acceptodds
Desk rejected at ICLR 2027

Understanding and Mitigating Hallucination Escape in Tool-Using LLM Agents

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

Abstract

Large language models (LLMs) increasingly serve as autonomous agents that invoke external tools. However, this capability introduces tool hallucination, selecting incorrect tools or generating invalid calls. Existing mitigation methods report substantial improvements, yet we identify a previously overlooked failure mode that we term **Hallucination Escape**. These methods reduce hallucination on the tool configuration they are tuned on but increase it on other configurations, canceling out the gain. We further investigate this phenomenon and find that hallucination rises sharply when a model's intrinsic tool-use tendencies conflict with the current tool configuration, and that existing methods reinforce rather than suppress these tendencies, which in turn contributes to hallucination escape. Building on these findings, we propose **EscapeGuard**, a training-free inference-time method that combines conflict-aware gating with configuration-derived attention enhancement to mitigate tool hallucination while preventing hallucination escape. Across six benchmarks on various models, **EscapeGuard** reduces tool-selection hallucination by 9.0 pp and suppresses hallucination escape, lowering the cross-configuration mean by 23.7 pp and achieving an 89.1% net improvement in paired-query evaluation. We hope this work can encourage evaluation beyond a single tool configuration and pave the way for more reliable tool-using LLM agents.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Tool Hallucination, LLM Agents, Hallucination, Inference-Time Intervention

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_SpYwOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper shows that existing tool-hallucination mitigation methods reduce errors on the tool definitions they were tuned on, but increase errors once tool names, descriptions, or argument names change, a failure the authors call hallucination escape. They link this to models’ built-in tool preferences, which existing methods reinforce, and propose EscapeGuard, a training-free method that uses a probe to detect conflicts and then boosts attention to the current tool definitions.

Critical strengths
  • Evaluating mitigation methods beyond the tool definitions they were tuned on, including the paired-query test for side effects on related queries, is a useful evaluation practice that most prior mitigation work skips.
  • The probe analysis includes reasonable controls, such as random labels, query-only inputs, and a novelty-matched condition, which make the conflict signal more credible than a plain probing result.
Critical weaknesses
  • The evaluated models are mostly older or mid-sized open models such as LLaMA-3.1, Mistral-7B, and Mixtral-8x7B, plus two Qwen-3.5 models. Strong open-weight models built for agentic tool use, such as DeepSeek-V4, GLM-5.2, and Kimi K3, are not evaluated, and neither are closed-source models. DeepSeek-V4 was released months before the deadline, and its Flash version activates only 13B parameters per token, so white-box access and compute cost do not explain this gap. Testing the phenomenon itself only needs API access. The paper’s own results show the problem shrinks with scale, as C2 errors drop from 50.4% for LLaMA-3.1-8B to 36.8% for the 70B model, so it is unclear whether hallucination escape still matters for current strong tool-use models.
  • Robustness to changes in tool names, descriptions, and argument names has been studied before. RoTBench, https://arxiv.org/abs/2401.08326, perturbs tool and parameter names at several noise levels and shows large drops even for GPT-4, and it also proposes RoTTuning, which trains on more diverse tool environments. RobustBench-TC, https://arxiv.org/abs/2605.11928, studies a broader set of tool-registry perturbations. The paper cites neither, so its claim that this failure mode is “previously overlooked” is overstated, and RoTTuning is a natural baseline that is missing.
  • The claim that existing methods cause hallucination escape by reinforcing tool preferences rests on weak evidence. It is based on only five methods, with no correlation statistic and no causal test, and the data in Table 3 are not consistent with it. On LLaMA, LinSteer increases preference strength by only 0.07 but has an escape ratio of 1.92, higher than Relign at 1.57 with a 0.11 increase. On Qwen, Gorilla has the largest increase but not the largest escape ratio.
Questions for authors
  • How do strong recent models such as DeepSeek-V4, GLM-5.2, Kimi K3, or closed-source models behave across C1 to C6? Does hallucination escape still appear for these models?
  • How does EscapeGuard compare with RoTTuning, which directly targets robustness to tool-name and parameter changes?
  • How does EscapeGuard behave when no suitable tool is available and the right behavior is to abstain? Does boosting attention to the tool list increase forced calls?
  • Can you report standard deviations or confidence intervals for the main tables, given that results are averaged over three seeds?
7 comments
AuthorsOpenReview ↗

A Big Thank You and Our Response to Reviewer SpYw (1/2)

Thank you so much for the careful and insightful review! The comments have helped us make the paper noticeably stronger, and we have addressed each of them. All changes are marked in blue in the revised PDF. Our point-by-point responses are as follows.


