Surgical Reinitialization

• Surgical reinitialization is a context-dependent process that selectively resets system components while preserving invariant structures like level-set interfaces and model references.
• It is applied across diverse domains, including mathematical PDE redistancing, intraoperative model updates in sinus surgery, transformer attention repair, and soft robotic mode switching.
• The approach balances targeted intervention with preservation constraints to restore functionality without discarding the overall accumulated system state.

Searching arXiv for recent and directly relevant papers on “surgical reinitialization” and closely related usage contexts. “Surgical reinitialization” is a context-dependent research term denoting a selective reset, correction, or rewriting step applied during an ongoing process in order to restore a desired operational state without discarding the entire system state. In current arXiv usage, the term appears in at least four distinct senses: intermittent restoration of the signed-distance property in first-order level-set evolution (Hamamuki et al., 2015), continuous intraoperative updating of a preoperative three-dimensional sinus model from endoscopic video (Mangulabnan et al., 2024), targeted repair of collapsed attention heads in ALiBi-based transformers (Schallon, 10 Mar 2026), and mid-procedure magnetic rewriting of functional modes in a miniature soft surgical robot (Ng et al., 19 Sep 2025). This suggests a shared procedural motif—localized intervention under preservation constraints—rather than a single standardized method.

1. Terminological scope and invariant-preserving reset

Across these usages, reinitialization is not a full restart. In the level-set setting, the object being corrected is the level-set function, while the evolving zero level set $\Gamma_t$ must remain unchanged. In navigated sinus surgery, the object being updated is the preoperative TSDF model, while unchanged regions are explicitly protected by region-of-interest-restricted fusion. In ALiBi transformers, only selected Q/K/V projections and the corresponding output projection are reset, while all non-surgical parameters are frozen. In the soft robot, only the reprogrammable module and remote heating component are magnetized or demagnetized on demand, while higher-coercivity NdFeB-based domains retain their magnetization profiles.

A common misconception is to equate reinitialization with global replacement of state. The recorded uses are instead selective and constraint-driven. In each case, the method is designed around a preserved structure: an interface, a navigation reference tied to current anatomy, a pretrained residual-stream backbone, or a locomotion-capable magnetic body. The technical problem is therefore not merely to “reset,” but to do so without violating the structure that must remain valid.

2. First-order level-set equations and signed-distance restoration

In the mathematical usage developed for first-order Hamilton–Jacobi level-set equations, surgical reinitialization addresses the fact that a level-set function typically loses the signed-distance property during evolution and may develop small gradients near the interface, degrading robustness. The baseline interface evolution is defined by

$$\begin{cases} w_t = H_1(x,t,\nabla w) & \text{in } \mathbf{R}^n\times (0,T),\ w(x,0)=u_0(x) & \text{in } \mathbf{R}^n, \end{cases}$$

with evolving interface $\Gamma_t=\{x:w(x,t)=0\}$ and signed distance

$$d(x,t)= \begin{cases} \mathrm{dist}(x,\Gamma_t), & x\in D_t^+\cup \Gamma_t,\ -\mathrm{dist}(x,\Gamma_t), & x\in D_t^-. \end{cases}$$

Surgical reinitialization means intermittently and possibly locally applying a corrector PDE to drive $|\nabla u|\to 1$ without moving $\Gamma_t$ (Hamamuki et al., 2015).

The rigorous framework introduces the parameterized augmented equation

$$\begin{cases} u_t^\theta = H_1(x,t,\nabla u^\theta)+\theta\,\beta(u^\theta)\,h(\nabla u^\theta) & \text{in } \mathbf{R}^n\times(0,T),\ u^\theta(x,0)=u_0(x) & \text{in } \mathbf{R}^n, \end{cases}$$

