arXiv subjects

Oliver Zahn

Publications and source records attributed to Oliver Zahn.

At least 19 recordsLinked to original sources

arXiv2026

Discovery by Dreaming: Cross-Domain Recombination in Artificial Memory

Dreams splice together people, places, and times that never met. Neuroscience suggests this recombination is not noise, but a function driving insight and creative discovery. This reframes memory consolidation: rather than merely defending against forgetting, its measurable value lies in recombining knowledge across experiences that have not yet co-occurred. We test this directly by isolating the recombinatory-replay mechanism and implementing it in two architecturally unrelated systems: a LoRA fine-tuning pipeline (DREAMS) and a symbolic engine replaying structured knowledge objects (SAPIENCE). Both systems converge on the same finding: cross-domain consolidation creates value, while within-domain rehearsal does not. The symbolic arm surfaces novel cross-domain connections at 85.7%, a +21 percentage point (pp) gain over baseline. The neural arm improves overall by +5.64 pp, but on subtasks explicitly requiring cross-domain transfer (like unseen math reasoning on GSM8K), gains reach +14.5 pp. This effect is a genuine property of the weights--not a prompt artifact--as prepending the same material in-context to a 671B-parameter model actually reverses the gain. We validate this prediction against documented discoveries across 50,000 real papers and state a falsifiable hippocampal-recording prediction to distinguish recombination from rehearsal. Ultimately, this principle is substrate-general, tracking real discovery at scale. Reading the literature teaches a model to recall what it has seen, but producing discovery requires a separate offline phase that recombines knowledge across domains--the computational analog of dreaming. Consolidation is not for remembering, but for discovering.

cs.LG↗

arXiv2026

Facts as First Class Objects: Knowledge Objects for Persistent LLM Memory

Large language models increasingly serve as persistent knowledge workers, with in-context memory - facts stored in the prompt - as the default strategy. We benchmark in-context memory against Knowledge Objects (KOs), discrete hash-addressed tuples with O(1) retrieval. Within the context window, Claude Sonnet 4.5 achieves 100% exact-match accuracy from 10 to 7,000 facts (97.5% of its 200K window). However, production deployment reveals three failure modes: capacity limits (prompts overflow at 8,000 facts), compaction loss (summarization destroys 60% of facts), and goal drift (cascading compaction erodes 54% of project constraints while the model continues with full confidence). KOs achieve 100% accuracy across all conditions at 252x lower cost. On multi-hop reasoning, KOs reach 78.9% versus 31.6% for in-context. Cross-model replication across four frontier models confirms compaction loss is architectural, not model-specific. We additionally show that embedding retrieval fails on adversarial facts (20% precision at 1) and that neural memory (Titans) stores facts but fails to retrieve them on demand. We introduce density-adaptive retrieval as a switching mechanism and release the benchmark suite.

cs.AI↗

arXiv2026

Selective Memory for Artificial Intelligence: Write-Time Gating with Hierarchical Archiving

Retrieval-augmented generation stores all content indiscriminately, degrading accuracy as noise accumulates. Parametric approaches compress knowledge into weights, precluding selective updates. Neither mirrors biological memory, which gates encoding based on salience and archives rather than deletes superseded information. We introduce write-time gating that filters incoming knowledge objects using composite salience scores (source reputation, novelty, reliability) while maintaining version chains that preserve prior states. Using real LLM evaluation without oracle access to quality labels, write gating achieves 100 percent accuracy versus 13 percent for ungated stores. The critical finding emerges under distractor scaling: at 8:1 distractor ratios, read-time filtering (Self-RAG) collapses to 0 percent while write gating maintains 100 percent, revealing a structural advantage of write-time over read-time curation. Validation on Wikipedia (20 entities), procedurally generated pharmacology data, and 2026 arXiv papers confirms these findings. The gating advantage scales inversely with parametric memory support: +25pp for Wikipedia, +48pp for post-cutoff arXiv, +65pp for procedural data with zero training knowledge. Signal ablation confirms the method does not depend on oracle-correlated metadata. Write gating matches Self-RAG accuracy at one-ninth the query-time cost.

