Solution space path planning for supporting en-route air traffic control
A new conflict-free path-planning approach for en-route ATC is built around “solution-space path planning” so controllers can see all feasible safe actions while shifting operational priorities like separation, maneuver limits, and routing practicality. The method integrates multiple intent-based conflict detection schemes and introduces node variants optimized for speed and solution quality, reporting ~3.69 ms average path computation in operationally relevant scenarios.
Constructive Alignment: Governing Preference Dynamics in Human-AI Interaction
This work challenges the idea that alignment is about locking onto fixed human preferences, arguing instead that preferences evolve through interaction as AI systems shape attention and values over time. It proposes a control-theoretic “Constructive Alignment” framework that treats alignment as managing value-trajectory formation – aiming for coherence, epistemic grounding, and bounded resistance to manipulation under uncertainty.
Constructing Epistemic AI Literacy: Detecting Epistemic Aims and Processes in Student-AI Co-Programming
The paper defines Epistemic AI Literacy as a process-oriented learning phenomenon, then operationalizes it by identifying observable epistemic aims and strategies in human-AI co-programming dialogues. Across a large dataset, most student interactions skew toward outsourcing and less reliable verification behaviors, with only a small share showing mastery-oriented aims combined with epistemic justification and stronger verification practices.
The MMM Data Model — A Normative Specification for Knowledge Interoperability in a Decentralisable Knowledge Commons
MMM proposes a portable knowledge documentation data model designed to replace rigid, document-centric structures with a normalized, interoperable representation that still allows free-text labels. The model is constrained by a small set of normative rules to support reuse and cross-system exchange without forcing semantic convergence, backed by a reference implementation and pilot deployment evidence.
A Contextual-Bandit Oversight Game with Two-Sided Informational Asymmetry
This research models real oversight situations where humans know the reward but the AI knows the quality of its proposed action, creating asymmetric information that can block safe shutdown. Using a contextual-bandit oversight game built on Cooperative Inverse Reinforcement Learning ideas, the paper identifies a “region of avoidable harm” tied to non-credible oversight communication and analyzes how passive learning and signaling can help over repeated rounds.
RareDxR1: Autonomous Medical Reasoning for Rare Disease Diagnosis Beyond Human Annotation
RareDxR1 targets open-domain rare disease diagnosis directly from unstructured clinical notes by prioritizing reasoning-centric training rather than pipeline steps like ontology-based phenotype extraction. It combines knowledge internalization and evolutionary learning plus a reflection-enhanced sampling strategy to learn from failures without requiring expert annotation, reporting state-of-the-art results across benchmarks.
From Signals to Structure: How Memory Architecture Drives Language Emergence in LLM Agents
This study explores how two-agent language coordination emerges under different memory architectures, showing that memory design can outweigh raw channel capacity. Agents with persistent private “notebooks” coordinate more reliably, while stateless agents degrade as vocab growth outpaces what a rolling context window can track – and a theoretical bottleneck hypothesis finds a fragility point rather than a simple optimum.
OpenAI floats giving Trump administration 5 percent cut of AI boom
Reporting says OpenAI has discussed offering the US government a 5 percent stake to reduce political friction and manage backlash, with CEO Sam Altman allegedly pitching the idea and a value tied to OpenAI’s reported funding round valuation. The move is framed as a way to share upside while easing tensions with the Trump administration.
Making Failure Safe: A Constrained, Verifiable Agent Framework for Open-Web Data Collection
Instead of letting an LLM generate brittle web-scraping code, this framework converts requirements into typed JSON collector configurations with constraints and verification steps aimed at preventing dependency errors and schema mismatches. By executing deterministic, structured Airflow DAGs with rule-based quality checking, the system reports zero execution-stage LLM tokens and faster wall-clock performance on many source-verified collection tasks.
Achieving operational excellence with AI
This piece focuses on how AI is being adopted as an operational layer in industries where continuity, infrastructure constraints, and safety matter more than consumer features. It highlights the practical shift from experimentation to operational excellence – emphasizing governance, monitoring, and reliability engineering as companies deploy AI into real workflows.
Teaching AI to run with the turbines
The article looks at the engineering challenge of using AI in environments where timing, physical constraints, and continuous control dominate – “running with the turbines” rather than doing standalone prediction. It underscores that deploying AI at the edge of operations requires robust control integration and safety-aware design, not just model accuracy.
Seed2.0 Model Card: Towards Intelligence Frontier for Real-World Complexity
Seed2.0 positions itself as a step toward handling complex, long-horizon real-world tasks by building evaluation around users’ needs and realistic scenario complexity. The model card claims improvements on long-tail knowledge and complex instruction following, supported by broad real-world use cases spanning reasoning, visual understanding, and search.
Bounded Morality: Defining the Space of Moral Computation
Bounded Morality reframes ethical reasoning as a constrained computational problem for finite agents, defining moral “breadth” and “depth” and mapping the feasible space of moral computation under limited resources. It argues ethical theories can be seen as locally efficient strategies in different demand regimes, and suggests that scaling and allocating moral reasoning capacity may be central to moral alignment.
Seed2.0 Model Card: Towards Intelligence Frontier for Real-World Complexity
Seed2.0 positions itself as a step toward handling complex, long-horizon real-world tasks by building evaluation around users’ needs and realistic scenario complexity. The model card claims improvements on long-tail knowledge and complex instruction following, supported by broad real-world use cases spanning reasoning, visual understanding, and search.