Tula in Technical Research: A Polysemous Term

• Tula is a polysemous technical term that encompasses distinct innovations in distributed deep learning, mmWave communications, language-model privacy, robotics, and stochastic sampling.
• In distributed training, Tula optimizes large-batch training by predicting optimal configurations and rescaling gradients to mitigate generalization gaps, while in mmWave systems it refines antenna beam patterns.
• In privacy and robotics, TULA exposes vulnerabilities in language-model unlearning and innovates actuator designs for flapping-wing robots, with its Langevin variants taming superlinear drift for stable sampling.

Tula is a polysemous technical term used in several unrelated research literatures. In recent arXiv usage it denotes an online service for optimizing distributed large-batch training of convolutional models (Tyagi et al., 18 Mar 2026), a twin-ULA antenna structure for composite beamforming in mmWave systems (Torkzaban et al., 2022), the Textual Unlearning Leakage Attack against language-model unlearning mechanisms (Du et al., 2024), a Tiny Ultrasonic Linear Actuator used in bat-inspired flapping-wing robotics (Ciampaglia, 20 Apr 2026), and the Tamed Unadjusted Langevin Algorithm together with later variants such as mTULA and TUSLA for stable sampling and stochastic optimization under superlinear drifts (Neufeld et al., 2022, Lovas et al., 2020). The shared label is therefore lexical rather than conceptual: each usage belongs to a distinct systems, communications, privacy, robotics, or stochastic-analysis lineage.

1. Principal technical senses of the term

Research area	Meaning of Tula	Representative paper
Distributed DL systems	Online service for large-batch configuration and optimizer adaptation	(Tyagi et al., 18 Mar 2026)
mmWave beamforming	Twin-ULA antenna structure	(Torkzaban et al., 2022)
LM privacy	Textual Unlearning Leakage Attack	(Du et al., 2024)
Flapping-wing robotics	Tiny Ultrasonic Linear Actuator, notably TULA-50	(Ciampaglia, 20 Apr 2026)
Langevin methods	Tamed Unadjusted Langevin Algorithm and variants	(Neufeld et al., 2022, Lovas et al., 2020)

The coexistence of these meanings matters because the same acronym labels objects of very different ontological types: a controller, an antenna geometry, an attack, an actuator family, and an MCMC discretization. In technical reading, disambiguation usually follows immediately from neighboring terms such as “large-batch training,” “twin-ULA,” “unlearning,” “TULA-50,” or “Langevin.”

2. Tula as a controller for distributed large-batch training

In distributed training, Tula is an online, black-box service for choosing how to train a deep network with large batches on a distributed cluster so as to obtain near-minimal training time or cost without paying an unnecessary generalization penalty (Tyagi et al., 18 Mar 2026). It combines parallel-systems modeling with statistical performance prediction, then outputs an optimal batch size and cluster configuration together with a modified optimizer based on Adaptive Gradient Scaling. The motivating setting is synchronous data-parallel SGD with global batch size $B = Nb$, where increasing $B$ can be achieved either by horizontal scaling or by increasing the per-GPU batch.

The system-side rationale is that larger configurations are not uniformly better. Per-step compute time grows roughly linearly with local batch size, synchronization time grows with worker count, and epoch time behaves approximately as

$$T_{\text{epoch}} \approx \frac{|D|}{B}\,(t_{\text{comp}}(b)+t_{\text{sync}}(N)).$$

This yields an initial speedup regime followed by diminishing returns and then a plateau, producing a knee-point in the time- or cost-versus-batch-size Pareto curve. The paper reports this behavior explicitly for CNN workloads: on a 16-node cluster the knee points of time versus batch size are roughly near $B \approx 8192$ for ResNet50 and VGG11, $B \approx 1024$ for AlexNet, and $B \approx 2048$ for MobileNetv3 (Tyagi et al., 18 Mar 2026).

The optimization-side rationale is the large-batch generalization gap. Tula models sensitive phases of training through the relative change in squared gradient norm,

$$\Delta(G^{(i)}_{(b)}) = \left| \frac{\|G^{(i)}_{(b)}\|^2-\|G^{(i-1)}_{(b)}\|^2} {\|G^{(i-1)}_{(b)}\|^2} \right|,$$

and uses this statistic to decide when large-batch gradients should be left unchanged and when they should be rescaled. AGS updates per-parameter scaling factors by comparing large-batch and small-batch gradients, while capping the scale by $s_{\max}=\sqrt{b_{\text{large}}/b_{\text{small}}}$ (Tyagi et al., 18 Mar 2026). The intent is not to inject synthetic noise, but to transform large-batch updates so that they more closely resemble the optimization behavior of smaller batches.

