acceptodds
Under review as a conference paper at ICLR 2027

One Task, Many Interfaces: Simple Cross-Interface Consistency Training Improves LLM Agent Robustness

Abstract

Post-training binds the competence of an LLM agent to the interface it was trained on. Semantics-preserving perturbations of that interface, such as renaming an action or changing the notation, leave the underlying task unchanged yet make the agent worse at it. Scoring each sample independently means interface randomization only broadens this dependence from a single interface to the training interface distribution, so out-of-distribution perturbations still degrade the agent. We propose __Cross-Interface Policy Regularization (CIPR)__, a self-supervised objective that replaces action names with uninformative identifiers and presents each training sample under two interfaces in which equivalent actions are identified by construction. CIPR penalizes the KL divergence between the action distributions induced under the two interfaces, so the policy behaves consistently across them. We evaluate CIPR in both supervised and reinforcement-learning post-training, where it improves robustness. On agentic function calling, held-out interface perturbations reduce the accuracy of a post-trained agent by nearly 25%, and CIPR recovers 10.5% while leaving accuracy on the unchanged interface intact.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Agentic, LLM, Post-training, robust

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_8HRaOpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the interface sensitivity of LLM agents that use tools, where semantically equivalent modifications to tool names, argument names, tool ordering, or the inclusion of similar tools can markedly harm task performance. The authors argue that standard post-training and multi-interface/domain-randomization methods optimize task performance only independently for each interface, and therefore do not explicitly promote equivalent policies for equivalent interfaces. They propose Cross-Interface Policy Regularization (CIPR), which canonicalizes tool and argument names, builds a paired perturbed interface with known tool correspondences, and regularizes the KL divergence between action distributions over equivalent tools. CIPR consists of a rescoring consistency term, which compares relabeled versions of the same response, and a rollout consistency term, which compares the average action distributions from independently sampled rollouts under the two interfaces. Experiments using GRPO post-training on BFCL with Qwen2.5-3B and Qwen3.5-4B show that CIPR generally improves robustness to held-out interface perturbations while preserving accuracy on the original interface, outperforming multi-interface training by a modest margin. A supervised self-distillation experiment on WebShop further indicates that CIPR can improve robustness to renamed actions.

Critical strengths

1.The paper addresses a real failure mode of tool-using agents: reliance on tool names, positions, and other interface-specific shortcuts. Robustness to changing tool registries and schemas is important for deploying agents in dynamic environments.

2.The distinction between independently scoring multiple interfaces and explicitly regularizing equivalent interfaces is well motivated. Comparing distributions over correspondence-matched actions is more appropriate than directly comparing output tokens when tool names change.

3.The paper evaluates both reinforcement-learning and supervised post-training, uses two different model backbones, tests multiple held-out perturbation types, and includes WebShop, τ²-bench, RoTBench, ablations, and error analysis.

Critical weaknesses

1.The method depends on a strong inference-time canonicalization assumption. CIPR does not make the policy fully robust to arbitrary names by itself. The deployment system must canonicalize every interface before the model sees it. This limits the claim from general interface robustness to robustness under a specific wrapper or adapter architecture, which may not be available in realistic third-party tool ecosystems.

2.Much of the improvement appears to come from canonicalization, training on perturbed interfaces, and especially degraded twins. On BFCL, CIPR improves over multi-interface training by roughly 1 point on Qwen3.5-4B and 1.9 points on Qwen2.5-3B in the aggregate held-out score, with several per-condition differences within the reported confidence intervals.

3.The empirical evidence is not consistently favorable across robustness benchmarks. On RoTBench, CIPR is not uniformly better than multi-interface training and can be substantially worse in the canonical condition for Qwen3.5-4B. This raises questions about whether the proposed regularizer generalizes beyond the specific perturbation family used to construct the training interfaces.

4.The experiments use models up to 9B parameters, mostly single-turn function calling, and one training run per configuration. The rollout consistency term is computationally expensive, but there is no systematic analysis of its cost-performance tradeoff, scaling behavior, or robustness across random seeds and larger multi-turn environments.

Questions for authors

1.How much of the reported gain is attributable specifically to the KL regularization rather than canonicalization, twin expansion, and interface randomization?

2.Can the authors evaluate CIPR without an inference-time canonicalization adapter, or explain how the adapter can be reliably deployed when tools are dynamically registered by external providers? In particular, how should the system handle new tools for which canonical names and argument mappings are not known in advance?