where $\beta$ is a smoothed sign function with $\beta(0)=0$, and $h$ is chosen so that it is positive for $$\begin{cases} w_t = H_1(x,t,\nabla w) & \text{in } \mathbf{R}^n\times (0,T),\ w(x,0)=u_0(x) & \text{in } \mathbf{R}^n, \end{cases}$$0 and either negative or zero for $$\begin{cases} w_t = H_1(x,t,\nabla w) & \text{in } \mathbf{R}^n\times (0,T),\ w(x,0)=u_0(x) & \text{in } \mathbf{R}^n, \end{cases}$$1. Concrete examples are $$\begin{cases} w_t = H_1(x,t,\nabla w) & \text{in } \mathbf{R}^n\times (0,T),\ w(x,0)=u_0(x) & \text{in } \mathbf{R}^n, \end{cases}$$2 and $$\begin{cases} w_t = H_1(x,t,\nabla w) & \text{in } \mathbf{R}^n\times (0,T),\ w(x,0)=u_0(x) & \text{in } \mathbf{R}^n, \end{cases}$$3. The parameter $$\begin{cases} w_t = H_1(x,t,\nabla w) & \text{in } \mathbf{R}^n\times (0,T),\ w(x,0)=u_0(x) & \text{in } \mathbf{R}^n, \end{cases}$$4 is not heuristic: it arises from homogenization of an alternating-time scheme in which the physical evolution and the corrector are applied over short subintervals, with $$\begin{cases} w_t = H_1(x,t,\nabla w) & \text{in } \mathbf{R}^n\times (0,T),\ w(x,0)=u_0(x) & \text{in } \mathbf{R}^n, \end{cases}$$5 equal to the ratio of time spent in the corrector to the time spent in the physical step.

The central convergence theorem states that as $$\begin{cases} w_t = H_1(x,t,\nabla w) & \text{in } \mathbf{R}^n\times (0,T),\ w(x,0)=u_0(x) & \text{in } \mathbf{R}^n, \end{cases}$$6, the solution $$\begin{cases} w_t = H_1(x,t,\nabla w) & \text{in } \mathbf{R}^n\times (0,T),\ w(x,0)=u_0(x) & \text{in } \mathbf{R}^n, \end{cases}$$7 converges to the signed distance function to the evolving interface. The convergence is locally uniform in space for each fixed time, and locally uniform in spacetime when the distance function $$\begin{cases} w_t = H_1(x,t,\nabla w) & \text{in } \mathbf{R}^n\times (0,T),\ w(x,0)=u_0(x) & \text{in } \mathbf{R}^n, \end{cases}$$8 is continuous. When $$\begin{cases} w_t = H_1(x,t,\nabla w) & \text{in } \mathbf{R}^n\times (0,T),\ w(x,0)=u_0(x) & \text{in } \mathbf{R}^n, \end{cases}$$9 may be discontinuous, the paper introduces a weaker directional convergence, expressed through half-relaxed limits from below in time. This is one of the technically distinctive features of the framework: the target object is not assumed to be everywhere continuous, and the notion of convergence is adapted accordingly.

The continuity of $\Gamma_t=\{x:w(x,t)=0\}$0 is governed by a geometric obstruction: extinction points of the interface. For $\Gamma_t=\{x:w(x,t)=0\}$1, an extinction point is one for which a whole neighborhood becomes strictly positive immediately after time $\Gamma_t=\{x:w(x,t)=0\}$2. The paper proves that $\Gamma_t=\{x:w(x,t)=0\}$3 is continuous at $\Gamma_t=\{x:w(x,t)=0\}$4 if and only if at least one nearest point to $\Gamma_t=\{x:w(x,t)=0\}$5 on $\Gamma_t=\{x:w(x,t)=0\}$6 is a non-extinction point. This criterion links analytical convergence directly to interface geometry and explains why topology change can limit spacetime-uniform redistancing.

The framework also establishes preservation of the zero level set for augmented equations of the form

$\Gamma_t=\{x:w(x,t)=0\}$7

under mild lower-boundedness assumptions on $\Gamma_t=\{x:w(x,t)=0\}$8. Thus the corrector can restore the distance structure without displacing the interface. A simpler preferred corrector is obtained with $\Gamma_t=\{x:w(x,t)=0\}$9, yielding

$$d(x,t)= \begin{cases} \mathrm{dist}(x,\Gamma_t), & x\in D_t^+\cup \Gamma_t,\ -\mathrm{dist}(x,\Gamma_t), & x\in D_t^-. \end{cases}$$0

This version does not act when $$d(x,t)= \begin{cases} \mathrm{dist}(x,\Gamma_t), & x\in D_t^+\cup \Gamma_t,\ -\mathrm{dist}(x,\Gamma_t), & x\in D_t^-. \end{cases}$$1, acts with the sign of $$d(x,t)= \begin{cases} \mathrm{dist}(x,\Gamma_t), & x\in D_t^+\cup \Gamma_t,\ -\mathrm{dist}(x,\Gamma_t), & x\in D_t^-. \end{cases}$$2 when $$d(x,t)= \begin{cases} \mathrm{dist}(x,\Gamma_t), & x\in D_t^+\cup \Gamma_t,\ -\mathrm{dist}(x,\Gamma_t), & x\in D_t^-. \end{cases}$$3, “prevents the gradient of the solution to approach zero on the zero level set,” and admits a simple monotone numerical scheme with a CFL condition. The formulation also rigorously connects to the classical Sussman–Smereka–Osher redistancing PDE by showing that alternating physical evolution and corrector steps converges, in the homogenization limit, to the averaged equation above.

