acceptodds
Desk rejected at ICLR 2027

Says Block, Still Acts: Why LLM Safety Judgments Fail to Govern Action in LLM Agents

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Large language model agents can correctly judge that an action should be blocked while still preferring to take it. We ask why this judgment–action disconnect arises, and whether explicit safety judgment causally governs subsequent action preference. Across three open-weight language models, safety-predictive information remains recoverable from action states, arguing against a simple information-loss account. Instead, the disconnect is better explained by weak coupling between judgment- and action-side causal control: interventions that reliably shift explicit safety judgments toward BLOCK produce much smaller changes in action preference than action-native interventions. This asymmetry persists within a shared judgment-to-action trajectory, where strong upstream control of judgment does not translate into comparably strong downstream control of action preference. Beyond individual intervention directions, judgment- and action-control subspaces overlap only partially, while effective action control remains available in directions orthogonal to the judgment-control subspace. Together, these results distinguish information availability from causal control: an LLM agent can retain the information needed to recognize an action as unsafe without the variables supporting that judgment reliably governing its action preference. For agent safety, this suggests that improving safety recognition or self-critique alone may be insufficient unless safety-relevant computations are also causally coupled to action selection. Code is available at https://anonymous.4open.science/r/SayBlockButAct-DE32/

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LLM agents, agent safety, AI alignment, mechanistic interpretability, causal interventions, judgment–action gap, action control

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 4 reviews, above 27% of the others.
Reviewer_7zVnOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates why an LLM agent may correctly judge an action as unsafe while still preferring it. Across three open-weight models and synthetic tool-use scenarios, the authors show that safety information remains readable in action representations. However, interventions that strongly control safety judgments transfer weakly to action preference compared with action-native interventions. Counterfactual state swaps and subspace analyses further suggest limited overlap between judgment and action control. The paper concludes that representing safety information does not necessarily grant it causal authority over action selection.

Critical strengths

The paper clearly separates information availability from causal control. This matters because correct safety judgments or self-critique are often treated as evidence of safer behavior.

The combination of probing, causal interventions, state swaps, subspace analysis, held-out scenario families, and three model architectures provides broader evidence than a single steering experiment.

Critical weaknesses

The high-level judgment–action gap is not itself novel, as prior work has already shown that agents may recognize safety, privacy, or abstention constraints while failing to act accordingly.

The controller comparison is not fully diagnostic. The action-native controller is directly optimized for the evaluated action margin and injected at the action position, whereas the judgment controller targets another endpoint and is sometimes injected earlier. Its weaker action effect may therefore reflect endpoint optimization, intervention position, or perturbation decay.

The shared-trajectory experiment fixes a plausible mediator. Teacher-forcing the same BLOCK token removes the pathway through which judgment may normally affect action. Indeed, explicit BLOCK text strongly changes action preference. The evidence therefore supports weak latent-state transfer conditional on a fixed judgment token, but the broader conclusion that safety judgments fail to govern action is somewhat overstated.

Questions for authors

Can the authors include a matched upstream action-relevant intervention to distinguish semantic decoupling from generic attenuation across positions?

What is the total causal effect when the ALLOW/BLOCK judgment is generated rather than teacher-forced? Can it be decomposed into latent direct and explicit-token-mediated effects?

Do the conclusions hold for free-form tool generation or actual execution in a stateful environment, rather than scored candidate actions?

4 comments
AuthorsOpenReview ↗

Response to Reviewer 7zVn

Response to Reviewer 7zVn

We thank the reviewer for emphasizing the distinction between information availability, latent judgment control, and action selection. Our response adds matched-upstream controls, factorial and natural-state interventions, and free-generation and sandbox evaluations to clarify the causal relationship and its behavioral scope.

1. Mechanistic contribution: from behavioral inconsistency to causal control

Prior work documents judgment–action gaps. We ask whether safety-relevant information remains readable in action-side states, whether variables controlling explicit judgment exert comparable causal control over action preference, and how the control structures underlying judgment and action relate. A judgment-trained probe transfers to action states without refitting (best leave-family-out J→A AUROC: Qwen 0.875, Gemma 0.840, Mistral 0.756; descriptive maxima across layers). Hidden-state interventions test causal control; subspace and orthogonal-component analyses characterize its structure. The orthogonal action component retains an effect of 0.0240 versus 0.0238 for the full action controller at Mistral L24, and 0.123 versus 0.0110 for the judgment-parallel component at Qwen L32. We therefore provide a causal analysis of information availability and the distinct control structures underlying judgment and action.

2. Matched-upstream controller comparison

To test whether weaker judgment-to-action transfer reflects injection position or perturbation decay, we derived an action-native controller, cAupc_A^{up}, from safe-oriented action gradients and injected it at the same upstream judgment position as gJsafeg_J^{safe}. Both use the same case-wise recipient-norm matching procedure and teacher-forced candidate-action endpoint; a matched-norm random direction serves as control.

Across nine semantic families, we use 20 training and 10 distinct evaluation variants per family (90 cases). Families overlap, but scenario variants do not; evaluation cases are not filtered by baseline KBD phenotype, judgment output, action margin, or generated behavior.

ModelgJsafeg_J^{safe}cAupc_A^{up}Paired difference (95% CI)Positive casesLOFO range
Qwen3-32B0.007200.07280.0656 [0.0537, 0.0783]90/90[0.0611, 0.0690]
Gemma3-27B-0.004220.04370.0480 [0.0339, 0.0634]90/90[0.0428, 0.0518]
Mistral-Small-3.2-24B0.000860.01790.0171 [0.0107, 0.0238]89/90[0.0150, 0.0186]

CIs are family-cluster bootstrap 95% intervals (20,000 resamples). All nine family means are positive in each model; the family-level exact sign-flip test gives two-sided p=0.00391p=0.00391 for each (512 configurations, assuming family-level sign-exchangeability). The action-native controller remains more effective under matched position and perturbation norm, so injection distance alone cannot explain the asymmetry. The controllers remain endpoint-native by construction—gJsafeg_J^{safe} uses judgment gradients and cAupc_A^{up} action gradients—so this comparison concerns the action-side efficacy of these tested directions, not all possible judgment representations.

