Degradation Concept Vectors

• Degradation concept vectors are specialized representations that isolate and manipulate degradation signals while preserving shared, useful model behaviors.
• They are applied across various models—from language models and diffusion models to vision-language systems—to enable precise, targeted editing.
• Techniques include task-vector decomposition, null-space projection, sparse autoencoder denoising, and disentangled latent encoding for robust, efficient control.

Searching arXiv for the cited papers to ground the article in current literature. arxiv_search.query({"^^^^^^^^11search_query11 OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11^^^^^^^^"," OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11search_query11^^^^,"^^^^^^^^11sort_by11 Searching by title as a fallback to confirm metadata. arxiv_search.query({"^^^^^^^^11search_query11 Task Vectors for Refined Model Editing\" OR ^^^^^^^^11ti:\11 Concept Editing in Diffusion Models without Performance Degradation\" OR ^^^^^^^^11ti:\11 Concept Vectors with Sparse Autoencoders for Improved LLM Steering\"","^^^^^^^^11max_results11 OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11search_query11^^^^,"^^^^^^^^11sort_by11 Degradation concept vectors are representational or parametric objects used to identify, isolate, erase, encode, or synthesize degradations, but the term is not uniform across current literature. In recent work it denotes, depending on the setting, a negated task-specific parameter update for model editing, a closed-form perturbation of diffusion cross-attention weights for unsafe-concept erasure, a denoised steering direction in LLMs, an embedding of a structured physical degradation description, or a disentangled latent code for image corruption synthesis (&&&^^^^11search_query11^^^^&&&, &&&^^^^11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11^^^^&&&, &&&^^^^11sort_by11^^^^&&&, &&&^^^^11sort_order11^^^^&&&). The common theme is selective manipulation of a degradation-relevant subspace while preserving a complementary subspace associated with general capability, safe semantics, or image content.

^^^^11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11^^^^. Parameter-space degradation vectors from task-vector decomposition

In parameter-space model editing, a task vector is defined as the parameter difference between a fine-tuned model and its base model,

PRESERVED_PLACEHOLDER_^^^^11search_query11^^^^

Given a collection of task vectors PRESERVED_PLACEHOLDER_^^^^11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11^^^^, "Decomposing Task Vectors for Refined Model Editing" separates each vector into a shared component and a unique component by identifying a shared subspace PRESERVED_PLACEHOLDER_^^^^11max_results11^^^^ and task-specific subspaces PRESERVED_PLACEHOLDER_^^^^11sort_by11^^^^ such that PRESERVED_PLACEHOLDER_^^^^11submittedDate11^^^^ and

PRESERVED_PLACEHOLDER_^^^^11sort_order11^^^^

The construction starts by reshaping each flattened task vector into a weight-difference matrix PRESERVED_PLACEHOLDER_^^^^11descending11^^^^, computing

PRESERVED_PLACEHOLDER_^^^^11search_query11^^^^

and defining the column-space projector

PRESERVED_PLACEHOLDER_^^^^11ti:\11^^^^

These projectors are chained into a joint operator

PRESERVED_PLACEHOLDER_^^^^11 OR ti:\11^^^^

followed by an eigendecomposition

PRESERVED_PLACEHOLDER_^^^^11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11search_query11^^^^

With eigenvalue threshold PRESERVED_PLACEHOLDER_^^^^11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11^^^^, the shared basis is PRESERVED_PLACEHOLDER_^^^^11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11max_results11^^^^ for PRESERVED_PLACEHOLDER_^^^^11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11sort_by11^^^^, giving the shared projector PRESERVED_PLACEHOLDER_^^^^11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11submittedDate11^^^^ and the decomposition

PRESERVED_PLACEHOLDER_^^^^11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11sort_order11^^^^

The degradation vector is then defined by negating only the unique component,

PRESERVED_PLACEHOLDER_^^^^11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11descending11^^^^

The stated rationale is that subtracting only the task-unique component avoids removing information from the shared subspace, which "typically encodes broadly useful, non-target behavior" (&&&^^^^11search_query11^^^^&&&).

