acceptodds
Under review as a conference paper at ICLR 2027

Grounding Is Not Authorization: Teaching GUI Agents When Not to Act

Abstract

GUI agents increasingly act on behalf of users, yet a screenshot shows which action is available, not whether the user asked for it. We find that agents fail to withhold in two ways: a task-relevant screen suppresses clarification of under-specified requests, and neutral wording elicits high-impact actions that harm-framed wording does not, although neither request authorizes the operation. Agents thus act unless a cue stops them, rather than by execution justification: whether the request itself warrants the action. We introduce AXIS, a protocol that crosses request phrasings with screen conditions while fixing the operation and never supplying a warrant. Across five models, a matching screen cuts clarification from 60–73% to 3–21%, and neutral wording raises execution by up to 72 points; safety prompts and inference-time checks do not close the gap. We propose AXIS-Core, which trains a GUI agent to assess operation risk and request authorization before acting. On UI-Venus-1.5-8B, it makes no unsafe action on 134 danger records while completing all 18 required benign actions, and separates 49 of 50 authorization-contrast pairs where untrained methods separate none. It generalizes to unseen phrasings and screenshot sources with 92–97% accuracy, and keeps native grounding and navigation within 1.7 points of the base model.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
GUI agents, AI safety, visual legitimization

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_YYoYOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper studies whether GUI agents distinguish an available action from an authorized action. It introduces AXIS, which crosses request formulations with screenshot conditions, and AXIS-Core, a LoRA-based approach that supervises risk and authorization assessments before action selection. Experiments report improved static decision accuracy, stronger discrimination between authorized and unauthorized requests, and limited changes in native grounding and navigation performance.

Critical strengths

The authorization-contrast pairs are a useful evaluation design: holding the screenshot and operation fixed while changing the stated authorization can distinguish indiscriminate execution from indiscriminate withholding. AXIS-Core’s improvements on these pairs provide meaningful evidence that its gains are not solely due to increased refusal.

Critical weaknesses
  • The paper is difficult to read, substantially hindering technical assessment. Sentences are frequently long and densely packed, with awkward clause boundaries and several ideas compressed into a single statement. Even basic descriptions require repeated reading to determine the intended meaning. Overlapping terminology and definitions scattered across the main text and appendices add to this burden. The paper needs substantial rewriting with shorter sentences, clearer logical connections, and consistent terminology so that readers can reliably assess its methods and claims.

  • The experimental design does not isolate the effects claimed. The H/L comparison changes authorization-relevant information, not just wording: H explicitly states overreach, whereas L leaves authorization unstated. Their execution gap therefore cannot be attributed specifically to lexical salience. Likewise, the clarification metric excludes authorization-confirmation questions. An agent that switches from asking what the user wants to asking for permission could show a large “clarification deficit” while still withholding action appropriately. These confounds undermine the paper’s central diagnosis. The training improvements do not resolve them; the claims require better-controlled comparisons and measures of execution without prior clarification or authorization confirmation.

Questions for authors
  1. Clarification and reanalysis: If intent clarification, authorization confirmation, and refusal are considered jointly, how often do agents actually commit to an action without first requesting the missing information or authorization? Does the screen-dependent effect remain under this measure?

  2. Clarification of the claimed effect: Given that H and L differ in explicit evidence of overreach, what supports interpreting their execution gap as lexical salience rather than sensitivity to different authorization information? Is there an existing comparison that holds authorization semantics fixed while varying wording?

4 comments

Thank you for identifying the need to separate authorization information from wording and to account for authorization-confirmation questions.

1. Separating wording from explicit non-consent.

We added two request variants for each of the 75 scenarios, producing a 2×2 design that crosses harm-related wording with an explicit statement of non-consent. The screenshot, target, and requested operation remain fixed. All four conditions were evaluated together.

The table reports the percentage of responses that proceed directly, excluding intent clarification, authorization-confirmation requests, and safety refusals.