3. Explicit judgment context and natural-state causal influence

A separate factorial intervention crosses the latent judgment intervention with externally supplied ALLOW/BLOCK text:

ModelText BLOCK − ALLOW SafeEffLatent under ALLOWLatent under BLOCK
Qwen3-32B0.1320.006960.00170
Gemma3-27B0.0989-0.01090.00093
Mistral-Small-3.2-24B0.1520.006720.00256

BLOCK text strongly shifts action preference; latent effects are smaller and architecture-dependent. The latent effect is attenuated under BLOCK for Qwen and Mistral, while Gemma changes sign across textual contexts.

Separately, within-trajectory interventions hold the explicit BLOCK judgment fixed. At Qwen L32, Gemma L32, and Mistral L24, gJsafeg_J^{safe} strongly shifts the judgment readout but has little effect on later action preference:

ModelJudgment effect of gJsafeg_J^{safe} (95% CI)Later action effectPaired vs. random (95% CI)Downstream +cA+c_A
Qwen+4.17 [+4.07, +4.28]+0.000924+0.00133 [-0.000134, 0.00276]+0.0435
Gemma+18.7 [+18.1, +19.3]-0.000065-0.00132 [-0.00346, 0.000772]+0.212
Mistral+6.03 [+5.98, +6.08]+0.00135-0.000127 [-0.00120, 0.000929]+0.0944

The paired effects relative to matched random perturbations include zero in all models, whereas downstream action-native effects are larger.

Natural-state counterfactual swaps replace the full residual state immediately before explicit judgment with a matched donor state (60 held-out pairs; both directions, 120 interventions per model), without gradients, learned directions, scaling, or norm matching. We measure action effects with BLOCK fixed and when the explicit label follows the model's ALLOW/BLOCK readout:

ModelJudgment alignment (95% CI)Action, judgment fixed (95% CI)Action, label unchanged (95% CI)
Mistral L24+0.0604 [+0.0505, +0.0698]+0.00004 [-0.00063, +0.00071]-0.00070 [-0.00147, 0.00007]
Qwen L32+0.0156 [-0.00417, 0.0359]-0.00027 [-0.00099, 0.00047]-0.00039 [-0.00122, 0.00041]
Gemma L32+0.410 [+0.318, +0.507]+0.00051 [-0.00090, 0.00196]+0.00013 [-0.00149, 0.00173]

Intervals are pair-cluster bootstrap 95% CIs (10,000 resamples). With the label unchanged, downstream effects are near zero. In a descriptive Mistral subset, 10 label-crossing interventions produce a donor-aligned action effect of +0.135 [0.117, 0.152]; crossings occur in 1/120 Qwen interventions and 0/120 Gemma interventions. This is descriptive evidence of coupling, not a formal mediation estimate. Explicit judgment context and latent judgment control are distinct causal inputs.

4. Free-generation and sandbox evaluations: behavioral scope

Free-generation baselines, without teacher-forcing BLOCK, are:

ModelGenerated BLOCKBLOCK outcomes: Unsafe / Safe alt. / Other tool / Non-toolALLOW: Unsafe / Total
Qwen3-32B80/900 / 11 / 0 / 6910/10
Gemma3-27B78/9041 / 17 / 20 / 012/12
Mistral-Small-3.2-24B76/900 / 20 / 48 / 814/14

Gemma shows a pronounced BLOCK-then-unsafe pattern (41/78); Qwen often produces non-tool/final responses, while Mistral often proposes other tools. These latter outcomes are not counted as successful safe execution.

Matched-upstream controllers were applied once at the final prompt token before the judgment stage; models then freely generated judgment and action under greedy decoding (temperature 0):

ModelControllerUnsafe-proposal change (pp, 95% CI)Unsafe → non-unsafe / reverseTool-call rate change (pp)
QwengJsafeg_J^{safe}0.0 [0.0, 0.0]0 / 00.0
QwencAupc_A^{up}0.0 [0.0, 0.0]0 / 0+4.4
GemmagJsafeg_J^{safe}-2.2 [-6.7, 0.0]2 / 00.0
GemmacAupc_A^{up}-1.1 [-3.3, 0.0]1 / 00.0
MistralgJsafeg_J^{safe}-2.2 [-6.7, 0.0]2 / 0-12.2
MistralcAupc_A^{up}-2.2 [-6.7, 0.0]2 / 00.0

Qwen's cAupc_A^{up} converts four non-tool/final responses into safe alternatives without changing unsafe proposals. In Mistral, both controllers convert two unsafe targets into other-tool proposals; gJsafeg_J^{safe} also reduces tool calls. Gemma changes are small. These interventions do not consistently reduce unsafe proposals across models.

Across Direct, Safety Reminder, Explicit Judgment, and Neutral Deliberation, the table combines generated proposal categories (first four outcome columns) with three-step in-memory sandbox outcomes:

ModelConditionUnsafe targetSafe alt.Other toolNon-tool/finalTool callsUnsafe transitionsConditional rate
QwenDirect1210365251248.0%
QwenSafety Reminder513360905763.3%
QwenExplicit Judgment1013859311032.3%
QwenNeutral Deliberation5610717705071.4%
GemmaDirect702513756890.7%
GemmaSafety Reminder83430908695.6%
GemmaExplicit Judgment641880906976.7%
GemmaNeutral Deliberation703107817390.1%
MistralDirect122670201260.0%
MistralSafety Reminder22233114792734.2%
MistralExplicit Judgment1827405852225.9%
MistralNeutral Deliberation3122361873540.2%

All counts are out of 90; conditional rate = unsafe transitions / tool calls. The sandbox extends evaluation to multi-step state transitions in an in-memory simulation, not external execution. Explicit Judgment yields parseable judgments in all cases, with BLOCK rates of 88.9% (Qwen), 72.2% (Gemma), and 88.9% (Mistral). Relative to Direct, Safety Reminder increases unsafe proposals by 43.3, 14.4, and 11.1 percentage points, respectively, and increases tool calls in all models. Safety reminders change both entry into the tool-use path and the resulting action distribution; they are not equivalent to latent judgment control or to the externally supplied BLOCK/ALLOW intervention in the factorial experiment. Mistral's conditional rate decreases despite more total unsafe transitions.

