--- title: 'Erasure Modality: A Multi-Domain Framework' url: https://www.emergentmind.com/topics/erasure-modality type: topic --- # Erasure Modality: A Multi-Domain Framework Searching arXiv for the cited works and recent uses of “erasure modality” across domains. Search query: "erasure modality" Erasure modality is not a single standardized term across contemporary research. In current usage, it denotes a structured way of specifying what is being erased, where the erasure acts, which access channel or phase is affected, and what guarantee accompanies the intervention. In multimodal learning it can mean certified deletion of one sensed channel from model parameters; in diffusion models it can mean either a mode of intervention or an input attack surface; in quantum information it can mean recoverable coherence under restricted access or a flagged loss channel; in thermodynamics it can mean a controlled regime of memory reset; and in type theory it can mean a phase or grade distinction between runtime-relevant and runtime-irrelevant data [2602.16144] [2505.17013] [1410.2313] [2512.05761] [2605.00655]. ## 1. Terminological scope The phrase appears in several technically distinct literatures. In each case, “erasure” is not merely absence. It is an operational condition: the system must continue to function, and the erased component must be made irrelevant in a specific sense defined by the surrounding formalism. | Domain | Erased object | Meaning of “modality” | |---|---|---| | Multimodal ML | One input channel | Certified deletion regime for model behavior and parameters | | Diffusion models | A concept or its access path | Intervention locus or input/attack surface | | Quantum information | Which-path information or a lost mode | Recoverable coherence or flagged-noise regime | | Thermodynamics | Memory content or low-cost reset capability | Strong/weak erasure or exclusive control | | Type theory | Runtime-irrelevant terms | Phase or grade governing compilation-time removal | A plausible unifying implication is that an erasure modality always couples an object of erasure to a control structure. The control structure may be a calibration batch and certificate, a text-conditioning interface, an accessible subsystem, a work budget, or a typing discipline. What varies from field to field is the formal criterion for saying that erasure has actually occurred. ## 2. Certified modality deletion in multimodal learning In revocable multimodal sentiment analysis, erasure modality is formulated most explicitly as deletion of one synchronized input stream—such as \(X^L\), \(X^A\), or \(X^V\)—from a model that was previously trained on all of them. The central claim of Missing-by-Design is that deletion must exceed “not reading” a modality at inference time. It introduces two stronger requirements: parameter-level erasure and certifiability. The target is a surgically modified model \(\mathcal{W}'\) that is hard to distinguish from a hypothetical model \(\mathcal{W}^{-m^\star}\) trained without the deleted modality, up to a formal \((\varepsilon_{\mathrm{mod}},\delta_{\mathrm{mod}})\)-style bound [2602.16144]. The framework prepares for deletion during representation learning. Frozen feature extractors yield \(d_L=768\), \(d_A=74\), and \(d_V=512\), missing modalities are zero-padded, and a fusion network produces a joint representation \(Z_i\). For each modality \(m\), a generator \(\mathcal{G}_m\) reconstructs \(\hat X_i^m\) from the remaining modalities and a property embedding \(P^m \in \mathbb{R}^{1\times d_p}\) with \(d_p=128\). A decomposition module splits each modality into a sample-specific component \(\Sigma_i^m\) and a sample-invariant component \(\mu_i^m\), constrained by orthogonality, invariance, and property-alignment losses. Back-translation networks \(\mathcal{B}_m\) and an NCE objective ensure that the fused representation retains modality-specific detail. Deletion itself is implemented as weight surgery. For each parameter \(w_q\), modality saliency is computed from reconstruction-loss gradients, \[ s_q^{(m^\star)}=\frac{1}{|\mathcal B|}\sum_{i\in\mathcal B}\left|\nabla_{w_q}\mathcal L_{\mathrm{re}}^{(m^\star)}(i)\right|, \] while a SwiftPrune-inspired proxy \[ L_q \approx \frac{1}{2}\frac{w_q^2}{1-\chi_q} \] approximates task-loss increase under removal. Candidates satisfy high saliency and low \(L_q\), after which only a fraction \(k=\lfloor r|\mathcal W|\rfloor\) is modified. The update is either deterministic zeroing for \(\varepsilon_{\mathrm{mod}}\le 1\) or Gaussian perturbation with \[ \sigma=\frac{\Delta\sqrt{2\ln(1.25/\delta_{\mathrm{mod}})}}{\varepsilon_{\mathrm{mod}}}. \] The framework outputs a Modality Deletion Certificate containing the target modality, modified parameter indices, seed and noise scale, privacy budget, SHA-256 digest, and diagnostic specifications. On CMU-MOSI audio deletion, the reported white-box and black-box attack success rates drop from approximately \(78\%\) and \(72\%\) without deletion to approximately \(49\)–\(52\%\) after surgery, while Acc2 remains approximately \(88.6\)–\(89.3\%\) versus a no-deletion baseline of approximately \(89.9\%\). Full retraining without audio takes \(2.9\) hours, whereas MBD surgical deletion takes \(39\) seconds total—\(17\) s for saliency computation, \(8\) s for sensitivity calibration, and \(14\) s for parameter surgery. In this literature, erasure modality is therefore a certified operating regime for revocable multimodal models, not merely a missing-input pattern. ## 3. Concept erasure in diffusion models In diffusion-model research, “erasure modality” has two distinct meanings. One is a mode of intervention inside the model. The other is an input modality through which an erased concept can re-enter generation. The intervention-based meaning is made explicit in work that distinguishes destruction-based removal from guidance-based avoidance. Destruction-based removal changes the model so that outputs containing the target concept become low likelihood across prompts; guidance-based avoidance leaves concept knowledge in place but disrupts or redirects the conditional guidance