The Morning Preprint

Monday 12 October

Showing the most recent batch — Monday 12 October. No papers have been ingested today yet.

893 papers ingested by the estate's stations that day — about 9 minutes to read. The counts and the chart are the estate's own ingest, not arXiv's submission figures.

This day's harvest looks short: 893 papers against a busiest day of 2027 in the week shown below. Weekends and a broken fetch look the same in the chart, and the chart counts the estate's ingest, not arXiv's.

The 60-second version

Five things from this day's batch, best first — everything below the fold expands on them.

  1. New AI framework tests autonomous agent risk

    The AutoControl Arena framework evaluates frontier AI risk by separating logic from generation, showing capable models can exhibit severe alignment illusion under stress. arXiv 2603.07427 arXiv 2609.11405 arXiv 2603.18101 arXiv 2603.06144

  2. Autonomous vehicle safety framework shows collision risk

    A new component-agnostic safety evaluation framework simulates eighteen thousand three hundred fifty six scenarios, revealing that certain perception failures significantly increase collision odds. arXiv 2510.07749

  3. Astrophysics links far-infrared spectra to structure

    Work on WASP-39b shows that the dependence of far-infrared polarization spectra on wavelength is more tied to column density than dust temperature. arXiv 2607.18212

  4. Genomics reveals no global blood reference patterns

    Similarity analysis across continents found no reproducible population or geography-linked structure in blood count reference intervals, suggesting personalized diagnostic thresholds. arXiv 2510.13154 arXiv 2606.04004 arXiv 2608.15336 arXiv 2508.16558

  5. Quantum methods stabilize complex physics schemes

    A remedy for misleading convergence in self-consistent diagrammatic approaches was found by redefining iterative procedures based on Jacobian eigendirections. arXiv 2609.10289 arXiv 2507.21209 arXiv 2609.11405 arXiv 2607.18486 arXiv 2512.10335

Across the stations

AI Radio

364 papers today

Today's work on AI Radio centers on testing the limits of autonomous agents and improving how we understand their reasoning. The AutoControl Arena framework is particularly important because it uses an automated system to evaluate frontier AI risk by separating logic from generation, showing that capable models can exhibit severe alignment illusion under stress.

Other papers contribute to this by exploring different aspects of AI reliability. Some focus on learning safe exploration in reinforcement learning through the Sampling-Based Safe Reinforcement Learning method, while others tackle challenges like X-hacking to detect manipulated explanations and how language models suppress cultural knowledge during safety training.

Overall, the day highlights a tension between building powerful systems and ensuring their safety and interpretability. The central question left open is whether current methods for testing emergent risks are sufficient when models become highly autonomous agents.

The day on AI Radio

Astrophysics Radio

76 papers today

Today's work on Astrophysics Radio centers on understanding the physics of various astrophysical systems. The paper Limb Asymmetries on WASP-39b: A Multi-GCM Comparison of Chemistry, Clouds, and Hazes investigates how temperature differences, clouds, and hazes cause observed asymmetries in a hot Jupiter's limb spectra.

Other papers contribute by exploring diverse areas. Some studies look at the dynamics of cosmic structure formation using spinodal decomposition models to describe the large-scale structure of the universe. Others focus on observational techniques, such as using a quadratic maximum likelihood power spectrum estimator for kinetic Sunyaev-Zel'dovich velocity reconstruction.

The day leaves open the question of how complex, multi-factor processes combine to explain observed physical anomalies across different scales.

The day on Astrophysics Radio

Genomics Radio

19 papers today

The day's work focused on examining how biological structures relate to external contexts, with papers like Similarity Analysis of Blood Count Reference Intervals Across Continents Reveals No Reproducible Population or Geography-Linked Structure and Supports Personalised Values showing that global blood reference intervals lack consistent geographical patterns. This suggests moving toward personalized diagnostic thresholds based on individual baselines rather than universal standards.

Other papers contributed different perspectives; Cultural evolution shaped human's uniquely sensitive disgust avoidance explored how cultural transmission might shape affective traits, while RetroTAC introduced a machine-learning tool to predict the synthetic accessibility of PROTAC molecules by using route-derived scores. TopoLIA advanced drug discovery by using topology to create attention mechanisms for protein-ligand interactions, and Robust Machine Learning for Regulatory Sequence Modeling tested model performance under biological distribution shifts.

