LETHE: An Interdisciplinary Research Overview

• LETHE is a multifaceted term representing diverse entities in computational mechanics, privacy systems, machine unlearning, and astronomy.
• In computational simulations, Lethe frameworks improve efficiency and accuracy in fluid dynamics, discrete element modeling, and additive manufacturing through innovative numerical methods and dynamic load balancing.
• In data privacy and machine learning, Lethe designs facilitate controlled deletions, persistent unlearning, and backdoor defense, while in astronomy, it identifies a stellar stream indicative of a globular-cluster origin.

Searching arXiv for the provided LETHE-related works to ground the article in current arXiv records. LETHE is not a single technical object but a recurrent name attached to several unrelated research entities on arXiv. In current usage it denotes, among other things, an open-source computational mechanics framework and its applications to implicit large-eddy simulation and discrete-element modeling (Saavedra et al., 2022, Golshan et al., 2021, Gaboriault et al., 30 Sep 2025); a server-side mechanism for concealing social-media deletions from persistent observers (Minaei et al., 2017); a delete-aware LSM-based key-value engine (Sarkar et al., 2020); a federated unlearning method designed for persistent erasure under continued training (Tan et al., 30 Jan 2026); a backdoor-defense framework for LLMs based on knowledge dilution (Chen et al., 28 Aug 2025); and a stellar stream in the Milky Way inner halo whose proper motion was measured with the Hubble Space Telescope (Sohn et al., 2016).

1. Referential scope

The name appears across multiple research domains with no single shared formal definition.

Domain	Referent	Principal source
Computational fluid dynamics	Lethe solver for incompressible multiphase flows and ILES	(Saavedra et al., 2022)
Discrete element method	Lethe-DEM module with load balancing	(Golshan et al., 2021)
Additive manufacturing simulation	Lethe used for multi-layer powder spreading DEM	(Gaboriault et al., 30 Sep 2025)
Social-platform privacy	Lethe intermittent withdrawals for deletion privacy	(Minaei et al., 2017)
Storage engines	Lethe delete-aware LSM engine	(Sarkar et al., 2020)
Federated unlearning	LETHE adapter-augmented dual-stream update	(Tan et al., 30 Jan 2026)
LLM security	LETHE purification by knowledge dilution	(Chen et al., 28 Aug 2025)
Galactic structure	Lethe stellar stream	(Sohn et al., 2016)

The commonality is nominal rather than architectural. Some instances are software infrastructures, some are algorithms or system designs, and one is an astronomical substructure. This makes disambiguation essential in technical writing, especially because several of the instances are active in areas where “forgetting,” deletion, or erasure is a primary concern, while others are entirely unrelated to those themes.

2. Computational mechanics: the Lethe framework, ILES, and DEM

In computational mechanics, Lethe is an open-source, deal.II-based code for incompressible multiphase flows. For the periodic-hills benchmark, it solves the unsteady, incompressible Navier-Stokes equations in nondimensional form with a fully coupled, implicit finite-element discretization, using first-order $Q_1$ isoparametric hexahedral elements for both velocity and pressure and the SUPG/PSPG approach to circumvent the Ladyzhenskaya-Babuška-Brezzi constraint and control spurious oscillations in advection-dominated regimes. Time integration is performed with BDF2, and the stabilization parameter $\tau_u$ couples $\Delta t$, convection, and diffusion scales (Saavedra et al., 2022).

The periodic-hills study treats the case as an implicit large-eddy simulation benchmark at $Re=5600$, with additional runs at $Re=10600$ and $Re=37000$. The computational domain is discretized by a curvilinear, uniformly refined hexahedral mesh with five global-refinement levels, from approximately $120\,\mathrm{K}$ cells to $4\,\mathrm{M}$ cells. In $x$ and $z$ the elements are uniform, whereas in $\tau_u$0 the spacing is graded to cluster points near both walls. The results show good accuracy for average velocities and Reynolds stresses using fewer degrees of freedom than the reference numerical solution. On coarse meshes, decreasing $\tau_u$1 can reduce numerical dissipation and degrade turbulent statistics; on finer meshes, the solution becomes rapidly time-step independent, and accuracy remains essentially unchanged for $\tau_u$2 as large as $\tau_u$3 with $\tau_u$4. The computed reattachment point $\tau_u$5 approaches the experimental benchmark value as the mesh is refined, with the fine mesh yielding $\tau_u$6 against an experimental $\tau_u$7 (Saavedra et al., 2022).