3. Intraoperative reinitialization of anatomical models in sinus surgery

In navigated functional endoscopic sinus surgery, surgical reinitialization denotes continuous updating of the preoperative three-dimensional anatomical model during tissue removal so that the navigation reference reflects current rather than preoperative anatomy. The method takes a CT-derived TSDF volume and mesh, known endoscope poses from optical tracking and registration, undistorted endoscopic RGB frames, and monocular depth estimates from an endoscopy-specific depth network fine-tuned on the preoperative sequence. The preoperative model is rendered at the current poses, intraoperative monocular depth is estimated at the same poses, and the two depth sources are compared after scale recovery and RANSAC-based ICP alignment of the back-projected point clouds (Mangulabnan et al., 2024).

The detection of anatomical change is depth-based. After registration, the aligned intraoperative depth $$d(x,t)= \begin{cases} \mathrm{dist}(x,\Gamma_t), & x\in D_t^+\cup \Gamma_t,\ -\mathrm{dist}(x,\Gamma_t), & x\in D_t^-. \end{cases}$$4 is compared to the rendered depth $$d(x,t)= \begin{cases} \mathrm{dist}(x,\Gamma_t), & x\in D_t^+\cup \Gamma_t,\ -\mathrm{dist}(x,\Gamma_t), & x\in D_t^-. \end{cases}$$5, and a discrepancy map

$$d(x,t)= \begin{cases} \mathrm{dist}(x,\Gamma_t), & x\in D_t^+\cup \Gamma_t,\ -\mathrm{dist}(x,\Gamma_t), & x\in D_t^-. \end{cases}$$6

is thresholded to obtain a binary change mask with $$d(x,t)= \begin{cases} \mathrm{dist}(x,\Gamma_t), & x\in D_t^+\cup \Gamma_t,\ -\mathrm{dist}(x,\Gamma_t), & x\in D_t^-. \end{cases}$$7. Missing values in the aligned depth are imputed from the rendered depth to ensure dense overlap. Only pixels in the change mask contribute new information to the volumetric update. This is a strict ROI-restricted fusion policy: unchanged geometry is not overwritten.

The volumetric integration step uses a truncated signed distance observation

$$d(x,t)= \begin{cases} \mathrm{dist}(x,\Gamma_t), & x\in D_t^+\cup \Gamma_t,\ -\mathrm{dist}(x,\Gamma_t), & x\in D_t^-. \end{cases}$$8

with weighted fusion

$$d(x,t)= \begin{cases} \mathrm{dist}(x,\Gamma_t), & x\in D_t^+\cup \Gamma_t,\ -\mathrm{dist}(x,\Gamma_t), & x\in D_t^-. \end{cases}$$9

where $|\nabla u|\to 1$0 only in the detected modified region. The updated TSDF is then converted to a mesh via Marching Cubes. The same cycle is repeated sequentially over five surgical steps.

The reported evaluation is ex vivo and uses five sequential “bites” of a lamella. Each step uses the longest tool-free sequence, with 70–90 frames. The metric is mean 3D point-to-point distance between correspondences in a constant bounding box around the modified region, with intraoperative CT rigidly registered to preoperative CT and residual registration error on unmodified regions of 1.832 mm. Without update, the mean error increases from $|\nabla u|\to 1$1 mm at Bite 1 to $|\nabla u|\to 1$2 mm at Bite 5. With the proposed update, the corresponding values are $|\nabla u|\to 1$3 mm at Bite 1 and $|\nabla u|\to 1$4 mm at Bite 5. A depth ablation using intraoperative CT-rendered depths instead of monocular estimates yields smaller errors, from $|\nabla u|\to 1$5 mm to $|\nabla u|\to 1$6 mm, while removing the RANSAC-ICP alignment produces substantially worse values, from $|\nabla u|\to 1$7 mm to $|\nabla u|\to 1$8 mm. Longer sequences reduce update error.