To W1 and Q1: Thank you for the suggestion. We have added an evaluation of four recent strong models in Appendix E.3 (new Table E.5) and a full method comparison on two recent models in Appendix H.12 (new Table H.15), and would like to clarify the scale results as follows:

  • Behavior across C1–C6. We added an evaluation of four recent strong models (DeepSeek-V4.1-Flash, GLM-5.2, Kimi K3, and GPT-6.1 Sol) via their official APIs. Their C1 error is low (3.9%–8.6%) but rises to 19.6%–31.4% under C2 (Table E.5, Appendix E.3).
  • Whether escape still appears. We added the full comparison on DeepSeek-V4.1-Flash and GLM-5.3-Flash: both training-free baselines still exhibit escape, while EscapeGuard reduces hallucination under all configurations (Table H.15, Appendix H.12).
  • Clarification on scale. The lower C2 error of LLaMA-3.1-70B reflects better general capability: its C1 error nearly halves (22.8% → 12.4%), while the renaming-induced increase (C2 − C1) barely changes (27.6 → 24.4 pp).

Table E.5. Hallucination rates (%) of base models.

ModelC₁C₂C₃C₄C₅C₆H̄(m)
DeepSeek-V4.1-Flash8.631.414.627.839.120.323.6
GLM-5.25.625.611.722.933.516.319.3
Kimi K34.926.710.122.634.317.819.4
GPT-6.1 Sol3.919.69.120.228.413.615.8

Table H.15. Hallucination rates (%); parentheses: change from Base.

ModelMethodC₁C₂C₃C₄C₅C₆H̄(m) -Var(m) -
DeepSeek-V4.1-FlashBase8.631.414.627.839.120.323.6105.9
LinSteer5.9 (-2.7)34.2 (+2.8)15.4 (+0.8)27.5 (-0.3)41.2 (+2.1)21.9 (+1.6)24.3 (+0.7)136.2
PRISMS5.2 (-3.4)33.6 (+2.2)15.3 (+0.7)27.3 (-0.5)40.3 (+1.2)21.1 (+0.8)23.8 (+0.2)134.3
EscapeGuard3.4 (-5.2)4.3 (-27.1)5.3 (-9.3)20.9 (-6.9)24.1 (-15.0)14.1 (-6.2)12.0 (-11.6)68.0
GLM-5.3-FlashBase7.127.912.825.635.818.221.292.6
LinSteer4.7 (-2.4)31.0 (+3.1)13.9 (+1.1)25.2 (-0.4)37.7 (+1.9)20.4 (+2.2)22.2 (+1.0)117.5
PRISMS4.1 (-3.0)29.7 (+1.8)13.4 (+0.6)24.9 (-0.7)37.0 (+1.2)19.3 (+1.1)21.4 (+0.2)115.4
EscapeGuard2.9 (-4.2)3.7 (-24.2)4.8 (-8.0)19.9 (-5.7)22.6 (-13.2)12.8 (-5.4)11.1 (-10.1)62.4

To W2 and Q2: Thank you for pointing us to this related work. In the revised PDF, we now cite and discuss both RoTBench and RobustBench-TC, and we have added RoTTuning as an additional baseline (Tables 6, 7, and H.7).

  • Related work. Thank you for these references. We have adjusted our claims and discussed both works in §1 and §2. RoTBench and RobustBench-TC measure a model’s robustness to tool-environment perturbations and improve it by training on them (RoTTuning, domain-randomized RL). We instead show that hallucination mitigation methods shift errors across configurations, which a robustness score does not reveal, trace this to reinforced tool-use tendencies, and address it without training by relying on the runtime configuration.
  • Comparison with RoTTuning. We added RoTTuning as a baseline (Tables 6, 7, and H.7). It improves the perturbations it is trained on (C2, C4, C5) but worsens the others (C3, C6), while EscapeGuard is better under every configuration. This complements our analysis: mitigation methods fail where reinforced tendencies conflict with the configuration, RoTTuning fails outside its training perturbations, and single-configuration evaluation hides both; EscapeGuard avoids both by relying on the runtime configuration.

Table 7 (excerpt). Hallucination rates (%); parentheses: change from Base (full C1–C6 results in Table 7 of the revised PDF).

ModelMethodC1C2C3C6H̄
LLaMA-3.1-8BRoTTuning14.5 (−8.3)35.3 (−15.1)35.2 (+3.6)43.1 (+5.1)35.8 (−5.6)
EscapeGuard7.4 (−15.4)8.6 (−41.8)8.0 (−23.6)24.4 (−13.6)17.2 (−24.2)
Qwen-3.5-9BRoTTuning12.0 (−7.4)33.3 (−14.3)32.4 (+3.0)40.6 (+4.6)33.6 (−5.2)
EscapeGuard5.6 (−13.8)6.8 (−40.8)6.2 (−23.2)22.8 (−13.2)15.5 (−23.3)
AuthorsOpenReview ↗

A Big Thank You and Our Response to Reviewer SpYw (2/2)

To W3: Thank you for this comment. We have revised Table 3 to report ΔH̄, and added a correlation analysis and an intervention experiment (Appendix F.3, new Table F.9) linking prior reinforcement to escape:

  • Escape severity and Table 3. Escape is defined by the change in the cross-configuration mean H̄ (Definition 3.1), so we now report ΔH̄ in Table 3. Ew = O(m)/I(m) is divided by the C1 gain, so a small gain inflates it: LinSteer on LLaMA has the smallest C1 gain, while its total deterioration on other configurations (O(m) = 12.1 pp, Table F.9) is below that of Relign and PALADIN.
  • Correlation. We added a correlation analysis: ΔS correlates with ΔH̄ (Spearman ρ = 0.90 over all ten method–model pairs, p < 0.001).
  • Causality. We added an intervention experiment that splits LinSteer's steering vector on LLaMA into its prior-aligned component and the remainder (Table F.9, Appendix F.3). Removing the prior component keeps most of the C1 gain (4.7 of 6.0 pp) but removes escape (ΔH̄: +1.0 → −0.5), while the prior component alone reproduces it (ΔH̄ = +0.9) with little C1 gain.