cs.AI↗

arXiv2026

ANML: Attribution-Native Machine Learning with Guaranteed Robustness

Frontier AI systems increasingly train on specialized expert data, from clinical records to proprietary research to curated datasets, yet current training pipelines treat all samples identically. A Nobel laureate's contribution receives the same weight as an unverified submission. We introduce ANML (Attribution-Native Machine Learning), a framework that weights training samples by four quality factors: gradient-based consistency (q), verification status (v), contributor reputation (r), and temporal relevance (T). By combining what the model observes (gradient signals) with what the system knows about data provenance (external signals), ANML produces per-contributor quality weights that simultaneously improve model performance and enable downstream attribution. Across 5 datasets (178-32,561 samples), ANML achieves 33-72% error reduction over gradient-only baselines. Quality-weighted training is data-efficient: 20% high-quality data outperforms 100% uniformly weighted data by 47%. A Two-Stage Adaptive gating mechanism guarantees that ANML never underperforms the best available baseline, including under strategic joint attacks combining credential faking with gradient alignment. When per-sample detection fails against subtle corruption, contributor-level attribution provides 1.3-5.3x greater improvement than sample-level methods, with the advantage growing as corruption becomes harder to detect.

cs.LG↗

arXiv2026

Attention Is Not Retention: The Orthogonality Constraint in Infinite-Context Architectures

Biological memory solves a problem that eludes current AI: storing specific episodic facts without corrupting general semantic knowledge. Complementary Learning Systems theory explains this through two subsystems - a fast hippocampal system using sparse, pattern-separated representations for episodes, and a slow neocortical system using distributed representations for statistical regularities. Current AI systems lack this separation, attempting to serve both functions through neural weights alone. We identify the Orthogonality Constraint: reliable memory requires orthogonal keys, but semantic embeddings cannot be orthogonal because training clusters similar concepts together. The result is Semantic Interference (connecting to what cognitive psychologists have long observed in human memory), where neural systems writing facts into shared continuous parameters collapse to near-random accuracy within tens of semantically related facts. Through semantic density (rho), the mean pairwise cosine similarity, we show collapse occurs at N=5 facts (rho > 0.6) or N ~ 20-75 (moderate rho). We validate across modalities: 16,309 Wikipedia facts, scientific measurements (rho = 0.96, 0.02% accuracy at N=10,000), and image embeddings (rho = 0.82, 0.05% at N=2,000). This failure is geometric - no increase in model capacity can overcome interference when keys share semantic overlap. We propose Knowledge Objects (KOs): structured facts with hash-based identity, controlled vocabularies, and explicit version chains. On Wikipedia facts, KO retrieval achieves 45.7% where Modern Hopfield Networks collapse to near-zero; hash-based retrieval maintains 100%. Production systems (Claude Memory, ChatGPT Memory) store unstructured text, causing schema drift (40-70% consistency) and version ambiguity. Knowledge Objects provide the discrete hippocampal component that enables reliable bicameral memory.

q-bio.NC↗

arXiv2017

Full covariance of CMB and lensing reconstruction power spectra

CMB and lensing reconstruction power spectra are powerful probes of cosmology. However they are correlated, since the CMB power spectra are lensed and the lensing reconstruction is constructed using CMB multipoles. We perform a full analysis of the auto- and cross-covariances, including polarization power spectra and minimum variance lensing estimators, and compare with simulations of idealized future CMB-S4 observations. Covariances sourced by fluctuations in the unlensed CMB and instrumental noise can largely be removed by using a realization-dependent subtraction of lensing reconstruction noise, leaving a relatively simple covariance model that is dominated by lensing-induced terms and well described by a small number of principal components. The correlations between the CMB and lensing power spectra will be detectable at the level of $\sim 5σ$ for a CMB-S4 mission, and neglecting those could underestimate some parameter error bars by several tens of percent. However we found that the inclusion of external priors or data sets to estimate parameter error bars can make the impact of the correlations almost negligible.