The technical significance lies in maintaining correspondence between virtual and physical anatomy as tissue is resected, thereby counteracting the failure mode in which preoperative-only navigation places the camera trajectory “inside bone” that has already been removed. The limitations are also explicit: the study is ex vivo; endoscopic depth remains sensitive to specularities, low texture, shadows, occlusions, and appearance change; tool-free frames are required; and runtime, voxel resolution, truncation distance, and memory or GPU requirements are not reported. The paper frames the resulting continuously updated patient model as a step toward a digital twin paradigm for sinus surgery.

4. Transformer repair: collapsed ALiBi heads and targeted parameter reset

In transformer LLMs, Surgical Reinitialization is a repair procedure for a specific pathology: attention collapse in ALiBi-based self-attention, where heads assign most of their probability mass to the beginning-of-sequence token and become functionally inert. Collapsed heads empirically have BOS mass above $|\nabla u|\to 1$9 and attention entropy near zero. In the BLOOM family, the pathology is systematic and head-index-dependent, aligning with the ALiBi slope schedule. Across BLOOM-560m, BLOOM-1b7, BLOOM-3b, and BLOOM-7b1, collapse rates lie between 31% and 44%, with concentration in upper head indices: H9–H15 in 16-head models and H20–H30 or H21–H30 in 32-head models (Schallon, 10 Mar 2026).

The diagnostic quantities are

$\Gamma_t$0

and

$\Gamma_t$1

Heads are classified as Healthy when BOS mass $\Gamma_t$2, BOS-sink when $\Gamma_t$3, and DEAD when BOS mass $\Gamma_t$4 and entropy is near zero. Operational head capacity is defined as the number of Healthy heads. In stock BLOOM-1b7, which has 384 heads, the counts are Healthy 242, BOS-sink 136, DEAD 3, and low-entropy 3.

The repair algorithm is explicitly targeted. First, collapsed heads are selected using BOS mass and entropy thresholds. Second, the Q, K, and V projections of each targeted head are reinitialized with Xavier normal initialization. Third, the head’s output projection $\Gamma_t$5 is set to zero at $\Gamma_t$6 and kept trainable, so the head can relearn without immediately injecting random values into the residual stream. Fourth, all non-surgical parameters are frozen by gradient masking:

$\Gamma_t$7

with mask $\Gamma_t$8 equal to 1 on surgical parameters and 0 elsewhere. Training is then performed only on the surgical subset, typically for 1–3 epochs.

The implementation reported for BLOOM-1b7 uses AdamW with learning rate $\Gamma_t$9 and weight decay $$\begin{cases} u_t^\theta = H_1(x,t,\nabla u^\theta)+\theta\,\beta(u^\theta)\,h(\nabla u^\theta) & \text{in } \mathbf{R}^n\times(0,T),\ u^\theta(x,0)=u_0(x) & \text{in } \mathbf{R}^n, \end{cases}$$0, bfloat16 precision, gradient clipping at 1.0, gradient accumulation 8, sequence length 512, linear warmup followed by cosine decay, and gradient checkpointing. On a single NVIDIA RTX 5070 Ti with 16GB VRAM, Pass 1 targets 108 heads in the H9–H15 band across layers 5–22 and trains 302M parameters, 17.5% of the model. By epoch 3, all 108 targeted heads recover, raising Healthy heads from 242 to 341. Pass 2 targets the remaining 39 collapsed heads outside the main band, trains 40M parameters, 2.3% of the model, and reaches 379 Healthy heads, 98.7% capacity, with BOS-sink reduced to 1 and DEAD reduced to 0.

The controlled comparison between a curated research corpus and the C4 validation split isolates method from corpus content. Head recovery is identical—108 out of 108 by epoch 3 in both conditions—showing that reinitialization rather than corpus composition drives recovery. Training perplexity differs markedly: curated epoch 3 is approximately 15.4, whereas C4 epoch 3 is 20.80, with C4 overfitting from about epoch 6 and rising to 36.31 and eventually above 85 by epoch 15. On 50 held-out C4 texts, the C4 surgical model improves over stock, 29.30 versus 32.42. The paper also distinguishes two post-surgical phenomena: early global functional redistribution, which improves training behavior, and late local degradation under noisy training signals, especially in the steep-slope H15 column. An extended experiment that additionally reinitializes the mostly healthy H5 column yields a transient training perplexity of 12.70, compared with 16.99 for stock BLOOM-1b7, suggesting that pretrained attention configurations can occupy suboptimal local minima.