Table F.9. Decomposing LinSteer's steering vector (LLaMA-3.1-8B); changes from Base. O(m): total deterioration on other configurations (pp). NER with 95% pair-bootstrap CI.

Steering vectorΔSRet.ΔC₁ΔH̄O(m)RepairDamagecNINER [95% CI]
Original v+.0786−6.0+1.012.155.338.334.112.6 [0.0, 23.7]
v<sub>∥</sub> only (prior component)+.0687−2.8+0.98.325.835.016.8−2.9 [−14.8, 7.4]
v<sub>⊥</sub> only (prior removed)+.0194−4.7−0.51.743.312.338.023.3 [13.3, 33.3]

To Q3: EscapeGuard reduces, rather than increases, forced calls. RelyToolBench (Table 4) includes 348 tasks with no matching tool, where any call counts toward TSHR; on this subset, EscapeGuard lowers the wrong-call rate from 41.3% to 29.6% (LLaMA) and from 37.4% to 26.3% (Qwen) (Table H.16, Appendix H.13). The bias only reweights attention toward tool definitions in the context and adds no term favoring a tool call, so it helps the model see that no tool fits rather than pushing it to call one; when the gate is not triggered, outputs are unchanged.

Table H.16. Wrong-call rate (%) on the 348 RelyToolBench tasks with no matching tool.

ModelBaseEscapeGuardGate firesGate on: Base → EGGate off: Base = EG
LLaMA-3.1-8B41.329.658.0%47.9 → 27.832.2
Qwen-3.5-9B37.426.356.1%44.6 → 24.828.2

To Q4: We have added standard deviations over the three seeds to Table 1. The seeds control sampling during decoding, and all methods are evaluated on the same query set (Appendix C.1). We also report 95% paired bootstrap confidence intervals over queries for the changes from the base model (Table H.17, Appendix H.6). All five methods significantly reduce C1 errors and, except PRISMS on Qwen, significantly increase C2 errors; none significantly lowers H̄.

Table H.17 (excerpt). H̄ (mean ± std over three seeds) and its change from Base with 95% paired bootstrap CI (full results in Table H.17 of the revised PDF).

ModelMethodH̄ ± stdΔH̄ [95% CI]
LLaMA-3.1-8BBase41.4 ± 0.4—
Relign42.4 ± 0.4+1.0 [−0.1, +2.1]
PALADIN42.2 ± 0.4+0.8 [−0.3, +1.9]
Gorilla43.4 ± 0.4+2.0 [+0.8, +3.2]
LinSteer42.4 ± 0.4+1.0 [−0.1, +2.1]
PRISMS41.7 ± 0.3+0.3 [−0.9, +1.5]
Qwen-3.5-9BBase38.8 ± 0.4—
Relign39.4 ± 0.4+0.6 [−0.5, +1.7]
PALADIN39.7 ± 0.4+0.9 [−0.3, +2.1]
Gorilla40.4 ± 0.4+1.6 [+0.5, +2.7]
LinSteer39.9 ± 0.4+1.1 [−0.1, +2.3]
PRISMS38.9 ± 0.4+0.1 [−1.1, +1.3]

We hope these revisions address the concerns and would be grateful if the reviewer would consider updating the assessment. We would be happy to answer any further questions during the discussion period. Thank you again!

Thank you for the detailed response and new experiments. Table H.17 shows that the increase in H̄ is significant only for Gorilla. Could you provide an analysis explaining this?

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

The paper studies hallucination escape, a failure mode in which mitigation reduces tool hallucination in the tuned setting but increases it under other tool configurations or on related queries, offsetting the gains. It proposes EscapeGuard, which uses learned conflict probes to gate attention enhancement toward the current tool definitions. Experiments cover six backbones and six benchmarks, with substantial reported improvements under renamed and otherwise modified tool configurations.

