- Rating
- 3
- Confidence
- 3
- Informativeness
- 3
- Presentation
- 3
- Soundness
- 3
Read review
Paper summary
This paper studies whether language models can use helpful external workflows while overriding misleading guidance. It introduces Box²-Bench, which holds the model, task, environment, and evaluator fixed across five conditions: no, good, bad, partial, and mixed workflows. Matched comparisons separate useful-guidance utilization, robustness to bad guidance, and recovery after guidance stops or becomes misleading. Experiments show that good workflows often help while bad workflows hurt. Training two open-weight models on bad-workflow inputs improves robustness, but SFT can also suppress useful-guidance utilization; outcome-based RL partly restores it. Peer-information and memory experiments provide preliminary evidence of transfer beyond workflows.
Critical strengths
The strongest aspect is the controlled evaluation design. Varying workflow quality while keeping the task and model fixed helps distinguish ordinary task failure from failure caused by misleading guidance. The Partial and Mixed conditions are especially useful because they share a helpful prefix and differ in whether guidance stops or continues incorrectly. Separating utilization from robustness also captures a practical distinction that raw accuracy misses: a model can benefit substantially from a good workflow and still be highly vulnerable to a bad one.
The training analysis identifies a meaningful failure mode of bad-workflow training. A model can improve by selectively overriding incorrect steps or by ignoring workflows altogether, and the held-out good-workflow condition distinguishes these behaviors. The AIME result, where SFT improves robustness but makes useful-guidance utilization negative before RL restores positive utilization, illustrates the trade-off clearly. The transfer probes also examine behavior beyond final accuracy; in particular, archive verification under corrupted memory offers a useful indication of how the trained model handles fallible context.
Critical weaknesses
The benchmark uses a specific workflow-error distribution produced by a fixed external generator and filtered by verification. This gives experimental control, but it leaves open whether good and bad workflows differ in surface style or structure in ways that models can exploit. Such cues could partly explain apparent selective reliance without requiring an assessment of guidance reliability. Evaluations with other generators, human-written workflows, or style-matched errors would help resolve this concern. The verified conditions also represent cleaner reliability interventions than the ambiguous, stale, or partially correct guidance encountered in deployment, so the conclusions should remain scoped to the controlled setting.
The statistical and training evidence is limited. Some evaluations use 30-problem subsets, making point estimates sensitive to a few examples, yet uncertainty reporting is sparse. Learning is demonstrated with Qwen3-4B on AIME 2026 and Qwen3.5-9B on WebShop, which supports a proof of concept more readily than a general training recipe. Cross-context transfer is also mixed: multi-agent episode success rises from 18.00% to 22.67% with SFT but reaches only 18.67% with SFT+RL, and the memory probe uses a 30-question LongMemEval-V2-Small subset with a modified archive protocol. These are informative probes, but broader claims need replication across models, tasks, and training seeds.
Questions for authors
Q1. How sensitive are the benchmark effects to the workflow generator and to surface differences between good and bad guidance? Could you evaluate a subset using another planner or human-written workflows, together with a style-matching control or a classifier that predicts workflow labels from surface form? Please include confidence intervals or paired uncertainty estimates for the utilization, robustness, and Partial-versus-Mixed effects, since small subsets can make these comparisons unstable.
Q2. Does the SFT-to-RL trade-off persist across model sizes, domains, and training seeds, and what explains the lower multi-agent success of SFT+RL relative to SFT alone? Trajectory-level diagnostics could help distinguish successful recalibration from renewed harmful reliance. Given the size and specialized protocols of the peer and memory experiments, could you clarify which cross-context claims are supported by the current evidence and frame the remaining results as preliminary transfer probes?
4 comments
Rebuttal (continued)
W2: SFT-to-RL trade-off and cross-context transfer.
Response: We address two aspects of this concern. First, we clarify the SFT-to-RL trade-off with additional empirical evidence from AIME and WebShop. Second, we provide additional behavioral diagnostics in multi-agent and memory settings to examine cross-context transfer.
Reason of SFT-to-RL trade-off:
- (1) Theoretical evidence. SFT trains the model to produce a prescribed correct trajectory given a misleading workflow (Eq. 4). RL, in contrast, optimizes task success without prescribing a target trajectory (Eq. 5). Through task success feedback, model may learn to selectively identify and use information from the workflow. This ability also useful for Good workflows. Similarly, using workflow information may introduce a trade-off: it can help the model benefit from useful guidance while increasing its exposure to misleading steps.
