acceptodds
Under review as a conference paper at ICLR 2027

UniTrust: Unifying Risk Detection and Intervention with Minimal Task Supervision

Abstract

Adapting language models to new safety tasks remains costly, as doing so often requires large task-specific datasets for learning both what to detect and how to respond. We introduce UniTrust, a unified approach for adapting language models to new safety tasks from a compact task specification consisting only of a natural-language description and a handful of positive and negative examples. We formulate task-specific safety adaptation as complementary read and write operations over a shared hidden-representation space: task-relevant risk is read from frozen representations for prediction, while task-specific behavioral changes are written through activation interventions. Starting from minimal task-specific supervision, UniTrust trains a lightweight detector over frozen hidden representations to identify task-specific risks and derives activation directions from matched pairs of safe and unsafe responses to guide generation. To provide broader task coverage without additional human annotation, our Corpus-Grounded Synthesis procedure retrieves diverse contexts from public corpora and rewrites them according to the target task specification, constructing supervision for both detection and intervention. With only ten positive and ten negative real examples per task, UniTrust consistently improves detection across ten safety domains and three model families, and reduces average unsafe compliance by more than 50% across aligned models and their safety-weakened counterparts. Experiments demonstrate that effective safety adaptation can be built from a minimal task specification with little task-specific human supervision. Code and data are available at https://anonymous.4open.science/r/UniTrust.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LLM Safety, Safety Adaptation, Risk Detection, Activation Steering, Minimal Supervision, Corpus-Grounded Synthesis

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_hHA6OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies task-specific safety adaptation from a blueprint, i.e., a simple task description, 10 positive examples, and 10 negative examples. The authors introduce a framework called UniTrust that enables frozen LLMs to perform this adaptation. UniTrust can be understood as having two main stages: data synthesis and read-write coupling. In the data synthesis stage, an external generator is used to generate thousands of synthetic inputs based on the 20 examples, and the Corpus-Grounded Synthesis (CGS) retrieves contextually similar unlabeled texts from an online corpus and uses the generator to rewrite them accordingly to the target task. These two types of synthetic data then merge with the 20 seeds to form the training pool. In the read-write stage, a lightweight MLP probe is trained on the frozen LLM’s hidden states to predict risk scores, and contrastive steering vectors are derived from matched response pairs. The score is used to gate the steering in inference. UniTrust is evaluated across 10 benchmarks and 3 LLMs. The results show that CGS improves macro-F1 by 4.2–15.6 points and lowers mean FPR from 32.25% to 12.78% compared with plain synthetic expansion. On 2 intervention tasks across aligned and safety-weakened model pairs, gated steering reduces average unsafe compliance from 41.71% to 20.17%.

Critical strengths
  1. The paper introduces a complete pipeline on a practical issue in task-specific safety adaptation for LLMs. The pipeline covers generating synthetic data, training a lightweight probe, extracting contrastive steering vectors, and gated steering at inference. It is practical and avoids backbone finetuning, achieving a 0.862–0.871 macro-F1 (vs. 0.931–0.938 full-supervision Oracle).

  2. The paper includes good evaluation and ablation studies that cover a wide range of benchmarks and models with different safety behaviors. Furthermore, the ablation studies show CGS's contribution over plain synthetic generation, which may be useful for other fields that need synthetic generation.

Critical weaknesses
  1. The detection gains are measured against weak baselines. Seed-based ICL and Prompt2Model use a fixed 0.5 threshold that collapses to 0.333 macro-F1 on several domains. That leaves Seed-only MLP as the only meaningful baseline, which UniTrust beats by 4.4–7.2 points despite using 3,020 training examples versus 20. The closest prior methods for each component, DataTune for data construction and CAST for gated steering, are not compared.

  2. The “minimal supervision” claim underestimates how much supervision UniTrust actually uses. The synthetic data is created by a stronger LLM, and CGS draws thousands of unlabeled texts from domain-specific Hugging Face pools, while the test sets are themselves public Hugging Face datasets. The paper describes excluding duplicates and original split texts, but no near-duplicate filtering, and Figure 3 shows CGS data shifting toward the original data distribution, from which the test sets are also drawn, so in-distribution leakage might be possible. Without the generator itself as a 20-shot classifier or off-the-shelf guards as references, it’s unclear how much of the gain comes from the method versus distilling a stronger model on near-test data.

