- 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
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 vector | ΔS | Ret. | ΔC₁ | ΔH̄ | O(m) | Repair | Damage | cNI | NER [95% CI] |
|---|---|---|---|---|---|---|---|---|---|
| Original v | +.07 | 86 | −6.0 | +1.0 | 12.1 | 55.3 | 38.3 | 34.1 | 12.6 [0.0, 23.7] |
| v<sub>∥</sub> only (prior component) | +.06 | 87 | −2.8 | +0.9 | 8.3 | 25.8 | 35.0 | 16.8 | −2.9 [−14.8, 7.4] |
| v<sub>⊥</sub> only (prior removed) | +.01 | 94 | −4.7 | −0.5 | 1.7 | 43.3 | 12.3 | 38.0 | 23.3 [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.
| Model | Base | EscapeGuard | Gate fires | Gate on: Base → EG | Gate off: Base = EG |
|---|---|---|---|---|---|
| LLaMA-3.1-8B | 41.3 | 29.6 | 58.0% | 47.9 → 27.8 | 32.2 |
| Qwen-3.5-9B | 37.4 | 26.3 | 56.1% | 44.6 → 24.8 | 28.2 |
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).
| Model | Method | H̄ ± std | ΔH̄ [95% CI] |
|---|---|---|---|
| LLaMA-3.1-8B | Base | 41.4 ± 0.4 | — |
| Relign | 42.4 ± 0.4 | +1.0 [−0.1, +2.1] | |
| PALADIN | 42.2 ± 0.4 | +0.8 [−0.3, +1.9] | |
| Gorilla | 43.4 ± 0.4 | +2.0 [+0.8, +3.2] | |
| LinSteer | 42.4 ± 0.4 | +1.0 [−0.1, +2.1] | |
| PRISMS | 41.7 ± 0.3 | +0.3 [−0.9, +1.5] | |
| Qwen-3.5-9B | Base | 38.8 ± 0.4 | — |
| Relign | 39.4 ± 0.4 | +0.6 [−0.5, +1.7] | |
| PALADIN | 39.7 ± 0.4 | +0.9 [−0.3, +2.1] | |
| Gorilla | 40.4 ± 0.4 | +1.6 [+0.5, +2.7] | |
| LinSteer | 39.9 ± 0.4 | +1.1 [−0.1, +2.3] | |
| PRISMS | 38.9 ± 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?
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:
Table E.5. Hallucination rates (%) of base models.
Table H.15. Hallucination rates (%); parentheses: change from Base.
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).
Table 7 (excerpt). Hallucination rates (%); parentheses: change from Base (full C1–C6 results in Table 7 of the revised PDF).