details an academy offering a structured path to learn how modern AI models operate under the hood. Additionally, some resources cover core concepts necessary for building and architecting complex systems effectively.
Joseph Krause discusses how Materials Science presents a challenging field for AI acceleration due to complex variables beyond just chemical formulas. Unlike simple molecular predictions, successful material creation involves factors like supply chains and manufacturing processes. Radical AI is developing "self-driving labs" utilizing an "AI scientist" to automate hypothesis generation and testing. This approach acknowledges that no single AI model can fully address the entire materials discovery pipeline. Accelerating materials discovery is crucial for advancements in electronics, aerospace, and defense technologies.
Solar geoengineering, an idea to deliberately cool the planet by intervening in the climate system, is moving from simulations to practical application. Researchers are developing systems involving aircraft and materials for this process. However, implementing solar geoengineering faces significant real-world hurdles. Early deployment would necessitate substantial new infrastructure, considerable time, and major investment. Separate content details advancements in interoception, the internal sense of bodily feelings.
Quince successfully won its legal dispute with Ugg maker Deckers regarding boot design. A jury found that Quince's shearling boot copied the look of Ugg's Classic Ultra Mini Boot. However, the ruling determined that the underlying patent for Ugg's design was invalid. This victory is highlighted within ongoing "dupe culture wars" concerning copycat fashion retailers.
Researchers are considering using massive, uncrewed aircraft capable of reaching the stratosphere to cool Earth via solar geoengineering. These planes would release materials designed to reflect sunlight back into space, mimicking natural cooling effects from major volcanic eruptions. Key challenges involve designing suitable high-altitude aircraft and perfecting the chemical dispersal of reflective aerosols. Scientists must also address unknown variables regarding safety, cost, and efficacy before any practical implementation. The technology presents a controversial approach with potential benefits against climate change alongside significant risks.
Universities are adapting curricula to prepare students for a workforce increasingly integrated with artificial intelligence. Educational institutions are focusing on developing core skills like critical thinking, creativity, and complex problem-solving—areas where human input remains vital. Integrating AI tools directly into coursework allows students hands-on experience with emerging technologies. Furthermore, many universities are emphasizing digital literacy and ethical considerations surrounding AI deployment. These comprehensive adjustments aim to ensure graduates can effectively collaborate with, rather than be replaced by, intelligent systems.
France announced comprehensive plans a year ago during NVIDIA GTC Paris to bolster its domestic AI capabilities. These plans focused on establishing new AI factories and increasing national compute capacity for artificial intelligence. Furthermore, France committed to developing open frontier models and robust industrial platforms supporting the sector. Currently, this planned infrastructure is becoming operational in practice. As a result, AI agents are now running in production environments, and startups are actively deploying applications throughout the French ecosystem.
Anthropic significantly overhauled Claude Design to transition it from a prototype tool into an enterprise platform focused on brand compliance. Key updates include importing design systems directly from GitHub repositories, allowing output validation against established company components and enabling admin lockdown features for corporate control. Furthermore, bidirectional integration with Claude Code establishes a seamless workflow between visual design and actual codebase implementation, aiming to resolve the persistent design-to-engineering handoff problem. The tool also addresses earlier usage concerns by sharing token limits across other Claude products and improving efficiency. These changes position Claude Design as an embedded worker within existing enterprise IT systems.
Integrating AI and reducing full-time staff signals a growing trend toward the gig economy. When Klarna cut customer service roles to implement an AI chatbot, initial consumer complaints led them to rehire human support. However, rather than restoring permanent positions, Klarna hired temporary workers through an outside agency. These new employees operate in a model described as "Uber type," where advanced queries are handled by these flexible gig workers. This pattern suggests AI integration may increase reliance on less secure, contract-based labor for essential corporate functions.
Efforts are underway for musicians to get paid when their work trains generative AI models, with startups like Sureel pioneering tracking systems. These initiatives explore methods beyond one-time buyouts, suggesting ongoing compensation based on the music's influence during AI output generation. Key challenges include technically attributing causality between specific training data and a model's output. Industry experts suggest this attribution process must balance rewarding creativity against avoiding new, exploitable economic structures. Ultimately, establishing fair payment requires careful design that accounts for cultural complexity and transparently involves creators.