Critical strengths
  1. Controlled tool-configuration changes expose a practically relevant limitation of evaluation under a fixed configuration. This diagnostic setting encourages assessing whether mitigation gains persist when tool interfaces change.
  2. EscapeGuard reports substantial cross-configuration gains, supported by component ablations in the appendix. These results provide meaningful empirical support for configuration-aware attention intervention.
Critical weaknesses
  1. The paired-query net-improvement metric omits some newly introduced errors. In Table 8, “Net Improvement” follows Equations D.11–D.16, which count collateral damage only on related queries whose initially incorrect targets have been repaired. The metric excludes regressions on initially correct targets and newly introduced related-query errors outside the repaired-target subset. Consequently, a positive value does not establish a reduction in total errors across the evaluated population. The authors should report complete before-and-after correctness transitions over all pairs, together with confidence intervals using the pair-level resampling described in Appendix H.6. The current metric should also be renamed to reflect its conditional scope. This concern affects the interpretation of the paired-query results, rather than the cross-configuration improvements in Table 7.
  2. The controlled configuration-shift evaluation primarily covers surface-form changes rather than substantive interface changes. Figure 2 and Appendix A.1 define transformations involving tool renaming, description rewriting, and argument renaming or reordering. Functionally equivalent tools in practical deployments may require different input representations, parameter structures, or units. The reported controlled experiments therefore leave unclear whether EscapeGuard’s attention enhancement supports these adaptations beyond surface-form changes. Evaluating such replacements, or explicitly limiting the generalization claims to the tested transformations, would clarify the method’s practical scope.
Questions for authors
  1. What is the change in total tool-selection errors across all 120 query pairs, including regressions outside the repaired-target subset? Please provide the full correctness-transition counts and pair-bootstrap confidence intervals.
  2. How does EscapeGuard perform when functionally equivalent tools require changes in input representation, parameter structure, or units, beyond renaming and reordering?
4 comments
AuthorsOpenReview ↗

A Big Thank You and Our Response to Reviewer EamH

Thank you so much for the careful and insightful review! The comments have helped us make the paper noticeably stronger, and we have addressed each of them. All changes are marked in blue in the revised PDF. Our point-by-point responses are as follows.


To W1 and Q1: Thank you for this point. We have renamed the metric and added full transitions (Table H.18).

  • Clarification on the metric. We have renamed it Conditional Net Improvement (cNI). The metric conditions on repaired targets so that new errors on the paired query can be attributed to the method; it does not measure the change in total errors.
  • Full transitions and total errors. We added the full correctness transitions over all 120 pairs, together with the Net Error Reduction (NER) over all 240 queries and pair-level bootstrap CIs (Table H.18, Appendix H.6). EscapeGuard reduces total tool-selection errors by 85.4% [77.9, 91.8] on LLaMA and 87.1% [79.5, 93.6] on Qwen, with only four new errors among the 240 queries; the baselines introduce 22–35 new errors and reduce total errors by 8.6%–18.4%. The abstract now reports NER.

Table H.18 (excerpt). Correctness transitions of EscapeGuard over the 120 pairs (C: correct; W: wrong; Tgt: target; Rel: related; all methods in Table H.18 of the revised PDF).

ModelMethodTgt C→WTgt W→CRel C→WRel W→CErrors (before → after)NER (%) [95% CI]
LLaMA-3.1-8BRelign536239103 → 8616.5 [2.9, 28.8]
PRISMS533189103 → 8418.4 [6.2, 29.4]
EscapeGuard152340103 → 1585.4 [77.9, 91.8]
Qwen-3.5-9BRelign43021893 → 8014.0 [0.0, 25.7]
PRISMS42818893 → 7915.1 [3.5, 24.8]
EscapeGuard14833793 → 1287.1 [79.5, 93.6]

To W2 and Q2: Thank you for the suggestion. We have added an evaluation on substantive interface changes (Table H.19).

  • Substantive interface changes. We added an evaluation of EscapeGuard beyond the surface-form changes of C2–C6 (which isolate the phenomenon while guaranteeing a valid solution). The tested tools are functionally equivalent but require changes in (a) input representation, e.g., a date string split into three integers; (b) parameter structure, e.g., flat arguments nested into an object; (c) units or encodings, e.g., Celsius vs. Fahrenheit; and (d) the allowed values of categorical arguments, e.g., "metric"/"imperial" replaced by "SI"/"US". EscapeGuard reduces OHR from 48.4% to 28.3% (LLaMA) and from 44.6% to 25.5% (Qwen), mainly through fewer tool-selection, argument-name, and schema-violation errors (Table H.19, Appendix H.14).
  • Clarification on scope. Wrong argument values barely change, and unit conversion remains the hardest case, consistent with the weak link between value errors and tendency conflict in §4. We now state the tested transformation types and list value conversion as a limitation in §6.

Table H.19. EscapeGuard on functionally equivalent tools with interface changes (%); parentheses: change from Base. AVHR is split into schema violations (missing required argument, invalid type or structure) and wrong values under valid argument names.

ModelTSHR ↓ANHR ↓AVHR (schema) ↓AVHR (value) ↓OHR ↓OHR by type (a) / (b) / (c) / (d)
LLaMA-3.1-8B8.7 (−15.9)5.0 (−6.2)10.3 (−9.1)12.1 (−0.7)28.3 (−20.1)27.9 / 29.6 / 31.4 / 24.2
Qwen-3.5-9B7.1 (−14.1)4.3 (−6.1)9.1 (−9.0)11.5 (−0.6)25.5 (−19.1)25.0 / 26.8 / 29.4 / 20.9

We hope these revisons address the concerns and would be grateful if the reviewer would consider updating the assessment. We would be happy to answer any further questions during the discussion period. Thank you again!

