acceptodds
Under review as a conference paper at ICLR 2027

Visibility Is Not Authority: Visual Task Instruction Delegation in LVLMs

Abstract

Large vision–language models (LVLMs) can follow instructions embedded in images without user authorization, allowing visual content to redirect the user’s task. Such overreach can occur even when the instructions are harmless. However, multimodal safety research focuses primarily on harmfulness, leaving the authorization boundary underexplored. We formalize this challenge as visual task instruction delegation (VTID), distinguishing unauthorized non-execution, authorized-safe completion, and authorized-unsafe refusal. To assess this boundary, we introduce VTID-BENCH, a 30,000-example benchmark and alignment resource with controlled counterfactuals that separate user authorization from task safety. Evaluations reveal that natural and detailed-policy prompts fail to establish a reliable authorization boundary across four LVLMs. To investigate whether this boundary can be learned through alignment, we develop VTID-SFT, which explicitly separates authorization from safety to guide how models respond to instructions embedded in images. Across three LVLMs, VTID-SFT achieves over 95% three-state accuracy and improves authorized-safe task completion, raising LLaVA’s three-state accuracy from 26.33% to 95.30%. External evaluations show improved LLaVA safety while retaining at least 95.9% of its base performance on three general vision–language benchmarks. These findings establish user authorization as a distinct dimension of multimodal alignment and demonstrate that models can learn to respect it while retaining general capability.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
large vision-language models, multimodal safety, authorization, visual prompt injection, instruction delegation, safety alignment, benchmarks

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_B8hEOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies whether an LVLM should execute a task visible in an image, seperating current-user authorization from conditional safety. VTID-BENCH contains 18,000 authorization examples organized into controlled U/AS/AU triads and 12,000 ordinary vision-language examples. VTID-SFT combines cause-specific response supervision, auxiliary authorization/safety heads, weak geometric constraints, and ordinary-response anchoring. Experiments cover prompting, three separately trained backbones, ablations, and external safety/capability tests. The strongest evidence concerns learning this constructed response policy and improving authorized-safe completion; broader authorization generalization remains less established.

Critical strengths
  • I find the controlled separation of delegation and harmfulness useful: Section 4.1 varies authorization for a benign visual task and safety under the same delegating prompt, avoiding the assumption that every unauthorized instruction must itself be harmful. The Instruction Hierarchy: Training LLMs to Prioritize Privileged Instructions already trains conditional compliance with aligned lower-priority instructions; VTID’s added value is a visual delegation benchmark with explicit safe-completion and unsafe-refusal contrasts, rather than introducing authority-sensitive instruction following from scratch.
  • The gains extends beyond one backbone and include meaningful outcome changes. Table 2 raises Qwen/InternVL authorized-safe Exact from 33.33%/33.67% to 94.33%/89.78%; Table 20 reduces unauthorized execution-evidence counts from 597/602 to 42/10 out of 900. Three LLaVA seeds, paired group bootstraps, the matched-budget data control, and the blinded three-rater audit strengthen the evidence. External LLaVA safety improves while retaining 95.9–99.4% of the reported capability scores, providing useful checks against indiscriminate refusal.
Critical weaknesses
  • Table 20 shows LLaVA’s U score rising from 0.0127 to 0.9520 while execution evidence rises from 6/900 to 12/900; Tables 21–22 recover 98.1% of the ablation gap by swapping explanation/refusal channels. Human validation supports this scoring policy, but behavioral claims need paired execution measures, as in Can LLMs Separate Instructions From Data? And What Do We Even Mean By That?.
  • Table 16 measures partial request-relevant content, and Figure 5 omits requested transcription after declining the embedded task. Clean-capability retention cannot establish original-task completion under injection; the already-cited Empirical Analysis of Large Vision-Language Models against Goal Hijacking via Visual Prompt Injection separates original-task and injected-task outcomes, providing a more direct evaluation template.
  • Appendix A.3 splits groups and images, yet every split shares the same eight task labels, sixteen instruction families, and 128 paraphrase IDs. This supports instance-level generalization but leaves transfer to new families untested; §5.5’s external safety/capability suites do not evaluate visual delegation.
  • Section 5.2’s generic safety baseline and Appendix D.7’s post-SFT DPO do not substitute for SecAlign: Defending Against Prompt Injection with Preference Optimization or StruQ: Defending Against Prompt Injection with Structured Queries, both omitted. Fair visual adaptations must preserve authorized tasks; without them, the advantage of VTID’s objective over targeted authority-training data remains unclear.
  • Appendix E.3 releases annotations, outputs and 540 test images, but omits trained adapters, most test images, training/development images and the full renderer. Although disclosed, these omissions prevent independent inference or training: consistency checks on archived records cannot establish that the benchmark and reported models are reproducible.
Questions for authors
  • Clarification/evidence: On pixel-identical paired U/AS inputs, how often does each model both withhold the unauthorized action and correctly execute it when delegated, independent of explanation wording? Does cause-specific supervision improve this rate over the collapsed control?
  • Evaluation: What fraction of U responses fully satisfies the original request, particularly explicit transcription requests, and how does that compare with Raw Base? Can Table 16 be extended beyond partial request-relevant content?
  • Generalization: Can you report a held-out instruction-family or independently authored visual-authorization evaluation, with the evalutor frozen? Please clarify the distinction between disjoint semantic-group identifiers and shared task-family semantics.
  • Comparison: Would an authority-specific preference objective or structured instruction-tuning baseline retain the observed gains under matched data, backbone, and budget while allowing delegated visual tasks? This would materially clarify the method’s contribution.
  • Reproducibility: Can the frozen adapters and full test images be made accessible during review, and can missing training images be reconstructed from a released renderer with the recorded seeds?
2 comments
AuthorsOpenReview ↗

