acceptodds
Desk rejected at ICLR 2027

DGCR: Learning from Local Corrections for Super-Tiny Function Calling Models

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

Abstract

Reliable function calling makes super-tiny models promising for efficient tool-based automation. However, local errors can invalidate otherwise useful calls, while unsuccessful rollout groups may be discarded without contributing to learning. Dense rewards give partial credit for correct decisions but do not specify how to repair the remaining errors. We introduce Discarded-Group Counterfactual Repair (DGCR) to recover explicit corrective targets from these rollouts. For collector-rejected groups that satisfy a low-reward condition, DGCR uses training references to propose bounded structural edits and verifies their gains under ordinary and fixed call assignments with consistency checks. It then applies gain-weighted supervision to the repaired spans under student-generated prefixes, without assuming that the complete response is correct. On Qwen3-0.6B, DGCR outperforms strong prior methods on both BFCL v3 and ACEBench, suggesting that even discarded rollouts can provide useful learning signals for super-tiny function calling models.

Primary area:
reinforcement learning
Keywords:
function calling, tool use, reinforcement learning, counterfactual repair, selective supervision, super-tiny function calling models

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_X2XQOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

In the context of ultra-small functional models, localized errors can cause an entire function call to fail. A model might correctly select the intended function and provide most of the required arguments, yet still fail due to a single incorrect value, a missing parameter, or the inclusion of extraneous fields. Existing approaches—such as the CKD+Sim-RL framework—attempt to address this by employing a "similarity reward" mechanism. While this rewards partially correct decisions, it fails to provide specific guidance for correcting the erroneous components. Furthermore, generation groups rejected by the data collector are typically discarded and rendered unusable for reinforcement learning; however, this paper argues that such rejected groups still hold significant value. To leverage these discarded groups, the authors propose "Discarded-Group Counterfactual Repair" (DGCR). Building upon the CKD strategy from the STAR framework—while retaining the Sim-RL reward—the authors identify that existing methods cannot pinpoint which specific fields are incorrect. DGCR addresses this by utilizing rejected generation groups that meet certain criteria: it matches the predicted output against the reference output, modifies a single AST (Abstract Syntax Tree) unit, and observes whether the score improves. To ensure that any observed improvement stems from the modification itself—rather than from a change in how the predicted and reference outputs are matched—the authors incorporate a "verifier." This verifier confirms that the improvement is genuine and persists even when the matching relationship remains constant; only modifications that pass this check are used as training targets. Ultimately, the model learns from these refined segments, incorporating two types of signals: the original CKD+Sim-RL signal and an additional, weighted supervisory signal derived from DGCR. In my view, the DGCR method proposed in this paper holds practical significance. Extracting effective corrective supervision from generation groups—which were rejected by the RL collector yet still meet the necessary criteria—represents a commendable innovation. Furthermore, the repair pipeline designed in the paper—integrating structural edits supported by reference answers, gain verification via controlled-invocation matching, and segment-level supervision weighted by these gains—offers a sound approach to localized error correction. Finally, the paper’s experiments demonstrate that incorporating DGCR learning into the CKD+SimRL framework yields a measurable improvement in accuracy.

Critical strengths

The article is well organized, logically coherent, and highly readable.On the basis of CKD+SimRL, a method of increasing DGCR learning is proposed. Training the model through the collector abandonment group is a good innovation. The experiment has also verified that it has achieved certain improvements. At the same time, the local correction in the paper has good practical significance.

Critical weaknesses

The abstract mentions that DGCR offers advantages; however, it does not provide specific data to show the actual extent of this improvement.

Questions for authors

Of the items in the discarded group, how many can actually be converted into effective supervision? Given that the verifier in Table 3 omits both gain checks and matching checks, is it still possible to determine the value of controlled matching?

3 comments

Thank you for your generous and encouraging assessment of our work. We are honored by your recognition of DGCR’s central motivation and practical value, and we appreciate your positive evaluation of the paper as a whole. Below, we clarify the recovery rate and the role of controlled matching, and explain how these points are addressed in the revised manuscript.

1. Numerical improvement in the abstract

We have added the numerical gains to the abstract of the revised manuscript so that readers can see the value of DGCR at a glance.