This decomposition is used as a model-editing primitive. The paper reports three domain-level outcomes: improving multi-task merging in image classification by ^^^^11sort_order11^^^^% using shared components as additional task vectors, enabling clean style mixing in diffusion models without generation degradation by mixing only the unique components, and achieving ^^^^11submittedDate11search_query11^^^^% toxicity reduction in LLMs by negating toxic information isolated to the unique component (&&&^^^^11search_query11^^^^&&&). In the toxicity case study, the base model is LLaMA-^^^^11max_results11^^^^-^^^^11search_query11 one toxic task vector is collected by fine-tuning on ToxiGen + RealToxicityPrompts together with a pool of PRESERVED_PLACEHOLDER_^^^^11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11search_query11^^^^ unrelated task vectors; the edit is applied as

PRESERVED_PLACEHOLDER_^^^^11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11ti:\11^^^^

with PRESERVED_PLACEHOLDER_^^^^11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11 OR ti:\11^^^^, yielding a reported relative reduction of ^^^^11submittedDate11search_query11^^^^% on ToxiGen toxicity while preserving GSM^^^^11ti:\11^^^^K, BBH, MMLU, and TyDiQA with average control-task drop below ^^^^11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11^^^^% (&&&^^^^11search_query11^^^^&&&).

A common misconception is that negation should be applied to the entire task vector. The decomposition result argues against that: only the isolated unique component is negated, precisely to avoid unintended amplification or diminution of shared behaviors. This suggests that, within parameter-space editing, a degradation concept vector is best understood as a selective anti-direction rather than a full reversal of fine-tuning.

^^^^11max_results11^^^^. Null-space projected degradation vectors in diffusion-model concept erasure

In diffusion-based text-to-image editing, ACE formulates degradation vectors as perturbations PRESERVED_PLACEHOLDER_^^^^11max_results11search_query11^^^^ and PRESERVED_PLACEHOLDER_^^^^11max_results11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11^^^^ added to cross-attention key and value matrices PRESERVED_PLACEHOLDER_^^^^11max_results11max_results11^^^^ and PRESERVED_PLACEHOLDER_^^^^11max_results11sort_by11^^^^ in order to erase unsafe concepts while preserving normal ones (&&&^^^^11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11^^^^&&&). The formulation begins with text-representation matrices:

• PRESERVED_PLACEHOLDER_^^^^11max_results11submittedDate11^^^^ for unsafe token representations,
• PRESERVED_PLACEHOLDER_^^^^11max_results11sort_order11^^^^ for normal text representations,
• PRESERVED_PLACEHOLDER_^^^^11max_results11descending11^^^^ for desired safe alignment targets.

Under the original model,

PRESERVED_PLACEHOLDER_^^^^11max_results11search_query11^^^^

If preservation is ignored, the minimal update sending PRESERVED_PLACEHOLDER_^^^^11max_results11ti:\11^^^^ to PRESERVED_PLACEHOLDER_^^^^11max_results11 OR ti:\11^^^^ is obtained by

PRESERVED_PLACEHOLDER_^^^^11sort_by11search_query11^^^^

Preservation is then enforced through null-space projection. ACE requires

PRESERVED_PLACEHOLDER_^^^^11sort_by11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11^^^^

so PRESERVED_PLACEHOLDER_^^^^11sort_by11max_results11^^^^ must lie in the null-space of PRESERVED_PLACEHOLDER_^^^^11sort_by11sort_by11^^^^. If

PRESERVED_PLACEHOLDER_^^^^11sort_by11submittedDate11^^^^

and PRESERVED_PLACEHOLDER_^^^^11sort_by11sort_order11^^^^ with PRESERVED_PLACEHOLDER_^^^^11sort_by11descending11^^^^ corresponding to zero singular values, then the projector onto PRESERVED_PLACEHOLDER_^^^^11sort_by11search_query11^^^^ is

PRESERVED_PLACEHOLDER_^^^^11sort_by11ti:\11^^^^

and the update is constrained by PRESERVED_PLACEHOLDER_^^^^11sort_by11 OR ti:\11^^^^, which guarantees PRESERVED_PLACEHOLDER_^^^^11submittedDate11search_query11^^^^ (&&&^^^^11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11^^^^&&&).