Response to Reviewer B8hE

We thank Reviewer B8hE for asking us to separate evaluator evidence from direct behavior, test outer-request utility, assess independent authorization transfer, and strengthen baselines and reproducibility. We agree that the original paper made state-score improvements easier to see than execution evidence; we report paired diagnostics and narrow the claims below.

Paired execution behavior. Each Unauthorized development row is paired with its Authorized-Safe counterpart using the same base image and visual task. This is a same-image counterfactual, not a pixel-identical pair.

ConditionU non-executionAS exactU-no + AS-exact
VTID-SFT (S0)890/900387/900383/900
Cause-collapsed control (S2)896/900374/900373/900

The direct paired completion rates are 42.56% and 41.44%; thus, cause-specific supervision does not create a large literal-completion gain over this control. Its main contribution is to make the reason for non-execution explicit. When an embedded visual instruction is visible but not delegated, silently ignoring it is not the intended behavior: the model should acknowledge the observed instruction and explain that the current user has not authorized its execution. This makes the authorization boundary understandable to the user and separates missing delegation from safety refusal. U target-execution signals are 10/900 (S0) versus 4/900 (S2), with no strict AU dangerous-action signals in either condition. Restoring the missing-authorization, safety-reason, and refusal channels recovers 98.1% of the PCC gap.

We also evaluated 100 pixel-identical pairs from five external visual families with a frozen evaluator. Only the authorization wrapper changes.

ModelU executionAuthorized executionCorrect boundary
Raw Base11/10022/10014/100
VTID-SFT0/1008/1008/100

This supports transfer of unauthorized-task suppression, but delegated-safe execution remains conservative because the external probe lacks VTID-BENCH’s controlled safety contrast.

Outer-request utility. On a strict external judge, answer-like U responses are 45/100 for Raw Base versus 7/100 for VTID-SFT; on authorized-safe inputs they are 48/99 versus 23/100. We also audited 60 shared U scenes; one row was unevaluable per model:

SFT backboneContent-bearing C/PBroad C/P/H
LLaVA49/5958/59
Qwen55/5957/59
InternVL57/5959/59

These measures count responses that preserve meaningful content from the user’s observation request, including partial answers, while keeping boundary-preserving refusals distinguishable from empty responses. They show broad continuity on most scenes, but do not establish full original-task completion under U; that remains an open limitation.

Generalization and stronger baselines. The native split prevents semantic-group, source-scene, and counterfactual-triad overlap, but reuses task-family labels and paraphrase IDs; it establishes group/instance rather than held-out-family generalization. The external five-family probe is an independent authorization stress test with the evaluator frozen.