2. Recovery rate and the role of controlled matching

Among the groups discarded by the collector at 0.6B, roughly 18.2% can be converted into an accepted local repair. This percentage is computed over the entire rejected-group pool and therefore measures group-level recovery rather than the fraction of individual responses. It shows that the current construction can turn a substantive subset of discarded groups into learning signal.

The verifier comparison in Table 3 is a joint ablation. The complete verifier requires both positive-gain verification and assignment-consistency checks, whereas the syntactic-only arm removes both the gain check and the matching check. Table 3 therefore establishes the value of the complete verification scheme. Controlled matching targets a specific failure mode: after one call is edited, ordinary greedy matching may rematch other predictions and create an apparent score increase that did not come from the edited call. DGCR therefore compares the ordinary gain with the gain under the original assignment, requires the two gains to agree, and keeps the assignments of unedited calls unchanged. Appendix C.2 gives a concrete example in which rematching produces a positive total gain while the fixed-assignment gain is negative. In our leave-one-out audit from the frozen starting point, removing each gate separately selected the same nine compiled candidates from two groups, ruling out one known form of spurious credit. Controlled matching is not a new training configuration. It is a verifier check that prevents rematching-based pseudo-gains.

Related Content in the Revised Manuscript. The numerical abstract update is in the abstract. The recovery-rate and ground-truth injection results are in Section 4.4, while the joint verifier ablation is in Section 4.3 and Table 3. The concrete rematching example is in Appendix C.2.

Thank you for clarifying the recovery rate and the role of controlled matching. I maintain my rating of 3 (Weak acceptance), although the joint ablation does not isolate the downstream benefit of the matching check.

AuthorsOpenReview ↗

Response to Reviewer X2XQ: Isolating the Downstream Contribution of Controlled Matching

Thank you for this follow-up. To isolate the incremental downstream contribution of controlled matching and directly address your concern, we conducted a factorized verifier ablation under the same unit-weight protocol as Table 3. Across all three settings, we keep the reference-guided candidate generator, parsing procedure, token round-trip check, local compiler, candidate-selection order, and training configuration fixed, varying only which verification checks are active.

Verification configurationGain checkControlled matchingBFCL Overall (%)ACEBench Summary (%)
Full verification✓✓51.5853.80
Gain-only✓✗50.8153.00
Syntactic-only✗✗46.1849.20

The factorized results reveal two patterns. First, the positive-gain check accounts for the larger share of the improvement. Moving from syntactic-only selection to gain-only verification increases BFCL Overall by 4.63 percentage points and ACEBench Summary by 3.80 percentage points, corresponding to 85.7% and 82.6% of the observed point-estimate lift from syntactic-only to full verification. Second, adding controlled matching on top of gain-only verification provides an additional 0.77-point improvement on BFCL and 0.80-point improvement on ACEBench, accounting for the remaining 14.3% and 17.4% of the observed lift. These results show that controlled matching has a measurable conditional downstream contribution. The positive-gain check provides the main performance improvement, while controlled matching serves as a targeted safeguard against rematching-induced pseudo-gains.

We sincerely appreciate your recognition of our paper, as well as your careful feedback. We hope that this factorized ablation clarifies the contribution of controlled matching. If any questions or concerns remain, we would be very happy to address them.

Related Content in the Revised Manuscript. We have added the corresponding factorized analysis in Appendix D.2.

Reviewer_a3wKOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

DGCR recovers local supervision from low-reward rollout groups rejected by an RL collector. It constructs reference-supported structural edits, checks their evaluator gains under ordinary and controlled call assignments, and trains on repaired token spans under student-generated prefixes. Accepted groups retain the relative RL objective. Across five training seeds at 0.6B, mean BFCL accuracy increases from 51.70% to 53.41%, and ACEBench Summary from 53.12% to 55.98%. Additional comparisons examine supervision construction, routing, verification, rewards, initialization, and model size.

Critical strengths
  1. The method provides explicit corrective targets from otherwise discarded outputs without requiring the complete response to be correct. Controlled assignment addresses a concrete ambiguity in structured rewards: rematching can raise the score without improving the edited call under its original correspondence. The paper clearly explains how this motivates its verification rule and local supervision.