ACE then introduces cross null-space projection because unsafe components in the value pathway can still re-couple with image tokens. With additional projectors PRESERVED_PLACEHOLDER_^^^^11submittedDate11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11^^^^ from PRESERVED_PLACEHOLDER_^^^^11submittedDate11max_results11^^^^ and PRESERVED_PLACEHOLDER_^^^^11submittedDate11sort_by11^^^^ from PRESERVED_PLACEHOLDER_^^^^11submittedDate11submittedDate11^^^^, the joint objectives are

PRESERVED_PLACEHOLDER_^^^^11submittedDate11sort_order11^^^^

PRESERVED_PLACEHOLDER_^^^^11submittedDate11descending11^^^^

yielding the closed-form degradation vectors

PRESERVED_PLACEHOLDER_^^^^11submittedDate11search_query11^^^^

The paper characterizes the key preservation property succinctly: because PRESERVED_PLACEHOLDER_^^^^11submittedDate11ti:\11^^^^ projects only onto directions orthogonal to PRESERVED_PLACEHOLDER_^^^^11submittedDate11 OR ti:\11^^^^, the subspace spanned by PRESERVED_PLACEHOLDER_^^^^11sort_order11search_query11^^^^ is exactly preserved (&&&^^^^11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11^^^^&&&).

Experimentally, ACE uses Stable Diffusion v^^^^11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11^^^^.^^^^11submittedDate11^^^^ and v^^^^11max_results11^^^^.^^^^11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11^^^^, edits ^^^^11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11^^^^,^^^^11search_query11search_query11search_query11^^^^ unsafe concepts, and evaluates on NSFW/I^^^^11max_results11^^^^P, CI, COCO, Imagenette, and UCE profession sets. Reported averages over all tasks and ^^^^11sort_order11^^^^ datasets are +^^^^11max_results11submittedDate11^^^^.^^^^11sort_order11descending11 semantic consistency by CLIP, PRESERVED_PLACEHOLDER_^^^^11sort_order11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11^^^^ LPIPS/FID image-quality degradation, maintained below ^^^^11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11^^^^% nudity or bias deviations, and runtime approximately ^^^^11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11^^^^% of UCE/RECE per concept (&&&^^^^11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11^^^^&&&). Unlike iterative editing methods, ACE requires no gradient steps and performs three fast matrix projections per concept.

In this line of work, a degradation concept vector is not a latent descriptor of corruption in an image. It is an edit operator in parameter space, designed so that erasure and preservation are solved jointly in closed form.

^^^^11sort_by11^^^^. Degradation through noise and format shift in language-model concept vectors

A distinct use of degradation concerns the deterioration of concept vectors themselves. "Denoising Concept Vectors with Sparse Autoencoders for Improved LLM Steering" models hidden states as

PRESERVED_PLACEHOLDER_^^^^11sort_order11max_results11^^^^

where PRESERVED_PLACEHOLDER_^^^^11sort_order11sort_by11^^^^ is concept-relevant signal and PRESERVED_PLACEHOLDER_^^^^11sort_order11submittedDate11^^^^ is noise arising from spurious features in the data (&&&^^^^11max_results11^^^^&&&). Two standard concept-vector constructions are considered.

For linear probing, given contrastive data PRESERVED_PLACEHOLDER_^^^^11sort_order11sort_order11^^^^, the concept vector is the logistic-regression or ridge weight

PRESERVED_PLACEHOLDER_^^^^11sort_order11descending11^^^^

For difference-in-means, with positives PRESERVED_PLACEHOLDER_^^^^11sort_order11search_query11^^^^ and negatives PRESERVED_PLACEHOLDER_^^^^11sort_order11ti:\11^^^^,

PRESERVED_PLACEHOLDER_^^^^11sort_order11 OR ti:\11^^^^

The paper argues that forming these vectors directly on noisy hidden states corrupts their direction and degrades steering performance (&&&^^^^11max_results11^^^^&&&).