Lethe-DEM is the discrete-element-method module of Lethe, implemented as two C++ template executables, dem_2d and dem_3d. It reuses Lethe’s mesh, MPI partitioning, load-balancing, I/O, and parameter-handling infrastructure. Its workflow is: initialize mesh, build cell-neighbour lists, insert particles, perform contact detection, evaluate forces, integrate in time, update mesh-particle mapping, load-balance, output or checkpoint, and repeat. The module solves Newton’s laws for translation and rotation of each particle, supports linear viscoelastic Hookean normal and tangential contact forces, Coulomb-limited tangential response, and a constant rolling-torque model, and provides both explicit Euler and velocity-Verlet integrators (Golshan et al., 2021).

A defining systems contribution of Lethe-DEM is dynamic load balancing. Each owned cell is assigned weight

$\tau_u$8

with default $\tau_u$9, and p4est repartitions the forest-of-trees to equalize $\Delta t$0 across MPI ranks. In the reported experiments, load balancing improves parallel efficiency by approximately $\Delta t$1, with the best scalability obtained for $\Delta t$2. Large-scale runs include a three-dimensional cylindrical silo simulation with $\Delta t$3 particles on 320 cores (Golshan et al., 2021).

The framework has also been used for additive-manufacturing powder spreading. Gaboriault et al. simulate up to $\Delta t$4 million spherical particles in a narrow domain with periodic boundary conditions in $\Delta t$5, representing the feedstock platform, measuring platform, vertical walls, and cylindrical recoater as rigid boundaries. In that setting, Lethe reproduces the experimentally observed densification beyond layers 5 to 10 and highlights significant particle displacement at low layer number together with static zones generated by the vertical walls surrounding the powder bed (Gaboriault et al., 30 Sep 2025).

3. Deletion privacy on social platforms

In the privacy-systems literature, Lethe is a lightweight, server-side mechanism intended to restore plausible deniability for user-initiated deletions. The setting is a public bulletin board where Data Owners publish posts and an Adversary can take arbitrary snapshots of the entire platform. The design assumes an honest platform operator and a fully informed adversary who knows the exact Lethe parameterization. The central idea is to interleave temporary withdrawals of non-deleted posts with their later resurrections, so that an absent post may be either permanently deleted or merely in a Lethe-imposed downtime (Minaei et al., 2017).

Formally, Lethe alternates publicly visible “up” states and hidden “down” states for each non-deleted post. Up-times $\Delta t$6 are drawn from a geometric distribution with mean $\Delta t$7, and down-times $\Delta t$8 are drawn from a negative binomial distribution with mean $\Delta t$9. Infinite support is required for both distributions to avoid adversarial exploitation of finite cutoffs. Availability is defined as

$Re=5600$0

typically set to $Re=5600$1. Deletion privacy is analyzed through a likelihood-ratio test over the observation history, while adversarial overhead is summarized by precision and recall under threshold-based flagging of absent posts (Minaei et al., 2017).

The empirical evaluation uses large-scale Twitter-derived measurements. Under $Re=5600$2 availability, mean down-time $Re=5600$3, and mean up-time $Re=5600$4, Lethe keeps adversarial precision below $Re=5600$5 for any decision threshold up to six months in both flag-once and flag-multi regimes. Even at $Re=5600$6 availability, adversarial precision never exceeds $Re=5600$7 for $Re=5600$8 days. The system is reported to provide deletion privacy for up to 3 months from the time of deletion while maintaining content availability as high as $Re=5600$9 and keeping adversarial precision to $Re=10600$0. On a corpus of $Re=10600$1 million tweets by $Re=10600$2 random users together with $Re=10600$3 million retweets, retained retweets exceed $Re=10600$4 even with only $Re=10600$5 availability, because engagement is heavily front-loaded in the first few hours after posting (Minaei et al., 2017).

The principal limitation identified in the paper is that the model treats posts independently. Correlational leakage across a user’s timeline, side-channel knowledge available to highly dedicated stalkers, and synchronization in geo-replicated deployments remain open problems (Minaei et al., 2017).

