Thursday 8 October

The 60-second version

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

1. 1

   Rectified flows reach optimal sample complexity

   Rectified flow models achieve Oe(epsilon minus two) sample complexity by constraining transport trajectories to be linear, enabling high-quality generation with a single Euler step. arXiv 2601.20250

2. 2

   Robot failure recovery learns from past mistakes

   The FAR framework learns from past mistakes at test time to adapt robot behavior, combining preference learning with lightweight action changes to boost success rates. arXiv 2607.01111

3. 3

   Symbolic regression models map large scale structure

   A symbolic regression emulator suite was developed for cold dark matter power spectra, providing fast and versatile fitting capabilities for cosmological parameters. arXiv 2511.05093

4. 4

   Agent safety tests fail on decomposed tasks

   The DECOMPBENCH benchmark shows that top agents often fail to refuse harmful tasks when they are broken down into smaller, harmless steps. arXiv 2606.13994

5. 5

   Quantum interference separates geometry from optimization

   A framework was developed to separate amplitude transport, interference, and classical postprocessing, demonstrating that quantum advantage requires coherent phase alignment. arXiv 2607.13630

Across the stations

The day, in brief

The work across these stations is mostly about advancing modeling efficiency, improving geometric supervision for vision tasks, controlling complex motion synthesis, and enhancing robotic resilience. Key results include achieving order-optimal sample complexity of Oe(epsilon minus two) for rectified flow models and developing a framework that learns from past mistakes at test time to adapt robot behavior.

The stations share an interest in understanding how complex systems predict outcomes, whether it is AI alignment or biological structural prediction. They diverge by focusing on different aspects of this challenge: one group concentrates on generative modeling efficiency, another on geometric supervision for vision and motion, and others explore safety testing for agents and physical resilience in robotics.

A concrete result worth taking away is the finding that rectified flow models can achieve order-optimal sample complexity of Oe(epsilon minus two) by constraining transport trajectories to be linear, a finding from AI Radio. This suggests a significant theoretical gain in generation quality achievable through structural restrictions on flow models.

The day leaves open the question of how to best balance the theoretical sample complexity gains from flow models with the practical needs of real-world visual understanding and physical motion control.