AuthorsOpenReview ↗

Gentle follow-up on our response

Dear Reviewer EamH,

Thank you so much for the time and care you have devoted to reviewing our paper. Your constructive comments have been very helpful in improving our work.

We understand that this is a very busy period, and we apologize for the additional message. As the discussion period is ending soon, we would be very grateful if you could let us know whether our response has addressed your concerns.

If anything remains unclear, we would be more than happy to provide further clarification. If you find that your concerns have been resolved, we would deeply appreciate it if you would kindly consider updating your assessment.

Thank you again for your time and consideration.

Best regards,

The Authors

AuthorsOpenReview ↗

Kind follow-up before the discussion closes

Dear Reviewer EamH,

Thank you again for your thoughtful review and specific suggestions. As today is the last day of the discussion period, we would be very grateful if you could take a brief look at our response and let us know if you have any further questions.

If you find our response satisfactory, we would sincerely appreciate it if you could kindly consider reflecting this in your final assessment.

Thank you very much for your time.

Best regards,

The Authors

AuthorsOpenReview ↗

Final follow-up before the discussion closes

Dear Reviewer EamH,

We apologize for writing once more as the discussion period closes today. In brief, we added the full correctness transitions over all query pairs and an evaluation on substantive interface changes (Tables H.18 and H.19), as you suggested.

Could you kindly let us know whether our response has addressed your concerns? If so, we would be truly grateful if you could consider raising your score.

Thank you so much for your time!

Best regards,

The Authors

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

This paper investigates hallucination escape, a failure mode in which existing tool-hallucination mitigation methods improve performance under their tuned tool configuration but introduce additional errors under other configurations or related queries. The authors evaluate five existing methods across six controlled tool configurations and paired-query settings. They further investigate the phenomenon through closed-book elicitation of intrinsic tool-use tendencies, behavioral analysis, and layerwise probing of model representations.

Based on these observations, the paper proposes EscapeGuard, an inference-time intervention that uses trained conflict probes to selectively enhance attention to the current tool definitions and argument schemas. Experiments across six benchmarks and six model backbones show improvements in tool-selection and argument-name hallucination, as well as cross-configuration robustness. The paper reports an average 9.0 percentage-point reduction in tool-selection hallucination and a 23.7 percentage-point reduction in the cross-configuration mean error rate.

Critical strengths

The paper identifies a meaningful limitation in evaluating tool-hallucination mitigation methods under fixed tool configurations. The cross-configuration and paired-query evaluations offer a useful perspective on whether localized improvements translate into broader reliability. The results in Table 1 demonstrate that several existing methods improve performance under the tuned configuration while worsening performance elsewhere. The additional TSHR-only analysis in Appendix D.4 helps address concerns about aggregating different error metrics. This evaluation perspective is a valuable contribution beyond the proposed mitigation method itself.

The paper connects behavioral observations, internal representation analysis, and inference-time intervention in a reasonably coherent framework. The conflict probes are supported by several control experiments, including random-label, query-only, and novelty-matched evaluations (Appendix G.4–G.5). EscapeGuard also achieves consistent improvements across the evaluated backbones and benchmarks. The component ablations and efficiency analysis provide additional evidence for the proposed design. In particular, the substantial improvements under tool-name replacement suggest that conditioning interventions on the current tool configuration can be effective.

Critical weaknesses

The observed deterioration in cross-configuration mean performance is relatively small for several baselines. For example, Table 1 reports increases of only 0.3 and 0.1 percentage points for PRISMS on LLaMA-3.1-8B and Qwen-3.5-9B, respectively. Although the experiments use multiple random seeds, the main results do not provide confidence intervals or significance analyses for these differences. This matters because the formal definition of hallucination escape requires the overall mean error rate to be no lower than that of the base model.

Furthermore, the configuration changes are systematically constructed, with a valid tool always available and relatively small tool sets (Appendix A). These controlled experiments are appropriate for isolating the phenomenon, but they do not fully establish how frequently or severely it occurs in realistic dynamic tool environments. The conclusions about the prevalence and generality of hallucination escape should therefore be qualified.

The behavioral correlations, probe results, and changes in prior strength provide evidence that intrinsic tool-use tendencies are associated with hallucination under configuration conflicts. However, they do not conclusively establish that reinforcement of these tendencies is the primary mechanism causing hallucination escape. The novelty-matched controls in Appendix G.5 address an important alternative explanation, but the analysis remains largely correlational. Moreover, the relationship between prior-strength changes and escape ratios in Table 3 is not consistently monotonic across methods.

The effectiveness of EscapeGuard is encouraging, but it does not independently validate the proposed causal mechanism. A configuration-aware attention intervention may improve tool selection for reasons other than suppressing the effects of reinforced intrinsic tendencies. The mechanistic claims would benefit from more cautious framing or more direct supporting evidence.