4. Delete-aware storage systems

In storage systems, Lethe is an LSM-based key-value engine that treats deletes as first-class operations rather than second-class tombstones. Standard LSM engines implement deletes by inserting tombstones that only become persistent once compaction pushes them to the largest level, so delete persistence latency is unbounded. Lethe addresses this with tunable delete persistence latency, early invalid-entry purging, and support for efficient range deletes on a secondary delete key (Sarkar et al., 2020).

The design adds very small in-memory metadata per file: $Re=10600$6, the age of the oldest tombstone in the file, and $Re=10600$7, an estimate of invalidations due to point and range deletes. Given a delete persistence latency threshold $Re=10600$8 and $Re=10600$9 disk levels, Lethe assigns each level a time-to-live $Re=37000$0 satisfying $Re=37000$1, with $Re=37000$2 growing exponentially with size ratio $Re=37000$3. To support secondary range deletes, the KiWi layout weaves the sort key order and delete key order: each file is partitioned into delete-tiles, tiles cover disjoint sort-key ranges, pages inside each tile are sorted by delete key, and records inside each page remain sorted by sort key (Sarkar et al., 2020).

The compaction subsystem, termed FADE, includes three policies: Saturation+Overlap, Saturation+Delete-Driven, and Delete-Driven+Delete-Driven. These policies use level saturation, tombstone age, and invalidation estimates to decide which files to compact so that every tombstone is made persistent within the application-defined threshold. The read path benefits because earlier tombstone removal reduces Bloom-filter false positives and shrinks the volume of invalid data still participating in query processing (Sarkar et al., 2020).

The evaluation is implemented atop RocksDB with $Re=37000$4, 10 bits per entry for Bloom filters, direct I/O, and no WAL. Lethe supports any user-defined threshold for delete persistence latency while offering higher read throughput of $Re=37000$5-$Re=37000$6 and lower space amplification of $Re=37000$7-$Re=37000$8, with a modest increase in write amplification between $Re=37000$9 and $120\,\mathrm{K}$0. It also supports efficient range deletes on a secondary delete key by dropping entire data pages without sacrificing read performance and without requiring a costly full-tree merge (Sarkar et al., 2020).

5. Machine unlearning and model security

In federated learning, LETHE is a federated unlearning method designed for the setting in which training continues after the unlearning request has been served. The motivating failure mode is termed knowledge resurfacing: after Phase U, continued training on retained data can reactivate unlearned knowledge, so the influence of the removed client-, class-, or sample-level data resurfaces in the global model. LETHE addresses this by de-correlating knowledge to be unlearned from knowledge to be retained through a Reshape-Rectify-Restore pipeline (Tan et al., 30 Jan 2026).

The method temporarily augments the backbone model with an adapter. In Reshape, the frozen backbone is paired with an adapter trained by gradient ascent on the unlearning data to obtain a magnified unlearning-direction probe. In Rectify, LETHE collects a forget-stream update from the unlearning set and a retain-stream aggregated update from the remaining clients, computes layer-wise cosine similarity, and rectifies each layer by subtracting the unlearning direction when the similarity is positive or replacing the retain update with the negative forget direction otherwise. In Restore, the adapter is discarded and a short FedAvg recovery stage is run on the retained clients. The final model has no extra parameters beyond the original backbone because the adapter is removed after rectification (Tan et al., 30 Jan 2026).

The reported experiments cover MNIST with LeNet-5, CIFAR-10 with ResNet-18, and Tiny-ImageNet with HSViT, at federation scales $120\,\mathrm{K}$1. LETHE supports client-level, class-level, and sample-level unlearning in a unified manner and achieves $120\,\mathrm{K}$2 in most cases, where RR is the Resurfacing Rate defined from pre-unlearning, post-unlearning, and post-continued-training accuracy on the unlearning set. Communication cost is typically $120\,\mathrm{K}$3 unlearning rounds plus $120\,\mathrm{K}$4 recovery rounds, versus more than $120\,\mathrm{K}$5 for full retraining (Tan et al., 30 Jan 2026).

A different LETHE appears in LLM security as a trigger-agnostic framework for purifying backdoored LLMs through knowledge dilution. Its threat model assumes that an untrusted ML-as-a-Service provider returns a potentially backdoored fine-tuned model $120\,\mathrm{K}$6, while the defender has white-box access to that model and up to $120\,\mathrm{K}$7 of clean training data. The method combines an internal mechanism, where a clean model trained with LoRA on the clean subset is merged with the backdoored model—defaulting to SLERP—with an external mechanism, where benign semantically relevant evidence is prepended to the prompt by extracting keywords with TextRank and retrieving definitions from WordNet (Chen et al., 28 Aug 2025).