5. Discrete action selection: top-1 ranking changes

At baseline, unsafe top-1 rates are 45/90 (Qwen), 30/90 (Gemma), and 29/90 (Mistral). In Qwen, cAupc_A^{up} changes 8/45 unsafe top-1 cases to non-unsafe, versus 3/45 for gJsafeg_J^{safe}; all eight action-native flips occur in the review family. Although case-paired McNemar gives p=0.0078p=0.0078, the family-cluster CI is [-26.7, 0.0] pp and the family-level exact test over 512 sign configurations gives p=1p=1, so the case-level result does not establish cross-family improvement. Neither controller produces beneficial flips in Gemma or Mistral.

Conclusion

Our results identify an asymmetric causal relationship between safety judgment and action selection. Safety information remains available and explicit judgment context can influence action preference, but the tested judgment-native direction exerts substantially weaker causal control over action than the action-native direction, even under matched upstream interventions. Representing a safety constraint is not equivalent to granting it control over action selection.

AuthorsOpenReview ↗

Follow-up on Judgment–Action Causal Control

Dear Reviewer,

Thank you again for your detailed and constructive feedback. We particularly appreciate your concerns regarding the diagnostic validity of the controller comparison, the role of explicit judgment tokens, and whether our findings extend beyond teacher-forced action-margin evaluations.

We conducted several additional experiments to address these concerns more directly.

1. Matched-upstream controller comparison

To address the concern that the weaker effect of judgment-native interventions might result from different intervention positions or perturbation decay, we derived an action-native controller and injected it at the same upstream judgment position as the judgment-native controller.

Both interventions use the same case-wise recipient-norm matching procedure, with a matched-norm random direction as a control.

The action-native controller retains a significantly stronger effect on downstream action preference across all three models:

  • Qwen: paired difference of 0.0656 (95% CI: [0.0537, 0.0783]).
  • Gemma: paired difference of 0.0480 (95% CI: [0.0339, 0.0634]).
  • Mistral: paired difference of 0.0171 (95% CI: [0.0107, 0.0238]).

All nine semantic-family means are positive for each model, with family-level exact sign-flip tests yielding p = 0.00391. We believe this provides a more direct comparison of judgment- and action-side causal control while controlling for intervention position and norm.

2. Separating latent judgment control from explicit BLOCK context

We added a factorial intervention that independently varies externally supplied ALLOW/BLOCK text and latent judgment interventions. The results show that explicit BLOCK text can strongly shift action preference, whereas latent judgment interventions have substantially smaller and architecture-dependent effects.

We also conducted natural-state counterfactual swaps and clarified the distinction between effects with the explicit judgment label held fixed and effects when the label is allowed to follow the model's judgment readout.

3. Extending the analysis to free generation and stateful behavior

We added free-generation experiments without teacher-forcing BLOCK, allowing us to measure spontaneous judgment generation and subsequent action distributions. We also extended the evaluation to multi-step in-memory sandbox trajectories.

These experiments reveal important model-dependent differences. For example, Gemma exhibits a pronounced BLOCK-then-unsafe pattern, while Qwen and Mistral produce different distributions of non-tool responses and alternative tool proposals. We also explicitly acknowledge that the sandbox evaluation is an in-memory simulation rather than external tool execution, and that our interventions do not consistently reduce unsafe proposals across all models.

Taken together, these additional results help refine the scope of our central claim: the tested judgment-native directions exert substantially weaker causal control over action preference than action-native directions, even under matched upstream intervention conditions. We do not claim that every possible representation of safety judgment is causally disconnected from action selection.

We sincerely hope these additional experiments address your concerns about intervention comparability and the broader behavioral interpretation of our findings.

We would greatly appreciate it if you could take another look at our responses when you have a chance. As the discussion deadline is approaching, we are concerned that there may not be sufficient time to address any further methodological questions or concerns that emerge closer to the deadline. If there are still aspects of the experimental design or causal interpretation that you find insufficiently clear, we would be very grateful if you could share them with us at your earliest convenience, so that we can provide further clarification before the discussion closes.

Thank you very much again for your time and thoughtful consideration!

Best regards,
The Authors

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

This paper presents a study on why LLMs tend to can judge an action be blocked "not taken" but still have a preference to take that action. I.e., does an agents judgement of the safety of an action actually influence its behavior in a positive way - is it less likely to take that action? The paper present experiments using a simulated workspace and the authors evaluate three models.

Critical strengths

The paper was introduced well. I found the examples clear and the motivation engaging. The goals of the paper are meritable and there is a lot of supplementary material to provide the reader with additional detail.

Critical weaknesses

Overall, I found the experiments and results underwhelming. The topic is well motivated in the paper, but the experiments and analyses do not seem sufficient to justify paper at a top conference. The experiments are somewhat limited and the presentation of the results is not very clear. This is in contrast with the presentation/motivation of the problem in the introduction of the paper which is clear.

Overall I found the experiments and results hard to follow and would appreciate if the authors could reiterate their main insights and highlight which numerical results specifically support them and show that they have performed statistical tests to back those claims up where necessary.

The authors say that AI was used to help with the readability of the paper, I might suggest a human pass could improve it somewhat more.

Questions for authors

Are there ways that you can better communicate the significant of the experiments and the results?

Is Figure 4 really the best way to present such a small amount of results? Is there truly zero variance in most of those results? The figure takes a lot of real-estate on the page but seems to communicate hardly any data.

Is presenting results to 6 significant figures really the best way to show those numbers?

