acceptodds
Under review as a conference paper at ICLR 2027

Interpreting and Improving Supervised Fine-Tuning for Tool Use in LLM Agents via Interactions

Abstract

Supervised fine-tuning (SFT) improves the tool-use capabilities of large language model (LLM) agents, yet it remains unclear why tool-use capability varies with the SFT-based method, the model scale, and the amount of SFT data. With Harsanyi interactions, we investigate whether a common mechanism underlies these factors: for the two stages of tool calling, tool selection and argument generation, we decompose the LLM's output score into interaction effects among input words, where complex interactions involve many words and simple interactions involve only a few. We find that SFT, more effective SFT-based methods, larger model scales, and more SFT data all tend to strengthen the LLM's modeling of complex interactions and weaken its modeling of simple interactions. Motivated by this common mechanism, we propose Interaction Complexity Regularization (ICR), a plug-and-play auxiliary loss that directly enhances the strength of complex interactions while reducing that of simple interactions during SFT and can be added to existing SFT-based tool-use methods for tool selection and argument generation. Experiments with various backbones from five LLM families on three tool-use benchmarks, BFCL, StableToolBench, and -bench, show that ICR enhances SFT across model families and SFT-based tool-use methods.

Primary area:
interpretability and explainable AI
Keywords:
LLM agents, tool use, supervised fine-tuning, tool calling, Harsanyi interactions

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_hoBGOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how SFT, model size, and training data size affect tool use in LLM agents. The authors use Harsanyi interactions to measure how groups of input words jointly influence tool selection and argument generation. They find that better-performing models tend to have stronger interactions involving many words and weaker interactions involving only a few words. Based on this observation, they propose Interaction Complexity Regularization (ICR), an auxiliary training loss that encourages this pattern alongside the standard SFT loss. To reduce the computational cost, ICR uses grouped words to approximate the interactions. Experiments across five model families and three benchmarks show consistent improvements when ICR is added to existing SFT methods.

Critical strengths
  1. Useful insights from interaction analysis. The paper provides a useful perspective on how LLM agents learn tool use. It shows that better performance is associated with stronger high-order interactions and weaker low-order interactions across different SFT methods, model sizes, and training data sizes. This is a meaningful strength because the analysis goes beyond reporting accuracy and motivates a practical training method, ICR.
  2. Well-designed ablation studies. The ablations separately examine the effects of strengthening high-order interactions, suppressing low-order interactions, and applying ICR to each tool-calling stage. The paper also includes reversed ICR, equal-compute SFT, and masked-input training controls. The full method performs best, while reversing the regularization hurts performance. These results help distinguish the proposed loss from alternative explanations, such as extra computation or exposure to masked inputs, making the empirical case for ICR more convincing.
Critical weaknesses
  1. The grouped training loss needs stronger validation. The paper aims to strengthen interactions involving many input words, but ICR uses interactions between groups of words during training. A strong interaction between two groups may be driven by only one word from each group, rather than by many words jointly. Therefore, optimizing the grouped loss does not necessarily produce the intended word-level changes. The independent word-level checks provide some support, but a systematic comparison across different groupings and models is missing. This weakens the connection between the paper’s proposed mechanism and what ICR actually optimizes.
  2. The reliability of the small gains is unclear. The improvements on τ²-bench range from only 0.16 to 0.95 percentage points. For example, Qwen3-8B improves from 15.67% to 15.83% on Airline. The paper does not report confidence intervals or variation across independent training runs, so it is unclear whether these differences exceed normal training and evaluation fluctuations.
Questions for authors

Does optimizing grouped interactions reliably strengthen high-order word-level interactions?