Questions for authors
  1. Which datasets make up the retrieval pool? Are there any procedures to ensure there were no near-duplicates or the same content under different dataset names?

  2. Although UniTrust keeps the LLM frozen, have you tested its effect against some safety alignment method (especially those using a small amount of data)?

  3. Have you tested/studies the effect of the generator in the pipeline?

  4. How does CGS compare with DataTune (or a similar method), and how does gated steering compare with CAST (or a similar method) under fair conditions using your setup and data?

10 comments

Question 1:

Which datasets make up the retrieval pool? Are there any procedures to ensure there were no near-duplicates or the same content under different dataset names?

Weakness 2:

The “minimal supervision” claim underestimates how much supervision UniTrust actually uses. The synthetic data is created by a stronger LLM, and CGS draws thousands of unlabeled texts from domain-specific Hugging Face pools, while the test sets are themselves public Hugging Face datasets. The paper describes excluding duplicates and original split texts, but no near-duplicate filtering, and Figure 3 shows CGS data shifting toward the original data distribution, from which the test sets are also drawn, so in-distribution leakage might be possible. Without the generator itself as a 20-shot classifier or off-the-shelf guards as references, it’s unclear how much of the gain comes from the method versus distilling a stronger model on near-test data.

Response:

Thank you for these important questions. We clarify UniTrust's minimal human supervision requirements, provide details on retrieval sources and data-overlap checks, and evaluate the contribution of the external generator.

Human annotation and additional resources.

Our “minimal supervision” claim refers to the human-provided task-specific seed annotations: ten risky and ten benign examples per task, together with a natural-language task description. The main configuration uses the following data:

ComponentExamples per taskSource
Labeled seeds20Human-provided task examples
Plain training examples2,000Generated from the task specification and seeds
CGS training examples1,000Generated using retrieved text as context
Generated validation examples800400 Plain + 400 CGS

The detector therefore trains on 3,020 examples, including the 20 supplied seeds. Validation inputs and their target labels are also constructed automatically.

UniTrust requires only 20 human-labeled examples per task. The stronger pretrained generator and retrieved unlabeled corpora automatically expand this limited human supervision into task-specific training and validation data. Importantly, CGS does not use annotations from retrieved datasets; the labels of generated examples are determined by the supplied task specification. Thus, UniTrust maintains minimal human supervision while leveraging automatically constructed data to support both risk detection and intervention.

Retrieval sources and content-overlap checks.

The anonymous repository submitted with our original manuscript already includes sources.json [1], which lists upstream identifiers, dataset links, and reference counts. This source inventory has not been modified during the rebuttal. Grounding sources include allenai/real-toxicity-prompts, Ayush-Singh/social-bias-frames-splits, Lakera/gandalf-rct, cybersectony/PhishingEmailDetectionv2.0, LabHC/moral_stories, and CAiRE/prosocial-dialog-eng_Latn-bloomz-rlhf. Per-task provenance.jsonl files identify retrieval-derived training examples and their upstream sources. The registry also attributes the original benchmark records, so these are distinguished from retrieval-derived examples by their recorded component.

Our procedure includes the following checks:

  1. Task-conditioned discovery and source exclusion. We combine task-specific search hints with short queries generated from the task description and supplied examples. Candidate sources undergo relevance and text-field suitability checks. Declared benchmark identifiers and name fragments are excluded through the publicly available source-exclusion function [2].

  2. Text-based exclusion across dataset names. Within each task, a shared text collection is initialized from the original benchmark split texts and updated across candidate sources. Case and whitespace normalization allow repeated content to be removed independently of its dataset identifier. Thus, identical text hosted under different names is checked at the content level. This implementation is available in retrieval.py [2].

  3. Source-label isolation and rewriting. The selected source text supplies context; its original annotation columns are excluded from retrieval and rewriting inputs. CGS constructs new examples using the task description, labeled seeds, and requested target label. The rewriting prompt [3] explicitly instructs the generator to use contextual attributes and avoid copying the anchor verbatim. Generated examples undergo another normalized-text exclusion check before acceptance, as shown in rebuild.py [4].

Question 1:

Which datasets make up the retrieval pool? Are there any procedures to ensure there were no near-duplicates or the same content under different dataset names?

