Wednesday 14 October

The 60-second version

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

1. 1

   Black hole energy swallowing is much less than expected

   Ultra-compact X-ray binaries show black holes swallow significantly less accreting energy than neutron stars.

2. 2

   AI search finds transferable nucleic acid design algorithms

   Agentic program search uncovered algorithms for nucleic acid design through the AutoNovo framework.

3. 3

   DelegateGuard creates boundaries for LLM agent actions

   DelegateGuard converts user authorization into constraints based on purpose and labels to block unauthorized moves.

4. 4

   Robotics planning time reduced by ten point three percent

   The ATOM framework cut planning error in estimating abstract transition costs by ten point three percent for vehicle navigation.

5. 5

   Thermodynamic consistency guides biochemical reaction network parameters

   RNAGenScape uses property-guided manifold Langevin dynamics to guide mRNA sequence generation while enforcing thermodynamic consistency.

Across the stations

The day, in brief

AI Radio focused on using Graph Neural Networks to control biological networks via the GATTACA Framework /papers/2505.02712 for managing complex biological systems. Astrophysics Radio found that the radiative efficiency for a black hole is about two to three orders of magnitude lower than that for neutron star systems by combining accretion rates and luminosities. Genomics Radio developed RNAGenScape which uses property-guided manifold Langevin dynamics to guide the generation of mRNA sequences. Quantum Radio demonstrated how time-frequency cat and Gottesman-Kitaev-Preskill states can act as an internal temporal ruler through fidelity revivals in quantum teleportation with single photons. Robotics Radio showed that the ATOM framework reduced the error in estimating abstract transition costs by ten point three percent during repeated navigation tasks. Security Radio introduced DelegateGuard which converts user authorization into constraints based on purpose, labels, operations, sinks, and delegation depth to block unauthorized actions.

