The Morning Preprint

Thursday 15 October

Showing the most recent batch — Thursday 15 October. No papers have been ingested today yet.

1425 papers ingested by the estate's stations that day — about 10 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: 1425 papers against a busiest day of 8455 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. Draft-Thinking cuts reasoning cost by 82.6 percent

    A new framework for draft chain-of-thought reduced response length on MATH500 by 82.6 percent while keeping accuracy high.

  2. JWST maps star formation from redshift zero to nine

    A study combining COSMOS and GOODS-N data reveals galaxy luminosity increases toward cosmic noon, linking it to halo structure.

  3. Diffusion models twice as efficient for longer sequences

    Comparing diffusion models against autoregressive ones showed they are about twice as efficient for longer sequences when using KV-cache.

  4. G cubed VLA structures vision-language models

    Introducing G cubed VLA provides a geometric inductive bias to better structure how vision-language-action models interpret their surroundings.

  5. Quantum systems show bulk superconductivity at ambient pressure

    Research realized bulk superconductivity in quintuple-layer square-planar nickelate La6Ni5O12, exhibiting a superconducting volume fraction of twenty-five percent.

Across the stations

AI Radio

734 papers today

The day's focus on AI Radio is heavily centered on mitigating forgetting in continual learning and improving reasoning efficiency. The work Draft-Thinking investigates a lightweight framework for learning model-adapted draft chain-of-thought that significantly reduces reasoning cost, achieving an 82.6% reduction in response length on MATH500 while maintaining high accuracy Draft-Thinking.

Other papers contribute by exploring various facets of LLM capabilities and efficiency across different domains. For instance, the study on energy consumption compares diffusion models against autoregressive ones, showing that diffusion models are about twice as efficient for longer sequences when using KV-cache Equilibrium Forcing: Adaptive Video Generation Without Noise Conditioning.

The collection of papers touches upon complex agent behaviors and knowledge structures. Concepts like DeLoop propose a self-supervised method to reduce redundant actions in long-horizon agents, while JockeyBench provides an evaluation protocol for forecasting decisions in horse racing JockeyBench: Evaluating LLM Agents on Horse-Racing Decision-Making.

The day on AI Radio

Astrophysics Radio

118 papers today

Today’s work at Astrophysics Radio focuses heavily on tracing cosmic star formation across vast epochs, particularly with new JWST data. The paper that stands out is a comprehensive study combining wide-area COSMOS and deeper GOODS-N data to map star formation from redshift zero to nine JWST View of Cosmic Star Formation from z about0 to 9 with Hydrogen Recombination Lines in the COSMOS and GOODS-N fields, which reveals how the characteristic luminosity of galaxies increases toward cosmic noon.

Other papers contribute by addressing different physical aspects. Some research looks at galactic structure, like Barflow: Modeling Gas Kinematics of Bar-Driven Flows, which uses simulations to explain how gas moves within barred galaxies, while others explore the physics of accretion in dark matter halos through The Contribution of Mergers and Accretion to Dark Matter Halo Assembly: An Excursion Set Theory Approach.

The remaining papers cover specialized topics like stellar evolution and high-energy phenomena. For instance, Diocotron Modes in Pulsar Magnetospheres: Charge Diffusion and Implications for Radio Emission Variability investigates plasma waves in pulsar environments, while Evidence for Suppressed Non-thermal Broadening in a Flare Ribbon from Stereoscopic Observations examines the magnetic field geometry governing solar flare ribbons.

The day on Astrophysics Radio

Genomics Radio

18 papers today

Today's work on Genomics Radio focuses heavily on developing mathematical frameworks for complex biological systems, particularly in cancer and neuroscience. The study by Modeling the Microenvironmental Dynamics of Adenoid Cystic Carcinoma in Lacrimal Gland with Treatment Intervention uses nonlinear differential equations to show how oxygen and nutrient depletion drive tumor growth, suggesting radiotherapy timing is key for control.

Other papers contribute to different modeling domains; Stochastic partial differential equation model for environmental DNA dynamics in river environments proposes a new stochastic PDE framework to model eDNA concentration in rivers using delayed source inputs, while Tracking large chemical reaction networks and rare events by neural networks presents a neural network method for simulating high-dimensional chemical reaction networks and rare events efficiently.