3.What is the computational overhead of the rollout consistency term in realistic training?

2 comments

We thank the reviewer for their feedback. In the following, we address their questions and concerns.

On canonicalization and newly registered tools.

The wrapper is a deterministic string mapping that needs no prior knowledge of the tools. For each request, it replaces the names of the available functions and their arguments, so a newly registered lookup_order(order_id) becomes function_01(arg_01) while its description and the argument values stay unchanged, and it translates the model's call back before execution. The provider's API is not modified, and the mapping costs negligible time compared with LLM inference. Correspondences between functions are needed only to construct training samples. Behind the wrapper, CIPR makes the model independent of tool identifiers, tool order and added tools. We keep the original names out because they can still bias the model. When the two WebShop actions are swapped, so that each name suggests the other action, reward falls to zero for both the released and the trained agent (Figure 3).

On the contribution of KL versus the training perturbations.

Both the training perturbations and the KL penalties are our contributions. We rename and reorder the functions and add a degraded copy of every function, such as one marked as deprecated, so the presence of a duplicate cannot identify the answer. The multi-interface baseline removes only the KL penalties, so it measures what the KL adds to our other contributions. The complete recipe raises mean accuracy under perturbation over standard post-training by 15.5 and 4.3 points, and by 6.6 and 7.3 points when standard post-training is evaluated with canonical names (Table 1). The intervals in Table 1 are at most $\pm$2.0 points, so any difference between two models above 2.8 points is significant even without pairing. The gain from the KL itself is consistent, as CIPR is more accurate than multi-interface training on 34 of the 38 held-out types across the two backbones (Tables 6 and 7). In Table 2, removing the KL lowers accuracy on explicit cost and latency by 1.9 points and removing the degraded copies by 5.5. Our AgentLM experiments is trained only by self-distillation with the CIPR KL terms and is evaluated with canonical names before and after training, and its WebShop reward rises from 75.7 to 81.0. The gain therefore comes from neither the wrapper nor further supervised fine-tuning.

On whether the gains require duplicating every tool (new experiment).

DR-Full [1] is trained with RL on interfaces randomized over the sixteen static perturbation types introduced in [1], which include the typo, distractor and explicit cost and latency perturbations of Table 1. DR-Full-CIPR adds the CIPR KL terms between the original and the perturbed interface of each DR-Full training sample, instead of our degraded tool interface used in Table 1. Over DR-Full, mean BFCL accuracy under perturbation rises from 57.2% to 69.4% with canonical names for both models, while accuracy on the original interface stays comparable (73.2% and 74.2%), so the drop under perturbation falls from 16.0 to 4.8 points (Table in the common rebuttal response). The largest gain is on the held-out implicit cost and latency perturbations, where accuracy rises from 46.1% to 67.2%. CIPR therefore improves robustness without our multiple interfaces, both on perturbation types seen in training and on a held-out one.

On generalization across robustness benchmarks.

Table 1 already tests generalization beyond the training perturbations, on thirteen perturbation types from RobustBench-TC [1] that are not used in training (Tables 6 and 7 give the results per type). τ2\tau^2-bench is an out of distribution benchmark that checks that robustness training does not reduce performance, and it does not. Each entry has a standard error of about $\pm$6 to 7 points on airline and $\pm$2 to 6 on retail, and on the original interface no difference between two models of the same backbone is statistically significant. On RoTBench with canonical names (Table 10), accuracy over the corrupted levels rises from 33.7 for standard post-training to 36.3 for CIPR on the 3B model, and from 31.3 to 39.7 on the 4B model.

On model scale, multi-turn tasks and training seeds.

On the multi-turn τ2\tau^2-bench, robustness training does not reduce performance, as shown above. The per type gains of the KL above hold on two independently trained backbones.

On training overhead.

Compared with multi-interface training, neither KL term requires additional generation. The rollout term compares responses that multi-interface training already generates under the original and the perturbed tool list, and the rescoring term reuses an existing response with the function names changed throughout. The only added cost is computing the probabilities and their gradients on these responses.

Reviewer_8HRaOpenReview ↗

Thanks for response.

The response solve my most concern, and I will increase the score to 2.