The week in the estate's ingest

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

8455 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. Early Signatures of Memorization in Diffusion Models via Basin Geometry and Cyclic Denoising [arXiv] [episode]
2. Multimodal Graph Retrieval-Augmented Sequential Recommendation via Collaborative Filtering Paths [arXiv] [episode]
3. Memorization and Malign Generalization in Conditional Diffusion Models with Random Features [arXiv] [episode]
4. Sample-Efficient Generative Conformal Prediction [arXiv] [episode]
5. Evi-VN: Hard Region Guided Virtual Node Evidence Injection for GNN-Based Fraud Detection [arXiv] [episode]
6. Compile the Table: Query-Calibrated Operator Compression for Tabular In-Context Learning [arXiv] [episode]
7. PRAXIS: Learning Dynamics of Self-Improving Models with Symbolic Archives [arXiv] [episode]
8. DADP: Dynamic Activity-Dependent Pruning, A Reverse Hebbian-Inspired Structural Pruning Method [arXiv] [episode]
9. Cost-Aware Mixture-of-Experts Coordination for Model Markets [arXiv] [episode]
10. SCORE: Spectral Correlation Estimation for Multivariate Gaussians [arXiv] [episode]
11. Invariant-Measure Reasoners: Stable Representations for Latent Reasoning [arXiv] [episode]
12. Low-rank tensor structure of precipitation and its application to satellite-reference merging [arXiv] [episode]
13. Efficient Multimodal Inference through Adaptive Acquisition and Sequential Fusion [arXiv] [episode]
14. Are We Recovering Mechanisms? Objective-Level Recovery Gaps in Mechanistic Interpretability [arXiv] [episode]
15. Preemptive LLM Unlearning against Forbidden Capability Acquisition via Gradient Sealing [arXiv] [episode]
16. Koa-decision: Fast and Consistent Structured Decision Making with Generative LLMs [arXiv] [episode]
17. ATLAS: A Foundation Neural Sampler for Amorphous Materials [arXiv] [episode]
18. FlexTab: A Flexible Encoder-Decoder Architecture for In-Context Learning Across Diverse Tabular Tasks [arXiv] [episode]
19. TraceCodec: A Compiler-Backed Neural Codec for Network Traffic Traces [arXiv] [episode]
20. Hierarchical Planning with Latent World Models [arXiv] [episode]
21. Differentiable Retrieval-Augmented Generation for Predicting Cellular Responses to Gene Perturbation [arXiv] [episode]
22. A Structural Theory of Position Bias in Transformers [arXiv] [episode]
23. PACC: Protocol-Aware Cross-Layer Compression for Compact Network Traffic Representation [arXiv] [episode]
24. ERIS: Enhancing the Privacy-Utility-Scalability Trade-Off in Federated Learning via Federated Shard Aggregation [arXiv] [episode]
25. Distance Marching for Generative Modeling [arXiv] [episode]
26. Beyond Prediction-Set Size: Quantifying Conformal Uncertainty through Credal Geometry [arXiv] [episode]
27. A Generalizable Machine Learning Framework for Half-Heusler Thermoelectrics [arXiv] [episode]
28. FIRE: Multi-fidelity Regression with Distribution-conditioned In-context Learning using Tabular Foundation Models [arXiv] [episode]
29. Refining Machine Learning Potentials through Thermodynamic Theory of Phase Transitions [arXiv] [episode]
30. Second-Order Optimization with Asynchronous Curvature [arXiv] [episode]
31. DaCe-DT: Data-Centric Offline Multi-Task Reinforcement Learning via Adaptive Prompts and Trajectory Correction for Heterogeneous Tasks [arXiv] [episode]
32. Randomized Transport Maps for Model-Free Policy-Gradient Mean-Field Control [arXiv] [episode]
33. Subspace Uncertainty and Sharp Sampling Thresholds on the Boolean Cube [arXiv] [episode]
34. Evaluating Rubric Generation with Interventional Transfer [arXiv] [episode]
35. ASPIRE: Saddle-Point Discovery through Set Prediction and Physical Refinement [arXiv] [episode]
36. Localized Operator Learning with Adaptive Partition-of-Unity Mixture-of-Expert Networks [arXiv] [episode]
37. QiYao-I: A Manifold Based Foundation Model for Irregular Multivariate Time Series Forecasting [arXiv] [episode]
38. Accelerated Algorithm for Sparse Regularized Partial Optimal Transport [arXiv] [episode]
39. QASP: Query-Adaptive Robust Vector Search Policy [arXiv] [episode]
40. Per-Context Reverse Curricula for Sparse-Reward Reinforcement Learning [arXiv] [episode]
41. Flexible Inverse Design of Optical Multilayer Coatings Using Discrete-Continuous Flow Matching [arXiv] [episode]
42. Tight Complexity of Muon under Heavy-Tailed Noise [arXiv] [episode]
43. Verifiers in the Loop: Decoding Time Verification for Code Translation [arXiv] [episode]
44. Diffusion-Based Posterior Sampling: A Feynman-Kac Analysis of Bias and Stability [arXiv] [episode]
45. Geometric Characterisation and Structured Trajectory Surrogates for Clinical Dataset Condensation [arXiv] [episode]
46. When to Retrain after Drift: A Data-Only Test of Post-Drift Data Size Sufficiency [arXiv] [episode]
47. Grables: When Graph Structure Helps Tabular Prediction [arXiv] [episode]
48. PISA: Prioritized Invariant Subgraph Aggregation for Out-of-Distribution Generalization on Graphs [arXiv] [episode]
49. SpellerSSL: Self-Supervised Learning with P300 Aggregation for Speller BCIs [arXiv] [episode]
50. Asynchronous Sharpness-Aware Minimization For Fast and Accurate Deep Learning [arXiv] [episode]
51. Cova-PINN: Cross-Domain Conservation Physics-Informed Neural Network for Fluid-Solid Conjugate Heat Transfer in Complex Geometries [arXiv] [episode]
52. QUILT: Rethinking Sparse-Attention Prefill through Shared Query Execution [arXiv] [episode]
53. When Do External Priors Help Low-Resource Structured Prediction? A Ranking-Level Alignment Study [arXiv] [episode]
54. CARE: A Lightweight Plug-in Gated Correction and Uncertainty-aware Module for Long-term Time Series Forecasting [arXiv] [episode]
55. SACQ: Structured Decoding with Memory-Conditioned Refinement for Long-Horizon Forecasting [arXiv] [episode]
56. MotiveMob: Motivation as Semantic Action for Closed-Loop Human Mobility Generation [arXiv] [episode]
57. TraceRelay: Attention-Aligned Recurrence over Rolling Traces [arXiv] [episode]
58. asdex: Automatic Sparse Differentiation in JAX [arXiv] [episode]
59. Ambient Discrete Diffusion: Using the Wrong Data at the Right Time for Data Efficient Learning [arXiv] [episode]
60. Democratizing MoE inference on commodity GPUs with CoMoE [arXiv] [episode]

Listen

Everything the estate published on 2026-10-14, in the player. Playing from this page is how the edition learns whether anyone did.

• Ambient Discrete Diffusion: Using the Wrong Data at the Right Time for Data Efficient Learning
• SplitJEPA: Learning Invariant and Variant Latent Worlds without Reconstruction
• Marformer: A Transformer for Predicting Missing Data Distributions
• Learning Kilometer-Scale Weather Prediction with Global-Regional Alignment
• Prospective Prediction of OOD Degradation from Source-Side Training Dynamics
• A Unified Bellman Operator for Safety-Critical Reinforcement Learning
• When Do External Priors Help Low-Resource Structured Prediction? A Ranking-Level Alignment Study
• Dynamics as Code: On Model Compression via Dynamic System
• asdex: Automatic Sparse Differentiation in JAX
• Rounding in Preconditioner Space: Redesigning 4-bit AdamW Optimizer-State Quantization
• Recovery Guarantees for Posterior Sampling of One-Bit Compressed Sensing
• HI3D 3.0 (Twinkle3D): Object-specific 3D Asset Generation with High Resolution

Previous editions

• Monday 12 October
• Saturday 10 October
• Thursday 8 October
• Wednesday 7 October
• Tuesday 6 October