astro-ph.CO↗

arXiv2016

Scientific Computing Meets Big Data Technology: An Astronomy Use Case

Scientific analyses commonly compose multiple single-process programs into a dataflow. An end-to-end dataflow of single-process programs is known as a many-task application. Typically, tools from the HPC software stack are used to parallelize these analyses. In this work, we investigate an alternate approach that uses Apache Spark -- a modern big data platform -- to parallelize many-task applications. We present Kira, a flexible and distributed astronomy image processing toolkit using Apache Spark. We then use the Kira toolkit to implement a Source Extractor application for astronomy images, called Kira SE. With Kira SE as the use case, we study the programming flexibility, dataflow richness, scheduling capacity and performance of Apache Spark running on the EC2 cloud. By exploiting data locality, Kira SE achieves a 2.5x speedup over an equivalent C program when analyzing a 1TB dataset using 512 cores on the Amazon EC2 cloud. Furthermore, we show that by leveraging software originally designed for big data infrastructure, Kira SE achieves competitive performance to the C implementation running on the NERSC Edison supercomputer. Our experience with Kira indicates that emerging Big Data platforms such as Apache Spark are a performant alternative for many-task scientific applications.

cs.DC↗

arXiv2016

POLARBEAR Constraints on Cosmic Birefringence and Primordial Magnetic Fields

We constrain anisotropic cosmic birefringence using four-point correlations of even-parity $E$-mode and odd-parity $B$-mode polarization in the cosmic microwave background measurements made by the POLARization of the Background Radiation (POLARBEAR) experiment in its first season of observations. We find that the anisotropic cosmic birefringence signal from any parity-violating processes is consistent with zero. The Faraday rotation from anisotropic cosmic birefringence can be compared with the equivalent quantity generated by primordial magnetic fields if they existed. The POLARBEAR nondetection translates into a 95% confidence level (C.L.) upper limit of 93 nanogauss (nG) on the amplitude of an equivalent primordial magnetic field inclusive of systematic uncertainties. This four-point correlation constraint on Faraday rotation is about 15 times tighter than the upper limit of 1380 nG inferred from constraining the contribution of Faraday rotation to two-point correlations of $B$-modes measured by Planck in 2015. Metric perturbations sourced by primordial magnetic fields would also contribute to the $B$-mode power spectrum. Using the POLARBEAR measurements of the $B$-mode power spectrum (two-point correlation), we set a 95% C.L. upper limit of 3.9 nG on primordial magnetic fields assuming a flat prior on the field amplitude. This limit is comparable to what was found in the Planck 2015 two-point correlation analysis with both temperature and polarization. We perform a set of systematic error tests and find no evidence for contamination. This work marks the first time that anisotropic cosmic birefringence or primordial magnetic fields have been constrained from the ground at subdegree scales.

astro-ph.CO↗

arXiv2013

Cosmology from cross correlation of CMB lensing and galaxy surveys

In recent years cross correlation of lensing of the Cosmic Microwave Background (CMB) with other large scale structure (LSS) tracers has been used as a method to detect CMB lensing. Current experiments are also becoming sensitive enough to measure CMB lensing without the help of auxiliary tracers. As data quality improves rapidly, it has been suggested that the CMB lensing-LSS cross correlation may provide new insights into parameters describing cosmological structure growth. In this work we perform forecasts that combine the lensing potential auto power spectrum from various future CMB experiments, the galaxy power spectrum from galaxy surveys, as well as the cross power spectrum between the two, marginalizing over a number galactic and non-galactic cosmological parameters. We find that the CMB lensing-LSS cross correlation contains significant information on parameters such as the redshift distribution and bias of LSS tracers. We also find that the cross correlation information will lead to independent probes of cosmological parameters such as neutrino mass and the reionization optical depth.

astro-ph.CO↗

arXiv2012

Measuring gravitational lensing of the cosmic microwave background using cross correlation with large scale structure

We cross correlate the gravitational lensing map extracted from cosmic microwave background measurements by the Wilkinson Microwave Anisotropy Probe (WMAP) with the radio galaxy distribution from the NRAO VLA Sky Survey (NVSS) by using a quadratic estimator technique. We use the full covariance matrix to filter the data, and calculate the cross-power spectra for the lensing-galaxy correlation. We explore the impact of changing the values of cosmological parameters on the lensing reconstruction, and obtain statistical detection significances at $>3σ$. The results of all cross correlations pass the curl null test as well as a complementary diagnostic test using the NVSS data in equatorial coordinates. We forecast the potential for Planck and NVSS to constrain the lensing-galaxy cross correlation as well as the galaxy bias. The lensing-galaxy cross-power spectra are found to be Gaussian distributed.

astro-ph.CO↗

arXiv2012

Millimeter-wave bolometer array receiver for the Atacama pathfinder experiment Sunyaev-Zel'dovich (APEX-SZ) instrument