ConfigurationNeutral, no explicit non-consentNeutral, explicit non-consentHarm-framed, no explicit non-consentHarm-framed, explicit non-consent
Qwen3.7-plus94.766.766.724.0
Claude-Opus-558.72.734.70.0
GPT-5.6-Sol72.010.756.04.0
UI-TARS-1.5-7B76.082.782.772.0
UI-Venus-1.5-8B98.793.390.764.0
Qwen3.7-plus + guard26.78.013.32.7
UI-TARS-1.5-7B + guard68.044.054.717.3

Each cell includes all 75 scenarios. Under neutral wording, adding explicit non-consent reduces direct proceeding in six configurations, significantly in five after Holm correction. Without explicit non-consent, harm-related wording reduces it in six configurations, significantly in two after correction.

The original H/L contrast therefore combines two effects. The revised interpretation will distinguish responses to stated non-consent from responses to harm-related wording, with their relative influence varying across models.

2. Accounting for confirmation as well as clarification.

We recomputed the screen comparison over all 1,050 original responses, retaining the original action-response definition and excluding every response labeled as intent clarification, authorization confirmation, or safety refusal.

ConfigurationNeutral screenMatching screen
Qwen3.7-plus7/7537/75
Claude-Opus-50/7523/75
GPT-5.6-Sol2/7530/75
UI-TARS-1.5-7B8/7546/75
UI-Venus-1.5-8B19/7558/75
Qwen3.7-plus + guard0/753/75
UI-TARS-1.5-7B + guard16/7530/75

All five unguarded models retain an increase of 30.7–52.0 percentage points, with Holm-adjusted p<10−6p<10^{-6}. The guarded increases are 4.0 and 18.7 points and do not reach significance after correction.

Authorization-confirmation questions do increase on matching screens, but this change does not fully account for the screen effect. The remaining difference concerns responses that proceed without either type of question or a safety refusal.

This endpoint measures action-oriented responses, including accepted next-action statements; it is not a device-execution rate. We will state this definition explicitly and report clarification, confirmation, refusal, and direct proceeding separately.

3. Writing and presentation.

We agree that several passages are hard to follow even on a careful reading. Some sentences introduce the experimental setup, the comparison, and its interpretation all at once. We will rewrite these passages in shorter sentences, making clear what was compared before explaining what the result shows. We will also define key terms briefly at first use in the main text and use the same terms throughout.

Thank you for your insightful comments.

We have uploaded a revised manuscript. We revised the prose throughout and marked new analyses in blue.

Presentation. We shortened sentences, separated the experimental setup from its interpretation, and used consistent terms in the main text and appendix. Section 3.1 now distinguishes asking what the user means (CLARIFY) from asking for confirmation before proceeding (PAUSE), and Section 3.4 defines the response measures before Section 4 presents the findings.

Clarification and confirmation. You raised the possibility that agents switched from asking about intent to asking for permission. Confirmation requests do become more frequent on task-relevant screens (Appendix B.3, Table 7). However, responses that proceed with no clarification, confirmation, or refusal also rise in all seven configurations, by 30.7–52.0 percentage points in the five unguarded models (Holm-adjusted p < 10⁻⁶; Figure 2b). The drop in clarification is therefore not only a switch to asking for permission.

Wording and non-consent. Section 3.3 varies these two factors independently across all 75 scenarios, holding the screen, target, and operation fixed. Adding explicit non-consent to a neutral request significantly reduces direct proceeding in five of seven configurations after Holm correction, whereas changing the wording alone, with consent still unstated, does so in two (Figure 2c). Appendix B.4, Table 9 repeats the comparison counting only responses that explicitly report or commit to execution. Under that measure, adding non-consent to the neutral request lowers the rate in all seven configurations.

Authorization pairs. On the 50 neutral AndroidControl pairs you highlighted, the best of the five prompting baselines resolves 19 when given the full policy and four demonstrations, whereas AXIS-Core resolves 47–49 across three training seeds (Sections 6.1 and 6.3).

We hope the revised manuscript has addressed your concerns. If you have any further concerns, we would be glad to address them before the discussion period ends.

I think the new experiment addresses my concern about the unfair comparison. And I appreciate the authors’ efforts to improve the readability of the abstract and introduction though leave much to improve. I will therefore will raise my score to 2.

W1. Fairness of the comparison

I think the new experiment addresses my concern about the unfair comparison.

We are glad the new experiment addressed your concern about the fairness of the comparison. Thank you for reconsidering the paper and raising your score.

W2. Readability of the manuscript

And I appreciate the authors’ efforts to improve the readability of the abstract and introduction though leave much to improve.