Taken together, these works question fixed biological or structural assumptions; while some studies look at molecular similarity and others at ecological community dynamics, a central unanswered question remains about how complex systems maintain stability amidst disorder.

The day on Genomics Radio

Quantum Radio

169 papers today

The day’s work focused heavily on stabilizing complex theoretical methods used in many-body physics. The paper Instabilities in self-consistent diagrammatic approaches and how to cure them demonstrated a remedy for misleading convergence by redefining iterative procedures based on Jacobian eigendirections. This work suggests that the physical fixed point of these schemes can be stabilized across challenging regimes.

Other papers contributed various aspects of quantum dynamics and material properties. For instance, Field-tuned quasiparticles and electronic structure in heavy-fermion YbNi4P2 explored how magnetic fields change electronic structure, showing a decrease in quasiparticle mass and suggesting abrupt changes near a putative Lifshitz transition. Lee--Yang Zeros and Pseudocritical Drift in J-Q N'eel Antiferromagnet--Valence-Bond Solid Transitions analyzed quantum phase transitions using Lee--Yang zeros, finding systematic drift consistent with an extended pseudocritical regime.

The collection of papers touches on fundamental limits and specific physical systems. The paper Beyond Robertson-Schr"odinger: A General Uncertainty Relation Unveiling Hidden Noncommutative Trade-offs establishes state-independent uncertainty bounds based on symmetry structure for multipartite systems. This contrasts with work on quantum information, such as Quantum Approximate Optimization for Decoding of Low-Density Parity-Check Codes, which uses QAOA to improve LDPC code recovery over Belief Propagation.

The day on Quantum Radio

Robotics Radio

126 papers today

The day's work on Robotics Radio focused heavily on safety and perception challenges. The paper Injecting Hallucinations in Autonomous Vehicles: A Component-Agnostic Safety Evaluation Framework introduced a framework that studies how false perceptions, or hallucinations, distort autonomous vehicle decision-making. This work involves running eighteen thousand three hundred fifty six simulations to show that certain perception failures significantly increase collision odds.

Other papers contributed diverse ideas across different domains. BrickCraft details a compositional framework for learning complex assembly skills using relative formulations and situated manuals for long-horizon tasks. Recovery-Driven Synthesis of Relational Concepts (ReSYNC) shows how robots can acquire abstract knowledge from failure experiences to avoid future problems. Furthermore, CoRE-VLA proposes a framework that lets vision-language models gracefully degrade when sensors are missing by routing action generation through modality-specialized experts.

Taken together, these papers address perception robustness in vehicles, skill composition in manipulation, and learning from failures for general robot intelligence. The day leaves open the question of how to create truly resilient embodied agents that can generalize skills learned from diverse sources while maintaining safety across various operational contexts.

The day on Robotics Radio

Security Radio

38 papers today

Today's work on Security Radio focused on advancing how we understand execution paths within software and LLMs. The paper Beyond Edge Coverage: Per-Task Data-Flow Extraction at Kernel Function Boundaries via LLVM introduces a framework that gives fuzzers state-aware feedback by extracting data flow from kernel functions, which is crucial for better testing guidance.

Other papers contributed different areas of concern. A comprehensive survey of Website Fingerprinting Attacks and Defenses in Tor systematically categorizes research into datasets, attack models, and defenses to map out existing solutions. LLMail-Inject provides a large benchmark for prompt injection attacks against LLMs that captures real-world adversary diversity better than previous small sets.

The day leaves open the question of how to reliably secure complex agent interactions. APEX-Harm evaluates prompt injection in shared workspaces where agents use tools, showing that simply completing a task does not guarantee safe behavior. This highlights the need for deeper security analysis beyond simple input checks when agents operate in shared environments.

The day on Security Radio

The day, in brief

AI Radio developed a Koopman operator approach for controlling a two-spool turbofan engine that resulted in improved thrust response when managing the engine pressure ratio. Astrophysics Radio observed that the dependence of far-infrared polarization spectra on wavelength is more strongly tied to column density than to dust temperature, suggesting that how finely structures are resolved matters significantly in what is measured. Genomics Radio found that cultural evolution shaped human disgust avoidance because culture transmits affective mechanisms like disgust sensitivity much faster than previously thought. Quantum Radio demonstrated that the rank of a Hamiltonian supports a maximum Clifford setting count related to two N plus one for all nonidentity Pauli strings on n qubits. Robotics Radio showed that the average classification accuracy for classifying five different gestures using signals from three muscles was ninety-four point eight percent when analyzing time domain electromyography signals.

