acceptodds
Under review as a conference paper at ICLR 2027

Disentangling Safety and Capability in Large Language Models via Causal Interventional Decomposition

Abstract

Safety alignment in large language models can deteriorate during downstream post-training because task learning and safety control remain coupled through shared Transformer computation. Existing safety-preserving methods primarily reduce geometric overlap between task-related and safety-related parameter directions, but geometric separation does not directly indicate whether the resulting functions are truly independent. To address this issue, we propose LCID, Linear-to-Causal Interventional Disentanglement, which decomposes post-training updates into shared, task-private, and safety-private components and uses controlled branch interventions to quantify their target and cross-functional effects. LCID directly suppresses task-to-safety and safety-to-utility cross-effects while preserving the intended contributions of each private branch and anchoring the complete model to its original aligned safety behavior. Experiments on Qwen3-4B, Llama-3.1-8B, and GLM-4-9B show that LCID substantially improves the safety-utility trade-off and consistently reduces unsafe responses across multiple safety benchmarks. Prompt-level paired analysis further demonstrates that these safety improvements are systematic across harmful inputs. Intervention analysis shows that near-orthogonal parameter updates can still exhibit substantial functional cross-talk, whereas LCID achieves markedly lower cross-effects while maintaining strong target effects. These results establish functional intervention as a direct criterion for disentangling safety and capability during post-training.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM Safety, Safety–Capability Disentanglement, Causal Intervention, Mechanistic Interpretability

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_pZosOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes LCID (Linear-to-Causal Interventional Disentanglement) for preserving safety during downstream fine-tuning of large language models. LCID decomposes the model into shared, task-private, and safety-private components, and uses branch-removal interventions to measure each private component’s intended contribution and its influence on the other function. Its training objective suppresses cross-effects, preserves target contributions, and anchors the complete model to the original safety-aligned model. Evaluation covers Qwen3-4B, Llama-3.1-8B, and GLM-4-9B, using GSM8K, MBPP, and IFEval for utility and AdvBench, BeaverTails-H, and StrongREJECT for safety. Additional analyses examine parameter geometry, intervention effects, and prompt-level safety changes. The authors report improved safety–utility trade-offs and reduced functional interference relative to the evaluated baselines.

Critical strengths
  1. A concrete formulation of functional disentanglement. The paper distinguishes parameter-space orthogonality from functional independence and operationalizes this distinction through controlled task/safety branch interventions. Measuring and explicitly suppressing cross-functional effects provides a useful framework for studying and reducing interference between downstream capability and safety.

  2. Promising reported safety–utility trade-off. On Qwen3-4B, LCID achieves a lower average unsafe-response rate than GuardSpace (4.09% versus 8.50%), with comparable average utility (71.39% versus 71.27%). These aggregate results suggest practical potential for preserving safety during downstream adaptation and provide the strongest empirical support for the proposed approach.

Critical weaknesses

1. Limited coverage of recent, closely related work. The related-work discussion and experimental comparisons do not adequately cover recent advances in safety-preserving fine-tuning. Two particularly relevant omissions are Safety Anchor: Defending Harmful Fine-tuning via Geometric Bottlenecks (SBR, ICML 2026) and Jailbreak to Protect (ICML 2026). Comparisons with these methods would help establish LCID’s additional benefit beyond safety anchoring and existing modular adaptation strategies, respectively. Please discuss these connections and include comparisons under matched task data, safety supervision, and comparable training budgets.

2. Incomplete baseline coverage across backbones. Table 1 provides the full baseline comparison only for Qwen3-4B. For Llama-3.1-8B and GLM-4-9B, the reported results compare only Orth. and LCID, omitting the external baselines evaluated on Qwen. This limits the evidence that LCID’s advantage over existing methods generalizes across model families. Please extend the comparison with at least one key external baseline to these two backbones, or qualify the corresponding claims.

3. Unclear and inconsistent figure reporting. Figure 5 requires clarification. Its middle-panel caption specifies a comparison from linear orthogonalization to LCID, whereas the annotated reductions repeat Figure 4’s reductions relative to standard fine-tuning. For example, Table 1 implies Orth.-to-LCID reductions of 68.67 and 57.66 percentage points for Llama and GLM, respectively, rather than the annotated 67.3 and 56.3. Furthermore, the right panel labels both target and cross-effects as JS effects, while Eq. (2) defines target effects as supervised-loss improvements. Please reconcile the comparison baselines and metric definitions, explain any differences in evaluation settings, and explicitly define the reported selectivity ratios.

