Monday 19 October

The 60-second version

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

1. 1

   New optimizer fits massive models onto one GPU

   TACO reduces optimizer state by 174 times compared to AdamW8bit, allowing 30 to 32 billion parameter models to fit on an eighty gigabyte GPU. arXiv 2610.02199

2. 2

   Human demonstration data improves robot success rates

   KODA improves success rates by forty five and forty six point two five percentage points on Unitree G1 tasks when using robot data fixed against human demonstrations. arXiv 2610.21152

3. 3

   Thermal attacks expose autonomous vehicle perception risks

   A survey compares thermal IR spoofing and telemetry leakage across privilege levels, highlighting physical adversary risks to perception systems. arXiv 2610.21034

4. 4

   Structural changes tune material superconductivity

   Applying specific compressive strain along a certain direction boosts electron-phonon coupling, significantly increasing the critical temperature for superconductivity. arXiv 2606.04033

5. 5

   Unifying learning from different data sources

   Cross-Behavior Policy Optimization unifies offline and online data sources, improving reasoning through shared regression targets. arXiv 2610.12830

Across the stations

The day, in brief

Did you know that scientists are using highly structured methods to predict deterioration in intensive care units without losing the vital link between a warning and the specific patient signals causing it? This is thanks to a new framework called SepsisLens. Instead of just giving clinicians a single risk score, SepsisLens preserves variable-indexed temporal states until the point of risk composition. This means it lets doctors see exactly which organ dynamics are driving the warning, rather than just a generalized alarm.

This approach is revolutionary because many existing temporal models collapse these variable connections into one big patient risk number. SepsisLens keeps each variable's dynamics separate while sharing a temporal encoder for the overall trajectory. It enables the creation of a Structured Risk Head that composes multi-horizon risk from explicit organ and variable components, which is a massive step toward actionable clinical insight.

Shifting to the world of artificial intelligence, researchers are tackling how we can make those massive language models fit onto hardware without losing accuracy. The breakthrough comes from an optimizer called TACO. This method drastically cuts down on the memory needed for the optimizer state—by about 174 times compared to AdamW8bit—while keeping accuracy on large models like OPT-13B nearly identical.

This efficiency jump is huge because it allows researchers to fine-tune much bigger models. With TACO, you can now fine-tune 30 to 32 billion parameter models on a single eighty gigabyte GPU. This means training massive AI systems is becoming dramatically more feasible for smaller labs and practitioners, making the next generation of large language models accessible.

In physics, we're seeing how subtle structural changes can fundamentally alter physical properties in materials. Research into strain-enhanced superconductivity in RuO two thin films shows that applying specific compressive strain along a certain direction pushes electronic bands right toward the Fermi level. This action boosts electron-phonon coupling, which significantly increases the material's critical temperature for superconductivity—and this enhancement scales almost linearly with how much you apply that strain.

This demonstrates a direct, tunable link between mechanical stress and quantum behavior in condensed matter systems. It suggests that materials could be engineered to become superconductors simply by applying precise physical pressure.