The week in the estate's ingest

Seven days of papers the estate ingested. Weekends are striped — the weekend is visible in the data.

2027Tue
1143Wed
1046Thu
1118Sat
893Mon

Show the answer

893 papers were ingested on Monday 12 October.

893 papers, three ways

Where today's batch landed

By the paper's primary arXiv category.
cs.CV172
quant-ph105
cs.RO97
cs.CL92
cs.LG46
cs.CR38
stat.ML30
eess.SY29

The index

The day's strongest 60, by score. Every title links its paper here and on arXiv.

  1. Agents' Last Exam [arXiv] [episode]
  2. Contrasting - and K-Valley Moir'e Physics in Twisted Monolayer/Bilayer WSe 2 [arXiv] [episode]
  3. The Alignment Veto: How Safety Training Suppresses Cultural Knowledge in LLMs [arXiv] [episode]
  4. On the estimation and validity of AI time horizons---a statistical look at the METR plot [arXiv] [episode]
  5. Instabilities in self-consistent diagrammatic approaches and how to cure them [arXiv] [episode]
  6. On the Limits of Quantum Multiparty Simultaneous Communication [arXiv] [episode]
  7. Avatar-Forever: Decoupled Parallel Training for High-Quality Real-Time Infinite Avatars [arXiv] [episode]
  8. Certified Bounds on the Measurement Cost of Molecular Hamiltonians [arXiv] [episode]
  9. The syntax of wh-agreement in Yemeni Ibbi Arabic [arXiv] [episode]
  10. VLA-FAIL: Efficient Task Failure Detection for Finetuned Vision-Language-Action Models [arXiv] [episode]
  11. Beyond Edge Coverage: Per-Task Data-Flow Extraction at Kernel Function Boundaries via LLVM [arXiv] [episode]
  12. BrickCraft: Visuomotor Skill Composition with Situated Manual Guidance for Long-Horizon Interlocking Brick Assembly [arXiv] [episode]
  13. Playing the network backward: A Game Theoretic Attribution Framework [arXiv] [episode]
  14. AnomalyAgent: Agentic Industrial Anomaly Synthesis via Tool-Augmented Reinforcement Learning [arXiv] [episode]
  15. Rethinking Entropy Allocation in LLM-based ASR: Understanding the Dynamics between Speech Encoders and LLMs [arXiv] [episode]
  16. Training-Only Heterogeneous Image-Patch-Text Graph Supervision for Advancing Few-Shot Learning Adapters [arXiv] [episode]
  17. AutoControl Arena: Synthesizing Executable Test Environments for Frontier AI Risk Evaluation [arXiv] [episode]
  18. Nonlinear dynamics in magnonic Fabry-P' e rot resonators: Low-power neuron-like activation and transmission suppression [arXiv] [episode]
  19. Scalable and telecom single-erbium system with record-long room-temperature quantum coherence [arXiv] [episode]
  20. Cascade of topological phase transitions and revival of topological zero modes in imperfect double helical liquids [arXiv] [episode]
  21. Single-Shuffle Full-Open Card-Based Secure Computation Protocols for Any Function [arXiv] [episode]
  22. A Comprehensive Survey of Website Fingerprinting Attacks and Defenses in Tor: Advances and Open Challenges [arXiv] [episode]
  23. Injecting Hallucinations in Autonomous Vehicles: A Component-Agnostic Safety Evaluation Framework [arXiv] [episode]
  24. Reducing Circuit Depth in Lindblad Simulation via Step-Size Extrapolation [arXiv] [episode]
  25. KSZ Velocity Reconstruction with ACT and DESI-LS using a Tomographic QML Power Spectrum Estimator [arXiv] [episode]
  26. Breaking the Cloak! Unveiling Chinese Cloaked Toxicity with Homophone Graph and Toxic Lexicon [arXiv] [episode]
  27. OARelatedWork: A Large-Scale Dataset of Related Work Sections with Full-texts from Open Access Sources [arXiv] [episode]
  28. Loop-Mamba: A Loop Mamba with Degradation-Aware and Shared Memory for Old Photo Restoration [arXiv] [episode]