that normally steers sampling toward it. This distinction maps onto intervention loci: global U-Net edits behave more destruction-like, while cross-attention or text-embedding edits are usually more guidance-like. Empirically, GA and STEREO are described as strongly destruction-based, UCE and ESD-x as guidance-based, and RECE as mixed [2505.17013]. A second line of work argues that erasure must be benchmarked across input modalities. M-ErasureBench evaluates text prompts, learned embeddings via Textual Inversion, and inverted latents via DDIM inversion, with white-box and black-box variants for the latter two. The benchmark reports that current methods often appear strong under text prompts yet fail badly under learned embeddings and inverted latents, with Concept Reproduction Rate exceeding \(90\%\) in the white-box setting. To mitigate this, IRECE localizes target concepts via cross-attention and perturbs the associated latents during denoising, reducing CRR by up to \(40\%\) under the hardest white-box latent inversion scenario while preserving visual quality [2512.22877]. A third formulation is Instant Concept Erasure, which defines erasure directly in semantic text-conditioning space and then compiles the edit into static weight updates. ICE constructs erase and preserve subspaces from text embeddings, uses anisotropic energy-weighted scaling to form \(\mathcal P_e\) and \(\mathcal P_p\), computes an explicit overlap projector \[ \mathcal P_{e\cap p}=2\mathcal P_e(\mathcal P_e+\mathcal P_p)^\dagger \mathcal P_p, \] and solves a convex Spectral Unlearning Objective whose analytical solution yields a dissociation operator \(\mathcal P_{ice}\). The permanent edit is \[ \mathcal W^\top_{ice}=(\mathcal I-\mathcal P_{ice})\mathcal W^\top. \] Because this acts at early text-conditioning layers, the method is presented as training-free, one-shot, runtime-free, and modality-agnostic across T2I and T2V models [2511.18684]. Taken together, these works make “erasure modality” in diffusion models fundamentally relational. It may denote where erasure is imposed, or which pathway an attacker uses to reconstitute the supposedly erased concept. ## 4. Quantum-information formulations In quantum information, erasure modality has at least three non-equivalent meanings: limited-access coherence recovery, flagged erasure noise, and continuous erasure detection. In the quantum-erasure problem with limited environmental access, the task is to maximize the average visibility of a signal qubit \(A\) by performing a non-selective measurement only on an accessible subsystem \(B\) of the environment, while another subsystem \(C\) remains inaccessible. The operational quantity is the limited-access coherence \[ \mathcal C_{A|B}=2\,\mathrm{Tr}|\tilde\chi_B|. \] For a two-dimensional accessible environment, this becomes \[ \mathcal C_{A|B}=2\sqrt{p_0p_1\,E(\hat\rho_{C|0},\hat\rho_{C|1})}, \] where \(E\) is the sub-fidelity of the conditional states of the inaccessible environment. Here the erasure modality is the maximal operational capability to recover coherence given restricted control; its limit is set by what the inaccessible environment still “remembers” [1410.2313]. In continuous-variable communication, erasure is instead a noise modality: a mode is either transmitted or completely lost and replaced by vacuum, \[ \rho'=(1-P_e)\rho+P_e|0\rangle\langle 0|, \] with the erasure assumed detectable. A three-mode CV code is built precisely for this flagged replacement channel. It protects one signal mode by encoding it with a bipartite entangled resource, and the paper states that in realistic deployments it can almost completely reverse a single erasure and improve fidelities for two erasures relative to direct transmission. The code is presented as the simplest known code protecting a single mode against erasures, and performance gains are reported up to erasure probabilities of approximately \(70\%\) [2210.10230]. In superconducting dual-rail qubits, erasure is a hardware-level loss from the logical manifold to \(|gg\rangle\), and the “erasure modality” is a mode of operation in which such loss is monitored continuously with minimal disturbance of logical dynamics. A single symmetrically coupled resonator enables single-shot erasure detection in \(384\) ns, with residual error per check \(6.0(2)\times 10^{-4}\), induced dephasing per check \(8(3)\times 10^{-5}\), and erasure error per check \(2.54(1)\times 10^{-2}\). The same architecture also realizes time-continuous erasure detection during gates, with median \(7.2\times 10^{-5}\) error per gate and less than \(1\times 10^{-5}\) added by detection itself. In this setting, erasure modality denotes an always-on monitoring regime compatible with fault-tolerant control [2604.16292]. ## 5. Thermodynamic and operational control of erasure In thermodynamic work on memory reset, erasure modality is defined by the physical embedding of a logically many-to-one map. Strong erasure requires both a single state-independent procedure and an environment that ends in the same macrostate regardless of the initial logical value. Weak erasure keeps the single-procedure condition but allows the environment to end differently, so information may persist outside the designated memory cell. Phase-space analysis yields no minimum entropy cost for weak erasure and a positive minimum entropy cost for strong erasure; in the symmetric case of equal initial phase-space volumes, the lower bound is \(k_B\ln 2\). The same work argues that at molecular scales the dominant source of entropy creation is often the extra entropy required to suppress thermal fluctuations, rather than the many-to-one map itself [2502.18231]. A different thermodynamic formulation treats erasure as an assisted operational primitive whose cost can be distributed asymmetrically across agents. If Bob’s memory \(B\) is locally erased, the baseline cost is \(W_0=S(B)\) in units with \(k_BT=1\). If Alice holds a correlated system \(A\), the minimal assisted cost is \[ W_{A\to B}=S(B)-J_0(B|A). \] Exclusive erasure is defined by a work budget \(W_{\max}\) such that \[ W_{A\to B}\le W_{\max}