The collection also touches on broader computational challenges; How Long Is the Coast of a Neuron? Resolution-Dependent Interaction Order and Computational Capacity examines how observational resolution affects neural computation, and State Space Models Naturally Produce Time Cell and Oscillatory Behaviors and Scale to Abstract Cognitive Functions suggests state space models can generate complex temporal dynamics resembling time cells for cognitive functions.

The day on Genomics Radio

Quantum Radio

179 papers today

The day focused on using quantum algorithms to simulate thermal properties of generic quantum field theories, specifically looking at fermion and scalar fields in one plus one dimensions. This work aims to understand how these systems behave at thermal fixed points before moving toward real-time simulations.

Other papers covered diverse topics, including methods for preparing multi-qudit states using Gray codes and modeling interaction Hamiltonians to resolve paradoxes about remote systems. We also saw research into nonergodic behavior in quantum many-body systems due to symmetry-protected zero modes, as well as analyzing transport in tunnel ferromagnetic Josephson junctions.

The collection suggests a broad effort across simulation, theoretical bounds, and experimental control. One key open question is whether the nonergodic dynamics seen in coupled chains persists under symmetry-breaking perturbations, especially when compared to the results from simulating thermal field theories.

The day on Quantum Radio

Robotics Radio

148 papers today

Today's focus on Robotics Radio centers on how robots can navigate and interact with complex, unstructured environments using language understanding. The work by Driving, Fast or Slow? Neuro-Symbolic Guidance for Motion Prediction in Multi-Modal Ground Mobility explores guidance for motion prediction in ground mobility systems, suggesting a neuro-symbolic approach is key to handling multi-modal scenarios. Similarly, G cubed VLA: Geometric Inductive Bias for Vision-Language-Action Models investigates geometric inductive biases within vision-language-action models, which seems central to how robots interpret their surroundings and plan actions.

The remaining papers contribute various technical avenues to this theme. Some look at improving manipulation stability, like Hierarchical Residual Policy Learning for Real-World Mobile Manipulation with Sparse Human Guidance, which fine-tunes complex mobile manipulation systems efficiently in the real world using expert guidance. Others address perception and planning challenges, such as What Matters for World Modeling in Vision-Language-Action Models?, which systematically evaluates different paradigms for integrating world models into vision-language-action models to determine what aspects matter most.

Overall, the day's papers suggest a move toward robust, guided decision-making in embodied AI. The combination of geometric biases and world model integration points toward building agents that can effectively reason about their physical surroundings and plan long horizons.

The day on Robotics Radio

Security Radio

38 papers today

The day's focus on Security Radio was heavily weighted toward understanding how agents can be manipulated through visual interfaces. The work on <a href="/papers/MMA-SafetyBench_2610.19384">MMA-SafetyBench</a> explored visual semantic hijacking, showing high success rates for adversarial trajectories across various models and domains. This is complemented by the security analysis of ChatGPT /papers/2508.09426, which examines threats and privacy risks associated with this LLM.

Other papers contributed to this theme by looking at agent security in different ways. The research on <a href="/papers/SecureClaw_2606.09549">SecureClaw</a> presents a dual-boundary architecture for LLM agents that mitigates unauthorized actions and sensitive data exposure simultaneously. Furthermore, the work on <a href="/papers/CyberArena_2610.19005">CyberArena</a> benchmarks AI agents in live attack-and-defense combat against dynamic challenges, revealing how they adapt under pressure.

The collection suggests a major open question regarding agent reliability when faced with complex, multimodal threats. While tools like <a href="/papers/SecureClaw_2606.09549">SecureClaw</a> offer strong architectural defenses, the findings from <a href="/papers/MMA-SafetyBench_2610.19384">MMA-SafetyBench</a> indicate that visual deception can still compromise agents at high rates. The challenge remains in developing robust defenses against these context-aware attacks when agents operate in rich visual environments where their safety boundaries are easily breached by semantic elements.