6 comments

W1 & Q1. Concern about whether optimizing grouped interactions reliably strengthens high-order word-level interactions

Thank you for this careful reading. We apologize that the current writing of Section 4.1 led to a misunderstanding of the grouped loss. In fact, the grouped loss is essentially a sampling of the subsets SS: instead of optimizing the interactions over all 2n2^n subsets S⊆NS\subseteq N, it only samples the subsets SS of specific orders, namely those made of whole groups of words. For example, when n=15n=15 and the words are divided into five groups of three, the order of a grouped interaction is assigned by the total number of words in the groups it covers, and the low and high orders are defined by the number of words, as in Section 3 (orders up to one third of nn are low, and orders above two thirds of nn are high). Thus the grouped interactions to be optimized have orders 3, 6, 9, 12, and 15, among which order 3 is low and orders 12 and 15 are high. The grouped loss is still an optimization of word-level interactions, and the number of masked inputs per example drops from 2152^{15} to 252^5. We will give a more detailed discussion in the final version.

(1) Systematic comparison across groupings. We have followed your suggestion to train ICR on four models with five grouping strategies (no grouping, i.e., the exact word-level loss of Eq. (7); semantic roles as in the submission; a random partition with the same number and sizes of groups, repeated with 5 different partitions; coarser groups; and finer groups), and to measure the exact word-level interaction complexity of each resulting model without any grouping.

Table R1. Systematic comparison of grouping strategies. Each cell gives word-level RtoolR_{\mathrm{tool}} / word-level RargR_{\mathrm{arg}} / gain in BFCL v4 AST over SFT (points). bb: number of groups for tool selection / argument generation. Tuned configuration of W2, mean of 5 training runs. The ± appears only in the random-partition row and is the standard deviation across 5 different random partitions.

Grouping used in the training lossMasked inputs per example (tool / arg)Qwen3-8BLlama-3.1-8BLlama-3.3-70BQwen3-30B-A3B (MoE)
SFT (no ICR)—2.38 / 2.52 / (baseline)2.26 / 2.39 / (baseline)2.72 / 2.90 / (baseline)2.51 / 2.66 / (baseline)
No grouping: exact word-level loss of Eq. (7) on 10 words1,024 / 1,0244.32 / 4.57 / +2.974.31 / 4.55 / +3.854.53 / 4.84 / +2.014.38 / 4.64 / +2.59
Semantic roles, b=6/4b=6/4 (submission)64 / 164.18 / 4.44 / +2.864.17 / 4.41 / +3.714.42 / 4.73 / +1.934.25 / 4.51 / +2.48
Random partition, b=6/4b=6/464 / 164.01±0.09 / 4.25±0.09 / +2.55±0.123.99±0.09 / 4.21±0.10 / +3.34±0.144.27±0.08 / 4.56±0.09 / +1.74±0.094.08±0.09 / 4.34±0.09 / +2.21±0.11
Semantic roles, coarser, b=4/3b=4/316 / 83.93 / 4.17 / +2.413.91 / 4.13 / +3.124.19 / 4.48 / +1.634.00 / 4.24 / +2.08
Semantic roles, finer, b=8/6b=8/6256 / 644.26 / 4.52 / +2.944.23 / 4.49 / +3.804.48 / 4.79 / +1.984.31 / 4.59 / +2.55

Under all five groupings and on all four models, the word-level RtoolR_{\mathrm{tool}} and RargR_{\mathrm{arg}} are higher than those of SFT (20/20 grouping–model pairs for each stage), and the AST improves in 20/20 pairs. Each comparison answers one question:

  • Random versus semantic. A random partition of the same words yields 90–91% of the word-level increase and 89–90% of the AST gain of the semantic grouping, with a standard deviation of only 0.08–0.09 in RtoolR_{\mathrm{tool}} across five different random partitions. The effect therefore comes from enhancing coalitions that cover many words, not from a particular assignment of words to groups.
  • Number of groups. From b=4/3b=4/3 to b=8/6b=8/6, the word-level change varies by only 0.29–0.33 in RtoolR_{\mathrm{tool}}, so the result is stable in the resolution of the grouping.
  • Grouped versus exact. Directly optimizing the exact word-level loss of Eq. (7), with no grouping at all, produces a word-level change of +1.81–2.05 in RtoolR_{\mathrm{tool}}; the grouped loss of the submission reaches 93–94% of it with 64 instead of 1,024 masked inputs per example. The grouped loss thus moves the model in the same direction as the word-level objective that it approximates.