Questions for authors
  • Could the authors report confidence intervals or paired significance analyses for the cross-configuration mean differences in Table 1, particularly for methods whose deterioration is below 0.5 percentage points? This would help determine how consistently the methods satisfy Definition 3.1.

  • Could the authors quantify the relationship between prior-strength reinforcement and escape severity across methods and model backbones? In particular, how should the non-monotonic relationships in Table 3 be interpreted? What evidence supports treating tendency reinforcement as a causal contributor rather than a correlated indicator?

  • The definitions in Appendix D.1 explicitly measure fabricated argument names and incorrect argument values. Are missing required arguments, invalid argument types, and other schema violations also counted? Clarifying this would help assess how comprehensively the reported TUHR and OHR reflect invalid tool calls.

  • The reported Net Improvement is defined as the repair rate multiplied by one minus the transfer damage rate (Appendix D.3). Could the authors additionally report the overall change in correctness across both members of the query pairs, including initially correct queries? This would make the practical significance of the reported 89.1% result easier to interpret.

5 comments
AuthorsOpenReview ↗

A Big Thank You and Our Response to Reviewer svuL (1/2)

Thank you so much for the careful and insightful review! The comments have helped us make the paper noticeably stronger, and we have addressed each of them. All changes are marked in blue in the revised PDF. Our point-by-point responses are as follows.


To W1 and Q1: Thank you for raising this point. We have added seed standard deviations and paired bootstrap CIs in Appendix H.6 (new Table H.17), and would like to clarify how they bear on Definition 3.1 and on PRISMS:

  • CIs and Definition 3.1. We have added standard deviations over the three seeds and 95% paired bootstrap CIs over queries for the changes in Table 1 (Table H.17, Appendix H.6). Condition (3) of Definition 3.1 requires that a method does not lower H̄; all ten method–model pairs satisfy it in point estimate, and the CIs rule out any net reduction larger than 1.1 pp, whereas the same methods reduce C1 errors by 4.1–10.3 pp.
  • PRISMS. PRISMS is the weakest case but still meets all three conditions: it reduces C1 errors, increases errors on C2, C3, C5, and C6, and leaves H̄ no lower than Base (+0.3 / +0.1 pp), although this last increase is not significant. Its cross-configuration variance, however, rises significantly (+72.0 [+38, +108] and +79.5 [+44, +117]), so errors are redistributed rather than reduced. This weak escape is consistent with its having the smallest prior reinforcement (ΔS = +.05 / +.04).

Table H.17 (excerpt). Changes from Base in H̄ (pp) and Var (pp²), with 95% paired bootstrap CIs. Base H̄: 41.4 / 38.8; seed std of H̄ ≤ 0.4 pp. Full results in Table H.17 of the revised PDF.

MethodΔH̄ (LLaMA)ΔVar (LLaMA)ΔH̄ (Qwen)ΔVar (Qwen)
Relign+1.0 [−0.1, +2.1]+99.7 [+61, +141]+0.6 [−0.5, +1.7]+112.2 [+71, +156]
PALADIN+0.8 [−0.3, +1.9]+85.0 [+49, +124]+0.9 [−0.3, +2.1]+76.0 [+41, +114]
Gorilla+2.0 [+0.8, +3.2]+149.6 [+104, +198]+1.6 [+0.5, +2.7]+133.7 [+90, +180]
LinSteer+1.0 [−0.1, +2.1]+70.8 [+36, +107]+1.1 [−0.1, +2.3]+54.9 [+22, +90]
PRISMS+0.3 [−0.9, +1.5]+72.0 [+38, +108]+0.1 [−1.1, +1.3]+79.5 [+44, +117]

To W2: Thank you for this comment. We have clarified the scope of our claims in the abstract and §6, and added three evaluations beyond the original setting:

  • Scope. Configurations are controlled so that errors can be attributed to the change: a valid tool is always present, and at most 10 tools avoid context-length confounds (Appendix A.1). The abstract now states that escape is observed under controlled changes, and §6 notes that its frequency in multi-turn settings, dynamic registries, and larger tool libraries remains open.
  • Beyond this setting. We added evaluations on recent strong models (Table H.15), tasks with no matching tool (Table H.16), and interface changes (Table H.19). Escape persists on recent models, and EscapeGuard reduces errors in the latter two settings.
AuthorsOpenReview ↗

A Big Thank You and Our Response to Reviewer svuL (2/2)

To W3, W4, and Q2: Thank you for these suggestions. We have added correlation and intervention analyses.

  • Table 3. Table 3 now reports ΔH̄, matching the definition of escape (Definition 3.1). Ew = O(m)/I(m) divides by the C1 gain, so a small gain inflates it, which explains the non-monotonic cases (e.g., LinSteer on LLaMA has the smallest C1 gain).
  • Correlation. Across all ten method–model pairs, ΔS correlates with ΔH̄ (Spearman ρ = 0.90, p < 0.001).
  • Causality. We split LinSteer's steering vector on LLaMA into its prior-aligned component and the remainder, and apply each alone (Table F.9, Appendix F.3). Removing the prior component keeps most of the C1 gain (4.7 of 6.0 pp) but removes escape (ΔH̄: +1.0 → −0.5), while that component alone reproduces it (+0.9).
  • EscapeGuard. We agree that its effectiveness alone does not validate the mechanism and now describe it as consistent with it: EscapeGuard leaves prior strength nearly unchanged while increasing attention to the current configuration (Appendix H.7).