Reviewer_Yma8OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Cross-Interface Policy Regularization (CIPR), a self-supervised objective that trains the agent to act consistently across interfaces offering the same tools, by applying KL regularization between canonical and perturbed interfaces. Experiments on function calling and WebShop demonstrate improved robustness compared to standard post-training and multi-interface training.

Critical strengths
  1. Simple and intuitive approach: CIPR introduces a simple consistency regularization objective that can be incorporated into existing post-training methods without additional annotations.
  2. Evaluation across different settings: The authors evaluate CIPR in both RL and supervised post-training settings, using multiple models and interface perturbations.
Critical weaknesses
  1. Limited diversity of interface differences: In practice, even within similar domains, different interfaces may have different prompt formats, response formats, and other characteristics. Thus, multi-interface training could consider more diverse interfaces, such as those of WebShop and WebArena [1]. However, the interface perturbations considered in this work are limited to modifications of the same underlying tool set, such as renaming, reordering, and adding degraded tools.
  2. Limited improvement: The multi-interface baseline also exhibits high performance comparable to that of the proposed method. In Table 1, CIPR improves the mean accuracy over multi-interface training by only 1.9 and 1.0 percentage points on Qwen2.5-3B and Qwen3.5-4B, respectively. Moreover, Table 8 shows that CIPR often underperforms standard post-training and multi-interface training in interactive environments. Therefore, the effectiveness of CIPR remains unclear compared to data augmentation alone.
  3. Limited baselines: The supervised post-training experiments lack a multi-interface training baseline, making it difficult to isolate the benefit of consistency regularization from exposure to relabeled interfaces. Including this baseline would strengthen the evidence for CIPR's effectiveness in supervised post-training.

[1] Zhou, Shuyan, et al. "Webarena: A realistic web environment for building autonomous agents." International Conference on Learning Representations. Vol. 2024. 2024.

Questions for authors
  1. Beyond robustness to tool names and menu structures, what aspects of an interface should an agent learn to generalize across? For example, would CIPR remain effective when interfaces differ in their interaction protocols, argument semantics, or response formats? While it is important for an agent to select appropriate functions based on the situation rather than relying on familiar function names, it would be interesting to discuss what additional factors should be considered to achieve more meaningful gains from multi-interface training.
  2. Data augmentation generally benefits from more diverse training data, but collecting or generating a large amount of data can be costly. How much diversity in interface perturbations does CIPR require to achieve strong performance? To investigate this, the authors could compare naive data augmentation (multi-interface training in this paper) and CIPR with varying numbers of perturbed training examples. If CIPR substantially outperforms data augmentation in low-data settings, this would provide stronger evidence that CIPR is a more data-efficient approach.
1 comment

We thank the reviewer for their feedback. In the following, we address their questions and concerns.

On interface diversity and scope.

Our aim is to train the model to choose a function by its semantics rather than by its presentation interface. In training, we rename and reorder the functions and add a degraded copy of every function, such as one marked as deprecated or one that lacks a required argument, so the presence of a duplicate cannot identify the answer. The training perturbations are narrow by design, but the evaluation is broad. Table 1 evaluates on thirteen perturbation types, none used in training, from RobustBench-TC [1], an independent benchmark built from documented tool calling failures. They change the query (typos), the available tools (a similarly named tool and five kinds of duplicates with the same name but corrupted descriptions or schemas) and tool metadata (six kinds of cost and latency annotations where the familiar name points to the costlier or slower tool). We evaluate the last group both in the explicit form of [1], where the query asks for the cheaper or faster tool, and in an implicit form, where the query states no preference. Each column of Table 1 averages one group (7, 6 and 6 types), and Tables 6 and 7 give the results per type. CIPR also applies to different prompt and response formats, as long as corresponding actions can be identified. If argument meanings or interaction rules change, we first need to establish which actions still accomplish the same task.

On the small gains over multi-interface training.

The multi-interface baseline in Table 1 is our training recipe without the KL penalties, so it already uses our canonical names and perturbed tool lists. The gains of 1.9 and 1.0 points therefore measure what the KL adds to our other contributions, not the gain over an existing method. The complete recipe raises mean accuracy under perturbation over standard post-training by 15.5 and 4.3 points while keeping accuracy on the original interface, and by 6.6 and 7.3 points when standard post-training is evaluated with canonical names (Table 1). These gains are statistically significant, since the intervals in Table 1 are at most $\pm$2.0 points and any difference above 2.8 points is significant even without pairing. The gain from the KL itself is consistent, as CIPR is more accurate than multi-interface training on 34 of the 38 held-out types across the two backbones (Tables 6 and 7) and in every held-out group of Table 1.