The week in the estate's ingest

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

1046 papers, three ways

Where today's batch landed

The index

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

1. Learning Perturbation Robust Policies for LLM Agents with Stable Optimization [arXiv] [episode]
2. Visible Touch: Rendering Contact for Visuomotor Policies [arXiv] [episode]
3. Separating Geometry From Interference in Constrained Quantum Optimization [arXiv] [episode]
4. LightCrafter: PBR-Conditioned Video Diffusion Refinement for Controllable and Consistent Relighting [arXiv] [episode]
5. Hallucination Self-Play: Bootstrapping Reinforced Detector via Evolved Generator [arXiv] [episode]
6. Quantitative Benchmarking of Spectroscopic Homogeneous and Inhomogeneous Linewidth Separation [arXiv] [episode]
7. FAR: Failure-Aware Retry for Test-Time Recovery and Continual Policy Improvement [arXiv] [episode]
8. Hand-4DGS: Feed-Forward 3D Gaussian Splatting for 4D Hand Reconstruction from Egocentric Videos [arXiv] [episode]
9. Hidden in Plain Sight: Benchmarking Agent Safety Against Decomposition Attacks with DECOMPBENCH [arXiv] [episode]
10. A nuclear clock synchronized to 229 Th [arXiv] [episode]
11. MAdam: Metric-Aware Multi-Objective Adam [arXiv] [episode]
12. How Far Do Auto-Interpretation Labels Generalize: A Controlled Study Across Languages, Scripts, and Rewordings [arXiv] [episode]
13. Residualized Temporal Sparse Autoencoders for Interpreting Diffusion Models [arXiv] [episode]
14. Quantum Interference Amplifies Weak Chirality into Giant Quantum Nonreciprocity [arXiv] [episode]
15. Which Way Did It Move? Diagnosing and Overcoming Directional Motion Blindness in Video-LLMs [arXiv] [episode]
16. A Few Steps Further: Why Defenses Against Malicious Finetuning Erode Under Continued Training [arXiv] [episode]
17. Med-StepBench: A Hierarchical Reasoning Framework for Evaluating Hallucinations in Medical Vision-Language Models [arXiv] [episode]
18. Steering Without Breaking: Mechanistically Informed Interventions for Discrete Diffusion Language Models [arXiv] [episode]
19. EduGage: A Multimodal Dataset and Benchmark for Sensor-Based Momentary Assessment of Engagement in Self-Guided Video Learning [arXiv] [episode]
20. MotionHiFlow: Text-to-motion via hierarchical flow matching [arXiv] [episode]
21. Thermodynamic signatures of spectral compression in weakly non-Hermitian Dirac fermions [arXiv] [episode]
22. Blind Catalytic Quantum Error Correction: Target-State Estimation and Fidelity Recovery Without A Priori Knowledge [arXiv] [episode]
23. Bi-CamoDiffusion: A Boundary-informed Diffusion Approach for Camouflaged Object Detection [arXiv] [episode]
24. CADReasoner: Iterative Program Editing for CAD Reverse Engineering [arXiv] [episode]
25. Order-Optimal Sample Complexity of Rectified Flows [arXiv] [episode]
26. Integration and Resource Estimation of Cryoelectronics for Superconducting Fault-Tolerant Quantum Computers [arXiv] [episode]
27. Analytical blueprint for 99.999% fidelity X-gates on present superconducting hardware under strong driving [arXiv] [episode]
28. Electronic Structure and Dynamical Correlations in Antiferromagnetic BiFeO 3 [arXiv] [episode]
29. Sym-EFT: Accelerating Effective Field Theory of Large Scale Structure with Symbolic Regression [arXiv] [episode]
30. Qubit-oscillator-based gate implementations for approximate Gottesman-Kitaev-Preskill codes [arXiv] [episode]
31. MAGIC: Learning from Visibility Asymmetry for Unsupervised Stereo Matching [arXiv] [episode]
32. Reasoning Enhances Robustness to Prompt Injection in LLM-Based Consensus [arXiv] [episode]
33. RoboFAC: A Comprehensive Framework for Robotic Failure Analysis and Correction [arXiv] [episode]
34. Lifting the maximally-entangledness assumption in robust self-testing for synchronous games [arXiv] [episode]
35. Revivals of Bell nonlocality require Schr"odinger and Heisenberg non-Markovianity [arXiv] [episode]
36. Curvature-Guided Module Localization for Low-Rank Detoxification of Backdoored Large Language Models [arXiv] [episode]
37. Beyond Captions: Context-Grounded Reconstruction for Biomedical Multimodal Continued Pretraining [arXiv] [episode]
38. StressDream: Steering Video World Models for Robust Policy Evaluation and Improvement [arXiv] [episode]
39. Seeing Is No Longer Believing: Frontier Image Generation Models, Synthetic Visual Evidence, and Real-World Risk [arXiv] [episode]
40. SpanVLA: Learning from Negative-Recovery Samples with Fast Action Bridging for Vision-Language-Action Model [arXiv] [episode]
41. Continuous Semantic Caching for Low-Cost LLM Serving [arXiv] [episode]
42. Document Optimization for Black-Box Retrieval via Reinforcement Learning [arXiv] [episode]
43. Tri-Prompting: Video Diffusion with Unified Control over Scene, Subject, and Motion [arXiv] [episode]
44. Torsional oscillation of carbon nanotubes driven by electron spins [arXiv] [episode]
45. Robotic Ultra-Long-Horizon Manipulation Skills via Human-guided Lifelong Code Generation [arXiv] [episode]
46. Geometry-Centered 3D Latent World Models for Growing Surfaces [arXiv] [episode]
47. ReToken: Improving Long-Context VLMs with Visual Retrieval Token [arXiv] [episode]
48. PersonaMem-v3: Toward Omni-Platform Personal Intelligence for Holistic User Understanding, Recommendation, and Agentic Tasks [arXiv] [episode]
49. Progressive Disclosure for LLM-Maintained Wiki Knowledge Bases: a Preregistered Ablation [arXiv] [episode]
50. BehaviorBench: Benchmarking Foundation Models for Behavioral Science Tasks [arXiv] [episode]
51. MixedPEFT: Combining Multiple PEFT Methods with Mixed Objectives for Unsupervised Domain Adaptation [arXiv] [episode]
52. The decay rate of metastable cosmic strings beyond the thin-string approximation [arXiv] [episode]
53. Nielsen complexity with multiple cost factors [arXiv] [episode]
54. Knowledge boundary probing and demand-guided intervention for LLM-based power system code generation [arXiv] [episode]
55. Cosmic gas accretion from filaments onto galaxy clusters using the IllustrisTNG simulation [arXiv] [episode]
56. U-CECE: A Universal Multi-Resolution Framework for Conceptual Counterfactual Explanations [arXiv] [episode]
57. Prototype-Based Knowledge Guidance for Fine-Grained Structured Radiology Reporting [arXiv] [episode]
58. Vectorizing the Trie: Efficient Constrained Decoding for LLM-based Generative Retrieval on Accelerators [arXiv] [episode]
59. Multi-encoder ConvNeXt Network with Smooth Attentional Feature Fusion for Multispectral Semantic Segmentation [arXiv] [episode]
60. Realisation of a Protected Cat-Qutrit Manifold via Engineered Quantum Tunnelling [arXiv] [episode]

Listen

Everything the estate published on 2026-10-08, in the player. Playing from this page is how the edition learns whether anyone did.

• WRIT: Write-Read Intensive Trajectory Synthesis for Multi-Turn User-Facing Agents
• How Far Do Auto-Interpretation Labels Generalize: A Controlled Study Across Languages, Scripts, and Rewordings
• Beyond Captions: Context-Grounded Reconstruction for Biomedical Multimodal Continued Pretraining
• An LLM-Native Psychometric Instrument Reveals a Self-Report--Behavior Gap Across 25 Models
• Who Brought Easter Eggs to Eid? Auditing LLM-Generated Cultural Translation of Math Word Problems Across Languages and Regions
• MixedPEFT: Combining Multiple PEFT Methods with Mixed Objectives for Unsupervised Domain Adaptation
• BehaviorBench: Benchmarking Foundation Models for Behavioral Science Tasks
• BiPO: Bidirectional Partial Occlusion Network for Text-to-Motion Synthesis
• Research on Deep Learning-Based Semantic Segmentation Algorithms for Subcortical Brain Structures
• FisheyeDistanceNet: Self-Supervised Scale-Aware Distance Estimation using Monocular Fisheye Camera for Autonomous Driving
• Reduction of Class Activation Uncertainty with Background Information
• Auto-Regressive Models Need Structural Registers: Semantic Foresight Improves Coherent Driving Video Continuation

Previous editions

• Wednesday 7 October
• Tuesday 6 October