Table F.9. Decomposing LinSteer's steering vector (LLaMA-3.1-8B); changes from Base. O(m): total deterioration on other configurations (pp). NER with 95% pair-bootstrap CI.

Steering vectorΔSRet.ΔC₁ΔH̄O(m)RepairDamagecNINER [95% CI]
Original v+.0786−6.0+1.012.155.338.334.112.6 [0.0, 23.7]
v<sub>∥</sub> only (prior component)+.0687−2.8+0.98.325.835.016.8−2.9 [−14.8, 7.4]
v<sub>⊥</sub> only (prior removed)+.0194−4.7−0.51.743.312.338.023.3 [13.3, 33.3]

To Q3: Yes. Given the correct tool, an undefined argument name counts toward ANHR; any other schema violation (missing required argument, invalid type or structure) or a wrong value for a valid argument counts toward AVHR. TUHR, and hence OHR, therefore cover all invalid calls, as now stated in Appendix D.1.


To Q4: Thank you for this question. We have added full correctness transitions and a total-error metric (Table H.18).

  • Total errors. We added full correctness transitions over both members of all 120 pairs and the Net Error Reduction (NER) over all 240 queries (Table H.18, Appendix H.6). EscapeGuard reduces total errors by 85.4% and 87.1%, with only four new errors.
  • Metric. The original metric conditions on repaired targets; we renamed it Conditional Net Improvement (cNI), and the abstract now reports NER instead of 89.1%.

Table H.18 (excerpt). Correctness transitions of EscapeGuard over the 120 pairs (C: correct; W: wrong; Tgt: target; Rel: related; all methods in Table H.18 of the revised PDF).

ModelMethodTgt C→WTgt W→CRel C→WRel W→CErrors (before → after)NER (%) [95% CI]
LLaMA-3.1-8BRelign536239103 → 8616.5 [2.9, 28.8]
PRISMS533189103 → 8418.4 [6.2, 29.4]
EscapeGuard152340103 → 1585.4 [77.9, 91.8]
Qwen-3.5-9BRelign43021893 → 8014.0 [0.0, 25.7]
PRISMS42818893 → 7915.1 [3.5, 24.8]
EscapeGuard14833793 → 1287.1 [79.5, 93.6]

We hope these revisons address the concerns and would be grateful if the reviewer would consider updating the assessment. We would be happy to answer any further questions during the discussion period. Thank you again!

AuthorsOpenReview ↗

Gentle follow-up on our response

Dear Reviewer svuL,

Thank you so much for the time and care you have devoted to reviewing our paper. Your constructive comments have been very helpful in improving our work.

We understand that this is a very busy period, and we apologize for the additional message. As the discussion period is ending soon, we would be very grateful if you could let us know whether our response has addressed your concerns.

If anything remains unclear, we would be more than happy to provide further clarification. If you find that your concerns have been resolved, we would deeply appreciate it if you would kindly consider updating your assessment.

Thank you again for your time and consideration.

Best regards,

The Authors

Comments on the paper

3 comments
AuthorsOpenReview ↗

Summary of Revisions for All Reviewers and ACs

Dear Area Chairs and Reviewers,

We thank all reviewers for their careful and constructive reviews and for recognizing hallucination escape as a "valuable contribution beyond the proposed mitigation method itself" (svuL), "a practically relevant limitation" (EamH), and "a useful evaluation practice that most prior mitigation work skips" (SpYw).

We have posted point-by-point responses to each reviewer and uploaded a revised PDF with all changes marked in blue. As the discussion period is drawing to a close, we summarize here the main concerns raised, what we changed, and where the changes can be found, so that the ACs and reviewers can verify them at a glance.

1. Summary of revisions

Detailed responses can be found in our replies to the individual reviewers. W / Q refer to the weakness and question numbers in each review; "In PDF" gives the location in the revised PDF.

1.1 New experiments

#What is addedRaised inIn PDF
E1Four recent strong models under C1–C6, with a full method comparison on two of themSpYw W1, Q1Tables E.5, H.15
E2RoTTuning added as a baselineSpYw W2, Q2Tables 6, 7
E3Steering-vector decomposition (prior-aligned vs. residual)svuL W3, W4; SpYw W3Table F.9
E4Tasks with no matching toolSpYw Q3Table H.16
E5Functionally equivalent tools with interface changesEamH W2, Q2Table H.19