The proposed remedy is Sparse Autoencoder-Denoised Concept Vectors. A single-layer sparse autoencoder maps hidden states to a high-dimensional latent space and reconstructs them:

PRESERVED_PLACEHOLDER_^^^^11descending11search_query11^^^^

The loss is

PRESERVED_PLACEHOLDER_^^^^11descending11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11^^^^

After reconstruction, the top-PRESERVED_PLACEHOLDER_^^^^11descending11max_results11^^^^ latents whose average activation difference between positives and negatives is largest are retained, other latents are zeroed, and the filtered representation PRESERVED_PLACEHOLDER_^^^^11descending11sort_by11^^^^ is used to construct denoised LP or DM vectors (&&&^^^^11max_results11^^^^&&&).

Steering success is measured as

PRESERVED_PLACEHOLDER_^^^^11descending11submittedDate11^^^^

where PRESERVED_PLACEHOLDER_^^^^11descending11sort_order11^^^^ counts outputs already exhibiting the target behavior without steering and PRESERVED_PLACEHOLDER_^^^^11descending11descending11^^^^ counts outputs after steering. On Llama-^^^^11sort_by11^^^^.^^^^11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11^^^^-^^^^11ti:\11^^^^B, layer ^^^^11max_results11sort_order11^^^^, the paper reports that SDCV-DM improves AI coordination from ^^^^11sort_by11ti:\11^^^^% to ^^^^11submittedDate11descending11^^^^%, sycophancy from ^^^^11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11descending11^^^^% to ^^^^11max_results11descending11^^^^%, survival from ^^^^11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11submittedDate11^^^^% to ^^^^11max_results11max_results11^^^^%, and corrigibility from ^^^^11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11submittedDate11^^^^% to ^^^^11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11ti:\11^^^^%; SDCV-LP improves refusal from ^^^^11sort_by11max_results11^^^^% to ^^^^11sort_by11submittedDate11^^^^%, survival from ^^^^11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11search_query11^^^^% to ^^^^11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11descending11^^^^%, myopic from ^^^^11ti:\11^^^^% to ^^^^11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11descending11^^^^%, and corrigibility from ^^^^11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11submittedDate11^^^^% to ^^^^11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11ti:\11^^^^% (&&&^^^^11max_results11^^^^&&&). Best PRESERVED_PLACEHOLDER_^^^^11descending11search_query11^^^^ is typically ^^^^11max_results11^^^^,^^^^11search_query11search_query11search_query11 in Llama, and varying PRESERVED_PLACEHOLDER_^^^^11descending11ti:\11^^^^ shows an inverted-U: too small misses signal, too large re-introduces noise.

A related robustness result appears in "Causality PRESERVED_PLACEHOLDER_^^^^11descending11 OR ti:\11^^^^ Invariance: Function and Concept Vectors in LLMs," which studies degradation under input-format shift rather than latent noise (&&&^^^^11submittedDate11^^^^&&&). There, Function Vectors are extracted from attention heads selected by Activation Patching and Concept Vectors from heads selected by Representational Similarity Analysis. The CV for prompt PRESERVED_PLACEHOLDER_^^^^11search_query11search_query11^^^^ is