Lazy Programmer Newsletter offers developers quick updates from over 50,000 subscribers. Recent articles cover system design resources, a new "dangerous" AI model, and project ideas for 2026. Content also addressed topics like tokenmaxxing's status and the importance of binary search. The newsletter provides regular educational material to help developers improve their skills efficiently. Examples of recent posts date from May 29 through June 17, 2026.
Joseph Krause of Radical AI contends that the primary advancement driver in materials science is not the AI model itself. He argues that material success depends on complex variables beyond mere chemical formulas, such as supply chains and manufacturing processes. The difficulty in achieving perfect, scalable material design with a single model was highlighted by past incidents like the LK99 controversy regarding reproducibility. To overcome this, Radical AI is building "self-driving labs." These labs integrate an "AI scientist" that autonomously combines scientific knowledge, computation, and intuition to generate and test hypotheses automatically.
Test-time compute (TTC) is the amount of computational power used by an AI model when generating a response after training. Recent advancements, exemplified by OpenAI's o1 model, prioritize deeper, slow reasoning over immediate output speed. Scaling TTC allows models to engage in detailed step-by-step processes, mimicking deliberate "System-2 thinking." The article discusses various open-source methods for effectively increasing this compute during inference. These scaling techniques are shown to significantly boost an AI model's overall reasoning capabilities.
Reinforcement Learning (RL), which involves agents learning through trial and error, has deep roots in psychology and mathematics. Early methods included Temporal-Difference (TD) learning, Monte Carlo techniques like REINFORCE, and actor-critic approaches. The field advanced significantly in the 2010s with Deep RL milestones like Google DeepMind's Alpha. Current advancements utilize RLHF and algorithms such as PPO and GRPO for modern applications, especially in LLMs. Future directions suggest that current LLM and RL capabilities will require further innovation beyond present methods.
Reasoning models utilize chain-of-thought to solve difficult problems and have sparked debate regarding their classification separate from general Large Language Models (LLMs). Proponents argue that specialized training methods, like reinforcement learning phases and increased compute, grant these models capabilities beyond standard LLMs. Examples include OpenAI's o1 and DeepSeek-R1, which demonstrate outperformance on complex tasks such as multi-step math problems. The discussion also addresses the model's limitations, including a potential for "overthinking." Overall, the text aims to define what constitutes a reasoning model and outline future AI possibilities in this domain.
A global search for dark matter is underway utilizing deeply underground laboratory sites. Massive detectors, specifically employing liquid xenon, are positioned beneath an Apennine massif, the Jinping Mountains in Sichuan, and a South Dakota mine. These instruments aim to achieve the first direct detection of dark matter particles. Dark matter is described as an invisible substance whose gravitational influence has shaped galaxies. The effort represents a significant effort in modern astrophysics and particle physics.
Watercolor washes blend across a digital canvas. Three-dimensional animated sea creatures are actively creating the artwork. These contributors include diverse marine life such as octopuses, a squid, and turtles. The process results in vibrant, colorful abstract paintings. The overall theme suggests an artistic collaboration from underwater inhabitants.
Amazon launched an intelligence stack featuring AWS Context, a new knowledge graph service that learns automatically from agent usage rather than manual curation. This service builds a self-learning graph by inferring relationships across existing enterprise datasets, business rules, and domain knowledge. Beyond the graph, AWS also announced S3 Annotations for rich storage-level context and skill assets in AWS Glue Data Catalog. The overall Context intelligence stack synthesizes these layers, enabling agents to query combined insights via semantic search and graph reasoning at runtime. This approach allows integration with existing AWS structures without requiring significant data movement or proprietary APIs.
As models process longer contexts, the Key-Value (KV) cache consumes more memory than the model weights themselves. Three techniques—TurboQuant, OSCAR, and EpiCache—address this escalating memory bottleneck. Instead of viewing these approaches as direct competitors, the article suggests they are more complementary solutions. Each method offers a different strategy for efficiently managing the large size of the KV cache during inference. Understanding their individual mechanisms helps optimize performance when deep context handling is required.
Vercel has released eve, an open-source AI agent framework under the Apache-2.0 license, which is currently in public preview. In this architecture, each individual agent functions as a directory containing files mapped to specific capabilities. The framework incorporates features like durable execution, built-in sandboxes, and approval mechanisms. Users can scaffold a new project using the command `npx eve@latest init`. Furthermore, deployed applications can remain unchanged after initialization through the `vercel deploy` process.
