acceptodds
Under review as a conference paper at ICLR 2027

Inoculation Adapters: Improved Selective Generalization of Capabilities with Fewer Surprising Backdoors

Abstract

Inoculation prompting (IP) is a selective-generalization technique used against emergent misalignment. We study inoculation adapters (IA), a family of methods that similarly reduce the optimization pressure to learn undesired traits by strengthening those traits during training. IAs are LoRAs used in three steps: (1) trained on undesired traits; (2) attached frozen while a separate task adapter is trained on data exhibiting both desired and undesired traits; (3) the IA is discarded at deployment, while only the task adapter is kept. Vanilla IA shares its mechanism with security vectors (Zhou et al., 2024); we introduce two gated variants. We compare IAs with four selective-generalization baselines: IP, preventative steering, Concept Ablation Fine-Tuning (CAFT), and KL regularization. Across nine setups and five model families, the IA family spans the observed Pareto frontier of desired trait retention vs. undesired trait suppression among the evaluated methods, although given wide confidence intervals the magnitude of improvement remains uncertain. IAs also avoid two drawbacks of IP: they can suppress capabilities and traits that cannot be reliably elicited by a prompt, and they introduce fewer surprising backdoors. However, no IA variant optimizes all objectives perfectly; gains in desired-trait generalization are generally accompanied by weaker suppression of the undesired trait and increased backdoor occurrence.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
selective generalization, selective learning, AI alignment, emergent misalignment, safe fine-tuning, inoculation adapters, backdoors, inoculation, misgeneralization

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_jYXNOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies Inoculation Adapters (IAs) for selective generalization, aiming to retain desired traits while suppressing undesired ones during fine-tuning. It introduces two gated variants, GIA and CGIA, and evaluates their effectiveness across nine setups and five model families. The results suggest improved suppression–retention trade-offs and fewer surprising backdoors compared with several existing methods.

Critical strengths

Strengths

  1. Clear presentation. The paper is well-written, easy to follow, and clearly motivated.
  2. Comprehensive evaluation. The experiments cover diverse traits, model families, and selective-generalization baselines.
  3. Systematic Pareto analysis. The paper provides a useful characterization of the trade-off between desired-trait retention and undesired-trait suppression.
  4. Interesting empirical findings. The results on hard-to-elicit traits and surprising backdoors offer valuable insights into adapter-based selective generalization.
Critical weaknesses

Major Weaknesses

1. Missing comparison with Security Vectors.

Vanilla IA closely resembles Security Vectors (Zhou et al., 2024). Although the authors acknowledge this connection, a direct empirical comparison is missing. Evaluating the original Security Vectors method, including its KL regularization and bi-level optimization, on representative setups would better establish the paper's incremental contribution.

2. Unclear contribution of dynamic gating.

It remains unclear whether GIA/CGIA improve the Pareto frontier through input-dependent gating or simply by adjusting IA strength. Two ablations would help clarify this: (1) sweeping fixed IA scaling coefficients and (2) replacing dynamic gates with trainable input-independent coefficients.

Minor Weaknesses

  1. Task complexity. Several desired traits involve relatively simple stylistic or formatting behaviors. Generalization to more complex capabilities remains unclear.
  2. Backdoor trade-offs. GIA and especially CGIA introduce more surprising backdoors than vanilla IA. The implications of this trade-off could be discussed more explicitly.

Robustness Concern

Each configuration uses only one training seed. Additional seeds on representative setups would strengthen the conclusions, particularly given the wide confidence intervals in the aggregate Pareto analysis.

Questions for authors

Questions for the Authors

  1. Comparison with Security Vectors: Could the authors compare IA/GIA/CGIA with the full Security Vectors method on representative tasks? This would better establish the empirical contribution beyond the closest prior work.

  2. Contribution of dynamic gating: Do GIA/CGIA outperform simpler alternatives, such as fixed IA scaling or trainable input-independent gating? This would clarify the independent methodological contribution of the proposed gating mechanisms.

  3. Robustness: Are the reported Pareto improvements consistent across multiple training seeds? This would strengthen the reliability of the empirical conclusions.

  4. Backdoor mechanisms: Why do GIA/CGIA introduce more surprising backdoors than vanilla IA? Is this due to weaker suppression or additional conditionalization induced by gating? Clarifying this would improve the mechanistic understanding of the method and its safety trade-offs.