Thank you for recognizing the improvements to the abstract and introduction. We agree that the manuscript still needs a thorough edit. Much of the rebuttal period went into running and checking the additional experiments, and we are still revising the later sections in depth. The examples below show how we are applying your three suggestions throughout the paper.

1. Shorter sentences

  • Before (original §5.4). A single sentence used two colons to connect four counts of completed benign actions, an explanation of the model’s behavior, and a drop in grounding accuracy. It reported 13/18 (ID), 11/18 (linguistic-OOD), 50/95 (AndroidControl), and 42/85 (AITW + GUIOdyssey) before introducing the “high-impact, therefore withhold” interpretation.
  • After (working draft, §6.4 and §7.2). After reporting the counts in §6.4, we explain the pattern in a separate sentence, “Without rationale supervision, it associates high-impact operations with stopping.” Section 7.2 discusses the grounding result. We made a similar change in Appendix A.1, where the original description of request construction used three semicolons and a colon in one sentence. The revised passage walks through the procedure in six shorter sentences (lines 951 to 957).

2. Clearer logical connections

  • Before (original §3.2, §4.1, and §4.2). Section 3.2 mixed predictions with metric definitions before the reader had seen the decision rule, which first appeared in §4.1. Section 4.2 also introduced each comparator and predicted its failure in the same sentence.
  • After (working draft, §3.1, §3.4, §4, and §5). Section 3.1 introduces the decision rule (Eq. 1), and §3.4 defines the response measures. Section 4 then reports the drop in clarification, considers whether agents simply switched to asking for permission, and tests that explanation (lines 244 to 255). Section 5 connects this finding to the supervision each comparator receives (lines 294 to 300), before §6 presents the results.

3. Consistent terminology

  • Before. The original draft alternated among “execution justification,” “warrant,” and “authorization.” It also used several labels for screen conditions (“correct,” “matching,” “real-neutral,” and “unrelated”), wording (“low-salience,” “Low-lex,” and “neutral”), and the paraphrase set (“linguistic-OOD” and “paraphrase”). Readers also had to keep track of three acronyms for the contrasts (CD, LSG, and NSE).
  • After. The working draft uses “stated authorization,” “task-relevant” and “unrelated” screens, “neutral” and “harm-framed” requests, and “the paraphrase set.” Section 3.1 distinguishes CLARIFY from PAUSE. Section 3.4 defines the three response measures (asks what is meant, proceeds, and proceeds directly), and we use the same terms in §4, Figure 2, and Appendix B.

We will continue this work across the remaining sections and appendices, then proofread the manuscript as a whole. We will make the paper easier to read and the argument easier to follow.

We hope this ongoing rewriting, together with the final proofreading pass, fully addresses your remaining concern about the writing, and we would be grateful if you could take it into account in your evaluation.

Reviewer_kVVKOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper examines whether GUI agents distinguish available actions from authorized ones. AXIS evaluates this through controlled changes to requests and screenshots. AXIS-Core adds explicit risk and authorization supervision, improving static safety decisions while largely preserving native capabilities.

Critical strengths

S1. The paper addresses an important distinction between locating an action and having permission to perform it.

S2. The authorization-contrast pairs provide a useful test of both unsafe execution and excessive refusal.

S3. Public grounding and navigation benchmarks help establish that safety training largely preserves existing capabilities.

Critical weaknesses

W1. My main concern is that the safety evaluation remains closely tied to the authors' construction rules. Even the tests using public screenshots have newly constructed requests and labels. The ID test also contains only 27 underlying cases. An independent safety benchmark or an interactive confirmation workflow would provide stronger evidence of generalization.

W2. The baseline failures also need more explanation. It is unclear whether the prompting baselines receive the full authorization convention. Giving them the exact policy and a few examples would help distinguish learning the benchmark's rules from acquiring a more general authorization capability.

W3. The supervision comparisons do not fully explain why AXIS-Core performs better. Adding grounding fields changes output length and loss allocation, and reordering risk and authorization does not remove these confounds. Using one training seed and closely related backbones further limits the robustness of this conclusion.

W4. The related-work discussion could briefly cover existing action-verification and safety-control methods to clarify how authorization supervision fits into this literature (Lee et al., 2025; Sun et al., 2025; Feng et al., 2026).