Research explores whether an artificial intelligence agent can independently execute the entire scientific method. The presented work focuses on developing generalist autonomous research capabilities through hypothesis-tree refinement. This methodology aims to allow AI systems to generate, test, and iteratively refine hypotheses without constant human oversight. By structuring investigation around a tree of potential theories, the agent can guide its own experimental design. Ultimately, the goal is to move toward more self-sufficient scientific discovery powered by advanced AI agents.
Explores the process of setting up image similarity search within the Milvus platform. It addresses the necessity of going beyond simple visual replication when performing such searches. Such vector-based image searching is a specialized technique in data retrieval. The guide provides practical steps for implementing this advanced search functionality. It highlights that mere visual matching does not guarantee accurate results.
ORPilot utilizes an Intermediate Representation (IR) to solve the critical need for reproducible and portable AI optimization modeling agents. Such representations are essential because production-level AI systems require consistent and transferable optimization capabilities across different environments. The IR serves as a standardized, abstract format that decouples the high-level optimization problem from its specific execution backend. This abstraction allows ORPilot's models to be optimized once and deployed reliably in various target platforms without significant rewriting. Consequently, the use of this structured IR ensures both portability across hardware and reproducibility of complex optimization results.
LLM applications often require a defined workflow rather than true autonomous agents. The content suggests building such workflows using standard Python code. This approach avoids the necessity of complex, dedicated agent frameworks for most use cases. It advocates for direct implementation to achieve necessary application logic. The original post was published on Towards Data Science.
An analysis details five distinct field families extracted by a question parser from user input strings. These fields are Keywords, Scope, Shape, Decomposition, and Clarification. The content explains the specific role of each family in understanding user intent. The parsing mechanism populates these five structures using underlying code logic. This information was originally published on Towards Data Science.
Researchers developed a new benchmarking suite to rigorously test open-source language models. This tooling evaluates whether large language models exhibit sufficient "agency"—the ability to take initiative and operate independently. The benchmark tests capabilities beyond simple question answering, focusing on complex, multi-step tasks requiring planning and tool use. Findings indicate significant variability in model performance across different agentic benchmarks. Overall, the research suggests a need for more robust evaluation methods to accurately gauge true autonomous capability in open models.
MolmoMotion is a novel framework designed for predicting human 3D movements guided by language descriptions. It leverages large language models to interpret complex textual instructions related to motion. The system then translates these linguistic prompts into accurate, physically plausible 3D skeletal trajectories. By integrating natural language understanding with generative modeling, MolmoMotion enhances the controllability of motion forecasting tasks. This allows users to guide virtual characters' actions simply by describing what they should do.
In game theory, evidence suggests generalist approaches can outperform specialized ones in specific gaming scenarios. Researchers have identified a particular category of algorithms with unexpectedly high performance levels. These findings challenge conventional assumptions about optimal strategic design within games. The research indicates that tailoring strategies narrowly may not yield the best results. Generalizing methods show significant advantages in certain competitive environments.
A guide details using llama.cpp to run the Qwen model on a Hugging Face platform. The process is optimized for small models, making deployment accessible. This approach allows users to leverage capable large language models without significant hardware overhead. Specific instructions are provided for implementation within a hackathon setting. Essentially, it offers a streamlined way to test advanced LLMs practically.
Developed a Q-learning agent from scratch on an Android phone using only pure Python, bypassing major libraries like TensorFlow or NumPy. The network architecture features 5 inputs, 8 hidden neurons (ReLU), and 3 outputs, utilizing a replay buffer and epsilon-greedy decay. After training with over 10,000 battles, the agent achieved notable success rates. Specifically, it attained a 92% win rate against both Exp3 and UCB1 strategies in games like Rock-Paper-Scissors. This demonstration proves that complex reinforcement learning models are feasible to implement on minimal hardware without cloud resources or external frameworks.
A proposal suggests creating a dedicated Discord server. This community aims to serve AI Agent Builders, MCP Developers, and Open-Source Contributors. The forum will act as a central hub for these specialized groups. It is intended to facilitate collaboration and knowledge sharing among practitioners in these fields. Details regarding the setup and purpose were submitted by user /u/Neither-Witness-6010.
Acquire over 100 expert ChatGPT hacks designed to boost workplace productivity across multiple departments. The "ChatGPT At Work Bundle" provides detailed model information, specific use cases, and actionable workflow tips for sales, marketing, management, and customer support. Included resources feature a full chatbot guide, 100 powerful prompts, a decision flowchart, and content refinement checklists. These tools aim to help users perfect their prompting skills and integrate AI strategically into professional tasks.