The Atacama pathfinder experiment Sunyaev-Zel'dovich (APEX-SZ) instrument is a millimeter-wave cryogenic receiver designed to observe galaxy clusters via the Sunyaev-Zel'dovich effect from the 12 m APEX telescope on the Atacama plateau in Chile. The receiver contains a focal plane of 280 superconducting transition-edge sensor (TES) bolometers instrumented with a frequency-domain multiplexed readout system. The bolometers are cooled to 280 mK via a three-stage helium sorption refrigerator and a mechanical pulse-tube cooler. Three warm mirrors, two 4 K lenses, and a horn array couple the TES bolometers to the telescope. APEX-SZ observes in a single frequency band at 150 GHz with 1' angular resolution and a 22' field-of-view, all well suited for cluster mapping. The APEX-SZ receiver has played a key role in the introduction of several new technologies including TES bolometers, the frequency-domain multiplexed readout, and the use of a pulse-tube cooler with bolometers. As a result of these new technologies, the instrument has a higher instantaneous sensitivity and covers a larger field-of-view than earlier generations of Sunyaev-Zel'dovich instruments. The TES bolometers have a median sensitivity of 890 uKcmb-sqrt(s) (NEy of 3.5e-4 sqrt(s)). We have also demonstrated upgraded detectors with improved sensitivity of 530 uKcmb-sqrt(s) (NEy of 2.2e-4 sqrt(s)). Since its commissioning in April 2007, APEX-SZ has been used to map 48 clusters. We describe the design of the receiver and its performance when installed on the APEX telescope.

astro-ph.IM↗

arXiv2012

New limits on Early Dark Energy from the South Pole Telescope

We present new limits on early dark energy (EDE) from the cosmic microwave background (CMB) using data from the WMAP satellite on large angular scales and South Pole Telescope (SPT) on small angular scales. We find a strong upper limit on the EDE density of Omega_e < 0.018 at 95% confidence, a factor of three improvement over WMAP data alone. We show that adding lower-redshift probes of the expansion rate to the CMB data improves constraints on the dark energy equation of state, but not the EDE density. We also explain how the small-scale CMB temperature anisotropy constrains EDE.

astro-ph.CO↗

arXiv2012

Constraints on neutrino and dark radiation interactions using cosmological observations

Observations of the cosmic microwave background (CMB) and large-scale structure (LSS) provide a unique opportunity to explore the fundamental properties of the constituents that compose the cosmic dark radiation background (CDRB), of which the three standard neutrinos are thought to be the dominant component. We report on the first constraint to the CDRB rest-frame sound speed, ceff^2, using the most recent CMB and LSS data. Additionally, we report improved constraints to the CDRB viscosity parameter, cvis^2. For a non-interacting species, these parameters both equal 1/3. Using current data we find that a standard CDRB, composed entirely of three non-interacting neutrino species, is ruled out at the 99% confidence level (C.L.) with ceff^2 = 0.30 +0.027 -0.026 and cvis^2 = 0.44 +0.27 -0.21 (95% C.L.). We also discuss how constraints to these parameters from current and future observations (such as the Planck satellite) allow us to explore the fundamental properties of any anomalous radiative energy density beyond the standard three neutrinos.

astro-ph.CO↗

arXiv2011

Reconstruction of Gravitational Lensing Using WMAP 7-Year Data

Gravitational lensing by large scale structure introduces non-Gaussianity into the Cosmic Microwave Background and imprints a new observable, which can be used as a cosmological probe. We apply a four-point estimator to the Wilkinson Microwave Anisotropy Probe (WMAP) 7-year coadded temperature maps alone to reconstruct the gravitational lensing signal. The Gaussian bias is simulated and subtracted, and the higher order bias is investigated. We measure a gravitational lensing signal with a statistical amplitude of $\mathcal {C}$ = $1.27\pm 0.98$ using all the correlations of the W- and V-band Differencing Assemblies (DAs). We therefore conclude that WMAP 7-year data alone, can not detect lensing.

astro-ph.CO↗

arXiv2011

Modeling Extragalactic Foregrounds and Secondaries for Unbiased Estimation of Cosmological Parameters From Primary CMB Anisotropy