- (2) Empirical evidence. We note that SFT+RL consistently outperforms SFT across all workflow conditions on WebShop (Fig3 in paper). On AIME 2025 and AIME 2026 with the GPT-6.1-Sol generator, we also observe improvements in mean accuracy on Bad and Mixed workflows (Tab1-2 in W1). These results suggest that the model can learn to identify and use useful workflow information. However, this remains a challenging capability. Performance may vary depending on how misleading the workflow is and how reliably the model can identify relevant information and use it correctly.
Robustness of cross-context transfer:
To further assess the robustness of our cross-context transfer results, we conduct additional evaluations with five random seeds and report mean ± SD.
Multi-agent transfer (EoM). Episode success rates remain relatively stable across training stages. However, under BAD conditions, SFT+RL achieves more repairs (7.20 ± 2.68 vs. 4.80 ± 3.35) and fewer corruptions (0.80 ± 1.10 vs. 3.20 ± 1.79) than SFT. This reveals a behavioral change in how agents handle misleading peer information, even when improvements in overall episode success are not consistent. Due to rebuttal budget, we use 10-problem subsets with five seeds.
Memory transfer. We observe a clearer change in verification behavior. After SFT+RL, archive-tool usage is higher under BAD than GOOD memory (72.67% vs. 66.67%), reversing the trend in the base model (64.67% vs. 73.33%). This suggests that the model more actively verifies potentially misleading memories.
These results provide additional evidence of behavioral transfer beyond workflow tasks. Given the rebuttal compute budget, rerunning the main experiments across independent training seeds was not feasible. We will add additional experiments to further consolidate the evidence in next version.
Multi-agent transfer (EoM):
| Condition | Model | Pass Rate | Repair | Corruption |
|---|---|---|---|---|
| RAW | Base | 18.80 ± 2.28 | 2.00 ± 1.41 | 1.60 ± 1.67 |
| RAW | SFT | 20.00 ± 2.45 | 6.40 ± 2.61 | 0.80 ± 1.79 |
| RAW | SFT+RL | 19.60 ± 2.19 | 3.20 ± 1.79 | 2.80 ± 1.79 |
| BAD | Base | 17.60 ± 2.19 | 2.00 ± 1.41 | 2.00 ± 1.41 |
| BAD | SFT | 19.60 ± 2.61 | 4.80 ± 3.35 | 3.20 ± 1.79 |
| BAD | SFT+RL | 17.60 ± 2.61 | 7.20 ± 2.68 | 0.80 ± 1.10 |
Memory transfer:
| Model | RAW Accuracy | RAW Tool Rate | GOOD Accuracy | GOOD Tool Rate | BAD Accuracy | BAD Tool Rate |
|---|---|---|---|---|---|---|
| Base | 24.00 ± 7.96 | 100.00 ± 0.00 | 94.67 ± 1.83 | 73.33 ± 6.24 | 10.67 ± 1.49 | 64.67 ± 9.89 |
| SFT | 27.33 ± 2.79 | 100.00 ± 0.00 | 92.67 ± 3.65 | 72.67 ± 7.23 | 12.00 ± 6.06 | 71.33 ± 6.91 |
| SFT+RL | 27.33 ± 4.94 | 100.00 ± 0.00 | 92.00 ± 4.47 | 66.67 ± 8.50 | 12.00 ± 2.98 | 72.67 ± 4.35 |
Thanks authors for the detailed response. They have largely address my concerns on my side. Thanks.
Thank you again for your time and thoughtful engagement with our submission! Your valuable suggestions have greatly helped us improve the quality of our work!
Rebuttal
Thank you for the encouraging feedback. We appreciate your recognition of the paper's
controlled evaluation design,training analysis, andtransfer studies. We address your concerns point by point.Response: We address this concern from two aspects: workflow generator robustness, controlled comparisons of workflow content.
(1) Sensitivity to workflow generators. To test whether our findings depend on the workflow generator, we conduct additional experiments using GPT-6.1-Sol to generate workflows and evaluate them on both AIME 2025 and AIME 2026. We report mean accuracy and standard deviation over eight rollouts.
Across both datasets, we observe a consistent pattern with our original results: SFT improves Bad-workflow performance while reducing Good-workflow performance, and subsequent RL further improves Bad and Mixed performance while maintaining comparable Good-workflow performance. These results suggest that our findings are not specific to the original workflow generator. The results are in Table 5 of the Appendix B.1 (Ablations of Workflow Generator) in the updated PDF.
AIME 2025 with GPT-6.1-Sol generator :
AIME 2026 with GPT-6.1-Sol generator :
(2) Surface-level confounds. To probe this, Box²-Bench includes Partial and Mixed conditions, which share the same helpful prefix but differ in whether misleading steps are appended. Thus, Δswitch=SM−SP measures the impact of misleading information while controlling for the helpful content, providing a finer-grained evaluation of selective workflow use.