PRESERVED_PLACEHOLDER_^^^^11search_query11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11^^^^

where heads are ranked by concept-RSA, namely the Spearman correlation between the head’s representational-similarity matrix and a concept design matrix. The paper reports that FVs are nearly orthogonal across formats, while CVs generalize better across open-ended, multiple-choice, and cross-lingual settings (&&&^^^^11submittedDate11^^^^&&&). In AmbiguousICL, FVs achieve PRESERVED_PLACEHOLDER_^^^^11search_query11max_results11^^^^–^^^^11search_query11 in-distribution on antonyms, but out-of-distribution they drop to about ^^^^11search_query11^^^^.^^^^11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11search_query11^^^^–^^^^11search_query11^^^^. for OE-FR and about ^^^^11search_query11^^^^.^^^^11search_query11search_query11 for MC, whereas CVs remain around ^^^^11search_query11^^^^.^^^^11submittedDate11search_query11 for OE-FR and ^^^^11search_query11^^^^.^^^^11sort_by11sort_order11 for MC. Lower PRESERVED_PLACEHOLDER_^^^^11search_query11sort_by11^^^^ for CVs indicates more consistent token distributions across extraction formats (&&&^^^^11submittedDate11^^^^&&&).

Taken together, these results establish two distinct failure modes for concept vectors in LLMs: degradation by noisy superposed features and degradation by surface-format dependence. The corresponding remedies—SAE denoising and RSA-based invariance selection—both aim to isolate a more stable concept-bearing subspace.

^^^^11submittedDate11^^^^. Structured degradation concept vectors in vision-LLMs

In "Understanding Degradation with Vision LLM," a degradation concept vector is neither a parameter update nor a steering direction. It is the learned embedding of a hierarchical, autoregressively generated description of image corruption (&&&^^^^11sort_by11^^^^&&&). A degradation instance for clean image PRESERVED_PLACEHOLDER_^^^^11search_query11submittedDate11^^^^ is represented as

PRESERVED_PLACEHOLDER_^^^^11search_query11sort_order11^^^^

where PRESERVED_PLACEHOLDER_^^^^11search_query11descending11^^^^ is the degradation type, PRESERVED_PLACEHOLDER_^^^^11search_query11search_query11^^^^ a type-specific physical parameter key, and PRESERVED_PLACEHOLDER_^^^^11search_query11ti:\11^^^^ the associated continuous physical value.

For degraded input PRESERVED_PLACEHOLDER_^^^^11search_query11 OR ti:\11^^^^, DU-VLM generates a token sequence

PRESERVED_PLACEHOLDER_^^^^11ti:\11search_query11^^^^

with each continuous value quantized to a discrete bin index by PRESERVED_PLACEHOLDER_^^^^11ti:\11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11^^^^. Each token is embedded through a shared token-embedding matrix to produce vectors such as

PRESERVED_PLACEHOLDER_^^^^11ti:\11max_results11^^^^

These are concatenated or summed to form the final degradation concept vector (&&&^^^^11sort_by11^^^^&&&).

The paper unifies degradation type prediction, parameter-key prediction, and continuous-value estimation under autoregressive next-token prediction. If the value space is partitioned into PRESERVED_PLACEHOLDER_^^^^11ti:\11sort_by11^^^^ uniform bins of volume PRESERVED_PLACEHOLDER_^^^^11ti:\11submittedDate11^^^^ and grid scale PRESERVED_PLACEHOLDER_^^^^11ti:\11sort_order11^^^^, the NTP loss is

PRESERVED_PLACEHOLDER_^^^^11ti:\11descending11^^^^

Under Proposition ^^^^11submittedDate11^^^^.^^^^11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11^^^^, assuming a locally Gaussian conditional distribution and small PRESERVED_PLACEHOLDER_^^^^11ti:\11search_query11^^^^, this becomes approximately a classification term plus a regression term:

PRESERVED_PLACEHOLDER_^^^^11ti:\11ti:\11^^^^

Proposition ^^^^11submittedDate11^^^^.^^^^11max_results11^^^^ further bounds classification error and regression mean-squared error by the average KL-excess risk PRESERVED_PLACEHOLDER_^^^^11ti:\11 OR ti:\11^^^^:

PRESERVED_PLACEHOLDER_^^^^11 OR ti:\11search_query11^^^^

This is a formal statement that the unified autoregressive formulation is bounded by the value-space quantization grid (&&&^^^^11sort_by11^^^^&&&).