The day on Security Radio

The day, in brief

AI Radio is exploring whether simpler methods like Sequential Fine-tuning with Averaging or replay-free Lookahead outperform complex rehearsal buffers for continual learning, showing that both methods achieve up to thirty six point three percent less training time than replay and adaptive-merging setups in Qwen3-14B adapter continual learning. Astrophysics Radio is demonstrating that the connection between massive galaxies and their dark matter haloes from z=0.25 to z about 4 links observed galaxy growth to halo structure, finding that the stellar-to-halo mass ratio peaks around a halo mass of about ten twelve solar masses at redshift one. Genomics Radio is developing a mathematical model using nonlinear ordinary differential equations to show that tumor cell growth in adenoid cystic carcinoma is linked to oxygen and nutrient depletion, suggesting postoperative radiotherapy offers better control if residual cells are radiosensitive and treatment interruptions are avoided. Quantum Radio is showing that the most significant development involves realizing bulk superconductivity at ambient pressure in quintuple-layer square-planar nickelate La6Ni5O12, which exhibits a superconducting volume fraction of twenty-five percent. Robotics Radio is introducing G cubed VLA, which introduces a geometric inductive bias for vision-language-action models to better structure their understanding of the world to make superior decisions. Security Radio is highlighting TrackShield because it introduces a runtime defense mechanism that aims to mitigate availability attacks in high-performance real-time multi-object tracking.

The week in the estate's ingest

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

1118Sat
893Mon
8455Wed
1425Thu

Show the answer

1425 papers were ingested on Thursday 15 October.

1425 papers, three ways

Where today's batch landed

By the paper's primary arXiv category.
cs.LG242
cs.CL173
cs.CV153
cs.AI136
cs.RO117
quant-ph104
astro-ph.GA44
cs.CR38