The evaluation spans GPT2-XL, GPT-J, Llama, Llama-2, and DeepSeek-R1 across classification, chat generation, and code generation tasks, and compares against eight baselines. Internal+External LETHE reduces the attack success rate of advanced backdoor attacks by up to $120\,\mathrm{K}$8 while keeping clean-data accuracy drops below $120\,\mathrm{K}$9. Training the LoRA clean model on $4\,\mathrm{M}$0 of the data requires approximately $4\,\mathrm{M}$1-$4\,\mathrm{M}$2 minutes on an A100 per model, model merging via SLERP adds negligible extra compute, and external dilution requires no training (Chen et al., 28 Aug 2025).

These two ML usages share an emphasis on persistent removal rather than one-shot correction: one targets resurging memorized influence under continued optimization, the other targets latent backdoor behaviors under adversarial prompting.

6. The Lethe stellar stream

In Galactic-structure studies, Lethe is a stellar stream rather than a computational method. A multi-epoch Hubble Space Telescope study measured proper motions of individual stars in four ACS/WFC fields along the Orphan Stream, with Lethe candidates found in field ORPH-F4 at $4\,\mathrm{M}$3, $4\,\mathrm{M}$4 (J2000). First-epoch imaging in 2004.36 and second-epoch imaging in 2014.28 provide a baseline of approximately $4\,\mathrm{M}$5 years, and background galaxies define the stationary astrometric reference frame (Sohn et al., 2016).

Candidate selection in ORPH-F4 used a Dartmouth isochrone of age $4\,\mathrm{M}$6, $4\,\mathrm{M}$7, and distance $4\,\mathrm{M}$8, combined with a tight proper-motion clump near $4\,\mathrm{M}$9. Three stars satisfy both CMD and PM criteria. Their error-weighted mean observed proper motion is

$x$0

equivalently

$x$1

After subtraction of solar reflex motion at the field position and distance, the net motion becomes

$x$2

with magnitude $x$3 and tangential speed $x$4 (Sohn et al., 2016).

The one-dimensional tangential dispersion is $x$5 after deconvolving measurement errors. Together with the old, metal-poor population, this strongly favors a globular-cluster progenitor. The corrected proper-motion vector follows the great-circle path mapped by Grillmair, and the adopted distance $x$6 is consistent with Grillmair’s distance gradient from $x$7 in the east to $x$8 in the west. The paper does not present a full orbit model because radial velocities are unavailable, but the 2D kinematics are reported as fully consistent with Lethe’s published track (Sohn et al., 2016).

7. Comparative significance

Across these literatures, LETHE names at least four different kinds of scientific object: a family of open-source simulation frameworks, system mechanisms for deletion or persistence control, machine-learning methods for durable erasure or purification, and a cold stellar stream in the Milky Way halo. That diversity has practical consequences for citation, indexing, and literature review, because keyword-only search is insufficiently specific.

A recurring pattern is that several computational uses of the name are organized around controlled disappearance or suppression. The social-platform mechanism attempts to conceal true deletion among temporary withdrawals (Minaei et al., 2017); the LSM engine bounds delete persistence latency and physically purges invalidated data (Sarkar et al., 2020); the federated unlearning method seeks persistent erasure under continued training (Tan et al., 30 Jan 2026); and the LLM defense neutralizes backdoor behaviors by diluting malicious parametric memory and distracting attention from backdoor features (Chen et al., 28 Aug 2025). This suggests a loose thematic affinity around durable forgetting, although the CFD/DEM framework and the stellar stream are semantically independent of that motif.

In the mechanics papers, by contrast, Lethe denotes a reusable open-source numerical infrastructure built on deal.II and p4est, with applications ranging from periodic-hills ILES to million-particle granular-flow and powder-spreading simulations (Saavedra et al., 2022, Golshan et al., 2021, Gaboriault et al., 30 Sep 2025). In astronomy, it denotes an observed halo substructure whose measured proper motions support a globular-cluster origin (Sohn et al., 2016). The name therefore functions less as a stable technical term than as a label reused across distinct research traditions.