Training proceeds in three stages. First, supervised fine-tuning with Multimodal Chain-of-Thought takes PRESERVED_PLACEHOLDER_^^^^11 OR ti:\11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11^^^^ plus a fixed prompt, generates a free-form rationale PRESERVED_PLACEHOLDER_^^^^11 OR ti:\11max_results11^^^^, then predicts the token sequence for PRESERVED_PLACEHOLDER_^^^^11 OR ti:\11sort_by11^^^^ with standard MLE. Second, offline structured reinforcement learning maximizes

PRESERVED_PLACEHOLDER_^^^^11 OR ti:\11submittedDate11^^^^

where the restoration reward is obtained by feeding the predicted tuple into a frozen diffusion prior PRESERVED_PLACEHOLDER_^^^^11 OR ti:\11sort_order11^^^^. Third, online self-supervised RL replaces the full-reference restoration reward with a no-reference IQA score such as MUSIQ for real-world adaptation (&&&^^^^11sort_by11^^^^&&&).

The same representation is also used for zero-shot control of pre-trained diffusion restoration models. Once PRESERVED_PLACEHOLDER_^^^^11 OR ti:\11descending11^^^^ is predicted, restoration is constrained by the known forward operator PRESERVED_PLACEHOLDER_^^^^11 OR ti:\11search_query11^^^^, with a pseudo-inverse step

PRESERVED_PLACEHOLDER_^^^^11 OR ti:\11ti:\11^^^^

followed by Denoising Diffusion Null-Space Model sampling

PRESERVED_PLACEHOLDER_^^^^11 OR ti:\11 OR ti:\11^^^^

In this setting, the degradation concept vector is an explicit physics-aware control variable: it pins down the physical forward model, while the diffusion prior fills in the null-space degrees of freedom (&&&^^^^11sort_by11^^^^&&&).

^^^^11sort_order11^^^^. Degradation concept vectors as disentangled latent codes for synthesis

A further formulation appears in "Towards a Universal Image Degradation Model via Content-Degradation Disentanglement," where the degradation concept vector is the pair of disentangled latents PRESERVED_PLACEHOLDER_^^^^11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11search_query11search_query11^^^^ extracted from a distorted image and used to synthesize degradations on arbitrary clean images (&&&^^^^11sort_order11^^^^&&&). The architecture contains two parallel encoders.

The Homogeneous Degradation Encoder PRESERVED_PLACEHOLDER_^^^^11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11search_query11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11^^^^ takes distorted image PRESERVED_PLACEHOLDER_^^^^11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11search_query11max_results11^^^^, processes it with a short-range branch at full resolution and a long-range branch on a downsampled copy, fuses the features, globally averages over spatial dimensions to produce PRESERVED_PLACEHOLDER_^^^^11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11search_query11sort_by11^^^^, and maps this through a two-layer MLP to a homogeneous latent PRESERVED_PLACEHOLDER_^^^^11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11search_query11submittedDate11^^^^ (&&&^^^^11sort_order11^^^^&&&). The Inhomogeneous Degradation Encoder PRESERVED_PLACEHOLDER_^^^^11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11search_query11sort_order11^^^^ uses a similar two-branch stem but retains spatial layout, producing a spatial feature map PRESERVED_PLACEHOLDER_^^^^11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11search_query11descending11^^^^.

To prevent either latent from leaking image-content information, the model introduces entropy-rate regularizers:

PRESERVED_PLACEHOLDER_^^^^11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11search_query11search_query11^^^^

PRESERVED_PLACEHOLDER_^^^^11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11search_query11ti:\11^^^^

The total loss is

PRESERVED_PLACEHOLDER_^^^^11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11search_query11 OR ti:\11^^^^

with PRESERVED_PLACEHOLDER_^^^^11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11search_query11^^^^ and additional contrastive, color, diversity, and adversarial terms (&&&^^^^11sort_order11^^^^&&&). The paper states that minimizing the entropy terms simultaneously forces each latent entry to be marginally small and drives the joint toward the product of marginals, thereby decorrelating channels.