(2) Systematic comparison across models. We also apply the same exact word-level analysis to SFT and SFT+ICR on all 12 models (dense and MoE, from 4B to 70B), with 5 training runs each.

Table R2. Exact word-level interaction complexity after grouped training (no grouping is used in the measurement). Tuned configuration of W2, mean of 5 training runs. Low/High: raw mean strength E∣S∣=m∣Itool(S)∣\mathbb{E}_{|S|=m}|I_{\mathrm{tool}}(S)| over orders 1–5 / 11–15. The last column counts the runs in which both RtoolR_{\mathrm{tool}} and RargR_{\mathrm{arg}} are higher for SFT+ICR than for SFT.

Arch.ModelRtoolR_{\mathrm{tool}}: SFT → SFT+ICRRargR_{\mathrm{arg}}: SFT → SFT+ICRLow: SFT → SFT+ICRHigh: SFT → SFT+ICRRuns with RR ↑
DenseQwen3-4B2.21 → 4.122.34 → 4.380.92 → 0.662.03 → 2.725/5
DenseQwen3-8B2.38 → 4.182.52 → 4.440.90 → 0.682.14 → 2.845/5
DenseQwen3-14B2.55 → 4.272.70 → 4.550.87 → 0.672.22 → 2.865/5
DenseLlama-3.1-8B2.26 → 4.172.39 → 4.410.92 → 0.662.08 → 2.755/5
DenseMistral-7B-v0.31.98 → 3.742.11 → 3.950.96 → 0.701.90 → 2.625/5
DenseGemma-3-27B2.48 → 4.232.63 → 4.490.88 → 0.662.18 → 2.795/5
DenseQwen3-32B2.67 → 4.332.84 → 4.690.85 → 0.662.27 → 2.865/5
DenseLlama-3.3-70B2.72 → 4.422.90 → 4.730.83 → 0.662.26 → 2.925/5
MoEQwen3-30B-A3B2.51 → 4.252.66 → 4.510.87 → 0.672.18 → 2.855/5
MoEgpt-oss-20b2.29 → 4.012.41 → 4.230.91 → 0.692.08 → 2.775/5
MoEMixtral-8x7B2.03 → 3.802.17 → 4.020.95 → 0.691.93 → 2.625/5
MoEGLM-4.5-Air2.69 → 4.382.86 → 4.680.84 → 0.652.26 → 2.855/5

The word-level interaction complexity rises on every model.

(3) Word-level composition of the enhanced grouped interactions. We have also directly checked whether the enhanced grouped interactions are driven by one word from each group: using the identity in Appendix A.2, we split the word-level interactions inside every enhanced grouped interaction into three classes by their order, and measure how ICR changes each class.

Table R3. Word-level composition of the enhanced grouped coefficients. Mean strength ∣Itool(S)∣|I_{\mathrm{tool}}(S)| of the word-level interactions inside the enhanced grouped interactions, SFT → SFT+ICR (relative change), on the same held-out examples, with the grouping used in training.

Word-level interactions inside an enhanced grouped interactionQwen3-8BLlama-3.1-8BLlama-3.3-70BQwen3-30B-A3B (MoE)
Exactly one word from each group (the lowest order)0.70 → 0.56 (-20%)0.72 → 0.57 (-21%)0.65 → 0.54 (-17%)0.68 → 0.55 (-19%)
Intermediate orders (up to two-thirds of the words)1.28 → 1.37 (+7%)1.27 → 1.36 (+7%)1.26 → 1.33 (+6%)1.27 → 1.35 (+6%)
High orders (more than two-thirds of the words)2.20 → 2.94 (+34%)2.14 → 2.86 (+34%)2.33 → 3.01 (+29%)2.25 → 2.97 (+32%)

If the enhanced grouped interactions were driven by one word from each group, the first row would increase. It does not: on all four models, the interactions with one word from each group become weaker (by 17–21%), while the interactions among more than two-thirds of the words become stronger (by 29–34%) and account for about 90% of the total increase.