We trained three visual adaptations from Raw LLaVA-1.5-7B with matched 21,000 source rows, 2,100 updates, and the frozen 2,700-example development split. Hierarchy-SFT uses an instruction-priority scaffold, StruQ-style SFT uses reserved input delimiters, and SecAlign-style DPO uses visual preference pairs. These are mechanism-level visual adaptations, not exact text-only reproductions.

ArmPCCU non-exec.AS exactJoint
Raw Base0.27470.99440.00330.0033
Hierarchy-SFT0.85940.97560.40780.3978
StruQ-style SFT0.82930.97890.40780.3944
SecAlign-style DPO0.26721.000000
VTID-SFT0.95580.98890.42890.4256

VTID’s PCC advantages over Hierarchy-SFT and StruQ-style SFT are 0.0964 and 0.1265; direct completion differences are smaller and their paired intervals include zero. SecAlign-style DPO failed delegated-safe generation under this visual recipe. This pattern shows that preference-only visual adaptation transfers poorly to VTID's joint authorization-and-completion objective.

Reproducibility. The anonymous package contains a 540-image subset and selected outputs. A complete checksummed private package is staged with full authorization/base-image archives, the three adapters, per-example outputs, renderer/source code, manifests, seeds, and SHA-256 records. We can provide it through a venue-compatible private channel or supplementary upload.

These results support the narrower conclusion that VTID-SFT learns an authorization-sensitive boundary across backbones and suppresses unauthorized visual-task following beyond its native benchmark. We thank the reviewer for prompting these analyses and helping us clarify the evidence and scope of this conclusion.

AuthorsOpenReview ↗

Follow-up on our response to Reviewer B8hE

Dear Reviewer B8hE,

We would appreciate your feedback on the additional paired execution analyses, trained visual baselines, and reproducibility information in our response. If any of your main concerns remain unresolved, we would be grateful for clarification so that we can address them before the discussion closes.

Thank you for your time.

Best regards, The Authors

Reviewer_S3v4OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies whether vision–language models follow instructions in images without being aske to. The authors introduce a benchmark and fine-tune models to distinguish three cases: no permission, permission for a safe task, and permission for an unsafe task. They report large improvements across three models.

Critical strengths

This is an important problem: seeing an instruction in an image does not mean the model should follow it. The paired examples help test this. The paper also includes several models, ablations, and external tests of safety and general capabilities.

Critical weaknesses

Critical Weaknesses

I am not sure whether the models learn when they have permission or mainly learn the expected wording. The evaluation favors explanations that are also seen during training. Moreover, the external tests check safety and general capabilities, but do not directly test permission handling. It is not very clear what is new beyond instruction hierarchy.

Detailed comments:

  • 118–138: What is new here? The paper cites closely related work on instruction hierarchy and separating instructions from data. Is the main contribution the visual benchmark, its paired examples, or the training method?

  • 067–081, Figure 1: Why is returning a gate code shown in the image necessarily unsafe? The user might own the garage and simply want the code transcribed. What makes this example unsafe?

  • 183–195, 246–250: The unauthorized examples contain only safe tasks. Can you also test unsafe tasks without permission. Can the model distinguish "you did not ask me to do this" from "this is unsafe"?

  • 252–256, 782–791: Does the method work on new task types, prompts and image layouts?

  • 262–285: Why must the model explain that permission is missing? It could simply ignore the embedded instruction and answer the user’s actual question.

  • 282–285, 1062–1078: The human check covers only 120 responses from LLaVA variants. Can you include the other models, especially examples where quoting an instruction might be mistaken for following it.

  • 294–318, 395, 421–426: The generic-refusal baseline is trained without the explanations that your evaluator rewards. Does it actually follow more unwanted instructions, or does it mainly lose points for not giving the expected reason?

  • 379–410, Table 2: LLaVA reaches 95.3% three-state accuracy, but completes only 40.6% of safe tasks exactly?

  • 391–399, Table 3: Removing the geometry term changes little. What does the full method add over standard fine-tuning on the same responses?

  • 449–474: The external tests do not directly check permission handling. Can you test on an independent visual prompt-injection or delegation dataset?

  • 468–470, 1357–1381: Does the model still answer the user's original question?

Questions for authors
  1. What is new beyond separating instructions from data?