2 comments

Thank you for a precise review, and for calling our results on hard-to-elicit traits and backdoors "interesting empirical findings". Your two ablations of the gated variants changed what the paper says about them. The changes are in blue or red in the revised PDF.

Q1 and weakness 1: comparison with security vectors

Done on seven setups; point 1 of the general reply gives the results.

Q2 and weakness 2: input-dependent gating or IA strength?

Good suggestion! We ran both ablations (Appendix E.5): the frozen IA scaled by a fixed factor from 0.2 to 0.8, on B1, B2, B3, U2, and E1, and GIA with trainable input-independent gates, on the same five setups.

Per setup, strength explains most of GIA's gain. GIA falls on the path traced by the fixed scales on four of five setups, and the input-independent gates on all five. E1 is the exception for GIA: desired trait 0.56 against 0.50 for the closest fixed scale and 0.46 for the input-independent gates, with overlapping confidence intervals. On B1, the input-independent gates learn a weaker strength and retain less than GIA, while a fixed scale matches GIA. GIA's gates also spread from fully closed to fully open, while the input-independent gates stay near their mean, and traces of specific gates show that they vary with the input, mostly from token to token (29 to 55% of the gate variance lies within gates; Appendix E.5). GIA thus learns an input-dependent policy, although this does not show by itself that the input dependence improves retention.

Averaged over the five setups, GIA still retains more of the desired trait than the fixed scale with similar suppression or the input-independent gates (normalized, coherence-aware: 0.87 against 0.71 and 0.75; over all answers: 0.81 against 0.71 to 0.75 and 0.73), as does CGIA against its fixed scale (0.86 against 0.68, coherence-aware), with overlapping confidence intervals. A fixed scale also has to be selected per setup, which the input-dependent or input-independent gates avoid.

The contributions now state both results. To the sentence on GIA and CGIA, we added: "Per setup, GIA mostly performs like the inoculation adapter scaled by a fixed factor, whose strength must be selected, while a variant of GIA whose gates ignore the input learns that strength itself." A new Discussion paragraph ("Does gating add more than a weaker adapter?", Section 5) and Appendix E.5 ("Fixed inoculation strength and input-independent gates") give the details.

Q3: robustness across seeds

Mostly yes, on the three setups we repeated (general reply, point 2). GIA retains more of the desired trait than IA at every seed on B2 and E1, and at four of five seeds on U2.

Q4 and minor weakness 2: why do GIA and CGIA introduce more backdoors?

The paper states the trade-off (Section 3.3): gating gives up part of IA's advantage on backdoors. Our analysis suggests the rise is mostly CGIA's. Its backdoor strength, averaged over the surprising prompt families, is +0.053 on B1 and +0.041 on B2, against +0.006 and +0.001 for IA, while GIA stays within 0.016 of IA.

On B1, weaker suppression does not explain the surprising backdoors. Without a system prompt, CGIA and IA express the trait equally (0.019 and 0.017), yet CGIA's backdoor is about eight times stronger, which points to a dependence on the context.

On B2 and B3, CGIA also expresses more of the trait without a system prompt (EM 0.083 against 0.016 for IA on B2), so there the two explanations are confounded.

This rests on one seed per setup and does not test either explanation directly (e.g., by evaluating IA at a fixed strength matched to CGIA's mean gate), so we left it out of the paper.

Minor weakness 1: task complexity

We agree. We chose stylistic and formatting traits because they can be measured reliably, and the Discussion now leaves more complex capabilities for future work.

====

If the security-vector comparison and the gating ablations answer your questions, we would be grateful if you would consider raising your score; otherwise, please tell us what remains unconvincing.

I appreciate the additional results on comparison with security vectors and contribution of dynamic gating. The concern of robustness has also been solved by running multiple seeds. I therefore raise my score.

Reviewer_UvWaOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The authors introduce inoculation adapters, a family of methods that reduces the optimization pressure to learn undesired traits by strengthening those traits during training. They also introduce 2 variants (Gated Inoculation Adapter and Complementary-Gated Inoculation Adapter) which trade off between undesired trait and desired trait retention. Finally, they show that inoculation adapters avoid two drawbacks of inoculation prompting: they can suppress capabilities that can't be elicited by a prompt, and introduce less surprising backdoors.

Critical strengths

The paper is very well written and the experiments are well designed.