Weakness 2:

The “minimal supervision” claim underestimates how much supervision UniTrust actually uses. The synthetic data is created by a stronger LLM, and CGS draws thousands of unlabeled texts from domain-specific Hugging Face pools, while the test sets are themselves public Hugging Face datasets. The paper describes excluding duplicates and original split texts, but no near-duplicate filtering, and Figure 3 shows CGS data shifting toward the original data distribution, from which the test sets are also drawn, so in-distribution leakage might be possible. Without the generator itself as a 20-shot classifier or off-the-shelf guards as references, it’s unclear how much of the gain comes from the method versus distilling a stronger model on near-test data.

Response:

The following table summarizes exact-text overlap checks against the corresponding held-out test sets:

Audited materialRecords examinedExact-text matches with corresponding test sets
Candidate corpus pools3,851,5920
CGS rewrite pools24,0000

We found zero exact-text matches. We further audited 9.6 million same-task input–test pairs using character and word 5-gram Jaccard similarity. Neither metric identified any matches at the ≥0.90 threshold, with a maximum word-level similarity of 0.6296.

Direct generator and safety-guard references.

We evaluated DeepSeek using the data-generator API configuration as a direct 20-shot classifier, with the same task descriptions and labeled examples. We also evaluated Qwen3Guard-Gen-8B in prompt-moderation mode using its fixed moderation policy. All methods were evaluated on the same ten-task test sets.

MethodTraining/context inputTen-task Macro-F1 (%) ↑Mean latency (ms/input) ↓
Direct generator reference, 20-shotSame 20 examples in context91.12787.84
Qwen3Guard-Gen-8BFixed moderation policy65.54420.49
UniTrust, Qwen2.520 seeds + 3,000 constructed training inputs87.1227.99
UniTrust, Qwen320 seeds + 3,000 constructed training inputs86.1643.39
UniTrust, Llama3.120 seeds + 3,000 constructed training inputs86.6531.20

Latency was measured separately on 32 shared test inputs covering all ten tasks, with single-request execution. Local methods used one A100 40GB GPU. UniTrust timing includes tokenization, the frozen-backbone forward pass, and detector inference.

The stronger generator achieves higher classification performance, and we acknowledge its knowledge as a substantive source of supervision. UniTrust obtains 86.16–87.12 Macro-F1 using local frozen 7–8B backbones, compared with 65.54 for the off-the-shelf Guard. Generator calls occur during offline construction; deployment uses the learned local detector, which also supplies the risk signal for conditional activation intervention.

Finally, we conducted a size-matched comparison with 2,020 training examples per condition. Replacing 1,000 Plain training examples with 1,000 CGS examples increases average Macro-F1 from 76.40% to 86.55% across three backbones and ten tasks. The GLM-5.3-Flash comparison follows the same training-size and improves average Macro-F1 from 81.92% to 85.84% (+3.92 percentage points). With training and validation counts matched, these comparisons support a benefit from corpus-grounded training-and-validation construction beyond access to the stronger generator or additional training examples.

Our intended claim is low human annotation requirements for task adaptation, with the pretrained teacher, external text, generated supervision, and construction resources explicitly acknowledged.

[1] UniTrust anonymous source release. data/sources.json: dataset-source registry. https://anonymous.4open.science/r/UniTrust/data/sources.json

[2] UniTrust anonymous source release. src/unitrust/retrieval.py, lines 85 and 288: source exclusion and normalized-text exclusion. https://anonymous.4open.science/r/UniTrust/src/unitrust/retrieval.py#L85; https://anonymous.4open.science/r/UniTrust/src/unitrust/retrieval.py#L288

[3] UniTrust anonymous source release. src/unitrust/prompts.py, line 169: rewriting prompt. https://anonymous.4open.science/r/UniTrust/src/unitrust/prompts.py#L169

[4] UniTrust anonymous source release. src/unitrust/rebuild.py, line 75: generated-text exclusion. https://anonymous.4open.science/r/UniTrust/src/unitrust/rebuild.py#L75

Question 2:

Although UniTrust keeps the LLM frozen, have you tested its effect against some safety alignment method (especially those using a small amount of data)?

Response:

Thank you for this important question. We conducted an additional comparison with LoRA-DPO, which trains low-rank adapters through direct preference optimization.

Both methods use the same 1,000 response pairs per task: 500 Plain and 500 CGS pairs. We evaluate Qwen3-8B and its Abliterated variant using the same target checkpoints, test prompts, decoding settings, and scoring criteria. LoRA-DPO trains for three epochs.