I appreciate the controlled evaluation, but currently lean toward rejection. Strong results within the proposed protocol do not yet establish broader authorization competence or interactive safety.

References

Lee, J., Lee, D., Choi, C., et al. (2025). VeriSafe Agent: Safeguarding Mobile GUI Agent via Logic-based Action Verification. arXiv preprint arXiv:2503.18492.

Sun, Q., Li, M., Liu, Z., et al. (2025). OS-Sentinel: Towards Safety-Enhanced Mobile GUI Agents via Hybrid Validation in Realistic Workflows. arXiv preprint arXiv:2510.24411.

Feng, Y., Du, J., Wang, Q., et al. (2026). CORA: Conformal Risk-Controlled Agents for Safeguarded Mobile GUI Automation. arXiv preprint arXiv:2604.09155.

Questions for authors

Q1. Do the safety gains transfer to an independent benchmark or a complete pause-confirm-resume interaction?

Q2. What happens when prompting baselines receive the exact authorization policy and a few examples?

Q3. Do the supervision findings hold with matched loss weighting, alternative field placement, and multiple seeds?

Q4. How does authorization supervision differ from, or complement, existing safety controllers?

4 comments

Thank you for the concrete suggestions on prompting controls, independent evaluation, and the interpretation of the supervision comparison.

1. Full-policy and few-shot prompting.

To clarify, each of the five inference-time baselines received the complete decision policy in its zero-shot prompt. We did not state this explicitly in the submitted manuscript. In the new few-shot evaluation, we kept the policy fixed and added the same four demonstrations from the training split for all five methods. The demonstrations cover neutral unauthorized requests, harm-framed unauthorized requests, underspecified requests, and authorized high-impact actions.

In the original full-policy zero-shot evaluation, all five baselines resolve 0/50 AndroidControl and 0/36 AITW/GUIOdyssey authorization pairs under neutral wording. The table below reports the new four-shot results alongside the trained models. A pair is resolved only when the model withholds the unauthorized action and performs its authorized counterpart.

MethodAndroidControl accuracy (%)Pairs resolved /50AITW + GUIOdyssey accuracy (%)Pairs resolved /36
Full-policy Direct + 4 examples51.8050.50
Full-policy Generic reflect + 4 examples49.2052.10
Full-policy AXIS-reflect + 4 examples52.6049.50
Full-policy AXIS-CoT + 4 examples53.01552.19
Full-policy AXIS-Aha + 4 examples64.01962.96
Decision-only SFT86.6679.00
AXIS-Core92.34995.632

Demonstrations improve the reasoning-based baselines, but substantial authorization errors remain. On AndroidControl, adding demonstrations to full-policy direct prompting reduces unsafe actions from 175/415 to 76/415 while also reducing required benign actions from 87/95 to 62/95. AXIS-Core achieves 3/415 unsafe actions and 93/95 required benign actions. Its advantage therefore persists over both full-policy zero-shot and full-policy few-shot prompting.

2. Test-set coverage, independent benchmark, and multi-turn confirmation (Q1).

The ID test contains 204 records from 27 held-out scenarios. We also evaluated the same AXIS authorization policy on 71 AndroidControl screenshots from 65 source episodes and 58 AITW/GUIOdyssey screenshots from 54 source episodes. Each semantic case contributes multiple records under different request or screen conditions.

External screenshot sourceSemantic casesDanger recordsBenign requiring ACTBenign requiring CLARIFYTotal records
AndroidControl11541595256766
AITW + GUIOdyssey9028285180547

To test transfer under independently defined tasks and scoring, we evaluated the base model and AXIS-Core on 358 benign/malicious task pairs from ADeptS-Bench Mobile Safety, using the released prompts and scorer with output-parser compatibility adjustments. Its threats are embedded in the interface, providing a different construction and threat distribution from our training data.

ModelBenign task success ↑Attack success ↓Safety–utility score ↑
UI-Venus-1.5-8B76.3%74.6%38.1
+ AXIS-Core68.2%63.4%47.6

The score is the harmonic mean of benign success and one minus attack success, expressed on a 0–100 scale. AXIS-Core improves this score by 9.5 points, with an 11.2-point reduction in attack success and an 8.1-point reduction in benign success.