The index

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

  1. EXPO: a quantum leap in fast, wide-band X-ray polarimetry for astrophysics [arXiv] [episode]
  2. Magnon Kerr Effect in a Magnetic Thin Film Strongly Coupled to a Microwave Resonator [arXiv] [episode]
  3. Current validation practice undermines surgical AI development [arXiv] [episode]
  4. Lotus: Learning to Understand Time-series by Learning to Forecast [arXiv] [episode]
  5. Restoring the Surface Magnetic Gap in MnBi 2 Te 4 [arXiv] [episode]
  6. When Privileged Guidance Misaligns: State-Matched Routing and Contextualized Self-Distillation for Multi-Turn Agents [arXiv] [episode]
  7. Formalizing Flag Algebras in Lean [arXiv] [episode]
  8. CausalDS: Benchmarking Causal Reasoning in Data-Science Agents [arXiv] [episode]
  9. RhyMix: A Lightweight Adaptive Multi-Rhythm Network for Long-Term Time Series Forecasting [arXiv] [episode]
  10. Verification-Gated Agentic Mission-State Governance for Intelligent Industrial Multi-Robot Systems [arXiv] [episode]
  11. Gibbs-state phases beyond microscopic symmetry breaking: A phase invariant for the 3D toric code [arXiv] [episode]
  12. DynFS-MoE: Dynamic Functional-Structural Mixture-of-Experts for Post-Traumatic Epilepsy Diagnosis [arXiv] [episode]
  13. SPARC: Reliable Spatial Annotations from Robot Demonstrations at Scale [arXiv] [episode]
  14. SecureClaw: Clawing Back Control of LLM Agents [arXiv] [episode]
  15. Ouvia: A User-centered Framework for Measuring Usability of Speech Translation in Real-World Communication Scenarios [arXiv] [episode]
  16. LEGS: Fine-Tuning Teleop-Free VLAs for Humanoid Loco-manipulation in an Embodied Gaussian Splatting World [arXiv] [episode]
  17. S-duality, boundary states, and higher-form symmetries on ALE spaces [arXiv] [episode]
  18. BrainDyn: A Sheaf Neural ODE for Forecasting and Generating Brain Dynamics [arXiv] [episode]
  19. Njord: A Probabilistic Graph Neural Network for Ensemble Ocean Forecasting [arXiv] [episode]
  20. Learning, Fast and Slow: Towards LLMs That Adapt Continually [arXiv] [episode]
  21. Remix the Timbre: Diffusion-Based Style Transfer Across Polyphonic Stems [arXiv] [episode]
  22. TabFORGE: A Tabular Foundation Model for Generative Modelling [arXiv] [episode]
  23. Few-Shot Truly Benign DPO Attack for Jailbreaking LLMs [arXiv] [episode]
  24. Intrinsic Floquet Generation and 1/I Quantum Oscillations in a Sliding Charge-Density Wave [arXiv] [episode]
  25. ContextWeaver: Selective and Dependency-Structured Memory Construction for LLM Agents [arXiv] [episode]
  26. Beyond Reconstruction: Reconstruction-to-Vector Diffusion for Hyperspectral Anomaly Detection [arXiv] [episode]
  27. Quantum simulation of traversable-wormhole-inspired quantum teleportation in a chaotic binary sparse SYK model [arXiv] [episode]
  28. Towards Secure Agent Skills: Architecture, Threat Taxonomy, and Security Analysis [arXiv] [episode]
  29. KiToke: Distribution-Aware Token Compression for Video Large Language Models [arXiv] [episode]
  30. The Reasoning Trap -- Logical Reasoning as a Mechanistic Pathway to Situational Awareness [arXiv] [episode]
  31. CHRONOS Science Program [arXiv] [episode]
  32. Heterogeneous entanglement between a trapped ion and a solid-state quantum memory [arXiv] [episode]
  33. Spatial Phase Control of Energy and Ergotropy in Quantum Batteries [arXiv] [episode]
  34. LingLanMedBench: A Unified Benchmark for Evaluating Large Language Models in Traditional Chinese Medicine [arXiv] [episode]
  35. CASHEW: Stabilizing Multimodal Reasoning via Iterative Trajectory Aggregation [arXiv] [episode]
  36. Video-Browser: Towards Open-web Agentic Video Browsing [arXiv] [episode]
  37. What If They Took the Shot? A Hierarchical Bayesian Framework for Counterfactual Expected Goals [arXiv] [episode]
  38. Perfect Particle Transmission through Duality Defects [arXiv] [episode]
  39. Reclaiming Low-Quality Instruction Data with Rule-Guided LLM Reconstruction [arXiv] [episode]
  40. LLMs Can Get "Brain Rot": A Pilot Study on Twitter/X [arXiv] [episode]
  41. EconCausal: A Context-Aware Economic Reasoning Benchmark for Large Language Models [arXiv] [episode]
  42. Sig-DEG for Distillation: Making Diffusion Models Faster and Lighter [arXiv] [episode]
  43. PRIMEdit: Probability Redistribution for Instance-Aware Multi-Object Video Editing With a New Benchmark Dataset [arXiv] [episode]
  44. GLL: A Differentiable Graph Learning Layer for Neural Networks [arXiv] [episode]
  45. Task-space model-based control of pneumatic soft actuators [arXiv] [episode]
  46. G cubed VLA: Geometric Inductive Bias for Vision-Language-Action Models [arXiv] [episode]
  47. 4DVLT: Dynamic Scene Understanding with Worldline-Centered Vision-Language Tracking [arXiv] [episode]
  48. PoLAR: Factorizing Extent and Mode in Latent Actions for Robot Policy Learning [arXiv] [episode]
  49. C-MIG: Multi-view Information Gain-based Retrieval-Augmented Generation for Clinical Diagnosis Reasoning [arXiv] [episode]
  50. ConRAG: Consensus-Driven Evidence Flow for Multi-Hop Question Answering [arXiv] [episode]
  51. Symmetry-Compatible Principle for Optimizer Design: Embeddings, LM Heads, SwiGLU MLPs, and MoE Routers [arXiv] [episode]
  52. Failing Forward: Adaptive Failure-Informed Learning for Vision-Language-Action Models [arXiv] [episode]
  53. Ideological Bias in LLMs' Economic Causal Reasoning [arXiv] [episode]
  54. Reward Design for Physical Reasoning in Vision-Language Models [arXiv] [episode]
  55. LLMON: An LLM-native Markup Language to Leverage Structure and Semantics at the LLM Interface [arXiv] [episode]
  56. LoopRPT: Reinforcement Pre-Training for Looped Language Models [arXiv] [episode]
  57. Schr"odinger's Navigator: Imagining an Ensemble of Futures for Zero-Shot Object Navigation [arXiv] [episode]
  58. Phase-preserving steady-state operations beyond Lindblad master equations [arXiv] [episode]
  59. Scaling Personality Control in LLMs with Big Five Scaler Prompts [arXiv] [episode]
  60. Explanation User Interfaces: A Systematic Literature Review [arXiv] [episode]

Listen

Everything the estate published on 2026-10-15, in the player. Playing from this page is how the edition learns whether anyone did.

Previous editions