LoRA-DPO trains 43.65 million adapter parameters per task. UniTrust trains a 0.525-million-parameter risk detector and constructs steering directions without gradient optimization of the LLM. Thus, UniTrust uses 98.80% fewer trainable parameters.

The following table reports unsafe compliance, averaged across the two model variants. All rates are percentages.

MethodTrainable parametersHarmful Request ↓Safe–Sensitive Boundary ↓Average ↓
Unsteered backbone045.5043.2544.38
LoRA-DPO, 5×10−65\times10^{-6}43.65M43.7543.0043.38
LoRA-DPO, 10−510^{-5}43.65M42.2543.2542.75
LoRA-DPO, 10−410^{-4}43.65M38.2529.7534.00
UniTrust0.525M12.2511.0011.63

UniTrust achieves lower unsafe compliance than these DPO configurations while training 83.2 times fewer parameters. We additionally evaluated a higher DPO learning rate to examine whether stronger policy updates could further improve safety. The complete learning-rate comparison, averaged across both models and tasks, is shown below.

Method / learning rateUnsafe compliance ↓Appropriate help ↑Over-refusal ↓
LoRA-DPO, 5×10−65\times10^{-6}43.3898.380.50
LoRA-DPO, 10−510^{-5}42.7597.880.63
LoRA-DPO, 10−410^{-4}34.0097.751.00
LoRA-DPO, 5×10−45\times10^{-4}4.0039.2559.88‾\underline{59.88}
UniTrust11.6393.755.50

Increasing the DPO learning rate from 10−410^{-4} to 5×10−45\times10^{-4} reduces unsafe compliance from 34.00% to 4.00%, while appropriate help falls from 97.75% to 39.25% and over-refusal rises from 1.00% to 59.88%. This demonstrates learning-rate sensitivity in the evaluated small-data setting. UniTrust achieves 11.63% unsafe compliance while retaining 93.75% appropriate help.

Crucially, UniTrust also provides explicit task-conditioned risk detection. Its detector produces a risk score that supports standalone classification and controls whether steering is applied. LoRA-DPO itself optimizes response preferences and supplies neither a dedicated risk detector nor a configurable detection-based gate. UniTrust therefore supports both risk assessment and selective intervention within one framework, in addition to keeping the LLM parameters frozen.

Reviewer_nnJBOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes UniTrust for adapting a frozen language model to a new safety task from a task description and ten positive and ten negative examples. Corpus-Grounded Synthesis (CGS) retrieves public texts and rewrites them according to the target task and label. A small MLP detects risk from prompt representations, while safe/unsafe response pairs define activation directions that are applied when the detector triggers. Experiments cover ten detection domains across three model families and two intervention tasks across six model variants. The results show improved detection and lower average unsafe compliance, with some loss of helpfulness on benign inputs.

Critical strengths

The paper addresses a useful safety adaptation problem with a practical supervision interface. A task description and a few labeled examples specify the safety boundary and support both detection and intervention. The detector and steering modules are straightforward, and adaptation does not require updating the backbone. Evaluating detection across ten domains and three model families provides useful evidence beyond a single benchmark.

The controlled detection experiment provides convincing evidence for the value of corpus grounding. Table 8 holds training and validation budgets fixed and improves average macro-F1 from 0.764 to 0.865, ruling out data quantity alone as an explanation. The intervention results also reduce average unsafe compliance from 41.71% to 20.17%. Reporting benign helpfulness and over-refusal alongside safety outcomes is valuable because it makes the cost of intervention visible.

Critical weaknesses

The incremental contribution over closely related methods is not sufficiently established. CGS is close to SynthesizRR (EMNLP 2024), which already uses seed examples to retrieve corpus texts, filters them by similarity, and rewrites them according to a target label. DataTune (ACL Findings 2024) also retrieves and transforms data for a target task and is acknowledged in the paper. On the intervention side, CAST (ICLR 2025) already conditions activation steering on prompt representations. UniTrust offers a useful combination, but the current comparisons do not clearly establish what it adds beyond these approaches. Table 8 supports corpus grounding over plain synthesis; it does not establish an advantage over existing retrieval-and-rewriting methods. Comparisons under the same supervision and generation budget would clarify the contribution.