Critical weaknesses

Critical weakness:

  • Currently the paper reads as if the authors found out about security vectors last minute and hastily added parts to the introduction and related work to address this. While it is unfortunate that the authors developed Inoculation Adapters without knowledge of Security Vectors, I think the paper needs a substantial rewrite before being fit for publication. I do think there are useful contributions but the paper as currently framed does not properly position itself relative to security vectors. I think the paper should be reframed as: -- introducing a variant of security vectors (although it is debatable how useful these variants are, especially since there is no comparison to the original KL-regularized version of security vectors) -- linking it to inoculation prompting -- introducing principled datasets to study this desired vs undesired capability tradeoff -- studying the pareto frontier of this tradeoff -- examining the effect on backdoors
Questions for authors

None

5 comments

We appreciated the concrete structure you proposed for the paper; it drove most of this revision.

You are right that the initial submission did not position itself relative to security vectors well enough: we developed inoculation adapters independently and learned of security vectors late. We moved the abstract, contributions, results, discussion, and conclusion closer to the structure you propose, and added an appendix comparing the two methods (Appendix C.5). The changes are in orange or red in the revised PDF.

1. A variant of security vectors, compared with the original

The abstract, contributions, Discussion, and conclusion now present IA as a variant of security vectors with a simpler training recipe. We compare it with the original KL-regularized recipe and its bi-level optimization (SV), and with a strengthened recipe (SV++), on seven setups; the general reply (point 1) summarizes how they compare.

On your doubt about how useful these variants are, we also added two analyses of the gated variants (Appendix E.5): the inoculation adapter scaled by a fixed factor, and GIA with input-independent gates. They partly support your doubt. Per setup, GIA mostly acts as a learned IA strength. Averaged over setups, GIA retains more of the desired trait than any fixed scale at similar suppression or the input-independent version, and CGIA more than any fixed scale at similar suppression, with overlapping confidence intervals (see the reply to jYXN, Q2). Traces of single gates also show that GIA's gates vary with the input, mostly from token to token (Appendix E.5): GIA learns an input-dependent policy, although this does not show by itself that the input dependence improves retention.

2. Link with inoculation prompting

It remains the starting point, and we added no new experiments here beyond the U2 correction (general reply, point 3). The appendix on the elicitation strength of the inoculation prompts (Appendix C.4; D.6 in the submitted version) now discusses U2. The claim is narrower: inoculation prompting is unreliable when eliciting the trait would require significant training (U1, U3), where inoculation adapters still suppress it, though only partly on U3 (Section 3.2).

3. Datasets

You suggested presenting our datasets as a contribution. To make them usable, the supplementary material now contains the synthesized task and inoculation datasets, the evaluation datasets, and the prompts and code that generate them. It leaves out two kinds of data:

  • for safety, the hate-speech corpora and the misaligned-assistant corpora that train the inoculation adapters against emergent misalignment (ethics statement); the included prompts regenerate them;
  • to respect their authors' protection against web scraping, the emergent-misalignment datasets of Turner et al. (2025) (bad medical advice, risky financial advice, extreme sports), which their authors distribute at https://github.com/clarifying-EM/model-organisms-for-EM, and our task datasets of E2, E3, B2, and B3, which are built from them; the included prompts rebuild these from the original datasets.

4. Pareto frontier

Figure 1 now includes SV and SV++, with IA's values on E2 and U1, where they cannot be trained (drawn half-filled). The inoculation-adapter family spans the observed frontier, SV++ lies almost on it (the inter-setup confidence intervals remain wide) (Section 3.1).

5. Backdoors

Figure 3 now includes SV and SV++ (Section 3.3); the last paragraph of point 1 in the general reply describes what it shows.

====

The paper has moved substantially toward the framing you proposed, while keeping the experiments of the submission. We would be grateful if you would reconsider your score; if concerns remain, could you tell us which ones?

I appreciate the authors' efforts in reframing their submission. I unfortunately still think there are substantial problems with the structure of the paper. I think it has to be framed as improving and simplifying security vectors, instead of introducing a new method that "shares its mechanism with security vectors", as how it is currently makes it very confusing to understand what contributions the paper is actually making.

Thank you for the follow-up. We agree that presenting inoculation adapters (IA) as an improvement and simplification of security vectors is clearer: it states how IA relates to security vectors. We would also like to keep the link with inoculation prompting (IP), which inspired IA. We suggest the following framing:

  • We present IA as a variant of security vectors that improves and simplifies them: IA needs neither the KL term nor the min-min optimization, so it has no KL coefficient to tune and is cheaper to train. The min-min stage updates the backbone weights, while IA only trains LoRAs, so IA also trains on 4-bit quantized models. On most setups, it also retains more of the desired trait than the published recipe (SV).
  • Like security vectors, IA can be seen as inoculation prompting in weight space, and it avoids two drawbacks of IP.
  • We add the gated variants GIA and CGIA, which retain more of the desired trait than IA, with weaker suppression.