  29. DC Conductance of X-shaped Majorana Interferometer reveals Non-Abelian Anyon Statistics [arXiv] [episode]
  30. Recover, Discover, Plan: Learning Skills and Concepts from Robot Failures [arXiv] [episode]
  31. Unsupervised Identification and Removal of Spurious Correlations During Fine-Tuning [arXiv] [episode]
  32. Oxygenation and spatial heterogeneity shape radiotherapy protocol ranking through phenotypic adaptation [arXiv] [episode]
  33. TokenSwap: Benchmarking and Reducing the Modality Gap in Multimodal LLMs [arXiv] [episode]
  34. How AI Writing Assistance Distorts Perceptions of AI Users and Their Opinions [arXiv] [episode]
  35. Electric field switching of chiral phonons [arXiv] [episode]
  36. Liouvillian interpolation of the self-energy of cluster dynamical mean-field theories [arXiv] [episode]
  37. Query-Guided Spatial-Temporal-Frequency Interaction for Music Audio-Visual Question Answering [arXiv] [episode]
  38. TexTailor: Inference-Time Textual Guidance Tailoring for Multimodal Diffusion Transformers [arXiv] [episode]
  39. Managing Self-Learning Experts under Per-Round Budget Constraints [arXiv] [episode]
  40. Anomalies of global symmetries on the lattice [arXiv] [episode]
  41. Euclidean-Monte-Carlo-informed ground-state preparation for quantum simulation of scalar field theory [arXiv] [episode]
  42. Understanding Transformer-based Vision Models through Inversion [arXiv] [episode]
  43. Manifold4D: Denoising on Point Cloud Rendered Manifolds for Video Re-shooting [arXiv] [episode]
  44. SAGE-OR: Semi-supervised Adaptive Scene Graph Generation for Operating Rooms [arXiv] [episode]
  45. Low, High, and Semantic-Level Vision-Language Learning for AI-Generated Image Detection [arXiv] [episode]
  46. GraphVerse: A Comprehensive Visual Graph Reasoning Benchmark for Multimodal Large Language Models [arXiv] [episode]
  47. Equivariant Continuous Normalizing Flows with Offline Sampling for Fermionic Ground State Estimation [arXiv] [episode]
  48. CoRE-VLA: Towards Scalable and Robust Vision-Language-Action Modeling via Conditional Routing of Experts [arXiv] [episode]
  49. VITAL: Visual-Semantic Dual Supervision for Enhanced and Interpretable Latent Reasoning in Medical MLLMs [arXiv] [episode]
  50. SURF: Steering the Scalarization Weight to Uniformly Traverse the Pareto Front [arXiv] [episode]
  51. Decoding Alignment without Encoding Alignment: A critique of similarity analysis in neuroscience [arXiv] [episode]
  52. Quantum Vision Theory Applied to Audio Classification for Deepfake Speech Detection [arXiv] [episode]
  53. MotionRFT: Unified Reinforcement Fine-Tuning for Text-to-Motion Generation [arXiv] [episode]
  54. Weight Clustering Reveals the Importance of Relative Rank in Large Language Models [arXiv] [episode]
  55. TGPO: Trace-Guided Policy Optimization for Robot Task Planning via Verifiable Subgoal Generation [arXiv] [episode]
  56. Assessment of Generative Named Entity Recognition in the Era of Large Language Models [arXiv] [episode]
  57. ABLEIST: Intersectional Disability Bias in LLM-Generated Hiring Scenarios [arXiv] [episode]
  58. Scalar nonstandard neutrino interactions in galactic supernovae [arXiv] [episode]
  59. LLMail-Inject: A Dataset from a Realistic Adaptive Prompt Injection Challenge [arXiv] [episode]
  60. UniConFlow: A Unified Constrained Flow-Matching Framework for Certified Motion Planning [arXiv] [episode]

Listen

Everything the estate published on 2026-10-12, in the player. Playing from this page is how the edition learns whether anyone did.

Previous editions