The intervention experiments do not isolate CGS's additional benefit or establish that it makes the policy more selective. The full method changes detector training and validation data while increasing steering supervision from 1,000 to 2,000 response pairs per task (Appendices A.3 and A.5). The reduction in gated unsafe compliance from 21.83% to 20.17% therefore does not separate the effects of detection, steering, and data quantity; the size-matched detection ablation does not resolve this. More directly, Table 2(a) and the missed-input counts in Section 5.2.2 imply that steering is applied to 50.79% of benign inputs with CGS, versus 21.33% without CGS, while harmful-input coverage rises from 84.08% to 97.54%. These calculated gate-activation rates are not refusal rates; the detection FPR uses a separate threshold and evaluation setting. Gating improves on always-on steering, but mean benign helpfulness still falls from 97.71% unsteered to 94.79%, and over-refusal rises from 0.62% to 4.00% (Table 9). Fixed-budget detector/direction swaps and comparisons at matched harmful coverage or helpfulness would clarify the contribution. Reduced steering counts also need runtime measurements before they establish computational savings.

Questions for authors
  1. What specifically distinguishes CGS from existing retrieval-and-rewriting methods? Could the authors compare it with SynthesizRR using the same seeds, corpus, generator, data budgets, and downstream probe? For the full pipeline, how does UniTrust compare with CAST under the same constructed supervision?

  2. Does CGS improve steering directions independently of the detector? Could the authors cross detectors and directions built from plain-only or CGS-enriched data in a 2×2 comparison, holding detector training sizes and response-pair counts fixed? Using a common validation set and the same gate-selection rule would separate the effects of data composition, detection, and steering. The existing systems use different budgets, so their results alone do not complete this controlled comparison.

  3. Does CGS improve the gate's tradeoff at matched harmful-input coverage? Table 2 implies that benign gate activation increases from 21.33% without CGS to 50.79% with CGS, alongside higher harmful coverage. Can CGS reduce benign interventions and preserve helpfulness when harmful coverage is matched? Figure 6 varies steering strength while holding the gate fixed; please also vary the gate threshold and report variability across independently sampled seed sets. Matching an operating point need not mean using the same numerical threshold for different detectors.

  4. How much of the detection gain comes from better ranking versus a better decision threshold? For Qwen2.5 Toxicity, macro-F1 rises from 0.560 to 0.962, while AUROC rises from 0.982 to 0.995 (Tables 1 and 7). This suggests that the operating point deserves closer analysis, although ranking also improves. Could the authors show validation/test score distributions and FPR at matched recall? For ICL and Prompt2Model, please also report results with thresholds selected under the same validation budget as UniTrust, alongside their default 0.5 thresholds (Table 6).

  5. How is the gated policy executed during online generation? Section 5.2.1 selects between pre-generated steered and unsteered answers offline. When is the gate available, which tokens receive the intervention, and is a second prompt pass needed? The reported 25.83% reduction in intervention counts needs measured latency or compute to establish a deployment benefit.

8 comments

Question 1:

What specifically distinguishes CGS from existing retrieval-and-rewriting methods? Could the authors compare it with SynthesizRR using the same seeds, corpus, generator, data budgets, and downstream probe? For the full pipeline, how does UniTrust compare with CAST under the same constructed supervision?

Weakness 1:

The incremental contribution over closely related methods is not sufficiently established. CGS is close to SynthesizRR (EMNLP 2024), which already uses seed examples to retrieve corpus texts, filters them by similarity, and rewrites them according to a target label. DataTune (ACL Findings 2024) also retrieves and transforms data for a target task and is acknowledged in the paper. On the intervention side, CAST (ICLR 2025) already conditions activation steering on prompt representations. UniTrust offers a useful combination, but the current comparisons do not clearly establish what it adds beyond these approaches. Table 8 supports corpus grounding over plain synthesis; it does not establish an advantage over existing retrieval-and-rewriting methods. Comparisons under the same supervision and generation budget would clarify the contribution.

Response:

Thank you for this important question. We clarify UniTrust's contributions relative to existing methods and provide direct comparisons with SynthesizRR [1], DataTune [2], and CAST [3] under controlled experimental settings.

Unlike approaches focused on either data synthesis or conditional steering, UniTrust provides a unified task-adaptation framework: from only a task description and 20 labeled examples, it constructs supervision for both risk detection and behavioral intervention, learns a task-specific risk detector, and uses its predictions to selectively steer model responses without backbone fine-tuning.