The brain processes massive amounts of sensory data—around 11 million bits per second—but only a small fraction reaches conscious awareness. Interoception is the term for sensing internal bodily signals, such as hunger pangs or changes in heartbeat. The study of interoception has recently gained significant attention following advancements and a 2021 Nobel Prize recognition. Researchers are now gaining clarity on how body signals communicate with the brain, informing our understanding of conditions like anxiety and chronic pain. This research challenges historical separations between emotional experience and cognitive thought processes.
MIT Technology Review Narrated provides high-quality journalism in an audio format weekly. Episodes feature ambitious stories on topics ranging from interoception to AI's impact on labor. Recent content has covered smart glasses development for warfare and the automation revolutionizing IVF. Other explored subjects include NASA's plans for nuclear reactor spacecraft to Mars. The platform promises sharp analysis and thought-provoking themes through professional voice narration.
New research published in "Nature" details the capabilities of a conversational medical AI system named AMIE. The study demonstrates that this AI is adept at supporting primary care physicians. Specifically, AMIE was shown to assist with complex disease management protocols. Its function mimics the expertise required for handling intricate patient health issues. This suggests the AI could significantly aid healthcare professionals in managing diverse conditions.
OpenAI has released LifeSciBench, a new benchmark designed to test frontier AI capabilities in real life-science research. This comprehensive evaluation incorporates 750 tasks structured across seven biological domains and seven distinct workflows. The benchmark was created with input from 173 PhD scientists and features 19,020 expert-written rubric criteria. Instead of assessing simple recall, LifeSciBench evaluates the reasoning and decision-making processes of AI models. Currently, the top-performing model, GPT-Rosalind, achieved a score of 36.1% on this rigorous test.
Demonstrates using NVIDIA SkillSpector to assess AI skills for security vulnerabilities prior to deployment. A corpus of both benign and vulnerable skills is constructed and scanned via SkillSpector's LangGraph workflow. Pandas is employed to organize the resulting risk scores and identified findings. The analysis outputs are exported in SARIF format, which allows registering a custom analyzer. An optional final step involves applying an LLM-based semantic pass for further evaluation.
Author suggests reviving the three-martini lunch for corporate settings. This mid-century ritual combined business discussions with enjoyment. In 1978, Gerald Ford humorously called it an "epitome of American efficiency." The author argues this blend is currently missing in modern workplaces due to developments like the AI boom. He posits that making this tradition a strategic plan should be reconsidered.
Experiencing severe calf pain for five days, the author initially assumed it was a simple muscle problem. A chiropractor treated the worsening symptoms accordingly, failing to diagnose the underlying issue. The situation highlights that seemingly minor ailments can mask serious medical conditions. While acknowledging AI is not a doctor's substitute and requires regulation, the piece suggests physician-AI collaboration could be beneficial. The personal experience underscores the danger of misdiagnosing potentially life-threatening issues like blood clots.
To remain irreplaceable in a changing professional landscape, focus on cultivating unique human attributes that artificial intelligence cannot replicate. Develop deep emotional intelligence and complex interpersonal skills, as these require nuanced human experience. Cultivate specialized knowledge in interdisciplinary fields, combining seemingly disparate areas of expertise. Prioritize creative problem-solving, involving novel thought processes rather than mere data processing. Ultimately, building an authentic, uniquely integrated skillset makes one invaluable beyond algorithmic capabilities.
In 2025, the economics of code production drastically changed, becoming instant and effectively free. Previously considered valuable resources requiring careful curation, lines of code became disposable and easily regenerated almost overnight. Charity Majors noted that this shift demands increased engineering discipline from developers interacting with AI tools. The rapid proliferation of generated code signifies a fundamental change in software development practices.
Determining a churn threshold is fundamentally a pricing decision rather than purely a data science one. Unit economics play a crucial role in correctly setting this classification cutoff. However, current practices often fail to properly integrate unit economic considerations into the modeling process. Successfully setting the churn threshold requires aligning technical models with core business financial metrics. Recognizing this connection shifts the focus from merely predicting churn to optimizing revenue strategy.
Researchers are developing and discussing the use of "sandboxes" as a defense mechanism against advanced AI models like Claude. These sandboxes aim to prevent sophisticated agents from cheating or exhibiting flawed performance when tested on standardized benchmarks. The concept involves creating controlled, isolated environments for evaluating AI behavior. Implementing such systems is crucial for maintaining the integrity and reliability of benchmark testing for large language models. This mitigation strategy seeks to validate model capabilities more accurately by restricting potential avenues for misconduct.