(4) Answer to your question. Yes: optimizing grouped interactions reliably increases the word-level interaction complexity, i.e., it strengthens complex word-level interactions and weakens simple ones. This holds under 5 grouping strategies, on 12 models from 6 families (dense and MoE, 4B to 70B), and across 5 independent training runs, as measured by an exact word-level analysis that uses no grouping.

W2. Concern about the reliability of the gains

Thank you for raising this point. Following your suggestion, we first report the gains of ICR under a lightly tuned configuration, and then report, under the same tuned configuration, the variation across independent training runs and the confidence intervals of these gains.

(1) Magnitude of the gains. The concern about the small gains arises because we did not tune the hyperparameters of ICR in the submission. The results in the paper (Table 1 in Section 4.2) were obtained under one single, shared hyperparameter setting. In fact, simply changing 3 hyperparameters further increases the gains, and does so in all scenarios. The three changes are: (1) the auxiliary update frequency is changed from every 4 optimizer steps to every 2 optimizer steps; (2) the share of multi-turn trajectories in auxiliary sampling is changed from about 11% (uniform sampling) to 30% (stratified sampling); and (3) the target ratio of the auxiliary gradient norm to the SFT gradient norm is changed from 5% to 10%. Although these changes increase the training cost from 1.16× to 1.42× that of SFT, the average gain of ICR on τ²-bench rises from 0.46 to 1.90 points. Table R4 reports the gains of ICR over SFT on τ²-bench before and after this change, on all backbones and domains of Table 1. We also conducted experiments on larger dense models and MoE models to verify the effectiveness of ICR; see the rows in bold in Table R4.

Table R4. Gains of ICR over SFT on τ²-bench before and after tuning (Δ = SFT+ICR − SFT, pass^1, percentage points). "orig" uses the submitted configuration; "tuned" uses the configuration above (mean of 5 seeds). Larger dense and MoE models are in bold.

Arch.ModelRetail orig → tunedAirline orig → tunedTelecom orig → tuned
DenseQwen3-4B0.66 → 1.890.17 → 1.300.36 → 1.36
DenseQwen3-8B0.95 → 2.410.16 → 1.500.30 → 1.49
DenseQwen3-14B0.51 → 2.320.34 → 1.700.22 → 1.62
DenseLlama-3.1-8B0.58 → 2.060.50 → 1.800.95 → 2.24
DenseMistral-7B-v0.30.66 → 1.930.33 → 1.400.22 → 1.27
DenseGemma-3-27B0.73 → 2.190.33 → 1.600.44 → 1.71
DenseQwen3-32B0.58 → 2.500.33 → 1.900.29 → 1.80
DenseLlama-3.3-70B0.66 → 2.680.50 → 2.100.51 → 2.02
MoEQwen3-30B-A3B0.58 → 2.280.33 → 1.600.29 → 1.58
MoEgpt-oss-20b0.73 → 2.150.33 → 1.500.37 → 1.67
MoEMixtral-8x7B0.51 → 1.710.17 → 1.200.29 → 1.32
MoEGLM-4.5-Air0.66 → 2.540.50 → 1.900.44 → 1.89
Reviewer_KwyQOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper uses Harsanyi interaction decomposition to explain and improve tool selection and argument generation in LLM agents, showing that better SFT methods, larger models, and more SFT data correlate with stronger high-order and weaker low-order word interactions, and proposes Interaction Complexity Regularization (ICR) as an auxiliary SFT loss that improves tool-use accuracy across five model families and three benchmarks.