5. Magnetic function rewriting in a miniature soft surgical robot

In miniature soft robotics, surgical reinitialization refers to mid-procedure rewriting of a robot’s functional mode by magnetizing or demagnetizing selected soft-magnetic components. The robot is millimeter-scale, 4.4 mm long, and supports five surgical functionalities—drug dispensing, cutting through biological tissues simulated with gelatin, gripping, storing biological samples, and remote heating—while retaining full six degrees of freedom for locomotion. Reinitialization operates by writing or erasing the magnetization of a reprogrammable module and a remote heating component, while the NdFeB-based tentacles, inner beams, and sixth-DOF enhancement module retain their magnetization because their coercivity exceeds the reprogramming fields (Ng et al., 19 Sep 2025).

Mode switching is quantitative. Magnetization uses a uniform field $$\begin{cases} u_t^\theta = H_1(x,t,\nabla u^\theta)+\theta\,\beta(u^\theta)\,h(\nabla u^\theta) & \text{in } \mathbf{R}^n\times(0,T),\ u^\theta(x,0)=u_0(x) & \text{in } \mathbf{R}^n, \end{cases}$$1 mT applied for 5 ms along the desired functional axis, while demagnetization uses an alternating field at 45 Hz whose amplitude decays linearly from 65 mT to 0 mT at 2 mT per period. Each reprogram operation takes less than 1 s. The functional mode is encoded by the magnetization direction angle $$\begin{cases} u_t^\theta = H_1(x,t,\nabla u^\theta)+\theta\,\beta(u^\theta)\,h(\nabla u^\theta) & \text{in } \mathbf{R}^n\times(0,T),\ u^\theta(x,0)=u_0(x) & \text{in } \mathbf{R}^n, \end{cases}$$2: $$\begin{cases} u_t^\theta = H_1(x,t,\nabla u^\theta)+\theta\,\beta(u^\theta)\,h(\nabla u^\theta) & \text{in } \mathbf{R}^n\times(0,T),\ u^\theta(x,0)=u_0(x) & \text{in } \mathbf{R}^n, \end{cases}$$3 for drug dispensing, $$\begin{cases} u_t^\theta = H_1(x,t,\nabla u^\theta)+\theta\,\beta(u^\theta)\,h(\nabla u^\theta) & \text{in } \mathbf{R}^n\times(0,T),\ u^\theta(x,0)=u_0(x) & \text{in } \mathbf{R}^n, \end{cases}$$4 for cutting, $$\begin{cases} u_t^\theta = H_1(x,t,\nabla u^\theta)+\theta\,\beta(u^\theta)\,h(\nabla u^\theta) & \text{in } \mathbf{R}^n\times(0,T),\ u^\theta(x,0)=u_0(x) & \text{in } \mathbf{R}^n, \end{cases}$$5 for gripping and storage, and demagnetized state for locomotion. In function mode, the aligned field component $$\begin{cases} u_t^\theta = H_1(x,t,\nabla u^\theta)+\theta\,\beta(u^\theta)\,h(\nabla u^\theta) & \text{in } \mathbf{R}^n\times(0,T),\ u^\theta(x,0)=u_0(x) & \text{in } \mathbf{R}^n, \end{cases}$$6 actuates internal beams: drug doors open when $$\begin{cases} u_t^\theta = H_1(x,t,\nabla u^\theta)+\theta\,\beta(u^\theta)\,h(\nabla u^\theta) & \text{in } \mathbf{R}^n\times(0,T),\ u^\theta(x,0)=u_0(x) & \text{in } \mathbf{R}^n, \end{cases}$$7 mT, the cutting tool extends when $$\begin{cases} u_t^\theta = H_1(x,t,\nabla u^\theta)+\theta\,\beta(u^\theta)\,h(\nabla u^\theta) & \text{in } \mathbf{R}^n\times(0,T),\ u^\theta(x,0)=u_0(x) & \text{in } \mathbf{R}^n, \end{cases}$$8 mT, and grippers extend when $$\begin{cases} u_t^\theta = H_1(x,t,\nabla u^\theta)+\theta\,\beta(u^\theta)\,h(\nabla u^\theta) & \text{in } \mathbf{R}^n\times(0,T),\ u^\theta(x,0)=u_0(x) & \text{in } \mathbf{R}^n, \end{cases}$$9 mT.

The robot’s actuation obeys the standard magnetic relations

$\beta$0

and six-DOF control is expressed through a full-rank control matrix:

$\beta$1

The sixth DOF is rotation about the net magnetic moment, and it is operationally important for steering in rolling and for anchor management in crawling. Reported angular velocities are 16.5 rad/s about $\beta$2, 16.1 rad/s about $\beta$3, and 1.56 rad/s about $\beta$4 through the sixth DOF. Rolling speeds range from 1.02 to 6.75 mm/s in oil and 1.29 to 12.7 mm/s in air along the length, while two-anchor crawling reaches 0.4–1.06 mm/s in oil and 0.322–0.78 mm/s in air.

