- Rating
- 4(+1)
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 3
Read review
Paper summary
This paper studies concept erasure in language model representations. The goal is to remove a sensitive attribute, such as gender or race, and keep other useful information. The authors propose a simple conditional autoencoder. The encoder maps a representation X to an erased representation Z. The decoder rebuilds X from Z and the concept label C. Training uses an MSE reconstruction loss plus an HSIC penalty between Z and C. The authors argue that reconstruction is a better way to keep information than the HSIC-based objective in Obliviator.
On the theory side, the paper shows that good reconstruction plus independence gives perfect nonlinear guardedness. It also gives a bound for the case of imperfect reconstruction. Under a symmetry assumption, the same decoder gives exact concept conversion by changing C. Experiments cover Bias in Bios and two Dial tasks, with representations from BERT, GPT-2, DeepSeek-7B, and LLaMA-1B. The method shows a clearly better utility-erasure trade-off than prior methods, including supervised Obliviator, without task labels.
Critical strengths
- I think the method is very simple and easy to use. The objective has only two terms. It is also reported to train faster than prior nonlinear methods (although it is expected).
- I think the empirical results are strong. On Bias in Bios with BERT, the method keeps about 74% profession accuracy at near-random gender accuracy. This is much higher than other methods. Similar gains appear across four language models and the Dial tasks.
- I like the explanation of why reconstruction works well for preservation. Under independence, maximizing I(X; Z) equals minimizing H(X | Z, C). The paper also explains why maximizing HSIC is different from maximizing mutual information. I think this is a useful insight for this area.
- I appreciate that the authors are clear about the scope of the theory (although a little wordy, I am afraid). They state that the guarantees do not cover finite-sample training. They also openly discuss the independence assumption between U and C, and study what happens when it only holds approximately with a controlled simulation.
Critical weaknesses
-
I think the paper should discuss the fair representation learning literature more directly. A conditional autoencoder with an independence penalty is close to earlier work, based on my knowledge. These methods also rebuild the input from the latent code and the attribute. They can also change the attribute at inference time. I would suggest comparing with at least one of them and softening claims such as “a capability no prior erasure method provides”.
-
I think the theoretical guarantees are narrower than the abstract suggests. Based on my understanding (please correct me if I am wrong), the erasure result mainly follows from the independence condition itself. There is no bound when HSIC is small and not zero, isn't it? The preservation results are stated given C, for example I(U; Z | C) = H(U). In practice, Z is used without C. Z alone is only guaranteed to be useful under Assumption 2. I think this assumption is hard to satisfy for continuous U. It would help to state these limits in the main claims.
-
I think the generation steering experiment does not fully match the theory. I think the update rule h′ = h + β[d(e(h), c′) − d(e(h), c)] treats every state as toxic. It also uses β > 1. So Corollary 3 does not directly apply here? I am not 100% sure here. Please correct me here if I am wrong too.
Questions for authors
Please refer to the Weaknesses section; I would appreciate it if the authors could address those points in the rebuttal.
4 comments
W3
W3 — Applicability of Corollary 3 to generation.
We thank the reviewer for this observation. The reviewer is correct that our original intervention assumes a toxic source concept at every generation step, and that is not covered by Corollary 3. To address this, we conducted an additional experiment with a gated intervention that follows Corollary 3 in the ideal regime, without requiring an external classifier.
Relation to the theory. For a state whose true concept is and , our residual update can be written as
Under perfect reconstruction, the residual vanishes, and the update reduces exactly to the conversion map in Corollary 3. In practice, retaining this residual helps preserve information not captured by the decoder. The remaining issue is that the original intervention also modifies states that may already belong to the target concept.
Gated steering. We address this by estimating the source concept of each state using our own trained autoencoder, without any external classifier, as the concept under which the decoder best reconstructs the state:
We apply the intervention only to states with , leaving all other states unchanged. Under the assumptions of Corollary 3, when the correct concept is uniquely identified by reconstruction, this recovers the true source concept. With , the gated intervention therefore implements the exact conversion map in the ideal regime.
Importantly, this introduces no additional decoder evaluations or separate model: the original residual update already computes both and . Gating only compares their reconstruction errors to the original state, adding negligible computational overhead.
Additional experiment. We evaluate this variant using the same Qwen2.5-7B setup and autoencoder as Table 1, on 1,000 toxic and 1,000 non-toxic held-out prompts. The latter allow us to test whether steering unnecessarily changes already non-toxic generations.
| Method | Toxic EMT ↓ | Toxic PPL ↓ | Non-toxic PPL ↓ |
|---|---|---|---|
| Untouched | 0.163 | 3.44 | 3.44 |
| Original, | 0.074 | 3.81 | 3.72 |
| Gated, | 0.074 | 3.74 | 3.47 |
| Original, | 0.037 | 4.77 | 5.14 |
| Gated, | 0.043 | 4.55 | 3.61 |
| Original, | 0.018 | 6.86 | 8.02 |
| Gated, | 0.027 | 5.99 | 3.93 |
The main finding is that the theoretically motivated intervention remains effective while better preserving unaffected generations. At , gating achieves the same toxicity reduction as our original intervention (EMT 0.074), with lower perplexity. Moreover, the gate activates on approximately 66–69% of generation steps for toxic prompts, compared with only 26% for non-toxic prompts. Consequently, even at , non-toxic perplexity remains at 3.93, compared with 8.02 without gating.