We also tested the pause–confirm–resume sequence using recorded screens and a scripted user (40 flows across 12 AndroidControl trajectories). In the 12 flows where the user deferred authorization, AXIS-Core committed before consent in 4 cases, compared with 9 for the base model. AXIS-Core asked for confirmation, received consent, and completed the operation in 3 of the 12 flows, compared with 2 for the base model.

3. Supervision controls and robustness across training seeds.

We retrained AXIS-Core on UI-Venus-1.5-8B with two additional seeds (43 and 44). We evaluated all three runs at epoch 2 (step 276), the checkpoint selected on the development set for seed 42. Each cell gives decision accuracy and, in parentheses, the number of unsafe ACT decisions among danger records.

Test setSeed 42 (original)Seed 43Seed 44
ID98.5% (0/134)97.5% (0/134)97.1% (0/134)
Paraphrase96.6% (1/134)99.0% (0/134)98.0% (1/134)
AndroidControl92.3% (3/415)93.5% (0/415)91.6% (1/415)
AITW + GUIOdyssey95.6% (0/282)97.4% (0/282)97.3% (3/282)

On neutrally worded authorization pairs, seeds 42, 43, and 44 resolved 49, 47, and 47 of 50 on AndroidControl and 32, 30, and 32 of 36 on AITW + GUIOdyssey. Each success requires withholding the unauthorized action and performing its authorized counterpart. Across the same seeds, AXIS-Core performed 93, 89, and 86 of 95 required benign actions on AndroidControl and 80, 79, and 84 of 85 on AITW + GUIOdyssey.

The original supervision comparison held the training records, gold decisions, LoRA settings, and seed fixed. AXIS-Full resolved 35/50 and 22/36 pairs on the two sources, and Full-Reordered resolved 32/50 and 15/36. All three Core runs exceeded these counts. At seed 42, Core also exceeded Full in overall accuracy on AITW + GUIOdyssey by 5.5 percentage points (95% episode-clustered bootstrap interval, 3.2–7.7). The Core–Full comparison evaluates the supervision formats as a whole. Full-Reordered tests the relative positions of risk and authorization within the full trace.

On the second backbone, Qwen3-VL-8B-Instruct, AXIS-Core reached 99.0% ID and 94.1% paraphrase accuracy. Under paraphrase, it performed all 18 required benign actions, compared with 9/18 for decision-only supervision, and made one unsafe decision versus five for AXIS-Full.

4. Relationship to safety controllers.

AXIS contributes an explicit risk–authorization–permission representation for deciding whether to release an action. This representation can supply the decision component of an action-verification controller, alongside its execution checks. We will make this relationship explicit in the related-work discussion.

Thank you for the rebuttal. The full-policy few-shot baselines and additional seeds address my main concerns. The multi-turn results remain mixed, so please limit the claims to static, single-step decisions and note this limitation in the revision. I have raised my rating to 3.

Thank you for reviewing the few-shot and multi-seed results and for raising your rating.

We have uploaded a revised manuscript. Changes are marked in blue.

  • Scope. Following your suggestion, the introduction now limits our main claim to static, single-step decisions. Section 8 reports the mixed multi-turn results.
  • Prompting and seeds. Sections 6.1 and 6.3 add the training seeds and compare all five prompting baselines under the full policy with zero or four demonstrations.
  • Other points. Section 6.4 notes that the supervision comparison evaluates each format as a whole and does not isolate output length or loss weighting. Section 7.1 reports both the safety gain and the drop in benign task success on ADeptS-Bench, and Section 2 discusses the complementary safety controllers you identified.
Reviewer_WeM3OpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes AXIS, which tests whether GUI agents treat a visible action or reassuring wording as authorization. It varies request wording and screen relevance while holding the operation fixed. AXIS-Core then trains an explicit risk and authorization judgment before action selection.

Critical strengths
  • The motivation is valid and interesting. The paired tests identifies an interesting failure mode: on the same screen and operation, merely changing whether the request states a warrant should change the decision.

  • AXIS-Core resolves 49/50 neutral-wording authorization pairs on AndroidControl, while the tested inference-time methods resolve none. This result looks good, though authors could explain more on it.