  2. The overall improvement is supported by five training seeds on two benchmarks. Reference-supervision and random-repair comparisons provide evidence beyond an RL-only baseline, while gains with other rewards and SFT initialization broaden the support. The paper appropriately distinguishes call-construction improvements from the still-low BFCL interactive multi-turn performance. These results support a useful, focused contribution.

Critical weaknesses

The supplied implementation leaves a concrete uncertainty about the claimed verification. Equation (3) defines gains using the complete Sim-RL reward, whose format gate includes schema-permitted argument keys (Appendix B.2). However, dgcr/matching.py computes candidate gains from greedy_match and fixed_assignment_score, which evaluate call similarity without that gate; the supplied routing and compilation path shows no subsequent complete-reward check. For example, deleting one of two schema-invalid extra keys can increase call similarity while both responses remain invalid and retain reward -1. This does not establish a problem with the reported benchmark scores, but it needs reconciliation because verified evaluator improvement is central to the method.

Questions for authors

Which reward and format-validation path were used to accept repairs in the reported experiments? Please explain whether an additional check excludes candidates that remain format-invalid, or whether verification actually uses call similarity alone. If the latter, clarifying how often accepted repairs fail to improve the complete Sim-RL reward would help assess the significance of this discrepancy.

1 comment

Thank you for carefully examining the implementation details and verification path of DGCR. We appreciate the opportunity to clarify how the released core package connects to the complete Sim-RL training pipeline.

1. What the code actually verifies

The released package contains the DGCR core, whereas the Sim-RL reward and collector are implemented in separate pipeline files. Specifically, dgcr/matching.py computes the local call-similarity gain under ordinary and controlled assignments and rejects rematching artifacts. The separate Sim-RL reward module parses responses and checks schema validity before assigning the total reward. Therefore, repaired candidates in the reported training runs pass through the complete Sim-RL format check. The reported training runs used these modules together through the full trainer pipeline. We will provide a one-command end-to-end entry point that makes this interface explicit. Thus, the similarity computation in the core package is not intended to replace the Sim-RL format gate.

2. Acceptance path and recovery rate in the reported experiments

Using all collector-rejected groups as the denominator, the 0.6B run converts about 18.2% of them into groups with an accepted local repair. In other words, the numerator is the number of rejected groups that yield a repair, and the denominator is the complete rejected-group pool. This indicates that the construction recovers usable learning signal from a nontrivial fraction of otherwise discarded groups.

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

The paper proposes DGCR, an auxiliary objective for GRPO-style training of very small function-calling models, built on STAR’s CKD initialization and Sim-RL reward. Some rollout groups are rejected by the collector and have a maximum answer-similarity score of at most τ_low = 0.01. For the parseable responses in these groups, DGCR enumerates single-AST-unit edits whose values come from the training reference: argument replacement, insertion, deletion, and single-call function-name replacement. An edit is accepted only if three conditions hold: (1) it must increase the Sim-RL score under both the ordinary greedy call assignment and a controlled one; (2) the two gains must agree, (3) the assignments of the unedited calls must stay the same.

Each group contributes at most one repair, the one with the largest gain. The repaired token span is then supervised with gain-weighted, length-normalized cross-entropy, conditioned on the student’s own prefix, and mixed into actor updates up to a 20% quota. On Qwen3-0.6B, DGCR improves BFCL v3 Overall from 51.76 to 53.45 and ACEBench Summary from 53.10 to 56.00 over CKD+Sim-RL. The paper also reports ablations on how supervision is built, which groups are routed to repair, the verifier, compatibility with different reward functions, initializing with SFT, and find that gains shrink at 1.7B and 4B.

