arXiv subjects

Neeraj Yadav

Publications and source records attributed to Neeraj Yadav.

5 recordsLinked to original sources

arXiv2026

Temporal Validity on Real Software Histories: Eliminating Stale-Fact Errors in Code-Assistant Memory over GitHub Fixes

Retrieval-augmented generation (RAG) has no model of time: when a fact changes across a coding session - a function is renamed, an endpoint moves, a dependency is bumped - RAG retrieves both the old and new value with near-identical similarity and cannot tell which is current, so it serves the superseded value. Paper 1 showed, on synthetic single-value benchmarks, that a deterministic (subject, relation, object) supersession memory eliminates this failure. Here we validate it end-to-end on real software history. From 707 real GitHub issues (SWE-bench Lite + Verified) we extract 130 clean atomic state transitions, a fix that changes one identifiable value from a pre-fix to a post-fix form, and render each marker-free (the stale and current statements differ only in the value). On this set, MemStrata reaches 0.91 answer accuracy versus RAG's 0.57-0.59; and, the structural result, when forced to answer RAG serves the superseded value 36-38% of the time (an LLM reranker does not help) while MemStrata drives this to ~0, at RAG retrieval latency (~2.1 s vs ~18 s for the reranker). We are explicit about scope: only ~18% of real fixes are clean atomic transitions; Paper 2 isolates the memory mechanism on that class, and extraction coverage of the remaining fixes is the orthogonal problem we defer to follow-on work. A real product bug surfaced and was fixed during the study (a case/punctuation-insensitive value comparison), with the moat property (deterministic-supersession accuracy on clean code mutations) preserved and verified.

cs.SE↗

arXiv2026

Temporal Validity in Retrieval Memory: Eliminating Stale-Fact Errors for AI Agents over Evolving Knowledge

Retrieval-augmented generation (RAG) gives agents access to accumulated knowledge, but has no model of time. When a fact changes (e.g., a function is renamed or API restructured), RAG retrieves both the stale and current value with near-identical embedding similarity. The agent then either abstains or serves the superseded fact. We show this is a structural problem: on a calibrated dataset, cosine similarity distinguishes a contradicted fact from a duplicated one with AUROC 0.59 (near chance), as contradictions are often more embedding-similar to the original than rephrased duplicates. We present MemStrata, a retrieval memory maintaining temporal validity. It stores facts like RAG, preserving static recall, but when a fact's value is contradicted, a deterministic (subject, relation, object) supersession rule retires the stale value in a bi-temporal ledger - with no similarity threshold and no LLM call. Across six benchmarks run locally with a 7B model, MemStrata ties RAG on static knowledge and reaches 0.95-1.00 accuracy on evolving knowledge (where RAG reaches 0.20-0.47). The central result is the stale-fact-error rate: when required to answer, RAG serves superseded values 15-40% of the time; MemStrata drives this to ~0%, a failure class RAG cannot avoid. MemStrata achieves this at retrieval latency (~2.1s) versus ~16-18s for LLM-reranking baselines. We release the harness, datasets, and a marker-free evaluation protocol for memory under knowledge evolution.

cs.CL↗

arXiv2023

Modeling the effect of wind speed and direction shear on utility-scale wind turbine power production

Wind speed and direction variations across the rotor affect power production. As utility-scale turbines extend higher into the atmospheric boundary layer (ABL) with larger rotor diameters and hub heights, they increasingly encounter more complex wind speed and direction variations. We assess three models for power production that account for wind speed and direction shear. Two are based on actuator disc representations and the third is a blade element representation. We also evaluate the predictions from a standard power curve model that has no knowledge of wind shear. The predictions from each model, driven by wind profile measurements from a profiling LiDAR, are compared to concurrent power measurements from an adjacent utility-scale wind turbine. In the field measurements of the utility-scale turbine, discrete combinations of speed and direction shear induce changes in power production of -19% to +34% relative to the turbine power curve for a given hub height wind speed. Positive speed shear generally corresponds to over-performance and positive direction shear to under-performance, relative to the power curve. Overall, the blade element model produces both higher correlation and lower error relative to the other models, but its quantitative accuracy depends on induction and controller sub-models. To further assess the influence of complex, non-monotonic wind profiles, we also drive the models with best-fit power law wind speed profiles and linear wind direction profiles. These idealized inputs produce qualitative and quantitative differences in power predictions from each model, demonstrating that time-varying, non-monotonic wind shear affects wind power production.

physics.flu-dyn↗

arXiv2023

Imprints of Einstein-Maxwell dilaton-axion gravity in the observed shadows of Sgr A* and M87*

Einstein-Maxwell dilaton-axion (EMDA) gravity provides a simple framework to investigate the signatures of string theory. The axion and the dilaton fields arising in EMDA gravity have important implications in inflationary cosmology and in addressing the late time acceleration of the universe. It is therefore instructive to explore the implications of such a model in explaining the astrophysical observations. In this work we explore the role of EMDA gravity in explaining the observed shadows of black holes (M87* and Sgr A*) released by the Event Horizon Telescope (EHT) collaboration. The Kerr-Sen metric represents the exact, stationary and axisymmetric black hole solution of EMDA gravity. Such a black hole is characterized by the angular momentum $a$ acquired from the axionic field and the dilatonic charge $r_2$ arising from string compactifications. We study the role of spin and the dilaton charge in modifying the shape and size of the black hole shadow. We note that black holes with larger dilaton charge cast a smaller shadow. We investigate the consequences of such a result in addressing the EHT observations of M87* and Sgr A*. Our analysis reveals that the shadow of M87* exhibits a preference towards the Kerr scenario. However, when 10% offset in the shadow diameter is considered, $0.1\lesssim r_2\lesssim 0.3$ is observationally favored within 1-$σ$. The shadow of Sgr A* on the other hand shows a preference towards the Kerr-Sen scenario since the central value of its shadow can be better explained by a non-zero dilaton charge $0.1 \lesssim r_2 \lesssim 0.4$. However, when the 1-$σ$ interval is considered the Kerr scenario is included. We discuss the implications of our results.

gr-qc↗

arXiv2022

Collective wind farm operation based on a predictive model increases utility-scale energy production

Wind turbines located in wind farms are operated to maximize only their own power production. Individual operation results in wake losses that reduce farm energy. In this study, we operate a wind turbine array collectively to maximize total array production through wake steering. The selection of the farm control strategy relies on the optimization of computationally efficient flow models. We develop a physics-based, data-assisted flow control model to predict the optimal control strategy. In contrast to previous studies, we first design and implement a multi-month field experiment at a utility-scale wind farm to validate the model over a range of control strategies, most of which are suboptimal. The flow control model is able to predict the optimal yaw misalignment angles for the array within +/- 5 degrees for most wind directions (11-32% power gains). Using the validated model, we design a control protocol which increases the energy production of the farm in a second multi-month experiment by 2.7% and 1.0%, for the wind directions of interest and for wind speeds between 6 and 8 m/s and all wind speeds, respectively. The developed and validated predictive model can enable a wider adoption of collective wind farm operation.

math.OC↗