We will make the following edits to the main body in the final version.

  • Abstract. "[...] We study inoculation adapters (IA), which improve and simplify security vectors (Zhou et al., 2024) by training the adapter with plain SFT, without their bi-level optimization and KL constraint. Like security vectors and IP, IAs reduce the optimization pressure to learn undesired traits by strengthening those traits during training. [...] We introduce two gated variants (GIA and CGIA). [...]"
  • Section 1 (Introduction and Related Work), paragraph "Security vectors". "[...] Vanilla IA improves and simplifies security vectors: the adapter is trained with plain SFT rather than bi-level optimization with a KL constraint, and a separate task LoRA is trained on a frozen backbone rather than updating the backbone itself. We developed inoculation adapters independently and learned of security vectors afterwards. Our contributions are this variant, the gated variants GIA and CGIA, and an empirical evaluation [...]"
  • Section 1 (Introduction and Related Work), paragraph "Inoculation adapters". "The vanilla Inoculation Adapter (IA) improves and simplifies security vectors. Like them, it can be seen as inoculation prompting in weight space. We introduce Gated Inoculation Adapter (GIA) and Complementary-Gated Inoculation Adapter (CGIA), which retain more of the desired trait than IA, with weaker suppression. [...]"
  • Section 1 (Introduction and Related Work), first contribution. "We introduce inoculation adapters, a variant of security vectors that improves and simplifies them, and evaluate them on selective generalization of co-occurring traits across nine setups and five model families, against our main baselines, inoculation prompting and security vectors, and against preventative steering, CAFT, and KL regularization. [...]"
  • Section 2 (Inoculation Adapters), paragraph "Inoculation-adapter family". "Vanilla IA simplifies security vectors (Zhou et al., 2024) and keeps their rationale, also used by inoculation prompting: supplying the undesired trait during training [...]"
  • Section 2 (Inoculation Adapters), paragraph "Relation to security vectors". We change the title to "From security vectors to inoculation adapters". And the text: "[...] IA simplifies this recipe: we train the inoculation adapter with plain SFT and a separate task LoRA with the backbone frozen. GIA and CGIA add input-dependent gates to IA, which trade suppression for retention."
  • Section 3.1 (Observed suppression-retention tradeoffs), list of methods. "[...] security vectors (Zhou et al., 2024), which inoculation adapters improve and simplify; as baselines, we use their published recipe (SV) and a strengthened variant (SV++) [...]".
  • Section 6 (Conclusion). "We studied inoculation adapters, which improve and simplify security vectors (Zhou et al., 2024), and introduced the gated variants GIA and CGIA. [...]"

With these edits, we think the sections can stay as they are, since they follow the outline of your first review: IA as a variant of security vectors (Section 2), the datasets (their design in Section 2, the nine setups in Section 3.1 and Appendix B), the Pareto frontier (Section 3.1), the link with inoculation prompting (Sections 3.2 and 3.3), and the backdoors (Section 3.3). We also do not think this framing requires new experiments: the revision already compares IA with security vectors on seven setups (Section 3.1, Appendix C.5).

Would this framing address your concern about the structure of the paper?

Reviewer_YfJWOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors introduce Inoculation Adapters where they train LoRA adapters on udnesired traits which is kept frozen while another task specific LoRA is trained. During inference, this is discarded. The paper also introduces two variants of the setting and an extensive evaluation across 9 setups and 5 model families. Multiple baselines like inoculation prompting and preventative steering are compared to study the tradeoff between suppressing the undesired traits and keeping the desired ones.

Critical strengths
  • The paper is well-written and easy to follow. I liked reading this.
  • The setting is clean and well presented. Desired and undesried traits co-occur in every data point and both can generalise.
  • I liked the non-elicitable setups where vanilla inocluation prompting would fail
  • Experimens are extensive and controls (random-norm adapter, irrelevant trait adapter, etc) are well chosen
  • Backdoor analysis is insightful
Critical weaknesses
  • Evaluation depends only on llm judges without any human validation of trait scores. General capability is only evaluated via coherence but authors should report general benchmarks like MMLU/GSM8K
  • I'm concerned about the novelty of the proposed approach compared to security vectors (which is not evaluated as a baseline too). Safe-LoRA and orthogonal projection are also not compared.
  • RL and full-finetuning is not tested at all. The authors briefly mention this but some experimental evidence towards this would be good to include
  • The configurations and reported statistics should be improved. Each config is trained with a single seed and the confidence intervals are wide.
Questions for authors

See above

2 comments

We are glad you found the backdoor analysis "insightful". Your requests were concrete, and most are now in the paper: the human check of the judges, MMLU and GSM8K, and the comparisons with Safe-LoRA, orthogonal projection, and full fine-tuning (changes in green or red in the revised PDF).