Critical strengths
  • The idea is practical. Rejected groups are otherwise thrown away, and turning them into targeted, on-prefix supervision needs no extra model calls (about 5 ms per searched group). Unlike plain reference SFT, the supervision is restricted to the 1 decision that was verifiably wrong, in a context that the policy actually visited. That's a nice middle ground between scalar rewards and full off-policy imitation.
Critical weaknesses
  • Novelty is limited and experiments don't isolate why local repair outperforms simpler supervised alternatives. The main parts of DGCR have precedent: reference-guided correction, learning from student-generated contexts, masked supervised learning, and weighting examples by estimated improvement.
  • Gains are small and only demonstrated in a narrow setting. Limited evidence about practical training efficiency. The main BFCL improvement is only +1.69%, and the strongest evidence concerns single-turn function-call construction. BFCL Multi-Turn Accuracy remains low, increasing only from 7.00% to 7.67%. ACEBench also shows a decline in Similar API accuracy from 62.00% to 58.00%. Although the method improves results at larger model sizes, the gains decrease substantially, reaching only 0.73 points on BFCL and 0.90 points on ACEBench at 4B. The approach also depends on reference answers, parseable outputs, a structured evaluator, and a bounded neighborhood of repairs. The paper doesn't report how many rejected groups yield accepted repairs or how much more compute is required across the entire training procedure. Furthermore, Appendix D.1 shows big sensitivity to the low-reward threshold and auxiliary loss coefficient. The default hyperparameters perform best, but nearby settings can erase the improvement over CKD+Sim-RL. It's uncertain whether the method is robust and practical.
Questions for authors
  1. What is the benefit of counterfactual repair over simpler targeted supervision?
  2. How frequently does DGCR recover usable supervision?
  3. How sensitive is the method to the collector and the supervision budget?
6 comments

Thank you for the careful and constructive review. We appreciate your recognition of the practical motivation and address the points below.

1. Novelty and simpler supervision

Our novelty lies in constructing learnable local repairs for rejected groups from which super-tiny models would otherwise receive no positive learning signal, thereby helping the model learn. Our work centers on how to train small models effectively. A small model often cannot produce a completely correct tool call in one attempt, yet a failed response can still contain correct local decisions. DGCR does not require the repaired response to become entirely correct at once. It first teaches the model to get one part right. This partial structural recovery matches the limited learning capacity of small models, where several decisions may be difficult to correct simultaneously.

As a direct example, under the same CKD initialization and Sim-RL reward, injecting ground truth into rejected groups yields 51.84% BFCL and 53.40% ACEBench, compared with 51.76% and 53.10% for CKD+Sim-RL and 53.45% and 56.00% for DGCR. Thus, in this 0.6B setting, directly injecting complete correct answers produces only a small improvement, indicating that a super-tiny LLM has difficulty learning these solutions directly. This supports the design of DGCR: starting from the AST structure of tool calls, DGCR constructs local corrections so that a super-tiny LLM can learn verified partial corrections as more effective training targets. The precedents you mention each address related ideas, but existing work does not show that any of them, alone or in combination, can provide this particular learnable signal. Our work is designed around constructing that signal.

2. Recovery rate, extra computation, and the collector or budget

At the 0.6B scale, accepted repairs arise from roughly 18.2% of the groups rejected by the collector. This is a group-level rate: the numerator counts rejected groups that yield an accepted repair, and the denominator is the full set of collector-rejected groups. The result shows that the current construction can recover a meaningful portion of the available learning signal.

In terms of additional computation, AST search, verification, and repair construction account for 0.6% of total training time, while auxiliary training on repaired examples accounts for 1.9%. These increases are acceptable in this setting. Regarding sensitivity to the collector and the supervision budget, we did not design a special collector. We reuse the STAR/Sim-RL collector, which is a standard configuration for reinforcement learning in this setting. The supervision budget prevents too many local-repair groups from entering training. On average, the 18.2% repair rate does not reach the 20% quota, and the added computation is small, so no additional selection of the supervision budget is needed.

3. Gain size, scope, and cost

Results from five random seeds support a reproducible overall improvement: BFCL increases from 51.70±0.27% to 53.41±0.42%, and ACEBench increases from 53.12±0.68% to 55.98±0.36%. The fact that gains decrease as the student model grows is part of our attempt to probe how strong the repairable signal remains in rejected groups for larger students. This trend is consistent with our goal of using partial corrections to train models with limited capacity. It does not challenge our results, because enabling sub-1B super-tiny LLMs to train effectively is the central problem we explore.

The lack of improvement on Similar API is also informative: it indicates that the ability to distinguish similar APIs is not something that can be learned from rejected groups simply by constructing local repairs. We thank the reviewer for pointing this out and have added this interpretation to the revised manuscript. The small Multi-Turn improvement is related to the limited long-context capacity of super-tiny LLMs. Since the Multi-Turn score for every method remains below 10%, even a small increase can reflect a contribution from the method. In addition, DGCR's advantage is broadly supported on other subsets and in the aggregate scores.