For data construction, SynthesizRR [1] requires a supplied retrieval corpus, limiting its ability to automatically assemble grounding resources for a newly specified task. DataTune [2] relies on a predefined collection of over 75,000 Hugging Face datasets, retrieving only 25 candidates and transforming at most four datasets per task in its reported setup. This restricts corpus selection to a predefined search space and a small set of selected datasets. In contrast, CGS automatically discovers and assembles task-relevant public-corpus contexts from each new safety specification, then rewrites them to generate both labeled detection inputs and matched safe/unsafe response pairs.

Corpus discovery and selection differ substantially across these methods:

MethodSearch spaceCorpus selection per task
SynthesizRR [1]5 predefined corpora1 task-assigned corpus
DataTune [2]Predefined Hugging Face dataset collection (75,000+)Fixed top-25 shortlist; at most 4 datasets transformed
CGS (Ours)Task-driven source discovery within Hugging Face datasetsNo predefined task-specific corpus

We compared the data-construction methods across ten tasks and three model backbones, averaging three training seeds. We fixed the 20 human-labeled seeds per task, generation teacher, detector architecture, and training procedure. The three corpus-based methods share the same 1,000 Plain training examples. The following results apply a common per-task token cap to the varied training-generation component. We reproduced SynthesizRR and DataTune on our safety-task datasets under the same training and evaluation protocol. To address your request for a fair comparison under consistent experimental settings, this comparison follows a different experimental protocol from Table 8, resulting in different reported values.

Data constructionMacro-F1 ↑AUROC ↑Benign FPR ↓
Plain-only79.4891.3427.67
CGS84.8893.3316.91
SynthesizRR84.3692.7022.42
DataTune80.2492.0618.00

In sample-count-matched comparison with 2,020 training examples per method, CGS further outperforms SynthesizRR by 0.85 percentage points in Macro-F1 (85.23% vs. 84.38%).

Overall, CGS improves Macro-F1 by 5.40 percentage points over Plain-only and 4.64 points over DataTune. Compared with SynthesizRR, CGS achieves higher Macro-F1 (+0.52 points) and AUROC (+0.63 points), while reducing benign FPR by 5.52 percentage points.

Question 1:

What specifically distinguishes CGS from existing retrieval-and-rewriting methods? Could the authors compare it with SynthesizRR using the same seeds, corpus, generator, data budgets, and downstream probe? For the full pipeline, how does UniTrust compare with CAST under the same constructed supervision?

Weakness 1:

The incremental contribution over closely related methods is not sufficiently established. CGS is close to SynthesizRR (EMNLP 2024), which already uses seed examples to retrieve corpus texts, filters them by similarity, and rewrites them according to a target label. DataTune (ACL Findings 2024) also retrieves and transforms data for a target task and is acknowledged in the paper. On the intervention side, CAST (ICLR 2025) already conditions activation steering on prompt representations. UniTrust offers a useful combination, but the current comparisons do not clearly establish what it adds beyond these approaches. Table 8 supports corpus grounding over plain synthesis; it does not establish an advantage over existing retrieval-and-rewriting methods. Comparisons under the same supervision and generation budget would clarify the contribution.

Response:

For intervention, CAST [3] relies on representation-similarity-based gating, whereas UniTrust learns a task-specific risk boundary directly from constructed supervision. We therefore compare the two methods to evaluate the effectiveness of learned risk gating.

We compared complete CAST, complete UniTrust, and CAST’s similarity gate combined with the same UniTrust direction. Unlike the main Table 2 configuration, this comparison uses matched supervision budgets of 2,020 detector-training examples and 1,000 response pairs per task, with consistent intervention settings. Results are averaged across six model variants and two tasks:

MethodTest risk coverage ↑Benign gate activation ↓Unsafe compliance ↓
CAST96.0057.8840.83
UniTrust97.4247.4619.88
CAST gate + UniTrust direction96.0057.8820.33

Holding the UniTrust direction and intervention settings fixed, its learned gate reduces benign activation by 10.42 percentage points relative to the CAST gate, while achieving higher test risk coverage and lower unsafe compliance. This isolates a contribution from the gate beyond the choice of steering direction.

Together, these comparisons clarify UniTrust’s incremental contribution: task-defined supervision connecting detection and intervention, and learned risk gating that improves intervention selectivity under the evaluated conditions.