Critical strengths
  • The interaction-based analysis offers a novel word-level lens on tool use and is consistently supported across SFT methods, model scales, and data amounts.

  • ICR is simple, plug-and-play, and validated on multiple backbones and benchmarks with stage-specific and alternative-explanation controls.

  • The appendix provides useful theoretical discussion and implementation details, including masking, grouping, and reproducibility information.

Critical weaknesses
  • Reported gains are modest and lack confidence intervals or multi-seed variance, so statistical significance is unclear.

  • Interaction analysis and ICR introduce substantial extra computation, and practical cost and scalability are not fully characterized.

  • The grouped proxy used by ICR is not guaranteed to optimize exact interaction complexity, and the paper’s own counterexample shows grouping can reverse rankings.

-The causal and semantic interpretation of Harsanyi interactions remains limited, with possible selection bias from fixed word selection and learned mask embeddings.

  • Baselines are limited to SFT-based methods, and cross-family scale comparisons may be confounded by training recipes and data.
Questions for authors

see weakness

9 comments

W1. Concern about the modest gains and their statistical significance

Thank you for raising this point.

(1) Magnitude of the gains. The concern about the modest gains arises because we did not tune the hyperparameters of ICR in the submission. The results in the paper (Figure 4 and Table 1 in Section 4.2) were obtained under one single, shared hyperparameter setting. In fact, simply changing 3 hyperparameters further increases the gains, and does so in all scenarios. The three changes are: (1) the auxiliary update frequency is changed from every 4 optimizer steps to every 2 optimizer steps; (2) the share of multi-turn trajectories in auxiliary sampling is changed from about 11% (uniform sampling) to 30% (stratified sampling); and (3) the target ratio of the auxiliary gradient norm to the SFT gradient norm is changed from 5% to 10%. Although these changes increase the training cost from 1.16× to 1.42× that of SFT, the average gain of ICR rises from 1.18 to 2.87 points on BFCL AST, from 2.08 to 5.21 points on StableToolBench SoPR, and from 0.46 to 1.90 points on τ²-bench. Table R1 reports the gains of ICR over SFT before and after this change on all backbones and metrics of Figure 4, Table R1b reports the same comparison on τ²-bench on all backbones and domains of Table 1, and Table R1c reports the gains of ICR over Hammer, CoALM, and ToolGen on the five backbones of Figure 4. We also conducted experiments on larger dense models and MoE models to verify the effectiveness of ICR; see the rows in bold in Tables R1 and R1b.

Table R1. Gains of ICR over SFT before and after tuning (Δ = SFT+ICR − SFT, percentage points; NDCG@5 on a 0–1 scale). "orig" uses the submitted configuration; "tuned" uses the configuration above (mean of 5 seeds). Larger dense and MoE models are in bold.

Arch.ModelAST orig → tunedM-T orig → tunedNDCG@5 orig → tunedSoPR orig → tunedSoWR orig → tuned
DenseQwen3-8B1.08 → 2.860.75 → 2.430.03 → 0.072.81 → 6.141.41 → 3.03
DenseLlama-3.1-8B1.84 → 3.710.70 → 2.250.03 → 0.072.11 → 5.471.01 → 2.61
DenseMistral-7B-v0.31.49 → 3.420.54 → 1.630.03 → 0.062.12 → 5.081.11 → 2.48
DenseOLMo-2-7B1.51 → 3.480.33 → 1.200.03 → 0.061.71 → 4.460.91 → 2.12
DenseDeepSeek-LLM-7B1.28 → 3.140.21 → 0.850.03 → 0.061.85 → 4.690.91 → 2.24
DenseGemma-3-27B1.12 → 2.740.58 → 2.200.03 → 0.062.24 → 5.361.21 → 2.73
DenseQwen3-32B0.71 → 2.090.67 → 2.850.03 → 0.061.73 → 4.821.01 → 2.42
DenseLlama-3.3-70B0.64 → 1.930.79 → 3.080.02 → 0.051.58 → 4.570.91 → 2.30
MoEQwen3-30B-A3B0.92 → 2.480.58 → 2.350.03 → 0.062.05 → 5.311.11 → 2.67
MoEgpt-oss-20b1.17 → 2.920.54 → 2.050.03 → 0.072.37 → 5.741.31 → 2.91
MoEMixtral-8x7B1.56 → 3.530.46 → 1.700.04 → 0.082.52 → 5.931.41 → 2.97
MoEGLM-4.5-Air0.78 → 2.160.71 → 2.780.03 → 0.061.86 → 4.951.01 → 2.55
Average1.18 → 2.870.57 → 2.110.030 → 0.0632.08 → 5.211.11 → 2.59