Degradation is injected into the synthesis U-Net through an IDA-SFT block. For the homogeneous component, Spatial Feature Transform uses

PRESERVED_PLACEHOLDER_^^^^11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11^^^^

and applies

PRESERVED_PLACEHOLDER_^^^^11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11max_results11^^^^

For the inhomogeneous component, the degradation-aware transform is

PRESERVED_PLACEHOLDER_^^^^11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11sort_by11^^^^

These are combined additively:

PRESERVED_PLACEHOLDER_^^^^11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11submittedDate11^^^^

At the top level, the joint degradation map is formed by tiling the global code and concatenating it with the local map:

PRESERVED_PLACEHOLDER_^^^^11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11sort_order11^^^^

The paper explicitly identifies the pair PRESERVED_PLACEHOLDER_^^^^11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11descending11^^^^ as the model’s degradation concept vector (&&&^^^^11sort_order11^^^^&&&).

Because the representation is disentangled from content, it can be transferred and interpolated. Given two degraded images PRESERVED_PLACEHOLDER_^^^^11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11search_query11^^^^, one computes PRESERVED_PLACEHOLDER_^^^^11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11ti:\11^^^^ and forms

PRESERVED_PLACEHOLDER_^^^^11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11 OR ti:\11^^^^

then synthesizes

PRESERVED_PLACEHOLDER_^^^^11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11max_results11search_query11^^^^

The reported applications are film-grain simulation and blind image restoration (&&&^^^^11sort_order11^^^^&&&).

^^^^11descending11^^^^. Comparative interpretation and recurrent design principles

Across these formulations, degradation concept vectors serve different computational roles. In task-vector decomposition and ACE, they are edit directions in parameter space or weight space, applied to suppress a targeted behavior while preserving a complementary subspace (&&&^^^^11search_query11^^^^&&&, &&&^^^^11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11^^^^&&&). In DU-VLM, they are discrete-continuous structured semantic encodings of physical degradations, used both for prediction and for zero-shot control of a restoration backbone (&&&^^^^11sort_by11^^^^&&&). In the universal degradation model, they are disentangled latent variables that support synthesis and transfer of corruptions across images (&&&^^^^11sort_order11^^^^&&&). In SDCV and the FV/CV literature, the emphasis shifts to conditions under which concept vectors themselves degrade—through noise or format mismatch—and to methods for recovering stable directions (&&&^^^^11max_results11^^^^&&&, &&&^^^^11submittedDate11^^^^&&&).

Several recurrent principles nevertheless unify the literature.

Subspace isolation: shared-versus-unique decomposition in task vectors, null-space projection in ACE, and top-PRESERVED_PLACEHOLDER_^^^^11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11max_results11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11^^^^ latent filtering in SDCV all isolate a target-relevant component from a preservation component (&&&^^^^11search_query11^^^^&&&, &&&^^^^11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11^^^^&&&, &&&^^^^11max_results11^^^^&&&).

Preservation constraints: ACE preserves the subspace spanned by PRESERVED_PLACEHOLDER_^^^^11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11max_results11max_results11^^^^ exactly; task-vector negation preserves the shared subspace PRESERVED_PLACEHOLDER_^^^^11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11max_results11sort_by11^^^^ by editing only PRESERVED_PLACEHOLDER_^^^^11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11max_results11submittedDate11^^^^; DU-VLM preserves consistency with the physical degradation operator PRESERVED_PLACEHOLDER_^^^^11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11max_results11sort_order11^^^^ during restoration; content-degradation disentanglement preserves image content by compressing degradation latents (&&&^^^^11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11^^^^&&&, &&&^^^^11search_query11^^^^&&&, &&&^^^^11sort_by11^^^^&&&, &&&^^^^11sort_order11^^^^&&&).

Closed-form or lightweight control: ACE requires no gradient steps and runs in approximately ^^^^11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11^^^^% of the time cost of UCE/RECE per concept; task-vector decomposition relies on SVD, projector chaining, and eigendecomposition; DU-VLM uses autoregressive generation rather than separate regressors for each degradation parameter (&&&^^^^11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11^^^^&&&, &&&^^^^11search_query11^^^^&&&, &&&^^^^11sort_by11^^^^&&&).