[1] Divekar, A., and Durrett, G. SynthesizRR: Generating Diverse Datasets with Retrieval Augmentation. EMNLP, 2024.

[2] Gandhi, S., Gala, R., Viswanathan, V., Wu, T., and Neubig, G. Better Synthetic Data by Retrieving and Transforming Existing Datasets. Findings of ACL, 2024.

[3] Lee, B. W., et al. Programming Refusal with Conditional Activation Steering. ICLR, 2025.

Question 2:

Does CGS improve steering directions independently of the detector? Could the authors cross detectors and directions built from plain-only or CGS-enriched data in a 2×2 comparison, holding detector training sizes and response-pair counts fixed? Using a common validation set and the same gate-selection rule would separate the effects of data composition, detection, and steering. The existing systems use different budgets, so their results alone do not complete this controlled comparison.

Weakness 2a:

The intervention experiments do not isolate CGS's additional benefit or establish that it makes the policy more selective. The full method changes detector training and validation data while increasing steering supervision from 1,000 to 2,000 response pairs per task (Appendices A.3 and A.5). The reduction in gated unsafe compliance from 21.83% to 20.17% therefore does not separate the effects of detection, steering, and data quantity; the size-matched detection ablation does not resolve this.

Response:

Thank you for highlighting this concern. We agree that the original end-to-end comparison does not by itself separate the contributions of detection, steering-direction construction, and supervision quantity. We therefore conducted the requested 2×2 ablation under matched supervision budgets.

The Plain detector uses 20 seeds + 2,000 Plain examples; the CGS-enriched detector uses the same 20 seeds + 1,000 Plain + 1,000 CGS examples. Both therefore receive 2,020 training examples.

The Plain direction uses 1,000 Plain response pairs, while the CGS-enriched direction uses 500 Plain + 500 CGS pairs, with 1,000 response pairs in both configurations.

We evaluated all four detector–direction configurations on Harmful Request and Safe-sensitive Boundary across three models: Qwen2.5-7B-Instruct, Qwen3-8B, and Llama3.1-8B-Instruct. The averaged results are summarized below.

Detector / DirectionUnsafe compliance ↓Benign gate activation ↓Appropriate help ↑Over-refusal ↓Test risk coverage ↑
A: Plain / Plain6.3314.3394.504.0884.67
B: Plain / CGS-enriched6.0014.3394.174.2584.67
C: CGS-enriched / Plain6.3310.0095.173.5084.08
D: CGS-enriched / CGS-enriched5.9210.0094.923.6784.08

Holding the detector and gate fixed tests the direction contribution. Replacing the Plain direction with the CGS-enriched direction reduces unsafe compliance from 6.33% to 6.00% with the Plain detector, and from 6.33% to 5.92% with the CGS-enriched detector. Appropriate help decreases and over-refusal increases in both comparisons, indicating a safety–helpfulness trade-off.

Holding the direction fixed tests the detector contribution. CGS-enriched detection reduces benign gate activation from 14.33% to 10.00%, increases appropriate help, and reduces over-refusal under both directions. Test risk coverage remains comparable (84.08% vs. 84.67%).

Within these three backbones, the combined configuration achieves the lowest unsafe compliance (5.92%) and higher appropriate help than the fully Plain condition (94.92% vs. 94.50%). These matched-budget comparisons separate the contributions of direction construction and detection.

Reviewer_jX3hOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

UniTrust adapts language models to new safety requirements using a task description and 10 risky and 10 benign examples. It expands these seeds through synthetic generation and rewriting retrieved text, then builds a lightweight classifier that detects risk from the model’s internal activations, and a steering vector that encourages safer responses when the classifier flags an input. The authors report improved detection across ten safety domains and three models and reduced unsafe compliance.

Critical strengths
  • The paper proposes a practical adaptation system with little human labeling through the combination of lightweight data construction, frozen-model probing, and activation steering.
  • The experiments covering ten domains and three backbones provides evidence of applicability across tasks.
Critical weaknesses
  • The effective diversity of the tasks is unclear. A lot of the safety tasks seems closely related and for example toxicity and hate/offensive language share the same Civil Comments dataset.
  • The necessity of task-specific steering is untested: the paper compares gated and always-on applications of its vectors, but does not compare separate task vectors against a pooled vector or vectors transferred from other tasks.
Questions for authors
  • How does a shared steering vector compare with task-specific vectors?
3 comments

Question 1:

How does a shared steering vector compare with task-specific vectors?