Questions for authors
  1. In Figure 5, are the unsafe-rate reductions computed relative to standard fine-tuning or Orth.? Please reconcile the caption and annotated reductions with Table 1.

  2. Why are target effects reported as JS divergence in Figure 5, whereas Eq. (2) defines them as supervised-loss improvements? Please clarify the metric definitions and how the selectivity ratios are calculated.

  3. Why does Table 1 compare only Orth. and LCID for Llama-3.1-8B and GLM-4-9B? Are comparable results against external baselines available to support the claimed advantage across model families?

Reviewer_ZXisOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

LCID splits post-training updates into shared, task-private, and safety-private components with gates that can switch each private component on or off. From these gated variants it defines four intervention effects, and it trains to suppress the cross-effects (task→safety, safety→utility) while preserving the target effects and anchoring the model to the original aligned model. The paper reports lower unsafe-response rates than baselines on Qwen3-4B, Llama-3.1-8B, and GLM-4-9B.

Critical strengths

The distinction between geometric orthogonality and functional independence is a useful framing.

Critical weaknesses
  1. The method and experiments are severely under-specified, and the reported numbers contradict each other.
  • There is no appendix, although the main text repeatedly defers ablations and counterfactual analyses to it.
  • how θ_sh is parameterized (full fine-tuning or LoRA);
  • the composition and size of D_U and D_S;
  • λ1–λ3 and the margins m_t and m_s;
  • the number of training steps and seeds.
  • No code is provided.
  1. Table 2 compares exactly the cross-effects that LCID directly minimizes. Showing that a method optimizing X achieves lower X does not show that functional disentanglement causes safety gains. More importantly, the shared parameters θ_sh receive both utility and safety gradients but are never intervened on. Task-induced safety degradation can therefore sit entirely in θ_sh without appearing in any of the four effects. If so, LCID’s safety gains likely come mainly from L_anchor and L_S. The paper lacks an ablation that removes L_cross and L_target and keeps only the anchor and L_S. It also lacks standard baselines that use safety data: safety-data mixing during fine-tuning, constrained SFT, Lisa, Vaccine, and similar methods. On Qwen3-4B, MBPP drops from 72.8 to 61.8, which contradicts the claim that downstream capability is preserved. For Llama and GLM, LCID is compared only against “Orth.”; neither the aligned base model nor standard fine-tuning is reported.
Questions for authors
  1. Please reconcile the inconsistencies above: the 72.0 pp reduction in Fig. 4 vs. the 48.53% unsafe rate in Table 1; the cosine and effect values in Fig. 5 vs. Table 2; and the difference between “Orth.” and “Shared–Private + Orth.”

  2. What results does L_U + L_S + λ3·L_anchor alone (without L_cross and L_target) achieve? Do the baselines also have access to D_S?

  3. How do you measure the task→safety influence carried by θ_sh, for example by comparing F_00 with the original aligned model on safety inputs?

  4. Are the intervention effects measured on training data or on held-out data? Please provide the full hyperparameter and data configuration.

Reviewer_4wNWOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper argues that geometric separation of task and safety updates (orthogonality, null-space projection) does not ensure the two functions are independent. It proposes LCID, which parameterizes post-training weights with binary gates. The author defines a 2x2 intervention matrix by removing each private branch and measuring effects on task performance and safety. The experiments finetune saveral model (Qwen, Llama, GLM) on math, IFEval task and measure unsafe response rates reported on Advbench, strongreject. LCID reaches 4.09% average unsafe rate vs. 8.5% SOTA and with similar task performance.

Critical strengths
  1. The question is well posed and well-motivated: does geometric separation of updates imply functional independence? The intervention matrix is an intuitive diagnostic that others could reuse.
  2. Prompt-paired analysis and explicit refusal measurement. The paper reports bootstrap intervals and Holm-corrected McNemar tests, and includes benign refusal rather than evaluating harmful prompts alone.
Critical weaknesses
  1. Reproducibility: could you include the experimental detail? There's no section about the safety dataset or whether finetuning on full-parameter or PEFT.
  2. Over-refusal. LCID's benign-refusal rate (18.0%) is the highest of all post-trained methods (GuardSpace 15.3%, SafeLoRA 14.2%, Standard FT 5.6%). That conflicts with the claim that the safety gain "is not obtained simply by shifting the model toward indiscriminate refusal"
  3. On Qwen3-4B, LCID drops MBPP by 11.0 points versus Standard FT (61.80 vs 72.80). Do you know what's the reason for causing this big drop?
  4. For Llama-3.1-8B and GLM-4-9B, Table 1 compares LCID only with "Orth.", which is never defined. Orth.
Questions for authors

as mentioned above

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.