General Motors is dramatically cutting vehicle development cycles by integrating advanced Artificial Intelligence and comprehensive simulation tools into its processes. Former Tesla executive Sterling Anderson spearheads this effort, allowing engineers to test designs in virtual environments that previously required extensive physical prototyping. This approach permits simultaneous optimization of hardware and software across thousands of simulated scenarios before building a single prototype. Specific advancements include reducing crash simulations from 15 hours to under one minute using AI probability methods. Ultimately, these tools enable GM to accelerate development for vehicles like the GMC Hummer, improving safety and refining designs faster than traditional methods allowed.
NAVI-Orbital successfully executed a groundbreaking in-orbit demonstration of an autonomous, multi-modal inference system aboard a Low Earth Orbit spacecraft. The software utilizes a local vision-language model (Gemma 3) to classify captured scenes and generate descriptive text about their content and relationships. This allows the system to respond to operator follow-up questions using natural language dialogue, replacing traditional command sequences. Benchmarking demonstrated high accuracy (88.16%) on curated datasets, validated by live processing of unseen, previously uncorrected Earth imagery in orbit. The achievement proves the feasibility of running large foundation models directly on satellite hardware for semantic compression of observations onboard.
Presents CaVe-VLM-CoT, a modular agentic-RAG framework designed to mitigate hallucinations in Vision-Language Models by enforcing evidence-grounded reasoning. This system operates through a five-stage closed-loop pipeline: Extractor, Retriever, Solver, Citation Injector, and Verifier. When ungrounded claims are detected, structured feedback loops back to the Extractor for targeted re-retrieval. Researchers introduce 23 component-wise metrics culminating in the CaVeScore to jointly measure retrieval quality, citation faithfulness, and cross-modal grounding. Tested on ScienceQA and MMMU, CaVe-VLM-CoT achieved strong performance without any necessary architectural or prompt modifications.
Examining shared workspaces, this study investigates when integrating human collaborators improves AI agent performance on complex tasks. Using 1,482 sessions in a Collaborative Gym environment, researchers found that adding collaborators can actually decrease performance if team structure is insufficient for coordination. To address this, they evaluated scaffolding combining shared memory with simulated human-in-the-loop gates requiring approval for certain actions. This proposed scaffolding significantly improved mean performance, particularly notable within three-person teams. The findings emphasize that the method of human-AI expertise integration is crucial, potentially as important as raw capability itself.
Language model agents excel at short, isolated tasks but struggle with complex real-world challenges like managing long-term goals amid uncertainty. Researchers introduce CEO-Bench, an evaluation suite simulating operating a startup for 500 days through a Python interface. This task requires agents to manage budgeting, marketing, and pricing using noisy, interconnected business data. Success necessitates coordinating diverse decisions while analyzing complex historical information to form sound strategies. Currently, only advanced models like Claude Opus 4.8 and GPT-5.5 show limited success in maintaining significant capital within this challenging environment.
Introduces DeFAb, a new benchmark dataset derived from four decades of public knowledge bases to test defeasible abduction—hypothesizing explanations by overriding defaults while respecting expectations. This benchmark requires hypotheses to pass rigorous, polynomial-time checks for derivation validity, conservativity, and minimality. Evaluation shows that frontier language models struggle significantly, with rendering-robust accuracy dropping as low as 7.8–23.5%. The resource provides over 372,000 instances across three difficulty levels, complete with verifiable gold standards. Additional resources include DeFAb-Hard and CONJURE, which test transformative creativity using Lean 4/Mathlib instances.
Optimal lithium production decision-making is complex due to uncertainties in geology, market pricing, and demand. Existing models often fail to integrate fluctuating prices or varied extraction technologies like direct lithium extraction versus hard rock mining. This research frames the challenge as a Partially Observable Markov Decision Process (POMDP) for robust optimization. Employing belief state planning methods allows the model to dynamically adapt to various stochastic pricing regimes. The resulting framework achieves superior demand fulfillment and balanced economic/environmental outcomes by optimally sequencing exploration, production, and technology choices.