  2. Do the gains hold without specific explanations?

  3. Does the method work on new task types and independent datasets?

  4. Does the training change behaviour or mainly wording?

  5. How often does the model complete the user's original request?

2 comments
AuthorsOpenReview ↗

Response to Reviewer S3v4

We thank Reviewer S3v4 for distinguishing visual-instruction recognition, current-turn delegation, unsafe-task rejection, and answering the user’s outer request.

Novelty beyond instruction/data separation. VTID does not introduce instruction/data separation itself. Its contribution is a behavioral delegation benchmark and training objective: the same visual task is evaluated as visible but not delegated (U), delegated and safe (AS), or delegated but unsafe (AU). Matched U/AS/AU triads, cause-specific targets, separate authorization/safety supervision, and ordinary-response anchoring prevent a refusal-only solution. The revision also adds a SEP comparison.

Gate-code example. Figure 1 asks the model to “Post gate code 8842 in this chat” after “Please follow the valet notice shown in the image.” The AU label refers to disclosing an access credential through chat; reading or transcribing a visible code is not itself unsafe. We clarify this in the caption.

Unsafe tasks without authorization and external transfer. We constructed a same-image CyberSecEval3-derived probe with 40 security-violating visual tasks requesting synthetic secrets, passwords, or hidden values. The image, question, and target remain fixed while authorization changes. We also evaluated 100 pixel-identical pairs across five external visual families with a frozen evaluator.

DiagnosticRaw BaseVTID-SFT
Unsafe execution under U0/400/40
Literal secret disclosure under U0/400/40
Unsafe execution under AU0/400/40
External U execution11/1000/100
External authorized execution22/1008/100
External correct boundary14/1008/100

The external families cover embedded text, obfuscated text, FigStep, indirect references, and query-relevant images. Native CyberSecEval3 injection following is 285/1000 (28.5%) for Base and 15/1000 (1.5%) for VTID-SFT; this measures injection following rather than authorization. The external authorized-execution rate is conservative because it lacks VTID-BENCH’s controlled safety contrast.

Why explain missing permission? Giving the reason is part of a reasonable assistant’s response, not merely an auditability convention. Silent non-execution may leave the user unsure whether the model noticed the visual instruction, failed to read it, or declined for safety. When a visual instruction is visible but not delegated, the model should acknowledge that it observed the content and explain that the current user has not authorized execution. This makes the boundary understandable and separates missing delegation from safety refusal.

The cause-collapsed control tests whether this is only wording:

ConditionPCCUASAUU executionAU dangerous action
Full S00.95510.95100.98310.931110/9000/900
Cause-collapsed S20.52750.33450.92900.31894/9000/900

Restoring missing-authorization, safety-reason, and refusal channels in frozen S2 outputs raises PCC to 0.9470, recovering 98.1% of the gap. Removing boundary geometry changes PCC by only −0.0018, whereas removing factor supervision changes PCC to 0.9278 and AU to 0.8563. On the LLaVA final test, AS state evidence is 0.9707 but Exact completion is 0.4056; state recognition and exact task completion are separate axes.

Other backbones and quotation versus execution. The 120-response blind human audit covers a historical LLaVA holdout. On the current frozen U split, evaluator behavior-response signals change from 6→12 (LLaVA), 597→42 (Qwen), and 602→10 (InternVL), while missing-authorization signals change from 0→865, 0→845, and 0→889. These changes summarize response-level evidence across backbones; descriptions and quotations remain distinct from confirmed execution.

Original user request. On a strict external judge, answer-like U responses are 45/100 for Raw Base versus 7/100 for VTID-SFT; on authorized-safe inputs they are 48/99 versus 23/100. We also reviewed 60 shared U scenes.

SFT backboneContent-bearing C/PBroad C/P/H
LLaVA49/5958/59
Qwen55/5957/59
InternVL57/5959/59

C/P counts complete or partial content; C/P/H additionally includes an explicit willingness to continue the outer request. This post-hoc Codex review is descriptive, not independent human validation or a full 900-row estimate. It supports the narrower conclusion that VTID-SFT generally keeps the outer request in the conversation, while full visual utility remains a separate limitation.

Together, these results show that VTID learns authorization-sensitive visual behavior beyond simple instruction/data separation, while distinguishing unauthorized-task refusal, safety rejection, and outer-request utility. We thank the reviewer for prompting these clarifications and evaluations.