4 comments
AuthorsOpenReview ↗

Response to Reviewer FwcM

Response to Reviewer FwcM

We thank the reviewer for the constructive comments on presentation and for recognizing the motivation and practical relevance of the work.

1. Reorganizing the main results

We will organize the main narrative around three claims:

  1. Information availability: judgment-trained probes retain above-chance cross-interface ranking information on held-out families. The best observed leave-family-out J→A AUROC values at the final prompt position are 0.875 for Qwen, 0.840 for Gemma, and 0.756 for Mistral. We will distinguish ranking information from calibration and causal use.
  2. Asymmetric causal control: judgment-native directions strongly control explicit judgment but have much weaker effects on action preference than action-native directions. The new same-upstream-position comparison will be promoted to the main results.
  3. Partially shared control structure: judgment- and action-control subspaces overlap above the random reference, but within-task reproducibility is much higher than cross-task overlap, and substantial action efficacy remains in the judgment-orthogonal component.

Free-generation and InertSandbox outcomes will be presented as a separate behavioral-boundary analysis, not merged with teacher-forced action-margin evidence.

2. Replacing Figure 4

We agree that the current Figure 4 uses too much space for too little visible information. We will replace it with a compact effect-size/forest-plot presentation showing, for all three models, the upstream judgment-controller effect on judgment, its effect on later action preference, the matched random-upstream control, and the downstream action-native positive control. Point estimates, 95% CIs, and N = 60 will be shown explicitly. Full protocol and detailed statistics will remain in the appendix.

3. Statistical reporting and numerical precision

We will standardize effect estimates and CI endpoints to approximately two or three significant digits rather than six. We will label the statistical unit and uncertainty procedure for each analysis:

  • matched-upstream controller differences: family-cluster bootstrap with 20,000 resamples; family-level exact sign-flip test over nine family means;
  • within-trajectory interventions: case-level bootstrap over 60 held-out blocked-valid scenarios;
  • natural-state swaps: pair-cluster bootstrap over 60 matched pairs, retaining both directional swaps;
  • free-generation paired behavior: family-cluster bootstrap, with case-paired exact tests reported separately where relevant.

We will not treat layer-by-case rows or multiple intervention conditions as independent observations. For Mistral, 89/90 positive cases and 9/9 positive family means are distinct summaries and will be reported separately.

4. Editing and readability

We will conduct a full human editing pass, improve transitions, reduce repetition, define information, causal directions, and control subspaces before presenting results, and ensure that each figure caption states the cohort, endpoint, intervention position, and uncertainty unit. The abstract, introduction, results, and conclusion will use the same operational definition of causal control.

AuthorsOpenReview ↗

Follow-up on Judgment–Action Causal Control

Dear Reviewer,

Thank you again for your constructive feedback and for pointing out the need to improve the clarity, organization, and presentation of our experimental results.

We took your comments seriously and have outlined substantial revisions to make the main findings easier to follow and their statistical support more transparent.

In particular, we would like to highlight the following changes:

1. Reorganizing the main findings

We have restructured the main narrative around three central claims:

  • Information availability: Safety-relevant information remains recoverable across judgment and action states.
  • Asymmetric causal control: Judgment-native directions have substantially weaker effects on action preference than action-native directions, even under matched upstream intervention positions.
  • Partially shared control structure: Judgment- and action-control subspaces overlap partially, while effective action control remains available in directions orthogonal to the judgment-control subspace.

We will also present free-generation and InertSandbox results as a separate behavioral analysis, rather than mixing them with teacher-forced action-margin evidence.

2. Improving the presentation of experimental results

We agree with your concern that the original Figure 4 does not communicate the results as effectively as it should. We have outlined a replacement using a more compact effect-size/forest-plot presentation, showing the relevant intervention effects, matched random controls, action-native positive controls, 95% confidence intervals, and sample sizes for all three models.

3. Improving statistical reporting and readability

We will standardize numerical precision to approximately two or three significant digits and explicitly identify the statistical unit and uncertainty estimation procedure for each analysis. We will also clarify how repeated measurements and family-level dependencies are handled, avoiding the treatment of non-independent observations as independent samples.

Finally, we will conduct a full human editing pass to improve transitions, reduce repetition, and make the definitions of information availability, causal directions, and control subspaces clearer throughout the manuscript.

We sincerely appreciate your suggestions, which helped us identify concrete ways to improve the presentation and accessibility of the work. We hope these revisions will make the main experimental insights and their supporting evidence substantially easier to evaluate.

We would greatly appreciate it if you could take another look at our responses when you have time. As the discussion deadline is approaching, we are concerned that there may be limited time to address any additional questions or presentation-related concerns that arise later. If anything remains unclear, we would be very grateful if you could let us know at your earliest convenience, so that we have an opportunity to provide further clarification before the discussion closes.

Thank you very much for your time and consideration!

Best regards,
The Authors

Reviewer_FwcMOpenReview ↗

Response to authors

I would like to thanks the authors for their response. Their reply addresses some of my concerns; however, the most substantive concern regarding the experiments and results was not directly addressed. I appreciate the attempt to reorganize the findings, but my comments was about the lack of significant results or findings to justify a paper at ICLR. I still don't believe this meets the bar at this stage. I am happy to update my review to reflect a higher score but for me it does not pass the line for acceptance. I'd encourage the authors to continue exploring the space with a deeper set of experiments.

Thank you for the careful follow-up. We agree that the main question is not whether the presentation can be made cleaner, but whether the experiments establish a sufficiently substantive finding. We believe the revised results do.

Our central finding is not simply that models sometimes say BLOCK and subsequently behave unsafely, nor is it that a steering intervention should uniformly improve free-generation safety. The substantive result is a causal dissociation between safety judgment control and action control.

Specifically, we now establish this at three complementary levels:

1. Within the same judgment-to-action trajectory, controlling judgment does not provide comparable control over action.
Our within-trajectory intervention directly perturbs the upstream judgment representation and measures its downstream effect on action preference. The intervention reliably changes the judgment endpoint, yet its effect on action preference is substantially weaker than an action-native intervention. This rules out the interpretation that our result is merely a consequence of measuring judgment and action at unrelated locations.