Introduces ForecastBench-Sim, a simulated-world forecasting benchmark developed using game rollouts from the turn-based strategy game Freeciv. Forecasters analyze a structured snapshot of a fixed game state and predict hidden future outcomes. Because it is simulated, the benchmark allows for diverse questioning types at various time horizons, including conditional/causal queries and rare event modeling. The system processes questions against an ongoing simulation to score accuracy immediately. ForecastBench-Sim aims to complement real-world benchmarks by providing controlled environments for studying probabilistic reasoning in dynamic states.
A formal account addresses what memory generalist agents require to perform near-optimally across various environments and goals. The research establishes a separation theorem showing that successful agents cannot rely solely on current state observations when domains share an observational bottleneck with incompatible optimal actions. Consequently, such agents must retain domain-relevant information in memory. Furthermore, the paper demonstrates that sufficient memory allows for approximating the agent's local transition dynamics for related goals. These findings characterize memory as crucial for supporting domain disambiguation, transition model reconstruction, and planning capabilities.
Robot soccer presents a complex testbed for multi-agent reinforcement learning due to partial observability, mixed cooperative/adversarial interactions, sparse rewards, and long-term tactical demands. Current platforms like RCSS2D have an architecture difficult to integrate with modern Python-based MARL workflows. R2D-RL solves this by connecting the mature RCSS2D platform and HELIOS player clients to a Python MARL interface using shared memory. This new environment supports full-field and scenario training, offering configurable opponents and various action spaces. Furthermore, R2D-RL incorporates reward shaping via Expected Possession Value (EPV) and allows for parallel execution, providing both specific scenarios and an 11-vs-11 benchmark.
ProfiLLM addresses the challenge of integrating complex behavioral signals from massive ride-hailing logs into low-latency dispatch systems using LLMs. It tackles constraints related to context window size, data sparsity in long-tail users, and ensuring profile utility for prediction. The system features two modules: Tool-Augmented Global Knowledge Mining, which uses 27 tools to extract global knowledge, and Utility-Aligned Profile Exploration, which iteratively refines candidate profiles using a lightweight utility proxy. Deployed at DiDi, ProfiLLM significantly improved outcome prediction (up to +6.14% AUC) and demonstrated gains in live A/B tests for GMV and completion rates.
Gaussian Mixture Attention (GMA) addresses the quadratic complexity of standard dot-product attention in long contexts. It replaces pairwise query-key comparison by routing interactions through $K$ learned Gaussian mixture components within a shared latent space. Queries and keys generate posterior responsibility vectors, whose overlap defines an implicit affinity. This design allows computation by leveraging two smaller matrices scaling as $\mathcal{O}(NK)$, avoiding the full $\mathcal{O}(N^2)$ affinity matrix. Empirically, GMA achieves linear memory scaling with respect to sequence length $N$ and performs competitively on long-context tasks, offering a probabilistic alternative.
Proposes CodeBlock, a structure-aware sparse supervision framework for fine-tuning code LLMs, addressing the limitation of token-level masking in code. Instead of selecting isolated tokens, CodeBlock partitions responses into syntactically complete coding items. It estimates utility by aggregating generalized cross-entropy over core logic tokens and further reranks blocks using data-flow reach and bridge signals. During training, loss is applied only to these selected code blocks and informative natural language tokens, while the full response remains in context. Evaluation on six benchmarks shows CodeBlock achieves superior average pass@1 compared to full-token SFT while utilizing significantly fewer supervised tokens (1.9%).
Multimodal neuroimaging using fMRI and DTI allows graph neural networks to analyze brain connectivity but is systematically confounded by demographic factors like age and sex. Existing causal GNN methods fail because they do not account for these specific, real-world clinical confounders or the region-specific sensitivity of brain atlases. Artemis proposes a novel, region-level causal framework that addresses this gap through independent causal intervention at each distinct brain region. This method learns lightweight, region-specific confounder representations while utilizing multimodal features for graph reasoning as an adaptable module. Testing on ADNI, OASIS, and HCP benchmarks showed Artemis significantly improves upon existing GNN baselines for tasks like diagnosis and classification.
A mathematical link is established between shock-wave theory and the symmetry-quotiented dynamics of stochastic gradient descent using differential geometry and Lie group theory. After applying local-entropy coarse-graining to parameter symmetries, the resulting effective dynamics conform to a viscous Hamilton--Jacobi equation on the quotient manifold. Furthermore, the loss function's gradient obeys a Burgers-type equation, allowing for rigorous establishment of shock formation. This framework has been applied to models including Transformers and CNNs, showing they follow these specific equations. The authors suggest using symmetry-corrected observables as a principled method for monitoring and controlling deep learning training transitions.