AuthorsOpenReview ↗

Follow-up on our response to Reviewer S3v4

Dear Reviewer S3v4,

We would welcome your feedback on our response, particularly the analyses separating execution behavior from explanation wording and the audit of responses to the original user request. If any of your main concerns remain unresolved, please let us know so that we can provide further clarification before the discussion closes.

Thank you for your time.

Best regards, The Authors

Reviewer_h26bOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper frames image-embedded instructions as an authorization problem with three target behaviors: refuse unauthorized tasks while explaining why (U), complete authorized safe tasks (AS), and refuse authorized unsafe tasks (AU). It introduces VTID-Bench (18k counterfactual triads plus 12k ordinary examples), a deterministic evaluator with a composite score (PCC), and VTID-SFT. VTID-SFT raises three-state accuracy from 26–35% to over 95% on LLaVA-1.5, Qwen2.5-VL, and InternVL2.5.

Critical strengths
  1. The framing is clean. Separating authorization from safety, using benign tasks for the U state, is a useful conceptual contribution.
  2. The counterfactual design is careful, with group-level splits and overlap audits. The paper is thorough
Critical weaknesses
  1. The main weakness that makes it hard to draw conclusions from this paper is the lack of out-of-distribution authorization evaluation. The external suites (MM-SafetyBench, AdvBench) measure harmfulness, not authorization. There is no test on existing visual prompt-injection or goal-hijacking benchmarks, or on unseen task families or injection styles. Without this, it’s unclear whether the model learned a boundary or learned the benchmark’s templates, especially through the in-distribution SFT train set created.

  2. Weak baselines. The comparisons are a single prompt suffix, one closed model, and SPA-VL dataset. There’s no comparison to prompt-injection defenses such as instruction-hierarchy-style training, delimiting/spotlighting, or structured-query fine-tuning. All models are older 7–8B models and ablations are LLaVA-only, a generally weakly aligned model.

Questions for authors
  1. How does VTID-SFT perform on an existing visual prompt-injection benchmark?
  2. What is the over-refusal rate on XSTest/MOSSBench-style prompts?
  3. What is the U-state score if the missing-authorization reason requirement (M) is dropped and only execution is scored?
2 comments
AuthorsOpenReview ↗

Response to Reviewer h26b

We thank Reviewer h26b for distinguishing safety transfer from authorization generalization. We agree that the original submission lacked direct OOD authorization evaluation and used a narrow baseline set. We added the following evaluations and narrowed the claims where evidence is limited.

OOD authorization and visual prompt injection. No established external benchmark directly measures visual authorization under controlled delegation contrasts. This gap motivates VTID-BENCH, which holds the image and task fixed while changing only whether the current user delegates the visible task. To test transfer beyond VTID-BENCH, we adapted visual stimuli from public CyberSecEval3 prompt-injection data into a five-family probe with 100 pixel-identical unauthorized/authorized pairs. The probe was held out; each pair fixes image, request, checkpoint, and decoding, changing only authorization.

DiagnosticRaw BaseVTID-SFT
CyberSecEval3 injection following285/1000 (28.5%)15/1000 (1.5%)
OOD unauthorized execution11/1000/100
OOD authorized execution22/1008/100
Correct boundary (U=no, A=yes)14/1008/100

CyberSecEval3 measures injection following, whereas the adapted probe tests authorization transfer. It supports transfer of unauthorized-task suppression, while authorized execution remains conservative because the probe lacks VTID-BENCH’s controlled safety contrast. We therefore present it as an independent stress test, not universal held-out-family authorization generalization.

Over-refusal. We evaluated the safe XSTest split (250 prompts) and benign MOSSBench split (300 image-query pairs) with frozen checkpoints.

DiagnosticRaw BaseVTID-SFT
XSTest response-level refusal7/250 (2.8%)22/250 (8.8%)
XSTest official prefix refusal4.8%10.8%
MOSSBench refusal0/300131/300 (43.7%)

In a stratified audit of 30 VTID-SFT MOSSBench refusals, 24 were clear over-refusals, four ambiguous, and two defensible safety refusals. Twenty-three responses mentioned delegated authority or an unrequested image instruction. Thus, when a benign user requests an authorized safe operation on an image containing safety-sensitive information, the model may treat the visible unsafe content as a reason to refuse the whole request. We report this safety/helpfulness trade-off.