1.2 Analysis and clarifications

#What is added or changedRaised inIn PDF
A1Seed std and paired-bootstrap CIs for Table 1, and why ΔH̄ is significant only for GorillasvuL W1, Q1; SpYw Q4, follow-upTable H.17, App. H.6
A2Full correctness transitions and Net Error Reduction; metric renamed cNIEamH W1, Q1; svuL Q4Table H.18, App. D.3
A3Correlation between reinforcement and escape (ρ = 0.90); Table 3 now reports ΔH̄svuL W3, W4; SpYw W3Table 3, §4, App. H.7
A4Scope and metric coverage clarified; RoTBench and RobustBench-TC citedsvuL W2, Q3; EamH W2; SpYw W2§1, §2, §6, App. D.1

2. What is new in the revised PDF at a glance

  • New results: six new tables (E.5, F.9, H.15, H.16, H.18, H.19) and one new baseline (RoTTuning).
  • Statistical rigor: seed std and 95% paired-bootstrap CIs for every change in Table 1.
  • Causal evidence: steering-vector decomposition in addition to the ρ = 0.90 correlation.
  • Generality: four recent strong models, substantive interface changes, and no-matching-tool tasks.
  • Clarified metrics and claims: cNI / NER, Table 3 reports ΔH̄, scope qualified in the abstract and §6.

None of the changes alters the paper's core findings; they strengthen the evidence and clarify its scope.

We hope the revised paper addresses all raised concerns and would be sincerely grateful if the reviewers could consider our responses and the new results, and reassess the paper in light of the revisions. We remain available to answer any further questions during the remaining discussion period.

Thank you again for your time and for helping us improve the paper.

Best regards,

The Authors

AuthorsOpenReview ↗

Thank you and a gentle follow-up

Dear Reviewers and Area Chairs,

As the discussion period is drawing to a close, we would like to sincerely thank all of you for the time and effort you have devoted to our submission. Your thoughtful and constructive comments have helped us substantially improve the paper.

Following your suggestions, we have carefully addressed all the points raised in the reviews and the follow-up question and have updated the manuscript accordingly.

We understand that this is a very busy period for everyone. If time permits, we would be very grateful if you could take a brief look at our responses and let us know if you have any further questions or suggestions. If you find our responses satisfactory, we would sincerely appreciate it if you could kindly consider reflecting this in your final assessment. We will be happy to respond promptly in the remaining time.

Thank you again for your valuable contributions to our work.

Best regards,

The Authors

AuthorsOpenReview ↗

Closing summary for ACs, SACs, and PCs

Dear ACs, SACs, and PCs,

Thank you so much for handling our submission! As the discussion period closes, we would like to briefly summarize where each reviewer's concerns stand, in the hope that this is useful for your recommendation.

All three reviewers recognized the importance of the failure mode we identify, hallucination escape, describing it as "a meaningful limitation" and "a valuable contribution beyond the proposed mitigation method itself" (svuL), "a practically relevant limitation" (EamH), and "a useful evaluation practice that most prior mitigation work skips" (SpYw). The reviewers also noted the "consistent improvements" (svuL) and "substantial cross-configuration gains" (EamH) of our EscapeGuard. To our knowledge, this is the first work to show that mitigation methods shift errors across configurations rather than reduce them, and to trace this to reinforced tool-use tendencies.

The reviewers' remaining concerns were about evidence and scope, and we believe each of them is now addressed with new results in the revised PDF (changes marked in blue):

  • Reviewer svuL (statistical significance; correlational mechanism): we added seed standard deviations and 95% paired-bootstrap CIs (Table H.17), and a causal intervention showing that removing the prior-aligned component of the steering vector keeps most of the C1 gain but eliminates escape (Table F.9). Prior reinforcement also correlates with escape severity (Spearman ρ = 0.90).
  • Reviewer EamH (conditional metric; surface-form changes only): we renamed the metric and report full correctness transitions over all 240 queries, where EscapeGuard reduces total errors by 85.4% and 87.1% (Table H.18). On functionally equivalent tools with substantive interface changes, it lowers OHR from 48.4% to 28.3% and from 44.6% to 25.5% (Table H.19).
  • Reviewer SpYw (recent strong models; baseline; causal evidence): on four recent strong models, errors rise from 3.9–8.6% under C1 to 19.6–31.4% under C2, and existing methods still exhibit escape on the two models with a full comparison (Tables E.5, H.15). We now discuss RoTBench and RobustBench-TC and add RoTTuning as a baseline, which EscapeGuard outperforms under every configuration (Table 7). The correlation and causal intervention address the concern about the mechanism (Table F.9). EscapeGuard also lowers the wrong-call rate when no tool matches (Table H.16).

None of these revisions changes the paper's core findings. No further objection has been raised to our responses, but despite 4 follow-up messages, we have not received any further feedback from the reviewers. We fully understand that this is a very busy period for everyone. Since the current scores (3, 2, 2) were given before the rebuttal, they do not yet reflect our responses and the new results, which resolve each of the concerns and questions raised by the reviewers. We would therefore be very grateful if you could kindly take these into account in your recommendation.

Thank you so much again for your time and effort on our submission. We truly appreciate it!

Best regards,

The Authors

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

Related papers

Discussion (0)

Sign in to comment.