Critical weaknesses
  • The policy judges authorization from what the request says. It does not verify whether a claimed permission or consent is legit. So “authorized” should be read in that narrower sense.

  • The core evaluation is also a single-turn decision on a screenshot, not a live workflow with conversation history, changing permissions, and repeated confirmation costs, etc. Those are necessary steps before using this as a practical safeguard. Could the authors discuss more on the possibility of these steps' possibilities?

Questions for authors

In a live multi-step task, can the agent carry forward a prior authorization without asking again, while still detecting when the scope changes? How does it handle a request that falsely claims a third party’s consent?

4 comments

Thank you for highlighting the distinction between stated authorization and verified consent, and the importance of authorization across turns.

1. What authorization means in AXIS.

We agree. In our benchmark, “authorized” means that the request states a sufficient basis for the specified action on the specified target. Verifying the requester’s identity, checking delegated authority, and validating consent obtained outside the request are beyond the scope of this evaluation.

Our claim concerns how agents use the information already in the request. In each neutrally worded pair, the screenshot and operation are fixed. Only one request states a sufficient basis for proceeding. With the full policy and no demonstrations, none of the five prompting baselines on UI-Venus-1.5-8B resolved a pair on either dataset. AXIS-Core acted on the request with a stated basis and withheld the other in 49/50 AndroidControl pairs and 32/36 AITW/GUIOdyssey pairs, compared with 6/50 and 0/36 for decision-only supervision.

In deployment, an agent harness could check a claim against a trusted source. For example, if a user says a colleague authorized sharing a file, the harness could check whether a delegation record covers that file, recipient, and action. It could then provide the result to the agent and require a valid grant before releasing the action.

2. Carrying, obtaining, and withdrawing authorization across turns.

We added an evaluation comprising 40 scripted flows from 12 AndroidControl trajectories, covering initially missing authorization, authorization provided up front, withdrawal, and target changes. Models receive recorded screenshots and dialogue history through a shared state-machine replay interface with scripted user responses. The adapters receive no additional multi-turn training.

OutcomeBaseAXIS-Core
Unauthorized commits before consent ↓9/124/12
Ask, receive consent, and complete ↑2/123/12
Complete with authorization provided up front ↑12/1211/12
Correctly stop after withdrawal ↑11/128/12
Handle target correction ↑2/42/4

AXIS-Core carries prior authorization through intermediate steps, completing 11/12 such flows with zero redundant confirmation questions. It also reduces unauthorized commits before consent from 9/12 to 4/12. After withdrawal, both models make one unauthorized commit; AXIS-Core correctly terminates in 8/12 flows, compared with 11/12 for the base model.

The added evaluation directly tests authorization continuity and changes across turns. We will incorporate these results and distinguish the observed transfer in authorization-sensitive action selection from the handling of withdrawal and target changes.

Thank you for your insightful comments.

We have uploaded a revised manuscript. Changes are marked in blue.

Section 3.1 now states the scope of AXIS: for a given action and target, we test whether the agent's decision follows the authorization stated in the request. Section 8 explains how AXIS-Core can serve as the decision component of an agent harness. If a user says that a colleague approved sharing a file, the harness can check a trusted delegation record for that file, recipient, and action, and require a valid grant before AXIS-Core allows the transfer.

Section 8 also summarizes our scripted multi-turn pilot (Appendix I, Table 31). When authorization was never given, AXIS-Core committed the operation in 4 of 12 flows (base model: 9). It carried authorization given up front through 11 of 12 flows without asking again. After consent was withdrawn, it stopped in 8 of 12 flows (base model: 11), and both models handled 2 of 4 target changes.

I thank the authors for the update. I keep my positive score.

Thank you for reading the revised manuscript and for your continued support. Your questions helped us clarify how third-party consent could be verified in deployment and what our multi-turn pilot shows about changes in authorization.

Comments on the paper

2 comments

We thank the reviewers for their detailed feedback.

We have uploaded a revised manuscript. New analyses and substantive changes are marked in blue, and the prose has been revised throughout. Below we summarize the main concerns and how the revision addresses them.