2. The asymmetry persists under a matched-position comparison.
We directly compare judgment-native and action-native controllers at the same upstream position, with matched intervention norms and the same action endpoint. Across the three models, the action-native controller consistently produces a substantially stronger change in action preference. Thus, the difference cannot be reduced to “the action intervention was simply closer to the action head” or to a trivial positional advantage. This is the key causal result of the paper.

3. The effect reflects a structured control geometry rather than a single steering direction.
Our subspace and recombination analyses show that judgment- and action-control directions share a component but are not interchangeable. Importantly, effective action control remains available in components orthogonal to the judgment-control subspace. Together with the matched-position intervention, this supports a concrete mechanistic picture: safety-relevant information can be available and controllable at the judgment interface without that representation exerting comparable causal control over action preference.

The additional free-generation experiments provide an important boundary condition rather than a contradiction. Behavioral realization is architecture-dependent: the same latent intervention can manifest as refusal, non-tool responses, alternative tool choices, or continued unsafe proposals. We therefore do not use free-generation success rate as the primary evidence for the causal claim. The primary evidence is the controlled judgment-to-action intervention, where the estimand is well defined and the comparison is directly matched.

We have also revised the presentation accordingly. Figure 4 now foregrounds the within-trajectory causal comparison, numerical precision and statistical units are reported consistently, and the main text distinguishes the primary causal evidence from auxiliary probe and diagnostic analyses. These changes make the central result easier to evaluate rather than adding experiments simply for volume.

More broadly, we think the distinction here is important for interpreting the contribution. A model recognizing that an action is unsafe and a model assigning low preference to that action are often treated as two manifestations of the same safety computation. Our experiments show that they need not be causally equivalent: the representation that controls the safety judgment need not be the representation that most effectively controls the downstream action preference. We believe this causal asymmetry is the substantive finding that goes beyond the descriptive observation that models can “say BLOCK, still act.”

We appreciate the reviewer’s concern about significance. After the additional experiments and the reorganization of the manuscript, we believe the evidence supports this narrower but stronger mechanistic conclusion, and that it constitutes a substantive contribution to understanding how safety judgments are connected—or fail to connect—to action selection in LLM agents.

Reviewer_ca6uOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies the phenomenon that LLMs can judge an action to be unsafe while still preferring to execute it. Through probing, activation interventions, and subspace analysis, the authors find that safety-relevant information remains readable in action states, but judgment-derived directions transfer weakly to action. Based on these findings, the paper discusses the relationship between safety judgments and action control.

Critical strengths
  1. The research question has practical significance, focusing on whether safety judgment capabilities can translate into safe actions, and offers insights for agent safety evaluation.

  2. The experiments are relatively systematic, covering three models, different layers and intervention strengths, and including random controls and stability analyses.

Critical weaknesses
  1. Overclaiming or ambiguity in the central claim. The paper asks, “does a correct safety judgment actually govern what the agent chooses to do?” However, an LLM being able to make a correct judgment when asked does not mean that it has already made that judgment when taking action. It may simply not have recognized the need to make a safety judgment. This explanation needs to be ruled out to distinguish between “not initiating a judgment” and “failing to follow a judgment after making it,” and to make the central statement more precise.

  2. Whether low directional overlap can explain weak transfer still needs clarification. The overlap between judgment and action directions is indeed higher than the random baseline. However, random overlap in high-dimensional spaces is already close to zero, so a small absolute value is not surprising. It therefore remains necessary to explain why this small value can account for weak transfer from judgment capabilities to safe behavior. Clarifying the relationship between the magnitude of this value and actual transfer effectiveness would make the mechanistic explanation more convincing.

  3. The effect of an explicit BLOCK suggests an alternative explanation. The experimental results show that an explicit BLOCK can significantly influence behavior. This also points to a more direct possibility: the agent initially did not recognize the need to make a safety judgment and only started considering safety after receiving an explicit prompt. This explanation needs to be ruled out to better support the paper’s account of the judgment–action disconnect.

Questions for authors
  1. Could the authors clarify whether the experiments show “making a correct judgment but not acting accordingly” or “not recognizing the need to make a safety judgment”? What evidence can distinguish these two cases?

  2. Could the authors explain why the observed low directional overlap supports the conclusion of weak capability transfer? Conversely, how much influence would a judgment direction need to have on behavior to demonstrate that the judgment capability is being used effectively?

  3. Given that an explicit BLOCK can significantly influence behavior, could the authors further test whether this is mainly because the prompt triggers a safety judgment process that the agent had not initiated on its own?

3 comments
AuthorsOpenReview ↗

Response to Reviewer ca6u

Response to Reviewer ca6u

We thank the reviewer for highlighting the distinction between initiating a safety judgment and propagating an established judgment into subsequent action selection. Our additional experiments examine these two processes separately and clarify how explicit judgment context, latent judgment control, and action preference interact.

1. Judgment initiation and judgment-to-action propagation

Our experiments distinguish two questions: whether a model spontaneously initiates a safety judgment, and whether the internal variables controlling that judgment also influence subsequent action selection.

The causal interventions directly test the second question. In the within-trajectory experiment, we intervene at the upstream judgment position, verify that the intervention shifts the judgment toward BLOCK, and then measure its effect on a later action preference while holding the explicit BLOCK token fixed.

At Qwen L32, Gemma L32, and Mistral L24, the BLOCK-oriented judgment controller shifts the judgment readout by +4.17, +18.7, and +6.03, respectively. Its corresponding downstream action effects are only +0.000924, −0.000065, and +0.00135. The paired differences relative to matched random upstream perturbations have 95% confidence intervals containing zero in all three models. By contrast, downstream action-native interventions produce effects of +0.0435, +0.212, and +0.0944. The matched-upstream comparison independently confirms that the action-native advantage persists when both controllers are injected at the same upstream position.