Weakness 1: judge validation

You noted that the evaluation relied only on LLM judges, without human validation.

We checked the judges in two ways (Appendix F.5). We re-scored a subset of our evaluations (1,920 completions) with five more capable LLM judges. Of the 98 differences between methods that are significant under our judge, gpt-6-luna, gpt-6.1-terra, and Claude Haiku 5.5 keep the sign of 95 to 98. gpt-6-astra and gpt-6.1-sol, which judged emergent misalignment only, keep 36 of their 38.

We scored 30 completions by hand. Our judge agrees with the human on harmfulness as well as the stronger judges do (Spearman 0.83, against 0.82 to 0.84); all judges agree less with the human on the binary EM label. This is weak evidence: one annotator and 30 answers. We first scored the answers blind, then revised 17 of the 30 scores after reading the comments of Claude Opus 5.5, which flagged scores it found inconsistent with the rubric, and after seeing the judges' scores. We also checked that the trait scores do not rest on incoherent answers (Appendix F.4).

Weakness 1: general benchmarks

We now report MMLU and GSM8K for every method we retrained, on seven setups (Appendix F.2). IA stays close to the base model on the Qwen setups, while GIA and CGIA lose some arithmetic ability on B1; on E1 (Llama-3.1-8B), all three lose some, CGIA the most. IA and GIA keep more than inoculation prompting and preventative steering on U2 and E1. SFT(Harmful) degrades both benchmarks on most setups, and on U3, a base model, SV gains less arithmetic ability from task training than IA.

Weakness 2: novelty and security vectors

The mechanism is that of security vectors, as the submission stated (general reply, point 1). What is new is a simpler training recipe for the same mechanism (plain SFT of the adapter, without bi-level optimization or a KL constraint), which is also cheaper on large models (Appendix C.5), the link with inoculation prompting, the evaluation on traits that a prompt cannot elicit without further training, the study of surprising backdoors, and the gated variants. For the comparison, we also train SV++, a stronger version of the published recipe (the full inoculation dataset, a KL coefficient tuned on B2, and a task LoRA); it is a baseline, not a contribution.

Weakness 2: Safe-LoRA and orthogonal projection

We implemented three such methods on B2, U2, and E1 (Appendix C.3): training-time projection, which keeps the task adapter orthogonal to the inoculation adapter; post-training projection, which projects the finished SFT(Harmful) adapter away from the inoculation adapter; and Safe-LoRA, which projects it onto the difference between the instruct and base checkpoints. None matches IA on all three setups. Training-time projection leaves the undesired trait close to SFT(Harmful), and post-training projection removes part of it.

Safe-LoRA reduces the undesired trait on the two Qwen setups, more at its stronger threshold, where it comes close to IA on U2 but stays short of it on B2. On E1 (Llama-3.1-8B), no threshold gives a usable model, and the threshold has to be chosen per setup.

Weakness 3: full fine-tuning and RL

With the IA frozen and the whole backbone trained for one epoch (Appendix C.5, on B1, B2, E1, and U2), IA keeps most of its protection on B1, E1, and U2, and part of it on B2 (EM rate 0.10, against 0.02 with a task LoRA and 0.21 without an adapter). Under SV's longer five-epoch stage, it loses part of its protection on E1, B1, and B2. We did not test RL, and the Discussion leaves it for future work.

Weakness 4: seeds

See the general reply, point 2.

Hard-to-elicit setups

You highlighted the setups where a prompt cannot elicit the trait. One of the three changed. On U2, inoculation prompting suppresses the trait once its prompt is applied correctly (general reply, point 3), but leaves a backdoor with the general prompt. Unlike U1 and U3, U2 is weakly elicitable: the trait-specific prompt raises hate speech from 0.00 to 0.11 before training. Inoculation prompting remains among the worst baselines on U1 and U3. Section 3.2 frames the condition as whether the trait can be elicited without further training, rather than by a prompt alone; we do not study this condition further.