Related Content in the Revised Manuscript. The distinction from prior supervision methods is discussed in Section 2, Sections 3.2--3.6, and Figure 1. The reference-supervision, random-repair, and routing comparisons are in Section 4.3 and Table 3. Section 4.4 reports the recovery rate, ground-truth injection comparison, Similar API and Multi-Turn results, and model scaling. Appendix D.1 gives the sensitivity results, Appendix D.4 and Table 7 give the five-seed results, and Appendix C.5 reports the additional computation.

Thanks for your time and effort in reviewing this paper. Your comments are valuable to improve the quality of our paper. With the discussion period closing in three days, we would appreciate any further feedback on whether our response addresses your concerns. We would be happy to provide further clarification.

AuthorsOpenReview ↗

Supplement to Reviewer cWyW: Multi-Turn Capacity and Similar API Scope

We further explain here why DGCR shows only modest performance on Multi-Turn and Similar API.

Multi-Turn. Existing work treats multi-turn function calling as a stateful capability requiring dialogue-state maintenance, cross-turn information retrieval, call revision, and multi-step dependency planning [1,2]. Compared with ACEBench’s Normal Multi-Turn subset, BFCL-v3 Multi-Turn places a substantially heavier burden on these abilities: it evaluates executable, stateful tool interactions in which later turns may depend on previous tool outputs and may involve missing functions, missing parameters, and long-context retrieval, whereas ACEBench’s Normal Multi-Turn subset mainly consists of fixed question-answer pairs involving topic switching or refinement of the original question [1,2]. LongFuncEval reports 13--40% degradation as tool-calling conversations become longer [3], while BUTTON and MAGNET improve multi-turn performance through dedicated compositional or graph-based trajectories [4,5]. This greater execution and state-tracking burden helps explain why BFCL Multi-Turn scores remain below 10%. For our 0.6B student, limited parameter capacity further constrains dialogue-state maintenance, retention of earlier tool outputs, and multi-turn call revision, supporting our interpretation that parameter capacity is a central bottleneck. DGCR provides verified local corrections under the student’s current generated prefix. The BFCL Multi-Turn score increases from 7.00% to 7.67%, corresponding to an absolute gain of 0.67 percentage points and a relative improvement of 9.6% over the baseline, which is a meaningful gain from improved local call construction.

Similar API. Several existing methods improve close-tool discrimination by adding explicit candidate-level supervision. TOOLVERIFIER uses contrastive questions to distinguish top candidate tools and reports an average improvement of 22% over few-shot baselines; in its tool-and-parameter verification setting, average success increases from 81.43% to 89.52% [6]. ToolExpNet constructs semantic-similarity relations between tools and uses contrastive trial-and-error guidance; removing these similarity relations increases tool-misselection from 4% to 10% [7]. These methods improve close-tool discrimination by explicitly modeling differences between competing APIs.

Given that DGCR does not construct API pairs, hard negatives, or contrastive candidate boundaries, the unchanged Similar API score identifies an orthogonal capability that the current local-repair signal does not directly target, and distinguishing similar APIs often requires a dedicated contrastive or retrieval-oriented component.

References

[1] Shishir G. Patil et al. (2025). The Berkeley Function Calling Leaderboard (BFCL): From Tool Use to Agentic Evaluation of Large Language Models. Proceedings of the 42nd International Conference on Machine Learning, 48371--48392.

[2] Chen Chen et al. (2025). ACEBench: A Comprehensive Evaluation of LLM Tool Usage. Findings of the Association for Computational Linguistics: EMNLP 2025, 12970--12998.

[3] Kiran Kate et al. (2025). LongFuncEval: Measuring the Effectiveness of Long Context Models for Function Calling. arXiv preprint arXiv:2505.10570.

[4] Mingyang Chen et al. (2025). Facilitating Multi-turn Function Calling for LLMs via Compositional Instruction Tuning. International Conference on Learning Representations (ICLR 2025).

[5] Fan Yin et al. (2025). MAGNET: Multi-turn Tool-use Data Synthesis and Distillation via Graph Translation. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics, 32600--32616.

[6] Dheeraj Mekala et al. (2024). TOOLVERIFIER: Generalization to New Tools via Self-Verification. Findings of the Association for Computational Linguistics: EMNLP 2024, 5026--5041.