Using the latest physical modeling and constrained by the most recent data, we develop a phenomenological parameterized model of the contributions to intensity and polarization maps at millimeter wavelengths from external galaxies and Sunyaev-Zeldovich effects. We find such modeling to be necessary for estimation of cosmological parameters from Planck data. For example, ignoring the clustering of the infrared background would result in a bias in n_s of 7 sigma. We show that the simultaneous marginalization over a full foreground model can eliminate such biases, while increasing the statistical uncertainty in cosmological parameters by less than 20%. The small increases in uncertainty can be significantly reduced with the inclusion of higher-resolution ground-based data. The multi-frequency analysis we employ involves modeling 46 total power spectra and marginalization over 17 foreground parameters. We show that we can also reduce the data to a best estimate of the CMB power spectra, and just two principal components (with constrained amplitudes) describing residual foreground contamination.

astro-ph.CO↗

arXiv2010

Delensing CMB Polarization with External Datasets

One of the primary scientific targets of current and future CMB polarization experiments is the search for a stochastic background of gravity waves in the early universe. As instrumental sensitivity improves, the limiting factor will eventually be B-mode power generated by gravitational lensing, which can be removed through use of so-called delensing algorithms. We forecast prospects for delensing using lensing maps which are obtained externally to CMB polarization: either from large-scale structure observations, or from high-resolution maps of CMB temperature. We conclude that the forecasts in either case are not encouraging, and that significantly delensing large-scale CMB polarization requires high-resolution polarization maps with sufficient sensitivity to measure the lensing B-mode. We also present a simple formalism for including delensing in CMB forecasts which is computationally fast and agrees well with Monte Carlos.

astro-ph.CO↗

arXiv2010

Simulations and Analytic Calculations of Bubble Growth During Hydrogen Reionization

We present results from a large volume simulation of Hydrogen reionization. We combine 3d radiative transfer calculations and an N-body simulation, describing structure formation in the intergalactic medium, to detail the growth of HII regions around high redshift galaxies. Our simulation tracks 1024^3 dark matter particles, in a box of co-moving side length 65.6 Mpc/h. This large volume allows us to accurately characterize the size distribution of HII regions throughout most of the reionization process. At the same time, our simulation resolves many of the small galaxies likely responsible for reionization. It confirms a picture anticipated by analytic models: HII regions grow collectively around highly-clustered sources, and have a well-defined characteristic size, which evolves from a sub-Mpc scale at the beginning of reionization to R > 10 Mpc towards the end. We present a detailed statistical description of our results, and compare them with a numerical scheme based on the analytic model by Furlanetto, Zaldarriaga, and Hernquist. We find that the analytic calculation reproduces the size distribution of HII regions and the 21 cm power spectrum of the radiative transfer simulation remarkably well. The ionization field from the simulation, however, has more small scale structure than the analytic calculation, owing to Poisson scatter in the simulated abundance of galaxies on small scales. We propose and validate a simple scheme to incorporate this scatter into our calculations. Our results suggest that analytic calculations are sufficiently accurate to aid in predicting and interpreting the results of future 21 cm surveys. In particular, our fast numerical scheme is useful for forecasting constraints from future 21 cm surveys, and in constructing mock surveys to test data analysis procedures.

astro-ph↗

arXiv2010

Comparison Of Reionization Models: Radiative Transfer Simulations And Approximate, Semi-Numeric Models

We compare the predictions of four different algorithms for the distribution of ionized gas during the Epoch of Reionization. These algorithms are all used to run a 100 Mpc/h simulation of reionization with the same initial conditions. Two of the algorithms are state-of-the-art ray-tracing radiative transfer codes that use disparate methods to calculate the ionization history. The other two algorithms are fast but more approximate schemes based on iterative application of a smoothing filter to the underlying source and density fields. We compare these algorithms' resulting ionization and 21 cm fields using several different statistical measures. The two radiative transfer schemes are in excellent agreement with each other (with the cross-correlation coefficient of the ionization fields >0.8 for k < 10 h/Mpc and in good agreement with the analytic schemes (>0.6 for k < 1 h/Mpc). When used to predict the 21cm power spectrum at different times during reionization, all ionization algorithms agree with one another at the 10s of percent level. This agreement suggests that the different approximations involved in the ray tracing algorithms are sensible and that semi-numerical schemes provide a numerically-inexpensive, yet fairly accurate, description of the reionization process.

astro-ph.CO↗