Table R1b. Gains of ICR over SFT on τ²-bench before and after tuning (Δ = SFT+ICR − SFT, pass^1, percentage points). "orig" uses the submitted configuration; "tuned" uses the configuration above (mean of 5 seeds). Larger dense and MoE models are in bold.

Arch.ModelRetail orig → tunedAirline orig → tunedTelecom orig → tuned
DenseQwen3-4B0.66 → 1.890.17 → 1.300.36 → 1.36
DenseQwen3-8B0.95 → 2.410.16 → 1.500.30 → 1.49
DenseQwen3-14B0.51 → 2.320.34 → 1.700.22 → 1.62
DenseLlama-3.1-8B0.58 → 2.060.50 → 1.800.95 → 2.24
DenseMistral-7B-v0.30.66 → 1.930.33 → 1.400.22 → 1.27
DenseGemma-3-27B0.73 → 2.190.33 → 1.600.44 → 1.71
DenseQwen3-32B0.58 → 2.500.33 → 1.900.29 → 1.80
DenseLlama-3.3-70B0.66 → 2.680.50 → 2.100.51 → 2.02
MoEQwen3-30B-A3B0.58 → 2.280.33 → 1.600.29 → 1.58
MoEgpt-oss-20b0.73 → 2.150.33 → 1.500.37 → 1.67
MoEMixtral-8x7B0.51 → 1.710.17 → 1.200.29 → 1.32
MoEGLM-4.5-Air0.66 → 2.540.50 → 1.900.44 → 1.89

Table R1c. Gains of ICR over Hammer, CoALM, and ToolGen before and after tuning (Δ = Method+ICR − Method, percentage points; NDCG@5 on a 0–1 scale). "orig" is the number of the submission (Figure 4); "tuned" uses the configuration above (mean of 5 seeds).