Mixture-of-Experts (MoE) models face deployment challenges due to large memory footprints and inference overhead. Prior compression techniques often prune entire experts, failing to capture fine-grained redundancy within them. The proposed structural pruning framework targets this by modeling prune-ratio allocation as a channel-score coverage maximization problem solved via attribution approximation. Testing on DeepSeek and Qwen MoE models showed the method maintains accuracy under 50% or 25% structured pruning when combined with 4-bit quantization. Specifically, it achieved a 5.27$\times$ memory footprint reduction on Qwen3-30B-A3B while surpassing existing state-of-the-art results.
Gaussian width measures set complexity using ambient Euclidean geometry, but this fails on statistically curved manifolds whose natural structure is Riemannian. The paper introduces Fisher width, a geometric measure that adapts Gaussian width by incorporating the local Fisher information metric tensor $G(\theta)^{1/2}$. This results in a quantity sensitive to local statistical curvature and invariant under smooth reparameterizations. Fisher width maintains key properties like concentration bounds while capturing anisotropic effects missed by Euclidean measures. This framework provides foundational tools for analyzing model complexity and learning theory on curved statistical manifolds.
Optimizing Supervised Fine-Tuning (SFT) data distribution is crucial for improving Large Language Model (LLM) capability upper bounds beyond what current methods can achieve. The paper introduces DRIFT to address shortcomings in standard Influence Function (IF) formulations used for data attribution. DRIFT overcomes proximity and bias issues by utilizing the model's own on-policy rollouts as validation targets instead of external reference data. It further employs signed weighting based on trajectory correctness to ensure reliable influence scoring across the dataset. Experiments confirm that DRIFT consistently raises performance ceilings on instruction and reasoning models compared to existing baselines.
Safe coordination in complex systems necessitates managing discrete/continuous actions, safety constraints, and physics dynamics, forming a difficult coupling bias. TRIDENT introduces a novel Multi-Agent Reinforcement Learning framework designed to address this coupled challenge simultaneously. It integrates three co-designed components: a gradient correction method, a Lyapunov-constrained update for per-iterate feasibility, and a physics-informed residual critic. The authors prove that TRIDENT converges to a constrained Nash equilibrium with an $O(1/\sqrt{K})$ rate while maintaining an $O(\sqrt{K})$ cumulative-violation bound. Empirical testing on multi-UAV systems demonstrated significant reductions in training violations (up to 95.5%) compared to existing methods like MADDPG and MACPO.
Achieving LLM unlearning requires balancing undesirable knowledge removal with preserving desired capabilities, an area marked by inherent tradeoffs in current methods. Researchers found that retention activation bias can quantify damage without needing to know the specific unlearning implementation. Proposing SAGE (Spectral Activation-Geometry Sanitization), a post-hoc approach corrects the final unlearning update vector. SAGE operates by collecting inputs from a small proxy model and determining their dominant activation geometry. This allows it to suppress unwanted updates aligned with retained knowledge while preserving necessary forgetting information across various methods and scales.
A multimodal graph neural network system was developed to forecast precipitation rates over a two-hour period across the Nordic radar domain using various data sources. The study systematically ablated different inputs—including historical radar, MEPS predictions, Netatmo observations, and satellite channels—to assess their impact. Findings indicated that each data source improves distinct aspects of the forecast; for instance, MEPS stabilizes extrapolation while Netatmo helps with local onset detection. While point observations are not uninformative, their benefit depends significantly on the model's training loss and uncertainty handling. Combining satellite inputs with CRPS-based losses yielded the best overall performance.
Modern AI agents pose a hidden risk because an acceptable final answer may mask errors in earlier steps like weak retrieval or incorrect tool outputs. ToolChain-CRC proposes a conformal risk-control method to address this for retrieval-augmented and tool-using agents experiencing drift. The system evaluates entire agent runs as trajectories, generating step-level scores combined into a total trajectory risk score. It establishes an accept/intervene rule calibrated against this trajectory risk and includes an anytime alarm to halt dangerous processes early. Experiments confirm that trajectory-level calibration successfully maintains acceptable risk levels where simply checking the final answer fails due to intermediate failures.