[7] Zijing Zhang et al. (2025). ToolExpNet: Optimizing Multi-Tool Selection in LLMs with Similarity and Dependency-Aware Experience Networks. Findings of the Association for Computational Linguistics: ACL 2025, 15706--15722.

Comments on the paper

3 comments

Dear Reviewers

The author-reviewer discussion period will close soon. The authors put much efforts for their rebuttal. Please read and respond to authors if you have not done yet.

Thank you.

AC

Dear Area Chair,

We have posted a detailed response to Reviewer cWyW in the rebuttal thread but have not yet heard back. The reviewer’s main concerns (novelty relative to simpler supervision, the recovery rate and added computation, sensitivity to the collector and supervision budget, and the scope of gains on Multi-Turn, Similar API, and larger models) are addressed in our response. Our response also directly answers the reviewer’s three questions. If possible, we would be grateful if you could encourage the reviewer to take a look at our response before the final decision. Thank you.

Best regards,

The Authors

AuthorsOpenReview ↗

Summary of Manuscript and Code Updates

Dear Area Chair,

We revised the manuscript and accompanying code following all reviewers’ comments. The substantive manuscript additions listed below are marked in blue in the revised PDF. We also made several author-initiated wording edits to improve clarity and consistency. Where a revision addresses multiple reviewers’ concerns, the relevant reviewer IDs are noted after that item.

For reference, the revised manuscript addresses the reviewers as follows:

· Reviewer cWyW:

Source of gain. Section 4.3 now explains the roles of target construction, targeted allocation, and joint verification. It compares student-prefix corrective supervision with reference supervision, targeted routing with random verified repair, joint gain and call-assignment checks with syntactic-only selection, and the default rejected-group route with all-group routing. The all-group route drops to 39.56% on BFCL Overall Acc and 45.30% on ACEBench Summary, clarifying that the source and structure of corrective examples are essential. (Reviewer cWyW.)

Multi-Turn and Similar API. Section 4.2 now explains the difficulty of sustained tool interaction for super-tiny models and reports that DGCR improves BFCL Multi-Turn from 7.00% to 7.67%. It also explicitly reports the Similar API decrease from 62.00% to 58.00% and clarifies that the strongest benefits concern general call construction. (Reviewer cWyW.)

Reward compatibility and model scale. Section 4 reports that DGCR remains effective with Binary, ToolRL, and Sim-RL rewards, as well as with SFT initialization, and clarifies why the gains are larger for smaller models. (Reviewer cWyW.)

· Reviewer a3wK:

Code and validation path. The accompanying code was revised to expose the interface between the DGCR repair verifier and the Sim-RL reward path. Candidate gains are computed from the complete Sim-RL reward, including response parsing and schema-permitted argument-key checks, and repaired candidates are revalidated before routing and span compilation. Repair construction, controlled-assignment verification, routing, reward evaluation, and auxiliary-loss components remain separately inspectable. (Reviewer a3wK.)

· Reviewer X2XQ:

Headline results. The abstract now reports that, on Qwen3-0.6B, DGCR improves BFCL v3 and ACEBench by 1.69% and 2.90%, respectively, over strong prior methods. (Reviewer X2XQ.)

Recovery rate and ground-truth injection. Section 4.4 now reports that approximately 18.2% of collector-rejected groups yield an accepted local repair. It also compares ground-truth injection with DGCR: ground-truth injection reaches 51.84% on BFCL Overall Acc and 53.40% on ACEBench Summary, whereas DGCR reaches 53.45% and 56.00%. (Reviewers a3wK and X2XQ.)

Controlled matching. Appendix D.2 adds a factorized verifier ablation separating the ordinary gain check from controlled matching. The syntactic-only, gain-only, and full-verification settings obtain 46.18%/49.20%, 50.81%/53.00%, and 51.58%/53.80% on BFCL Overall Acc/ACEBench Summary, respectively. This isolates the contribution of controlled matching and its role in filtering rematching-induced pseudo-gains. (Reviewer X2XQ.)

We thank you for your efforts throughout the review process and for helping facilitate communication during the discussion. Thank you for handling the paper.

Best regards,

The Authors

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

Related papers

Discussion (0)

Sign in to comment.