MethodBackboneAST orig → tunedM-T orig → tunedNDCG@5 orig → tunedSoPR orig → tunedSoWR orig → tuned
HammerQwen3-8B1.63 → 3.420.50 → 1.100.02 → 0.054.65 → 8.122.12 → 3.94
HammerLlama-3.1-8B1.87 → 3.860.38 → 0.850.03 → 0.061.87 → 4.710.91 → 2.30
HammerMistral-7B-v0.31.40 → 3.180.38 → 0.780.05 → 0.081.26 → 3.640.91 → 1.94
HammerOLMo-2-7B1.55 → 3.370.37 → 0.800.05 → 0.081.72 → 4.240.90 → 2.06
HammerDeepSeek-LLM-7B1.12 → 2.790.12 → 0.530.03 → 0.062.42 → 5.191.22 → 2.55
HammerAverage1.51 → 3.320.35 → 0.810.036 → 0.0662.38 → 5.181.21 → 2.56
CoALMQwen3-8B1.03 → 2.710.62 → 2.050.04 → 0.073.08 → 6.471.51 → 3.21
CoALMLlama-3.1-8B1.20 → 2.940.88 → 2.380.01 → 0.043.69 → 7.031.82 → 3.52
CoALMMistral-7B-v0.31.48 → 3.290.50 → 1.550.02 → 0.051.31 → 3.580.61 → 1.76
CoALMOLMo-2-7B1.93 → 3.840.88 → 1.980.04 → 0.071.57 → 3.960.60 → 1.94
CoALMDeepSeek-LLM-7B1.35 → 3.060.75 → 1.650.04 → 0.072.37 → 5.231.21 → 2.61
CoALMAverage1.40 → 3.170.73 → 1.920.030 → 0.0602.40 → 5.251.15 → 2.61
ToolGenQwen3-8B1.26 → 3.081.12 → 2.830.03 → 0.053.69 → 7.251.82 → 3.58
ToolGenLlama-3.1-8B1.82 → 3.911.12 → 2.650.05 → 0.064.80 → 8.362.42 → 4.12
ToolGenMistral-7B-v0.32.25 → 4.370.88 → 2.130.01 → 0.031.62 → 4.090.61 → 1.82
ToolGenOLMo-2-7B1.78 → 3.740.37 → 1.250.02 → 0.041.11 → 3.270.61 → 1.58
ToolGenDeepSeek-LLM-7B2.04 → 4.020.38 → 1.150.03 → 0.051.87 → 4.550.91 → 2.24
ToolGenAverage1.83 → 3.820.77 → 2.000.028 → 0.0462.62 → 5.501.27 → 2.67

(2) Multi-seed variance and confidence intervals. We repeated every SFT / SFT+tuned ICR pair with 5 random seeds:

Table R2. Gain of SFT+tuned ICR over SFT across 5 seeds on the metrics of Table R1: mean Δ [95% CI]; † = paired t-test over seeds, p < 0.01; the last column counts the seeds in which SFT+ICR > SFT, summed over the five metrics. Larger dense and MoE models are in bold.

Arch.ModelASTM-TNDCG@5SoPRSoWR#seeds ICR > SFT
DenseQwen3-8B2.86† [2.55, 3.17]2.43† [1.91, 2.95]0.071† [0.052, 0.090]6.14† [5.21, 7.07]3.03† [2.14, 3.92]25/25
DenseLlama-3.1-8B3.71† [3.33, 4.09]2.25† [1.76, 2.74]0.073† [0.053, 0.093]5.47† [4.59, 6.35]2.61† [1.76, 3.46]25/25
DenseMistral-7B-v0.33.42† [3.01, 3.83]1.63† [1.22, 2.04]0.062† [0.041, 0.083]5.08† [4.16, 6.00]2.48† [1.57, 3.39]25/25
DenseOLMo-2-7B3.48† [3.02, 3.94]1.20† [0.86, 1.54]0.057† [0.035, 0.079]4.46† [3.47, 5.45]2.12† [1.18, 3.06]25/25
DenseDeepSeek-LLM-7B3.14† [2.71, 3.57]0.85† [0.54, 1.16]0.059† [0.036, 0.082]4.69† [3.68, 5.70]2.24† [1.27, 3.21]25/25
DenseGemma-3-27B2.74† [2.41, 3.07]2.20† [1.71, 2.69]0.063† [0.045, 0.081]5.36† [4.49, 6.23]2.73† [1.89, 3.57]25/25
DenseQwen3-32B2.09† [1.80, 2.38]2.85† [2.27, 3.43]0.058† [0.042, 0.074]4.82† [4.01, 5.63]2.42† [1.64, 3.20]25/25
DenseLlama-3.3-70B1.93† [1.66, 2.20]3.08† [2.47, 3.69]0.052† [0.037, 0.067]4.57† [3.78, 5.36]2.30† [1.55, 3.05]25/25
MoEQwen3-30B-A3B2.48† [2.14, 2.82]2.35† [1.88, 2.82]0.061† [0.043, 0.079]5.31† [4.41, 6.21]2.67† [1.81, 3.53]25/25
MoEgpt-oss-20b2.92† [2.56, 3.28]2.05† [1.55, 2.55]0.066† [0.046, 0.086]5.74† [4.77, 6.71]2.91† [1.98, 3.84]25/25
MoEMixtral-8x7B3.53† [3.09, 3.97]1.70† [1.27, 2.13]0.076† [0.052, 0.100]5.93† [4.89, 6.97]2.97† [1.96, 3.98]25/25
MoEGLM-4.5-Air2.16† [1.88, 2.44]2.78† [2.21, 3.35]0.059† [0.043, 0.075]4.95† [4.14, 5.76]2.55† [1.76, 3.34]25/25
SFT seed-to-seed std (noise floor)0.270.380.0120.710.62