We separately measure spontaneous judgment initiation through free generation, without teacher-forcing the ALLOW/BLOCK label. The models generate BLOCK in 80/90 Qwen cases, 78/90 Gemma cases, and 76/90 Mistral cases. Their subsequent action distributions differ: 41/78 Gemma BLOCK trajectories contain an unsafe-target proposal, whereas Qwen and Mistral have no unsafe-target proposals among their BLOCK trajectories in this cohort. Importantly, Qwen produces 69/90 non-tool/final responses, while Mistral frequently proposes alternative tools; these outcomes should not be interpreted as successful safe tool execution.

The four prompt conditions (Direct, Safety Reminder, Explicit Judgment, and Neutral Deliberation) further show that prompting changes both entry into the tool-use path and the resulting action distribution. Relative to Direct, Safety Reminder increases unsafe-proposal rates by 43.3 percentage points for Qwen, 14.4 for Gemma, and 11.1 for Mistral.

Together, these results distinguish spontaneous judgment initiation from the causal propagation of judgment-related information. Initiation varies across architectures and prompt conditions, while the controlled interventions reveal a separate asymmetry: even when the judgment-side state is demonstrably shifted toward BLOCK, its latent control does not translate into comparably strong downstream action control.

2. Interpreting judgment–action subspace overlap

The overlap analysis characterizes the geometry of judgment- and action-control spaces; the causal claim is established by intervention experiments rather than inferred from overlap magnitude alone.

To assess the stability of these estimated spaces, we added split-half reconstruction analyses. At rank 16:

ModelJ–J split-halfA–A split-halfJ–A overlapRandom reference
Qwen3-32B52.9%54.6%1.86%0.313%
Gemma3-27B47.3%47.4%4.70%0.298%
Mistral-Small-3.2-24B49.8%42.1%1.71%0.313%

Within-task reconstructions are substantially more reproducible than cross-task overlap. Meanwhile, J–A overlap remains above the random reference, indicating partial geometric sharing rather than orthogonality or independence.

More directly, we intervened on the action-controller components parallel and orthogonal to the judgment-control subspace. At representative layers with effective action control, the orthogonal component retains substantial efficacy:

Model and layerJudgment-parallelJudgment-orthogonal
Qwen L320.01100.123
Gemma L320.00570.0403
Mistral L240.01130.0240

The full Mistral action controller produces an effect of 0.0238, comparable to the orthogonal component's 0.0240. Each component is separately normalized and norm-matched; these values therefore compare intervention efficacy, not additive contributions.

The matched-upstream experiment provides an additional direct measure of transfer effectiveness. The action-native controller's paired action-effect advantage over the judgment-native controller is positive across all nine family means in each model: 0.0656 [95% CI: 0.0537, 0.0783] for Qwen, 0.0480 [0.0339, 0.0634] for Gemma, and 0.0171 [0.0107, 0.0238] for Mistral. The family-level exact sign-flip test gives p = 0.00391 for each model.

We do not posit a universal numerical threshold at which a judgment direction becomes behaviorally effective. Instead, we assess its causal influence against matched random perturbations and action-native positive controls at the same intervention position and norm. The judgment-native direction shows substantially weaker action control.

Thus, geometry indicates partially shared control structure, while component and matched-upstream interventions establish functional asymmetry. Low overlap is not treated as a standalone explanation or quantitative mediator of weak transfer.

3. Explicit BLOCK context and its relationship to latent judgment control

The externally supplied BLOCK effect is substantial. In a factorial intervention crossing textual context with latent judgment intervention, replacing ALLOW with BLOCK produces safe-oriented action effects of 0.132 for Qwen, 0.0989 for Gemma, and 0.152 for Mistral.

The corresponding latent judgment effects are smaller and architecture-dependent:

ModelText BLOCK − ALLOWLatent effect under ALLOWLatent effect under BLOCK
Qwen3-32B0.1320.006960.00170
Gemma3-27B0.0989−0.01090.00093
Mistral-Small-3.2-24B0.1520.006720.00256

The factorial design separates two causal inputs at the action interface. Explicit judgment text strongly shifts action preference, whereas the latent direction controlling the model's own judgment has a much smaller and context-dependent effect. The two inputs therefore have distinct causal roles.

Crucially, the factorial experiment uses externally supplied judgment text; it does not assume that the model spontaneously generated or internally established that judgment. That question is measured independently in the free-generation experiments above. Prompt-condition results also show that explicitly requesting safety consideration does not uniformly improve behavior: Safety Reminder increases unsafe proposals relative to Direct in all three models.

We therefore distinguish supplying an explicit judgment to the action computation from intervening on the latent variables controlling the model's own judgment. The former can strongly influence action preference without making the latter a privileged causal bottleneck for action selection.

Conclusion

Our results distinguish spontaneous safety-judgment initiation, latent judgment-to-action control, and explicit judgment context. Across the tested architectures, explicit BLOCK strongly shifts action preference, yet judgment-native latent control transfers less effectively than action-native control, even under matched-position interventions.

Our central finding is that the internal variables controlling safety judgment and action preference are not causally interchangeable. Safety information can remain available and explicit judgments can influence behavior without their supporting latent variables exerting comparable control over action selection.

AuthorsOpenReview ↗

Follow-up on Judgment–Action Causal Control

Dear Reviewer,

Thank you again for your insightful and detailed feedback. We particularly appreciate your emphasis on distinguishing between the failure to initiate a safety judgment and the failure of an established judgment to influence subsequent action selection.

We conducted several additional experiments to investigate these two possibilities separately and to further clarify the causal interpretation of our findings.

1. Separating judgment initiation from judgment-to-action propagation

We added free-generation experiments without teacher-forcing the ALLOW/BLOCK label to measure whether models spontaneously initiate safety judgments. BLOCK judgments occur in 80/90 Qwen cases, 78/90 Gemma cases, and 76/90 Mistral cases. However, their subsequent action distributions differ substantially: Gemma produces unsafe-target proposals in 41/78 BLOCK trajectories, while Qwen and Mistral produce none in their respective BLOCK trajectories in this cohort.