====

If you have any remaining concerns, we are happy to answer further questions. Otherwise, please consider reassessing your score. Thank you!

Thanks for your clarifications! I like the work and believe my current score reflects my judgement accurately. All the best.

Comments on the paper

2 comments

We thank the three reviewers for careful, specific, and constructive reviews. Each engaged closely with the paper, and their suggestions led to the main changes of this revision: the comparison with security vectors and the reframing around them, the ablations of the gated variants, the seed study, and the checks of the judges and of general capabilities. All three find the paper clearly written, and they describe the experiments as well designed (UvWa), extensive (YfJW), and comprehensive (jYXN); between them, YfJW and jYXN highlight the hard-to-elicit setups, the controls, the Pareto analysis, and the backdoor analysis.

We uploaded a revised PDF in which the changes for each reviewer are highlighted: blue for jYXN, green for YfJW, orange for UvWa, and red for the points shared by several reviewers and our corrections. We will remove the colors and the "(Updated.)" caption markers in the final version.

1. Security vectors (all reviewers)

All three reviewers asked how inoculation adapters compare with security vectors (Zhou et al., 2024), which use the same mechanism and were not a baseline in the submission. We agree that this comparison was missing. The abstract, contributions, Discussion, and conclusion now present IA as the mechanism of security vectors with a simpler training recipe, and the paper compares the two on seven of the nine setups (Section 3.1, Figure 2 bottom right; Appendix C.5). The other two setups train a 4-bit Llama-3.1-70B, whose quantized weights the min-min step of security vectors cannot update.

We compare two recipes with IA. SV is the published recipe: KL regularization, bi-level min-min optimization, and a full-backbone task stage. SV++ is a strengthened recipe: a security vector trained on the full inoculation dataset, a KL coefficient tuned on B2, and a task LoRA as in IA.

Normalized means over the seven setups where security vectors can be trained, scores over all answers (desired trait: higher is better; undesired trait: lower is better; confidence intervals are wide, Appendix C.5):

  • IA: 0.65 desired, 0.12 undesired
  • GIA: 0.76 desired, 0.17 undesired
  • SV: 0.28 desired, 0.15 undesired
  • SV++: 0.62 desired, 0.12 undesired

SV retains much less of the desired trait than IA, at a similar level of the undesired trait. SV++ performs about as well as IA on six of the seven setups, but learns neither trait on E1, so on average it falls slightly behind IA.

Since SV++ comes close to IA, most of the gap between SV and IA comes from the recipe, not the mechanism. Controls on four setups show that SV's loss of desired trait comes from its security vector combined with its full-backbone task stage (Appendix C.5). The published recipe trains that security vector on 100 examples, which Zhou et al. report to be enough in their setting; on our setups, training it on the full inoculation dataset retains more of the desired trait.

As in the submission, we credit the mechanism to security vectors. The revision states more clearly that IA's advantage over SV++ is mainly a simpler training procedure, which is cheaper on large models and also trains on 4-bit quantized models (Appendix C.5).

On backdoors (Figure 3), we still observe no surprising backdoor attributable to IA. SV++ shows none either, while SV shows some on B1 and B3.

2. Training seeds (jYXN, YfJW)

jYXN and YfJW noted that each configuration was trained with a single seed. We repeated six methods with five seeds on B2, U2, and E1 (Appendix F.1). A single seed misses some differences between methods, but it did not create any:

  • Of the 40 differences significant at the single reported seed, all 40 stay significant with five seeds, with the same sign.
  • Of the 50 that are not, 17 become significant with five seeds, with the same sign.

3. A correction of ours: inoculation prompting on U2

A configuration error trained the submitted U2 runs of inoculation prompting without the inoculation prompt. Applied correctly, inoculation prompting suppresses hate speech on U2. With the general prompt, however, it leaves a backdoor that IA and GIA do not.

The claim is therefore narrower. Section 3.2 said that "inoculation prompting performs among the worst baselines" on U1 to U3; it now says so "on U1 and U3, where the trait is a new capability or the model follows the prompt unreliably." Inoculation adapters still suppress the trait on all three, though only partly on U3.

We uploaded the final revision. It includes the framing changes announced in our last reply to Reviewer UvWa. We also removed the colour coding that showed which changes addressed each reviewer, along with the “(Updated.)” markers. We made a few small wording and layout corrections; no result changed.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.