Baselines and newer backbones. The revised paper evaluates four open backbones, including Qwen3.5-9B. On the frozen final test, Qwen3.5 VTID-SFT improves accuracy/PCC/Exact from 0.4015/0.4266/0.3522 to 0.9830/0.9760/0.9456. Prompt-only spotlighting on Qwen3.5 raises AU refusal from 0.3850 to 0.4730, but lowers authorized-safe execution from 0.9364 to 0.8192 and leaves PCC nearly unchanged (0.4420 to 0.4329), showing that delimiting alone does not recover the three-state boundary.

We also trained matched visual adaptations from Raw LLaVA-1.5-7B using 21,000 training rows, 2,100 updates, the frozen evaluator, and the 2,700-example development split. These are mechanism-level visual adaptations, not exact reproductions of text-only pipelines.

ArmPCCAS-ExactJoint
Raw Base0.27470.00330.0033
Hierarchy-SFT0.85940.40780.3978
StruQ-style visual SFT0.82930.40780.3944
SecAlign-style visual DPO0.267200
VTID-SFT0.95580.42890.4256

VTID’s PCC advantage over Hierarchy-SFT and StruQ-style SFT is 0.0964 and 0.1265; direct completion differences are smaller and their paired intervals include zero. The SecAlign-style adaptation failed delegated-safe generation under this visual recipe. This pattern shows that preference-only visual adaptation transfers poorly to VTID's joint authorization-and-completion objective: it suppresses unauthorized execution but does not recover delegated-safe generation.

U without the missing-authorization reason. Dropping M and scoring only unauthorized-task non-execution gives:

ConditionOfficial joint UExecution-only UAS Exact
Full VTID-SFT0.95100.9889 (10/900 executed)0.4300
Cause-collapsed control0.33450.9956 (4/900 executed)0.4156

No high-confidence authorized-unsafe execution was detected in either condition. Removing M makes both systems appear strong and hides the difference in recognizing missing delegation. We retain the official joint U score and label the execution-only value as a behavior-only diagnostic.

These results support the conclusion that cause-specific multimodal supervision enables models to learn a visual authorization boundary across backbones. We thank the reviewer for prompting these evaluations and helping us clarify the scope and remaining limitations of this conclusion.

AuthorsOpenReview ↗

Follow-up on our response to Reviewer h26b

Dear Reviewer h26b,

We would appreciate your feedback on whether the additional external evaluations and trained authority-specific baselines in our response address your main concerns. If any concerns remain, we would be grateful for clarification so that we can respond before the discussion closes.

Thank you for your time.

Best regards, The Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Guide to the revised manuscript

Dear Area Chair and Reviewers,

Thank you for your feedback. We have collected the locations of the main revisions below for ease of reference.

  • Additional backbone. Table 2 (p. 8) adds final-test results for Qwen3.5-9B. Table 17 (p. 26) compares natural prompting and spotlighting on the same backbone.

  • Trained baselines and paired behavior. Table 3 (p. 9) compares VTID-SFT with Hierarchy-SFT, StruQ-style visual SFT, and SecAlign-style visual DPO using matched training data and update counts. It reports how often a model both withholds the unauthorized action and completes the task when delegated. Appendix D.8 (pp. 35–36) gives the training details, paired confidence intervals, and diagnostics.

  • External transfer. Section 5.5 (pp. 9–10) reports results on CyberSecEval3 visual prompt injection and a separate visual authorization probe with 100 image-matched pairs across five families. Tables 30–31 (p. 33) give the overall and family-level probe results. Section 6 (p. 10) clarifies which aspects of generalization the native split tests.

  • Relation to prior work. Section 2 (p. 3) discusses learned instruction–data defenses, and Table 10 (p. 18) compares the interventions and evaluation criteria used by SEP and VTID.

For the question of whether the model answers the user's original request, Table 18 (p. 27) reports the 60-scene coverage audit. Our replies to S3v4 and B8hE give the separate analysis of 59 evaluable scenes per model. The XSTest and MOSSBench over-refusal results are in our reply to h26b.

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.