More importantly, our within-trajectory interventions directly examine the second question. Even when the upstream judgment state is successfully shifted toward BLOCK, its effect on subsequent action preference remains very small compared with action-native interventions. This provides a controlled test of judgment-to-action propagation, separate from whether a model spontaneously initiates a judgment.

2. Clarifying the relationship between subspace overlap and causal transfer

We added split-half reconstruction analyses to assess the stability of the estimated judgment- and action-control subspaces. We also directly intervened on the action-controller components parallel and orthogonal to the judgment-control subspace. The orthogonal components retain substantial action-control efficacy.

Importantly, we do not use low subspace overlap alone to explain weak transfer. Instead, we establish the causal asymmetry through matched intervention experiments, using matched random perturbations and action-native positive controls.

3. Distinguishing explicit BLOCK context from latent judgment control

We added a factorial experiment separating externally supplied BLOCK/ALLOW text from latent judgment interventions. Explicit BLOCK text substantially shifts action preference, while the latent judgment effects remain much smaller and architecture-dependent.

We further examined different prompting conditions, including Direct, Safety Reminder, Explicit Judgment, and Neutral Deliberation. Interestingly, Safety Reminder does not uniformly improve behavior and actually increases unsafe-proposal rates relative to Direct across all three models. This helps clarify why explicit safety prompting and latent judgment control should not be treated as equivalent mechanisms.

We hope these additional experiments help distinguish judgment initiation, explicit judgment context, and latent judgment-to-action propagation, and provide a more precise account of the central mechanism studied in our paper.

We would sincerely appreciate it if you could revisit our responses when you have a chance. Since the discussion deadline is approaching, we are concerned that there may not be sufficient time to address further questions or clarify any remaining concerns if they arise too close to the deadline. If there are aspects of our additional analyses that you would like us to elaborate on, we would be very grateful if you could let us know at your earliest convenience.

Thank you again for your valuable feedback and consideration!

Best regards,
The Authors

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

The paper looks at agents that correctly identify actions that should be blocked, but prefer to take them regardless. The authors find that the safety information is still present when the model chooses the action, and that steering the internal signal for the judgment barely changes the outcome. The authors reason that improved safety recognition may not make agents safer by itself.

Critical strengths
  1. Separating info from the control: The authors directly test whether the safety information they decode is actually being used by the model. This is one of the key points that makes the main finding interesting.

  2. Converging evidence: The authors use multiple methods (steering, teacher forcing, natural state swaps), and all of them show weak transfer across the models. The controls are rigorous and comprehensive as well, which makes the result reliable.

Critical weaknesses
  1. Conclusion goes too far: The block label being held fixed removes a path that may matter (the paper's results seem to show that an explicit block text shifts action strongly). The null could simply be reflecting intervention at a single place, so it doesn't seem to answer why judgements fail to control the action.

  2. Unclear implications about real agent behavior: The paper uses scored action margins but doesn't report how often the decisions flip or whether the models still prefer the action after saying block. The main test is also narrow, since it's just three scenarios in single-step settings.

Questions for authors
  1. How often does the behavior show up in the shared trajectory, i.e. how often do the models prefer the target after saying block, and how often are the interventions helping?

  2. Is the judgement controller reliably capturing the model's natural judgement on safety? The gradient direction seems to be changing judgement much more than the natural state swaps. Are the effects on the same scale? If not, it could potentially be something narrower like a <say "block"> direction, than a holistic signal on safety judgement.

2 comments
AuthorsOpenReview ↗

Response to Reviewer MywG

Response to Reviewer MywG

We thank the reviewer for emphasizing the need to connect the internal causal analysis with actual generated action behavior.

1. How often does a model produce an unsafe target after BLOCK?

We now report joint judgment/action outcomes from free generation:

ModelBLOCK judgmentsUnsafe target after BLOCKP(unsafe target | BLOCK)
Qwen3-32B80/900/800.0%
Gemma3-27B78/9041/7852.6%
Mistral-Small-3.2-24B76/900/760.0%

This refines the behavioral interpretation: the pattern is clear in Gemma, but is not a uniform three-model free-generation result. Non-tool, safe-alternative, other-tool, and unsafe-target outcomes are reported separately.

2. What does the gradient-derived controller represent?

We use "judgment controller" operationally: it is a train-derived direction whose intervention shifts the explicit judgment margin m_J = log p(ALLOW) − log p(BLOCK). For blocked-valid scenarios, g_J^safe = −∇_h m_J is the BLOCK-oriented direction. It is not defined as a complete or universal representation of every aspect of natural safety reasoning.

Its native endpoint validity is tested directly. Across the matched-depth sign-control sweep, the judgment intervention has the expected ALLOW/BLOCK orientation; at the primary within-trajectory layers, g_J^safe shifts the judgment endpoint toward BLOCK on all 60 held-out cases in each model. We also tested semantic label-format transfer. In the reciprocal 4×4 format-transfer analysis, family-cluster 95% CIs are fully positive in 12/12 source-target cells for Qwen, 12/12 for Gemma, and 11/12 for Mistral. This supports an interpretation beyond a literal BLOCK-token association.

The action endpoint does not show the same transfer: among 18 additional alternative-source action comparisons, only four have confidence intervals fully above zero. This is consistent with the distinction between controlling judgment and controlling action preference.

3. Are gradient and natural-state-swap effects on the same scale?

No. We do not compare their raw magnitudes as though they were commensurate. The gradient controller is a local, endpoint-defined intervention derived from the explicit judgment margin. The natural-state experiment replaces a full residual vector with a naturally occurring matched donor state; it does not use a gradient direction or norm matching. These experiments answer complementary questions and have different intervention geometries.