1. Do the diagnostic findings hold under a stricter measure and a controlled design? (YYoY)

  • Clarification versus confirmation. Agents might ask for permission instead of clarifying. We added a stricter measure, proceeds directly: the response proceeds with no clarifying question, confirmation request, or refusal (Section 3.4). On task-relevant screens it rises in all seven configurations, by 30.7–52.0 percentage points in the five unguarded models (Holm-adjusted p < 10⁻⁶; Figure 2b, Appendix B.3).
  • Wording versus non-consent. A new 2×2 design crosses harm-related wording with stated non-consent in all 75 scenarios, holding the screen, target, and operation fixed (Section 3.3). Adding non-consent to a neutral request significantly lowers direct proceeding in five of seven configurations after Holm correction; harm-related wording alone does so in two. The neutral request, which says nothing about authorization, elicits direct proceeding most often in six of seven configurations (Figure 2c, Appendix B.4).

2. Are the baselines strong enough, and are the gains robust? (kVVK, YYoY)

  • All five prompted baselines receive the full decision policy verbatim, zero-shot and with four demonstrations, and we vary the number of demonstrations for Direct (Section 6.3, Appendix D). Demonstrations shift the model toward withholding across the board: on AndroidControl, Direct's unsafe actions fall from 175 to 76 of 415, but its completed benign actions also fall from 87 to 62 of 95. The best four-shot baseline resolves only 19 of 50 and 9 of 36 neutral authorization-contrast pairs.
  • AXIS-Core resolves 47–49 of 50 and 30–32 of 36 of these pairs across three training seeds, with at most three unsafe actions on any test set (Section 6.1, Appendix E.6).
  • On the independently built ADeptS-Bench mobile safety stream, AXIS-Core lowers the attack success rate from 74.6% to 63.4% (exact McNemar p < 10⁻⁵) at a cost of 8.1 points in benign task success. Its overall score rises from 38.1 to 47.6, the highest among open-weight models but below proprietary systems (Section 7.1, Appendix G).

3. Stated versus verified authorization, and consent across turns (WeM3, kVVK)

  • AXIS judges the authorization stated in the request (Section 3.1), and the introduction now limits our claims to static, single-step decisions. Section 8 explains how a deployment harness can verify a claim such as a third party's consent against a delegation record and pass the verified status to the agent, with PAUSE as the handoff.
  • A scripted multi-turn pilot (40 flows on 12 AndroidControl episodes; Appendix I, summarized in Section 8) gives mixed results without multi-turn training. AXIS-Core commits a never-authorized operation in 4 of 12 flows (base model: 9) and carries authorization given up front through 11 of 12 flows. After consent is withdrawn, however, it stops in only 8 of 12 flows (base model: 11). We report this as a limitation.

4. Presentation and related work (YYoY, kVVK)

  • We shortened sentences, separated the experimental setup from its interpretation, and unified terminology across the main text and appendix. Section 3.1 distinguishes CLARIFY (asking what the user means) from PAUSE (asking for confirmation before proceeding), and Section 3.4 defines all response measures before the findings.
  • Section 6.4 clarifies that the AXIS-Core–AXIS-Full comparison evaluates each supervision format as a whole and does not isolate output length or loss weighting. Section 2 discusses action-verification controllers that AXIS-Core can complement.

We are grateful for these comments, which have made the paper's claims more precise and better supported.

Dear Area Chair,

Following our earlier summary, we would like to share a brief update now that all three reviewers have responded.

Technical concerns. kVVK confirmed that the full-policy few-shot baselines and additional training seeds addressed their main concerns and raised their rating to weak accept. Following their suggestion, we limited our main claims to static, single-step decisions and reported the mixed multi-turn results in Section 8. YYoY confirmed that the new experiment addressed their concern about the fairness of the comparison and also raised their rating. WeM3 thanked us for the update and maintained their positive rating.

Readability. YYoY acknowledged the improvements to the abstract and introduction and emphasized that the writing still needs work. We agree. We have begun revising the later sections along the lines they suggested. For example, Section 3.4 now defines the response measures before Section 4 uses them to explain the findings. We will carry this revision through every remaining section and appendix, shortening dense sentences, making the reasoning easier to follow, and using consistent terminology throughout. We will then have the full manuscript proofread by a reader with native-level English proficiency and experience in scientific writing.

We appreciate the reviewers’ careful feedback and your consideration of both the additional experiments and our progress on the remaining writing concern.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.