Regarding , we agree that Corollary 3 applies to . Larger values extrapolate along the learned conversion direction, similarly to the strength parameters used by other steering methods. They are an empirical extension rather than part of the exact-conversion guarantee. Importantly, the results show that the intervention covered by the theory is already effective, without requiring such extrapolation.
We thank the reviewer again and hope these clarifications, additional theoretical results, and new experiments address the reviewer's concerns. We would appreciate it if the reviewer could reconsider their assessment and consider raising their score in light of these additions. We are happy to clarify any remaining questions.
Thanks for the rebuttal. I appreciate the additional experiments provided as well as the response " Applicability of Corollary 3 to generation". I would like to hear the response from other reviewers and now I keep my positive score.
Thank you for your positive response and for taking the time to consider our rebuttal and additional experiments. We appreciate your thoughtful feedback throughout the review process and are glad that you found our clarifications regarding the applicability of Corollary 3 and the additional experimental evidence useful.
W1, W2
We thank the reviewer for the thoughtful review and for recognizing the simplicity of our method, its strong empirical results, and the motivation for using reconstruction rather than HSIC for preservation. We address each concern below.
W1 — Relation to prior autoencoder-based methods.
We acknowledge the earlier work on conditional autoencoders and independence constraints, particularly Shaham et al. (2022), which we already discuss in Section 2. We deliberately build on the same underlying objective, but our contribution goes well beyond applying an existing architecture to another task. Shaham et al. focus on latent-variable recovery and do not establish guarantees for nonlinear concept erasure, complementary-information preservation under imperfect reconstruction, or exact concept conversion. We extend this theoretical framework to establish precisely these guarantees, addressing fundamental questions in nonlinear concept erasure and steering that were not studied in the earlier work.
To our knowledge, ours is the first framework in concept erasure and steering to jointly establish these theoretical guarantees, providing a principled foundation for nonlinear erasure, maximal information preservation, and exact concept conversion. Despite substantial progress in the field, these questions had not been addressed together theoretically.
Our contribution is therefore not a new autoencoder objective, but the theoretical framework and its implications for concept erasure and steering. Since we deliberately adopt the same underlying objective as Shaham et al., a direct empirical comparison would be less informative, as our contribution is not an improvement over their autoencoder objective or architecture. Instead, we compare against existing concept-erasure methods, including Obliviator, KRaM, and FaRM, to evaluate the practical benefits of our framework in the setting it is designed to address. Our statement that this is "a capability no prior erasure method provides" refers specifically to these existing concept-erasure methods, rather than to conditional autoencoders in general.
W2 — Scope of theoretical guarantees.
1. Erasure and preservation. While independence alone guarantees nonlinear erasure, it also permits trivial solutions such as constant representations. Reconstruction is essential to guarantee maximal information preservation alongside erasure, with I(X;Z)=H(X∣C)=H(U) under our assumptions. The contribution is establishing these guarantees jointly, rather than independence alone.
2. Imperfect independence. To keep the manuscript simple and elegant, we originally focused on the main theoretical insights and left out the more technical analysis of imperfect independence. However, given the reviewer's concern, we have now added quantitative erasure guarantees for small but nonzero population HSIC, together with full proofs in the revised Appendix A.5.
The main result is that approximate independence provides approximate protection against nonlinear concept recovery. Specifically, for any bounded-norm RKHS adversary attempting to predict the sensitive concept from the erased representation Z, the difference in its expected prediction scores across concept groups is bounded by O(HSIC(Z,C)).
Importantly, we extend this result beyond RKHS adversaries to every fixed bounded continuous adversary, including the MLP probes used throughout our experiments. Using the universality of the Gaussian kernel on compact representations, we show that their expected prediction scores become indistinguishable across concept groups as population HSIC approaches zero, directly connecting the theoretical guarantee to our practical evaluation.
Intuitively, the smaller the dependence between Z and C, the less an adversary can distinguish the concept groups from the erased representation. While the convergence rate depends on the adversary's RKHS approximability and the guarantee is not uniform over arbitrary classifiers, it applies to the bounded continuous prediction scores of the nonlinear probes used in practice. Our result thus directly connects the independence term we optimize to quantitative nonlinear concept erasure, extending the theory beyond the ideal setting.
Together, these results strengthen our theoretical guarantees for both ideal and approximate settings, complementing our empirical erasure and steering results, including the additional comparisons presented in our response to ELSp (W3).