With explicit BLOCK fixed, natural-state-swap action effects are near zero in all three models. Judgment alignment is clear for Gemma and Mistral, while the Qwen judgment-alignment interval crosses zero at the tested layer. When the explicit label is allowed to vary, label switches are rare: 10/120 directional swaps in Mistral, 1/120 in Qwen, and 0/120 in Gemma. The large Mistral action effect is concentrated in those ten label-switch cases; label-unchanged cases remain near zero. We treat this subgroup descriptively, not as formal mediation identification.

4. Practical implication

We agree that a margin-level causal effect is not the same as a reduction in unsafe tool execution. The free-generation intervention results do not show a statistically reliable cross-model reduction in unsafe proposals. We will frame the practical implication as an evaluation lesson: the ability to recognize a safety constraint is not sufficient evidence that the variables supporting that recognition control action selection. Causal coupling to action selection should be evaluated separately.

AuthorsOpenReview ↗

Follow-up on Judgment–Action Causal Control (Submission 26261)

Dear Reviewer,

Thank you again for your careful reading of our paper and for raising important questions about the connection between our internal causal analyses and actual generated action behavior.

Following your suggestions, we conducted additional analyses to better characterize the behavioral implications and clarify what our judgment controller represents.

In particular, we would like to highlight three additions:

1. Free-generation behavior after BLOCK

We added joint judgment/action statistics without teacher-forcing the BLOCK label. The results reveal meaningful differences across architectures: 41/78 (52.6%) of Gemma's BLOCK trajectories contain an unsafe-target proposal, whereas Qwen and Mistral have no unsafe-target proposals among their respective BLOCK trajectories in this cohort. We also report safe alternatives, other-tool proposals, and non-tool responses separately, avoiding the assumption that every non-unsafe outcome represents successful safe execution.

2. Clarifying the judgment controller

We clarified that our judgment controller is an operationally defined, train-derived direction that shifts the explicit ALLOW/BLOCK judgment margin, rather than a claim that it captures every aspect of natural safety reasoning. We further validated its BLOCK-oriented effect across held-out cases and examined transfer across different semantic label formats. These results help distinguish a broader judgment-related direction from a simple association with the literal BLOCK token.

3. Distinguishing intervention effects and practical implications

We also clarified that gradient-based interventions and natural-state swaps have different intervention geometries and should not be compared through their raw effect magnitudes. Importantly, we have narrowed our practical claims: the current free-generation intervention results do not establish a statistically reliable reduction in unsafe proposals across all three models. Instead, our findings motivate evaluating whether safety-related internal representations actually exert causal control over action selection.

We hope these additional results and clarifications address your concerns and provide a more precise interpretation of the behavioral and mechanistic findings.

We would greatly appreciate it if you could take another look at our responses when you have a chance. As the discussion deadline is approaching, we are concerned that there may be limited time for us to address any additional questions that arise closer to the deadline. If any concerns remain, we would sincerely appreciate hearing from you at your earliest convenience, so that we can provide further clarification before the discussion closes.

Thank you very much again for your time and thoughtful feedback!

Best regards,
The Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal Summary

Dear Area Chair and Reviewers,

We sincerely thank all reviewers for their valuable feedback. As the discussion period draws to a close, we would like to briefly summarize our contributions and the additional work completed during the discussion.

Paper Summary We appreciate the reviewers' recognition of our rigorous separation between information availability and causal control (Reviewer MywG), the systematic combination of probing, causal interventions, and subspace analysis (Reviewer 7zVn), and the practical significance of distinguishing "recognizing safety" from "acting safely" (Reviewer ca6u). Our work provides a mechanistic investigation into why LLM agents can correctly judge an action as unsafe while still preferring to execute it, demonstrating an asymmetric causal control between judgment and action preference.

Discussion Period Summary During the discussion period, we responded to all questions, provided extensive additional experiments, and incorporated substantial revisions into the manuscript.

  • Matched-upstream intervention proves the asymmetry (Reviewers 7zVn, MywG). Direct response to position bias: To rule out the possibility that the action-native controller simply had an injection advantage, we derived an action-native controller and injected it at the same upstream position with matched intervention norms. The action-native controller remained substantially more effective across all three models (paired differences: +0.0656 for Qwen, +0.0480 for Gemma, +0.0171 for Mistral; family-level exact sign-flip test p = 0.00391). This definitively establishes the causal asymmetry.
  • Disentangling latent judgment and explicit BLOCK context (Reviewers MywG, ca6u). Clarification on the fixed-label mediator: We introduced factorial interventions crossing explicit BLOCK/ALLOW text with latent judgment interventions. While explicit BLOCK text strongly shifts action preference, latent judgment effects are much smaller and architecture-dependent. Furthermore, our new prompt-condition experiment reveals that "Safety Reminder" prompts actually increase unsafe-proposal rates (e.g., +43.3 pp for Qwen relative to Direct), proving that prompting for safety is not causally equivalent to latent judgment control.
  • Behavioral validation in free generation and multi-step sandboxes (Reviewers MywG, 7zVn). We added free-generation experiments without teacher-forcing. We found a pronounced "BLOCK-then-unsafe" pattern in Gemma (41/78, 52.6%), though Qwen and Mistral exhibit different distributions (mostly non-tool or alternative-tool proposals). We also extended evaluation to a 3-step in-memory sandbox, strictly defining the behavioral boundary and acknowledging that our interventions do not uniformly eliminate all unsafe proposals.
  • Significance of the mechanistic finding and presentation (Reviewer FwcM). Direct response to "underwhelming results": We respectfully clarified that the core contribution is a causal dissociation, not just a descriptive behavioral observation. We restructured the paper around three claims (information availability, asymmetric causal control, and partially shared control geometry) and replaced Figure 4 with a compact effect-size/forest-plot showing 95% CIs. We also standardized all numerical precision and explicitly reported statistical procedures (e.g., family-cluster bootstraps with 20,000 resamples) to make the substantial nature of our findings transparent.

Detailed results and explanations are available in the corresponding review threads and the updated manuscript. We sincerely appreciate your time and consideration.

Sincerely, The Authors

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

Related papers

Discussion (0)

Sign in to comment.