The week in the estate's ingest

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

1196 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. Geometrical and Optical Design of the GUEST Laser-Ranged Test Masses for the Detection of Gravitational Waves in the mu Hz range [arXiv] [episode]
2. Between Spoilers and Guesswork: Avoiding Over-Specification in Coding Benchmarks [arXiv]
3. Quantumness of Entanglement in Schwarzschild de Sitter Spacetime [arXiv]
4. Shape-Aware Neural Time Fields for Rigid-Body Motion and Manipulation Planning [arXiv]
5. Acquisition Function-Guided Generation for Bayesian Optimization over Discrete Spaces [arXiv]
6. Pedicel-Targeted Strawberry Harvesting via Imitation Learning on a Low-Cost Teleoperated Robotic Arm [arXiv]
7. Dust in Galaxies at z=10-15: Attenuation and Re-emission Predictions from the FirstLight Simulations [arXiv]
8. StoBasket: Joint Upsell and Cross-Sell Optimization for Contribution Margin [arXiv]
9. One Communicator, Three Vendors: Extending the MoA Predictive Theory to a Real NVIDIA+AMD+CPU MPI Ensemble [arXiv]
10. Unit-level aggregation breaks coverage in causal foundation models [arXiv]
11. Reusing Calibration Gains across Model Representations [arXiv]
12. Development of a fabrication-to-benchtop process for SiN-based quantum devices [arXiv]
13. Direct Conditional Parameterization for N-Dimensional Splatting [arXiv]
14. When Can You Trust an Agent's Own Verifier? Stress-Testing Counterfactual Rationale Verification Under Search Pressure and Oracle Access [arXiv]
15. Anytime Contracts for AI Services [arXiv]
16. MultiPASE: Shape-Aware Comparative 3D Genome Analysis [arXiv]
17. Filter-Augmented Reinforcement Learning for Battery Fast Charging with State of Health and Plating-Aware Reward Design [arXiv]
18. Next Token or Vector? Learning How Far to Decode [arXiv]
19. A Multiwavelength View of SPLUS J102139.31-282247.5: an X-ray Luminous and Eclipsing Dwarf Nova [arXiv]
20. Self-Evolving Coding Agents via Harness-Policy Optimization [arXiv]
21. Prune the Leaky Neurons: Mitigating Private Data Leakage in Fine-Tuned LLMs via Joint-Embedding-Guided Pruning [arXiv]
22. Ori: Drift-Aware Streaming k-PCA for Very Low-Rank LLM Training [arXiv]
23. STRIDE: Evaluating Spatiotemporal Reasoning in Driving Edge Cases [arXiv]
24. Causal Latent Workspaces for Looped Transformers [arXiv]
25. Reasoning Reranking with Semantic IDs as the Sole Item Representation [arXiv]
26. Veil-Then-Reveal: Protecting PII Privacy via Reversible Pseudo-Anonymization in Zero-Trust User--LLM Interaction [arXiv]
27. Programming the Nonlinear Optics of a Nanoporous Photonic Metamaterial by Confined Crystallization [arXiv]
28. Gaia in Technicolor. I. A Simple Variability-Encoded Color-Magnitude Diagram [arXiv]
29. Compression Precedes, Content Decides: Emergent Misalignment in Reward Models [arXiv]
30. On the Position Bias of Large Language Models: A Comprehensive Study [arXiv]
31. Watermarkability: Measuring and Improving Language-Model Compatibility with Logit-Based Watermarks [arXiv]
32. Replay Pruning and Few-Bit Integer Inference for Hardware-Efficient Speed-Sign Recognition [arXiv]
33. PASS-VLN: Passability-Aware Local-Command Vision-and-Language Navigation [arXiv]
34. Simultaneous Force and Motion Control in Generalised Coordinates [arXiv]
35. Hybrid Cones: Set Representation with Sharp Unions of Minimal Overhead for Nonlinear and Switched Linear Systems [arXiv]
36. A Unified Solution to the Low-Mass X-ray Binary Diversity Puzzle: Saturated Magnetic Braking Coupled with Irradiation-Driven Winds [arXiv]
37. Analytic and numerical study of dark matter halos in the Horndeski theory [arXiv]
38. Causal Edit Serialization [arXiv]
39. VISTA: A Multimodal Egocentric Benchmark for Goal-Oriented Perception and Assistance to Blind and Low-Vision Users [arXiv]
40. On the relationship between the POVM framework and the no-signaling hypothesis [arXiv]
41. World Action Critic (WAC) Model: Joint Foresight and Criticism for Policy Learning [arXiv]
42. DefFlow: Goal-Conditioned Deformable Object Manipulation using 3D Point Flows [arXiv]
43. GRAD: Graph-Structured Antibody Defense for LLM Agents [arXiv]
44. Partial-Revelation Panel Forecasting: Predictive Information Gain, Allocation, and Coarsening-Consistent Operators [arXiv]
45. Too Emotional to Be Fake? Intra-Utterance Emotion Shifts Expose Failure Modes in Speech Deepfake Detectors [arXiv]
46. MIDSSAR: MId Dispersion Spectrograph for Stellar Activity Research [arXiv]
47. Image Moment Estimation for Crowded Datasets [arXiv]
48. The Reliance Hypothesis: Chain-of-Thought Faithfulness Rises as Language Resource Level Falls [arXiv]
49. VTrajEmbed: Aligning Text with Video Object Trajectories via Spatiotemporal Embeddings [arXiv]
50. Coexistence of Griffiths phase and superparamagnetic-like behavior in La-doped antiferromagnetic CaMnO3 [arXiv]
51. Variational and Decomposed method for Mitigating Bimodal Bias in Robust Visual Question Answering [arXiv]
52. Benchmarking and Enhancing Agentic Causal Discovery [arXiv]
53. Measurement-Adaptive Light-Cone Transform for Robust Non-Line-of-Sight Imaging [arXiv]
54. Sample-Efficient Reinforcement Learning for Policy Improvement via Flow-Reversed Action Anchors [arXiv]
55. Uncertainty-Aware Budget Allocation for Efficient LLM Tree Search [arXiv]
56. GASTRO-BH: Effects of a Central Black Hole in GSE-like Dwarf-galaxy Mergers Around Milky Way-like Hosts [arXiv]
57. Evo-B&B: Behavior-Aware LLM-Guided Joint Evolution of Branching and Node Selection Policies for MILP [arXiv]
58. Chirality-resolved imaging of spin waves with a quantum spin [arXiv]
59. Molecular genetics of Aedes aegypti: historical advances and oversights [arXiv]
60. MS-Eval: Benchmarking Sub-Agent Delegation and Orchestration in MCP-Intensive Working Scenarios [arXiv]

Listen

Everything the estate published on 2026-10-19, in the player. Playing from this page is how the edition learns whether anyone did.

• One-way electron mirrors from single-layer stacking shifts in rhombohedral graphite
• Edge state reconstruction in proximitized graphene and its probe using RKKY interaction
• Imaging defects in two-dimensional crystals by convergent beam electron diffraction
• PANDA-D: Interface-free determination of two dynamic contact angles
• Adding Temperature to the Adiabatic Connection: Hybrid Functionals for Static and Dynamic Correlation
• Classical Capacity of Single-Mode Bosonic Gaussian Channels
• Local wave-packet spectroscopy of the Drude weight in Tomonaga-Luttinger liquids
• Microwave spectroscopic evidence of a moir'e-commensurate electron crystal in monolayer graphene
• Parafermions in polarized nu=2/3 Fractional Quantum (Anomalous) Hall Heterostructures
• Weekly Summary for the week of 2026-10-12
• Unique Charge-density-wave Order and Fluctuation-induced Phonon Softening in the Optimally-doped Trilayer Cuprate Bi 2 Sr 2 Ca 2 Cu 3 O 10+ delta
• Magna: simulating magic state factories as cheaply as Clifford circuits

Previous editions

• Sunday 18 October
• Saturday 17 October
• Friday 16 October
• Thursday 15 October
• Wednesday 14 October
• Tuesday 13 October
• Monday 12 October