Geometric representation learning for data often uses dual encoders mapping items into a shared embedding space, which is extended to handle hierarchical retrieval based on DAG ancestor-descendant relationships. Existing methods struggle with deep hierarchies, requiring embeddings whose dimension scales with the total number of nodes. This paper introduces compact reachability embeddings that offer theoretical guarantees dependent only on graph structure parameters. Specifically, it proves a constant dimension (three) embedding exists for any directed tree, regardless of size. For general graphs of treewidth $t$, the authors construct embeddings in $O(t \log n)$ dimensions while establishing necessary lower bounds.
U-calibration studies online forecasting algorithms for unknown downstream agents, traditionally guaranteeing sublinear regret for all proper loss functions. Existing algorithms perform well in worst-case scenarios but fail by incurring high regret ($\Omega(\sqrt{T})$) even for smooth losses where $O(\log T)$ is achievable. The presented work overcomes this limitation by designing a single forecast algorithm. This new method simultaneously achieves $\tilde O(\sqrt{T})$ regret for all bounded proper losses and optimal $O(\log T)$ regret for all bounded smooth proper losses. The technique builds on a novel variant of Follow-the-Perturbed-Leader using self-concordant noise applied directly in the prediction space.
Developed a statistical theory for optimizing policies in offline reinforcement learning using only trajectory-level outcomes instead of typical process-level rewards. Introduced OPAC, a pessimistic actor-critic algorithm that uses latent reward models to learn from single scalar labels per trajectory. Proved performance guarantees and sharp lower bounds for replacing expected cumulative rewards with one label per trajectory. Extended the framework to handle preference-based feedback while maintaining key complexity dependencies. Identified structural coefficients defining a tractable regime where generalized outcome-based RL remains sample-efficient.
Introduces a transfer learning framework designed to adapt structural fragility modeling under challenging conditions like domain shift and class imbalance while maintaining engineering interpretability. The methodology demonstrates four strategies: instance-based, parameter-based, hierarchical Bayesian, and multi-source. Case studies applied this framework to coastal bridge (Hurricane Katrina), residential building (Hurricane Ian), and seismic bridge (Nisqually earthquake) fragility assessments. Results show that direct transfer of existing models fails when faced with domain shift or severe class imbalance. Targeted adaptation significantly enhances failure detection and predictive stability, emphasizing the need for guidelines on diagnostics and uncertainty reporting.
Data valuation is crucial for principled curation of time series in domains like healthcare and finance, but current methods are often model-dependent or fail to capture temporal dependencies. TimeLAVA introduces a learning-agnostic framework designed to value temporal segments by measuring their marginal contribution to distributional discrepancy against reference data. The core methodology involves a novel Selective Wavelet-based Wasserstein discrepancy, which combines multi-scale wavelet transforms with unbalanced optimal transport for robustness. Segment values are efficiently calculated using sensitivity analysis without requiring model training and yield point-wise scores. Experiments across anomaly detection, data pruning, and label noise consistently show TimeLAVA produces more informative value scores than existing techniques on diverse real-world datasets.
Presents a unified geometric representation, termed Kernel of Partition Paths (KPP), that indexes a feature map by nodes within a forest structure. This embedding uses a path metric to create coordinates forming a squared-Euclidean path-isometric space. The KPP framework unifies four key components: prediction capability, exact additive attribution, deterministic Lipschitz robust radius, and uniform Rademacher risk bounds for both regression and classification. All stated probabilistic guarantees are conditional upon the representation and are provided under three distinct conditioning regimes. Notably, the robust-radius guarantee is shown to be deterministic within the KPP metric, rather than solely dependent on the raw input norm.
FOSC-X is a new framework designed to extract multiple optimal "top-M" flat clusterings from the local, non-horizontal cuts of a hierarchical cluster tree. It addresses limitations in existing methods that typically find only a single best solution. The framework can solve this problem efficiently using dynamic programming, achieving polynomial time complexity with respect to nodes and dataset size. A key feature is its ability to incorporate optional constraints on the desired number of clusters. FOSC-X ensures optimal ranking of alternative clusterings, which can reveal structures missed by simpler extraction techniques.
Testing conditional independence is difficult because general methods struggle with controlling Type I errors without strong assumptions. Existing sequential tests are fragile when the required conditional distribution must be estimated rather than known exactly (the "Model-X" paradigm). A new proposed approach enhances robustness to estimation error by using testing-by-betting on an adaptively optimized Kernel Conditional Independence statistic. This method incorporates a normalization scheme and a truncate-and-shift calibration strategy for improvement. These combined modifications significantly reduce Type I error inflation while maintaining high power across various benchmarks.