Weakness 2:

The necessity of task-specific steering is untested: the paper compares gated and always-on applications of its vectors, but does not compare separate task vectors against a pooled vector or vectors transferred from other tasks.

Response:

Thank you for this valuable question! We compared task-specific and shared steering directions on Harmful Request and Safe-sensitive Boundary, with cross-task transfer as an additional comparison.

We evaluated three configurations:

  • Task-specific: Each task uses a direction constructed from 1,000 response pairs: 500 Plain and 500 CGS pairs.
  • Pooled: A single shared direction uses 1,000 pairs in total: 250 Plain and 250 CGS pairs from each task.
  • Cross-task transferred: The task-specific direction constructed for one task is directly applied to the other.

The following results are averaged across Qwen2.5-7B-Instruct, Qwen3-8B, and Llama3.1-8B-Instruct on both tasks.

Direction configurationUnsafe compliance ↓Safe alternative ↑Appropriate help ↑Over-refusal ↓
Task-specific5.3374.5892.176.50
Pooled5.5074.0892.256.42
Cross-task transferred5.8374.4292.426.33

Task-specific directions achieve lower unsafe compliance (5.33% vs. 5.50% pooled and 5.83% transferred), while shared directions offer comparable safety with better benign-input utility, indicating a modest safety–helpfulness trade-off.

Weakness 1:

The effective diversity of the tasks is unclear. A lot of the safety tasks seems closely related and for example toxicity and hate/offensive language share the same Civil Comments dataset.

Response:

Thank you for this helpful observation! We agree that the diversity of our evaluated safety requirements deserves clarification.

Although Toxicity and Hate/Offensive share the Civil Comments source corpus, they use distinct annotation targets: toxicity\mathtt{toxicity} and identityattack\mathtt{identity_attack}, respectively. These are separately annotated fields in the official Civil Comments dataset documentation [1].

Among 1,999,514 Civil Comments entries, only 8.56% of toxicity-positive comments were identity-attack-positive (threshold 0.5). Across all ten benchmarks, the 4,000 test examples are unique, with zero exact-text overlap across tasks or with the supplied seeds.

Our goal is to evaluate adaptation across different safety requirements. We acknowledge that some tasks are semantically related, while emphasizing their distinct annotation objectives and non-overlapping held-out test inputs.

[1] Google. Civil Comments dataset documentation, “Data Fields.” Hugging Face Datasets (google/civil_comments).

Dear Reviewer jX3h,

Thank you again for your helpful suggestions. We have revised the manuscript to explain how the Toxicity and Hate/Offensive tasks use different labels from the same dataset. We also added a comparison of task-specific, shared, and transferred steering directions.

Best regards

The Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Overall Response to All Reviewers

Dear Reviewers,

Thank you for your careful reading and constructive feedback. We appreciate your recognition of the practical value of UniTrust and its evaluation across diverse safety tasks and model families. Your comments have helped us strengthen our experimental evidence and clarify the contributions of our framework.

The central goal of this work is to enable task-specific safety adaptation of frozen LLMs with minimal human supervision, unifying risk detection and selective behavioral intervention within a single framework. To better substantiate this goal, we have conducted extensive additional experiments addressing four key aspects:

  1. Novelty and comparisons: We provide direct comparisons with SynthesizRR, DataTune, CAST, and LoRA-DPO to clarify UniTrust's contributions relative to existing approaches.

  2. Component contributions: We conduct controlled 2×2 ablations and matched-coverage evaluations to disentangle the effects of CGS on detection and steering, and demonstrate improved gate selectivity.

  3. Experimental rigor and robustness: We strengthen baseline comparisons through common validation calibration, evaluate different generators and steering configurations, and provide additional analyses of task diversity and data overlap.

  4. Supervision and deployment: We clarify the role of human annotations, synthetic supervision, and external resources, and provide direct generator comparisons and online inference measurements.

We address each concern in the individual responses below. We believe these additional results provide stronger empirical support for UniTrust's effectiveness and clarify its distinctions from prior work. We also appreciate the reviewers' suggestions for improving experimental rigor and practical evaluation, which we will incorporate into the revised manuscript.

Thank you again for helping us improve the clarity and empirical grounding of this work.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.