Implementation is reported for PyTorch 2.6 with DistributedDataParallel, NCCL ring all-reduce, and up to 16 V100 GPUs. Tula profiles a small number of configurations, predicts memory with 8–14% median error, predicts time with 2–5% median error under full search and 7.5–14% under partial search, and then selects the time-optimal, cost-optimal, or knee-point configuration (Tyagi et al., 18 Mar 2026). Across multiple vision workloads it predicts training time and cost within 7.5–14% error, achieves up to 20x overall speedup, and improves test accuracy by about 8.8–9% on average over vanilla large-batch training at the same batch size (Tyagi et al., 18 Mar 2026). A common misconception in this setting is that “max batch, max nodes” is necessarily optimal; Tula formalizes why that heuristic fails once communication overhead, memory growth, and generalization degradation are jointly modeled.

3. TULA as a twin-ULA antenna structure for composite beamforming

In mmWave communications, TULA denotes the twin-ULA antenna structure introduced to address a limitation of conventional uniform linear arrays when constructing composite beams with multiple disjoint lobes (Torkzaban et al., 2022). The target use case is codebook design for beams that cover several angular coverage intervals simultaneously while maintaining high and uniform in-band gain, low out-of-band leakage, and sharp edges.

The structural definition is explicit: TULA consists of two identical ULAs of size $M_t/2$ each, oriented parallel to the $x$-axis and separated in the $B$0-direction by $B$1 (Torkzaban et al., 2022). Its steering vector is written as

$B$2

with $B$3 induced by the vertical displacement. The codeword uses the same beamforming vector on both half-arrays up to a phase offset $B$4,

$B$5

so that the resulting gain factorizes into the ULA pattern for half the aperture times an additional shaping factor $B$6 (Torkzaban et al., 2022).

The motivation is that the ULA antenna structure inherently generates two-sided lobes and hence inefficient beams due to having beam lobes in undesired scopes and lower effective beam gain in desired ACIs (Torkzaban et al., 2022). TULA uses the extra phase degree of freedom $B$7 to suppress the mirrored lobe. The paper formalizes this through an isolation factor $B$8, where smaller $B$9 indicates better isolation. The design parameter $$T_{\text{epoch}} \approx \frac{|D|}{B}\,(t_{\text{comp}}(b)+t_{\text{sync}}(N)).$$0 in the closed-form ULA precursor controls a three-way trade-off between smoothness, gain, and leakage, and $$T_{\text{epoch}} \approx \frac{|D|}{B}\,(t_{\text{comp}}(b)+t_{\text{sync}}(N)).$$1 is used in most simulations because it gives the smoothest beam gain with an acceptable in-band gain (Torkzaban et al., 2022).

The resulting codebook construction is analytic rather than purely numerical. It first derives a closed-form ULA beam that approximates an ideal constant-gain pattern over the desired ACIs, then embeds that beam into the twin-ULA architecture. Numerical results in the paper show that TULA produces single-sided composite beams with high, stable gain in desired ACIs and significantly reduced parasitic lobes relative to a conventional ULA (Torkzaban et al., 2022). The paper also states that TULA can improve beam gain by almost 3 dB, citing earlier TULA work through its own reference chain. Hybrid and quantized-hybrid realizations are further constructed with OMP; the reported quantized-hybrid design at $$T_{\text{epoch}} \approx \frac{|D|}{B}\,(t_{\text{comp}}(b)+t_{\text{sync}}(N)).$$2 bits nearly matches the unquantized hybrid case (Torkzaban et al., 2022). A plausible implication is that TULA should be understood less as a beamforming algorithm in isolation than as an antenna-architecture intervention that changes the feasible beam-pattern manifold.