Each function mode has separate quantitative actuation behavior. Drug dispensing reaches opening width up to 1.2 mm at $\beta$5 mT and carries solid drug payload up to 0.230 mm$\beta$6. The cutting tool is a retractable tungsten-carbide element of diameter 0.25 mm and tip area approximately $\beta$7 mm$\beta$8; its extension reaches 464 $\beta$9m at $\beta(0)=0$0 mT, and cutting through 10 wt% gelatin is performed with spatial gradients up to 1.5 T/m, producing a pulling force of approximately 53.4 $\beta(0)=0$1N and an observed cutting rate of approximately 2 mm/min. Grippers open up to 709 $\beta(0)=0$2m at $\beta(0)=0$3 mT. The storage chamber holds 0.064 mm$\beta(0)=0$4 of sample. Remote heating uses $\beta(0)=0$5 mT at $\beta(0)=0$6 kHz, raising the robot temperature from 26.3°C to 40°C within 35 s, with estimated mean heat generation of approximately 1.36 mW. The safety guideline for magnetic hyperthermia, $\beta(0)=0$7 A/(m·s), is satisfied, with the highest value used reported as $\beta(0)=0$8 A/(m·s).

The practical importance of reinitialization here is sequential mode switching without replacement of the robot. A typical workflow is locomotion in the demagnetized state, magnetization to drug mode for local release, demagnetization back to locomotion, reprogramming to cut mode for incision, and reprogramming to grip or store mode for collection and extraction. The paper emphasizes that the robot remains stationary and controllable during reprogramming, that actuation fields are relatively uniform and weak—at most 65 mT and 1.5 T/m—and that this supports theoretical tissue penetration and safe remote operation. Limitations include the current 4.4 mm scale, restricted lab hardware for multi-axis reprogramming and high gradients, the fact that two-anchor crawling is available only in locomotion mode, and the need for future in vivo validation, surface coatings, sterilization protocols, and imaging-guided automation.

6. Comparative interpretation and recurrent design logic

The four usages show that “surgical reinitialization” is a polysemous technical label rather than a domain-invariant formalism. In the level-set and transformer cases, “surgical” is metaphorical and denotes targeted intervention on a mathematical or learned subsystem. In sinus surgery and magnetic soft robotics, the term is tied to literal intraoperative function: updating an anatomical model during resection, or switching a robot’s operative mode during a procedure. Treating all four as interchangeable would therefore be misleading.

At the same time, a recurrent design logic is visible. The intervention is narrow, not global: a corrector term is applied intermittently and possibly locally in level-set evolution; only masked depth discrepancies are fused into the TSDF in sinus surgery; only targeted heads are reset in BLOOM while the remainder of the model is frozen; and only low-coercivity components are rewritten in the soft robot. A plausible implication is that the value of reinitialization in these domains lies in restoring functionality while preserving accumulated state that would be expensive, unsafe, or counterproductive to discard.

A second commonality is the centrality of invariants and failure modes. In the PDE setting, the invariant is the zero level set and the main obstruction is discontinuity of the signed distance at extinction points. In sinus navigation, the invariant is unchanged geometry outside the modified region, while major failure modes arise from pose error, depth error, and limited visibility. In transformer repair, the invariant is the pretrained backbone outside the surgical subset, while major failure modes are iatrogenic drift and late local degradation under noisy corpora. In the soft robot, the invariant is the hard-magnetic locomotion structure, while safety depends on thresholds, hidden-tool defaults, stationary reprogramming, and decoupling of heating from locomotion by frequency.

These contrasts also delimit the term’s conceptual range. In one usage, reinitialization is rigorously analyzed through viscosity solutions, homogenization, half-relaxed limits, and interface geometry (Hamamuki et al., 2015). In another, it is a TSDF-based computer-vision pipeline operating on tracked endoscopic video (Mangulabnan et al., 2024). In another, it is a training intervention defined at the level of attention-head parameters and gradient masks (Schallon, 10 Mar 2026). In the last, it is a materials-and-actuation protocol grounded in coercivity, remanence, magnetic torque, and gradient force (Ng et al., 19 Sep 2025). The shared name therefore encodes an operational principle—selective restoration under preservation constraints—while the mathematical objects, guarantees, and evaluation criteria remain entirely domain-specific.