Robustness as a central criterion: robustness appears as preservation of general knowledge under toxicity reduction, preservation of image quality under concept erasure, resistance to noisy hidden features, and invariance across prompt formats or languages (&&&^^^^11search_query11^^^^&&&, &&&^^^^11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11^^^^&&&, &&&^^^^11max_results11^^^^&&&, &&&^^^^11submittedDate11^^^^&&&).

A plausible implication is that "degradation concept vector" now functions as a family resemblance term rather than a single canonical construct. What links the family is not shared parameterization, but shared intent: representing degradation-specific information in a way that is manipulable without collapsing unrelated functionality.

^^^^11search_query11^^^^. Misconceptions, limitations, and open directions

One misconception is that degradation concept vectors always encode visually degraded images. The literature shows otherwise. In ACE and task-vector decomposition, the vector is an intervention on model parameters or cross-attention weights; in DU-VLM, it is a token-embedding representation of a structured physical description; in Yang et al., it is a latent code extracted from an already degraded image (&&&^^^^11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11^^^^&&&, &&&^^^^11search_query11^^^^&&&, &&&^^^^11sort_by11^^^^&&&, &&&^^^^11sort_order11^^^^&&&).

A second misconception is that degradation removal is equivalent to full concept removal. The task-vector decomposition result argues for removing only the unique component, precisely because shared directions may carry useful non-target behavior (&&&^^^^11search_query11^^^^&&&). ACE makes a parallel point algebraically: erasure must be projected into the null-space of safe concepts to avoid disturbing them (&&&^^^^11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11^^^^&&&).

A third misconception is that a concept vector that is causal for task performance is necessarily format-invariant. The FV/CV study explicitly rejects this equivalence: FVs can causally drive in-context learning while remaining strongly tied to extraction format, whereas CVs selected by concept-RSA are more invariant but induce different steering behavior (&&&^^^^11submittedDate11^^^^&&&). This suggests that "robust" degradation vectors may need to satisfy both causal and invariant criteria, not only one of them.

Current limitations also follow directly from the published formulations. Task-vector decomposition requires a pool of task vectors defining what should remain in the shared subspace and uses a fixed eigenvalue threshold PRESERVED_PLACEHOLDER_^^^^11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11max_results11descending11^^^^ that was reported to work robustly across experiments (&&&^^^^11search_query11^^^^&&&). SDCV requires a trained or existing sparse autoencoder and selection of PRESERVED_PLACEHOLDER_^^^^11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11max_results11search_query11^^^^, with performance following an inverted-U rather than monotonically improving with more retained latents (&&&^^^^11max_results11^^^^&&&). DU-VLM quantizes continuous value spaces, and its regression bound includes the quantization term PRESERVED_PLACEHOLDER_^^^^11id:(Damirchi et al., 27 Dec 2025) OR id:(Wang et al., 11 Mar 2025) OR id:(Zhao et al., 21 May 2025) OR id:(Lan et al., 4 Feb 2026) OR id:(Opiełka et al., 25 Feb 2026) OR id:(Yang et al., 19 May 2025)11max_results11ti:\11^^^^ (&&&^^^^11sort_by11^^^^&&&). The universal degradation model leaves some architectural details, such as exact channel counts, as implementation choices (&&&^^^^11sort_order11^^^^&&&).

Open directions are implied by the convergence of these lines of work. One plausible direction is cross-pollination between subspace-preserving editing and structured physical degradation modeling: the former offers precise algebraic control, while the latter supplies grounded parameterizations of corruption. Another plausible direction is to combine denoising and invariance criteria when constructing editable concept vectors for LLMs, since recent evidence shows that noise suppression and format-independence address different degradation mechanisms (&&&^^^^11max_results11^^^^&&&, &&&^^^^11submittedDate11^^^^&&&).