Summary (Overview)

• Proposes Manifold Constraint Hypothesis (MCH): Natural representations lie on a lower-dimensional manifold; constraining concept-erasure interventions to this manifold better preserves other encoded information while still effectively removing the target concept.
• Introduces MANCE (MANifold aware Concept Erasure): An iterative algorithm that uses local tangent-space estimates from natural-representation neighborhoods to project concept-probe gradient updates onto the manifold.
• Empirical validation across 119 settings (39 NLP: 13 LLMs × 3 concepts; 80 vision: 40 CelebA-CLIP attributes × 2 control-correlation regimes) shows consistent improvements in leakage–surgicality tradeoffs.
• MANCE variants outperform prior methods: MANCE applied on top of INLP, LEACE, and IGBP reduces leakage by 8–17 pp at the same surgicality budget. MANCE⁺⁺ achieves state-of-the-art nonlinear erasure, reaching near-chance leakage on 35/39 NLP settings at ≤10 pp control degradation—far exceeding Obliviator (17/39).
• Ablation confirms MCH: An otherwise identical full-space variant (AmbCE⁺⁺) leaves 6–10 pp leakage and often violates the budget, isolating the gain to the manifold constraint.

Introduction and Theoretical Foundation

Concept erasure aims to remove a target concept (e.g., gender, toxicity) from a neural representation while preserving all other information encoded in it. This is challenging because concepts are often entangled due to data correlations and superposition. Existing methods such as INLP (linear nullspace projection), LEACE (closed-form linear erasure), IGBP (iterative boundary projection), and Obliviator (kernel-based nonlinear erasure) treat the representation space as Euclidean, making updates that can damage unrelated concepts.

The authors propose the Manifold Constraint Hypothesis (MCH): If natural representations concentrate on a structured, lower-dimensional manifold (\mathcal{M} \subset \mathbb{R}^d), then interventions constrained to that manifold will be equally effective for concept erasure while better preserving other encoded concepts. This is formalized in §3.1: among interventions with matched effect on the target concept, manifold-constrained edits (moving along (\mathcal{M})) preserve control concepts better than unconstrained edits.

Methodology

MANCE Algorithm (Algorithm 1)

MANCE performs iterative updates to representations using a nonlinear concept probe (f_t) (MLP) that predicts the target concept. At each round (t):

1. Local manifold estimation (Step 1): For each current representation (x^{(t-1)}i), find its (k) nearest neighbors among the natural (unedited) representations (X^{(0)}). Compute a local PCA via SVD of the centered neighbor positions: [ \text{SVD}(S_i) = L_i ,\text{diag}(\sigma{i,1},\dots,\sigma_{i,k}), V_i^\top ] Retain top (r) right singular vectors as tangent basis (B_i = [v^{(1)}_i,\dots,v^{(r)}i] \in \mathbb{R}^{d \times r}), spanning the estimated tangent space (T{x_i}(\mathcal{M})).

2. Build tangent erasure direction (Step 2): Normalize the probe’s gradient: [ u_i = \frac{\nabla f_t(x_i)}{|\nabla f_t(x_i)|2} ] Project onto the tangent basis: (c_i = B_i^\top u_i). Apply spectral weighting (with exponent (\alpha=1)): [ d_i = B_i,\text{diag}(\sigma_i^\alpha),c_i = \sum{\ell=1}^r \sigma_{i,\ell}^\alpha c_{i,\ell} v_i^{(\ell)} ] Normalize to (\hat{u}_i = d_i / |d_i|_2) and subtract the component along (\hat{u}_i): [ \tilde{x}_i = x_i - \lambda_i \langle x_i, \hat{u}_i\rangle \hat{u}_i ]

3. Per-sample step-size cap (Step 3): Limit displacement to an (\varepsilon) fraction of the local neighborhood radius: [ r_i = \frac{1}{k}\sum_{j \in \mathcal{N}_k(x_i)} |x_j - x_i|2, \qquad \lambda_i = \min\left(\lambda{\max},; \frac{\varepsilon \cdot r_i}{|\langle x_i,\hat{u}_i\rangle|}\right) ]

The probe is refit every (\tau=8) rounds. The loop runs for (H) rounds.

MANCE Variants

• MANCE: Core algorithm (no preprocessing).
• MANCE⁺: Prepends LEACE (removes first-moment linear signal).
• MANCE⁺⁺: Prepends both LEACE and CovMatch, a rank-2 specialization that removes second-moment class-conditional covariance asymmetry ((\Delta\Sigma = \Sigma_+ - \Sigma_-)). Both are one-shot affine projections of effective rank ≤3.

Empirical Validation / Results

Evaluation Metrics

• Target leakage (S): Accuracy of a newly trained nonlinear probe on the erased concept; reported as (D_S = S - S_{\text{floor}}) (pp above the majority baseline).
• Surgicality (\Delta Y): Change in control-concept accuracy from clean to edited representations. Degradation (D_Y = \max(0, -\Delta Y)). Budgets (D_Y \leq b) for (b \in {1,3,5,10}) pp.
• Coverage: Number of settings where the method has at least one within-budget step.

Settings

• NLP: 13 LLMs (0.5B–27B parameters), 3 concepts (sycophancy, gender, safety), 39 settings.
• Vision: CelebA-CLIP (CLIP ViT-B/32), 40 binary attributes, two control-correlation regimes (5 most/least correlated attributes).

Key Results

Table 1: MANCE Complements Prior Erasers (39 NLP settings, mean leakage (|D_S|) in pp)

Table 2: Main NLP Erasure Result (39 settings): Leakage (pp) and “At Chance” counts

• Vision results (Fig. 3): MANCE⁺⁺ achieves the best Pareto frontier in coverage vs. leakage. At (D_Y \leq 1) pp (high-correlation regime), MANCE⁺⁺ covers 19/40 attributes (vs. Obliviator 2/40) with near-floor leakage.
• Ablation AmbCE⁺⁺ (full-space gradient, same effective step size): leaves 6–10 pp leakage and low coverage, confirming that the manifold constraint drives the gains.

Theoretical and Practical Implications

• Support for MCH: The consistent improvement from manifold-constrained updates across text and vision provides strong empirical evidence that natural representations are structured on a low-dimensional manifold, and that respecting this structure is key to surgical interventions.
• Orthogonality to prior methods: MANCE can be applied as a post-processing step on representations from any erasure method, immediately improving its leakage–surgicality tradeoff.
• State-of-the-art nonlinear erasure: MANCE⁺⁺ sets a new bar, especially on hard settings where target and control concepts are highly entangled (e.g., gender vs. profession; correlated CelebA attributes).
• Generalizability beyond erasure: The authors argue MCH applies to other representation-editing tasks such as activation steering.

Conclusion

The paper proposes the Manifold Constraint Hypothesis and instantiates it with MANCE, a family of manifold-aware concept erasure algorithms. By estimating the local tangent space from natural representations and projecting concept-probe gradient updates onto it, MANCE achieves superior leakage–surgicality tradeoffs compared to unconstrained baselines. Extensive evaluation across 119 settings (text and vision) shows consistent improvements; MANCE⁺⁺ sets state-of-the-art results. The findings support the importance of geometry-aware interventions while leaving better manifold estimation and formal guarantees for future work.