4. TULA as a privacy attack on textual unlearning

In language-model privacy research, TULA is the Textual Unlearning Leakage Attack, introduced to show that approximate textual unlearning can itself create a strong side channel (Du et al., 2024). The central claim is that when an original model $$T_{\text{epoch}} \approx \frac{|D|}{B}\,(t_{\text{comp}}(b)+t_{\text{sync}}(N)).$$3 is transformed into an unlearned model $$T_{\text{epoch}} \approx \frac{|D|}{B}\,(t_{\text{comp}}(b)+t_{\text{sync}}(N)).$$4, the difference between the two models can reveal which texts were unlearned and, in white-box settings, can reveal the text content itself.

The black-box variant, TULA-black, is a membership-inference attack based on loss differences. For a candidate sample $$T_{\text{epoch}} \approx \frac{|D|}{B}\,(t_{\text{comp}}(b)+t_{\text{sync}}(N)).$$5, the decision rule is

$$T_{\text{epoch}} \approx \frac{|D|}{B}\,(t_{\text{comp}}(b)+t_{\text{sync}}(N)).$$6

The intuition is operational rather than abstract: if a sample was explicitly targeted by gradient-ascent-style unlearning, its loss should increase substantially in the updated model, while losses on unrelated samples should change far less (Du et al., 2024). On GPT-2-1.5B, Pythia-1.4B, and OPT-1.3B fine-tuned on SST and Yelp tasks, the paper reports that TULA-black improves membership inference AUROC from near-random baselines around 0.42–0.51 to roughly 0.64–0.72, with examples of 0.723 on SST and 0.722 on Yelp for GPT-2-1.5B (Du et al., 2024).

The white-box variant, TULA-white, uses the unlearning gap in parameter space,

$$T_{\text{epoch}} \approx \frac{|D|}{B}\,(t_{\text{comp}}(b)+t_{\text{sync}}(N)).$$7

and solves a gradient-matching optimization in continuous embedding space:

$$T_{\text{epoch}} \approx \frac{|D|}{B}\,(t_{\text{comp}}(b)+t_{\text{sync}}(N)).$$8

After optimizing the continuous representation, the attack projects back to discrete tokens through nearest-neighbor lookup in the embedding matrix (Du et al., 2024). Reported reconstruction quality is often substantial: on SST with Pythia-1.4B, ROUGE-1 is 66.374 and ROUGE-L is 58.497, while on Yelp with the same model ROUGE-1 is 57.931 and ROUGE-L is 51.855 (Du et al., 2024). The paper interprets this as evidence that important words, semantic content, and sentence outline can remain recoverable even after unlearning.

A notable point of terminology is that the abstract also names U-LiRA+, while the detailed exposition centers on TULA-black and TULA-white. This suggests that the paper’s main privacy argument is carried by loss-difference auditing and weight-difference reconstruction rather than by a formal LiRA-style likelihood-ratio derivation. The controversy here is intrinsic to the subject: textual unlearning is often assumed to mitigate privacy risk, whereas TULA argues that exposing pre- and post-unlearning models can instead amplify membership inference and data reconstruction leakage (Du et al., 2024).

5. TULA-50 as a piezoelectric slip-stick actuator in flapping-wing robotics

In the Aerobat thesis, TULA refers to the TULA-50, a Tiny Ultrasonic Linear Actuator used as a piezoelectric slip-stick linear actuator for linkage-length modulation in a bat-inspired flapping-wing robot (Ciampaglia, 20 Apr 2026). The actuator was evaluated as part of an attempt to achieve independent wing thrust control by modifying the effective length of the first radius link $$T_{\text{epoch}} \approx \frac{|D|}{B}\,(t_{\text{comp}}(b)+t_{\text{sync}}(N)).$$9 in the robot’s computational structure.

The experimental motivation is precise. Static experiments with FDM-printed $B \approx 8192$0 links at lengths 28.58, 29.33, and 30.08 mm across 3, 4, and 5 Hz flapping frequencies showed that a 1.5 mm length increase produced a 37% increase in peak lift force and shifted peak force timing within the downstroke (Ciampaglia, 20 Apr 2026). At 5 Hz, the shortest $B \approx 8192$1 mm yielded peak lift around 23% into the downstroke, whereas the longest $B \approx 8192$2 mm shifted the peak to about 40% into the downstroke (Ciampaglia, 20 Apr 2026). This established that small geometric perturbations in $B \approx 8192$3 materially alter both force magnitude and force timing.