On whether more diverse training interfaces are enough (new experiment).

DR-Full [1] is trained with RL on interfaces randomized over the sixteen static perturbation types introduced in [1], which include the typo, distractor and explicit cost and latency perturbations of Table 1. DR-Full-CIPR adds the CIPR KL terms between the original and the perturbed interface of each DR-Full training sample, instead of our degraded tool interface used in Table 1. On BFCL, mean accuracy under perturbation rises from 57.2% to 69.4% over DR-Full, with canonical names for both models (Table 1 in the common rebuttal). Accuracy on the original interface stays comparable (73.2% and 74.2%), so the drop under perturbation shrinks from 16.0 to 4.8 points. The largest gain is on the held-out implicit cost and latency perturbations, where accuracy rises from 46.1% to 67.2%. CIPR therefore improves over training on diverse interfaces alone.

On the results in interactive environments (Table 8).

Table 8 evaluates our RL models on τ2\tau^2-bench, an out of distribution benchmark, to check that robustness training does not reduce performance there. It does not. With 50 airline and 74 to 93 retail simulations, each entry has a standard error of about $\pm$6 to 7 points on airline and $\pm$2 to 6 on retail, and on the original interface no difference between two models of the same backbone is statistically significant.

On supervised experiment

The AgentLM experiment is self-distillation with the CIPR KL terms, and no additional supervised fine-tuning is performed. The released model is the teacher, the student matches it under canonical names, and relabeled interfaces enter only through the consistency KL, so the recipe is not supervised training on multiple interfaces plus the KL, and it generates no new trajectories. WebShop reward rises from 75.7 to 81.0 with canonical names at inference, and from 5.5 to 76.0 when the actions are renamed with synonyms (Figure 3). The RL experiments compare training with and without the KL on the same perturbed interfaces.

Reviewer_SuetOpenReview ↗
Rating
3
Confidence
2
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

This paper tackles the problem of AI agents not picking the correct tools to use when the interface to invoke those tools changes. The state-of-the-art solution for this is 'multi-interface training' which randomises tool interfaces during training. The proposed solution adds a KL divergence (between a set of perturbed tools and the set of tool choices) to the objective function of the post training. Evaluation was done on Qwen2.5-3B-Instruct/Qwen3.5-4B with BFCL using GRPO, and AgentLM-7B with WebShop using self-distillation. The authors' results show improved performance on most perturbations outside of the training set, comparing to the baseline multi- interface training. The strength of the new method is its generality which makes it applicable to many tasks, but the authors also correctly discuss the downside of increased training cost.

Critical strengths

The self-supervised objective does not require additional labels and can be generalised to different training paradigms and task families without modification. The method is simple and effective within the experiment settings.

Critical weaknesses
  1. It is not clear if the proposed consistency KL penalty terms will be necessary for larger models (as shown with the Qwen3.5-4B model), or effective in long multi-turn tasks where the KL of two action distributions may not regularise the tool-use behaviour well enough.
  2. The canonicalization and perturbed interface capture only a subset of possible tool interface changes.
Questions for authors

Q1: Do rescoring consistency term and rollout consistency term always agree with each other?

Q2: How to choose hyper-parameters lambda_1 and lambda_2? How sensitive are they?

1 comment

We thank the reviewer for their feedback. In the following, we address their questions and concerns.

On larger models and multi-turn tasks.

High task accuracy does not remove interface sensitivity. With Qwen3.5-4B, standard post-training reaches 84.8% on the original interface but 60.2% when the tool list offers a cheaper or faster alternative and the query does not ask for it, while CIPR reaches 85.0% and 73.0%. The KL still helps at this scale. CIPR is more accurate than multi-interface training, which differs from it only in the KL, on 16 of the 19 held-out perturbation types (Table 7) and in every held-out group of Table 1. For multi-turn interaction, Table 8 evaluates the RL models on τ2\tau^2-bench, an out of distribution benchmark where the agent works with a simulated user. Robustness training does not reduce performance there, since no difference between two models of the same backbone is statistically significant on the original interface.

On the restricted training perturbations.