The TULA-50 was selected after two alternatives failed. A string-tension regulator with a Dynamixel servo failed dynamically because of structural compliance, friction, and synchronization problems, while a 0.5 g micro-servo failed due to gear fragility, PCB damage, and unreliable PWM behavior (Ciampaglia, 20 Apr 2026). The TULA-50 was attractive because it is a piezoelectric slip-stick drive with linear output, very low mass, passive holding through static friction, and direct compatibility with the required link-length adjustment. The mechanism design assumed a 6 mm actuator stroke, needed a 1.5 mm regulator stroke, and derived a required axial regulator force of approximately $B \approx 8192$4, whereas the thesis reports the tested TULA-50 at about 20 grams-force maximum dynamic output (Ciampaglia, 20 Apr 2026).

This large force mismatch structured the rest of the design. A first force-amplifying triangular mechanism achieved an ideal mechanical advantage greater than 4 at minimum and greater than 13 near the start of the stroke, which in theory could bring nominal output into the required range, but internal friction in the SLA-printed prototype prevented actuation (Ciampaglia, 20 Apr 2026). A later direct-drive variable-length mechanism reduced friction and successfully traversed the full stroke in free space, yet when integrated into Aerobat Delta it could not reliably cover the full stroke under linkage load and did not have enough force for dynamic flapping tests (Ciampaglia, 20 Apr 2026). The thesis therefore concludes both that the specific TULA-50 implementation was force-limited and that linkage-length modulation via embedded slip-stick actuation remains a viable approach to independent wing thrust control (Ciampaglia, 20 Apr 2026). This suggests that the core feasibility claim concerns the control architecture, while actuator sizing and compliant-mechanism design remain open engineering bottlenecks.

6. TULA as the Tamed Unadjusted Langevin Algorithm and its descendants

In stochastic sampling and optimization, TULA denotes the Tamed Unadjusted Langevin Algorithm, an explicit Langevin discretization designed for target distributions with superlinearly growing gradients (Neufeld et al., 2022, Lovas et al., 2020). The continuous target is the overdamped Langevin SDE with invariant density proportional to $B \approx 8192$5, and the classical ULA update is unstable when the drift grows faster than linearly. TULA addresses this by replacing the raw gradient with a tamed drift, thereby preventing Euler-type explosions under non-globally-Lipschitz dynamics.

Two later developments in the supplied literature clarify the lineage. The paper on mTULA proposes the modified Tamed Unadjusted Langevin Algorithm,

$B \approx 8192$6

and analyzes it under a non-convex but convex-at-infinity regime with polynomial Lipschitz gradient and Hessian assumptions (Neufeld et al., 2022). The main non-asymptotic results are explicit Wasserstein bounds:

$B \approx 8192$7

and

$B \approx 8192$8

so the discretization bias is $B \approx 8192$9 in Wasserstein-1 and $B \approx 1024$0 in Wasserstein-2 (Neufeld et al., 2022). High-dimensional experiments on a Gaussian, a Gaussian mixture, and a double-well potential support these rates and illustrate stability in a superlinear, non-convex case.

The TUSLA paper extends the same taming philosophy from deterministic gradients to stochastic gradients in non-convex neural-network learning (Lovas et al., 2020). Its update is

$B \approx 1024$1

where $B \approx 1024$2 includes a high-order regularization term $B \approx 1024$3 (Lovas et al., 2020). TUSLA is therefore “SGLD + TULA-style taming + high-order regularization,” and the paper gives finite-time Wasserstein and excess-risk guarantees for non-convex learning problems with superlinear gradient growth (Lovas et al., 2020). Empirically, it is reported as robust on synthetic superlinear objectives where unadjusted SGLD diverges, and competitive with ADAM, AMSGrad, and RMSprop on Fashion-MNIST (Lovas et al., 2020).

Across these papers, the core meaning of TULA is consistent: taming is a stability device for explicit Langevin schemes when classical Euler discretizations are invalidated by superlinear growth. The differences lie in scope. Original TULA is the MCMC ancestor, mTULA improves non-asymptotic convergence rates under non-convex “convex at infinity” assumptions, and TUSLA transports the same technology to stochastic-gradient learning in neural networks (Neufeld et al., 2022, Lovas et al., 2020).