We designed the training perturbations so that the model has to read what each function does. We change the names and the order of the functions and add a degraded copy of every function, such as one marked as deprecated or one that lacks a required argument, so the presence of a duplicate cannot reveal the answer. The KL penalty asks the model to assign the same probabilities to corresponding choices across these perturbations, which together with task training discourages choices based on names, positions or duplicates. Table 1 shows improved robustness on thirteen perturbation types from RobustBench-TC [1], none of which is used in training. Each column averages one group of these types, and Tables 6 and 7 give the results per type.

CIPR also works with a broader set of training perturbations. DR-Full [1] is trained with RL on interfaces randomized over the sixteen static perturbation types introduced in [1]. Adding the CIPR KL terms between the original and the perturbed interface of each DR-Full training sample raises mean BFCL accuracy under perturbation from 57.2% to 69.4% over DR-Full (Table 1 in the common rebuttal). On the held-out implicit cost and latency perturbations, accuracy rises from 46.1% to 67.2%.

On whether the two KL terms agree.

They need not agree. The rescoring term holds the model's reasoning fixed and renames the functions consistently in the prompt and the response. The rollout term lets the model generate its own reasoning under each interface. A model can select the right function when it is given the translated reasoning and still overlook that function when it reasons from scratch. A low rescoring term therefore does not make the rollout term redundant.

On coefficient choice and sensitivity.

We chose λ1=0.02\lambda_1=0.02 and λ2=0.005\lambda_2=0.005 by grid search on held-out evaluation data points and use the same values on both main backbones (Appendix A, Table 3). Training is sensitive to large values. If the coefficients are too large, training collapses to never calling a tool, since a model that never calls a tool makes the same choice under every interface and minimizes the KL terms.

Comments on the paper

1 comment

We thank the reviewers for their feedback. Below we address the main concerns raised in the reviews, and we answer the remaining questions and concerns in the reply to each reviewer.

On canonicalization and newly registered tools. The wrapper is a deterministic string mapping. For each request, it replaces the function and argument names of the available tools with neutral identifiers, such as function_01(arg_01), and maps the model's call back before execution. It needs no prior knowledge of the tools, so a newly registered tool is renamed in the first request that includes it. It leaves the provider's API unchanged, adds no model calls, and costs negligible time compared with inference.

On whether CIPR gains mostly depend on our multiple interfaces (new experiment). In the paper, CIPR is more accurate than our multi-interface training, which differs from it only in the KL, on 34 of the 38 held-out perturbation types across the two backbones (Tables 6 and 7 in the appendix) and in every held-out group of Table 1.

To test this further, we include another experiment using DR-Full [1], a model trained with RL on interfaces randomized over the 16 static perturbation types introduced in [1]. These include the typo, distractor and explicit cost and latency perturbations of Table 1, so only the implicit cost and latency perturbations are held out. DR-Full-CIPR is trained on the same data and adds the CIPR KL terms between the original and the perturbed interface of each training sample, using the DR-Full perturbations and not our degraded tool interface used in Table 1. We evaluate on BFCL as in Table 1, with all 2,501 items for each perturbation type. Both models are evaluated with original and with canonical names, where a wrapper replaces function and argument names with neutral identifiers before the model sees them and translates the generated call back afterward.

With canonical names, mean accuracy under perturbation rises from 57.2% to 69.4% over DR-Full, while accuracy on the original interface stays comparable (73.2% and 74.2%), so the drop under perturbation shrinks from 16.0 to 4.8 points. The largest gain is on the held-out implicit cost and latency perturbations, where accuracy rises from 46.1% to 67.2%. With original names, mean accuracy rises from 39.7% to 53.7%. CIPR therefore improves over training on diverse interfaces alone, and its gains do not depend on our multiple interfaces. We answer the remaining questions and concerns in the reply to each reviewer.

ModelTool NamesOriginal interfaceTypos / distractorsExplicit cost / latencyImplicit cost / latencyMean
DR-Fulloriginal76.173.825.819.439.7
DR-Fullcanonical73.270.555.046.157.2
DR-Full-CIPRoriginal76.374.843.443.053.7
DR-Full-CIPRcanonical74.270.970.167.269.4

References

[1] X. Zhou et al. AgentRobustBench: A Sim-to-Real Benchmark and Domain-Randomized RL Recipe for Tool-Use Agents. arXiv:2605.11928v3, 2026.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.