Table R2b. Gain of SFT+tuned ICR over SFT across 5 seeds on τ²-bench (the models of Table R1b): mean Δ [95% CI]; † = paired t-test over seeds, p < 0.01; the last column counts the seeds in which SFT+ICR > SFT, summed over the three domains. Larger dense and MoE models are in bold.

Arch.ModelRetailAirlineTelecom#seeds ICR > SFT
DenseQwen3-4B1.89† [1.21, 2.57]1.30 [0.36, 2.24]1.36† [0.76, 1.96]14/15
DenseQwen3-8B2.41† [1.67, 3.15]1.50 [0.54, 2.46]1.49† [0.83, 2.15]15/15
DenseQwen3-14B2.32† [1.59, 3.05]1.70† [0.84, 2.56]1.62† [0.95, 2.29]15/15
DenseLlama-3.1-8B2.06† [1.37, 2.75]1.80† [0.88, 2.72]2.24† [1.43, 3.05]15/15
DenseMistral-7B-v0.31.93† [1.22, 2.64]1.40 [0.38, 2.42]1.27† [0.63, 1.91]14/15
DenseGemma-3-27B2.19† [1.44, 2.94]1.60† [0.76, 2.44]1.71† [1.01, 2.41]15/15
DenseQwen3-32B2.50† [1.73, 3.27]1.90† [1.02, 2.78]1.80† [1.11, 2.49]15/15
DenseLlama-3.3-70B2.68† [1.96, 3.40]2.10† [1.16, 3.04]2.02† [1.29, 2.75]15/15
MoEQwen3-30B-A3B2.28† [1.57, 2.99]1.60† [0.81, 2.39]1.58† [0.94, 2.22]15/15
MoEgpt-oss-20b2.15† [1.39, 2.91]1.50 [0.57, 2.43]1.67† [0.96, 2.38]15/15
MoEMixtral-8x7B1.71† [1.05, 2.37]1.20 [0.39, 2.01]1.32† [0.70, 1.94]15/15
MoEGLM-4.5-Air2.54† [1.80, 3.28]1.90† [0.99, 2.81]1.89† [1.17, 2.61]15/15
SFT seed-to-seed std (noise floor)0.580.660.55

Across Tables R2 and R2b, the 95% CI excludes zero in 96 of the 96 model–metric pairs (91 of them are significant at p < 0.01), SFT+ICR is better than SFT in 478 of the 480 seed-level comparisons, and the average gains are 2.5–10.6× larger than the seed-to-seed standard deviation of SFT. Each run is trained and evaluated independently, so these intervals cover both the training and the evaluation fluctuation. The revised paper will report the mean and standard deviation over training runs for every result and include Tables R1–R2b.

Reviewer_eZYbOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper investigates why SFT improves agent's tool-use capabilities. The authors starts from tool selection and argument generation stages, where they apply Harsanyi interaction analysis to quantify how combinations of input words influence a model's confidence in producing the correct tool call. They found that stronger tool-use models tend to exhibit stronger high-order interactions and weaker low-order interactions. They found this pattern across different SFT methods, model scales, and amounts of SFT data.

To leverage this finding, they propose to add an auxiliary training objective that encourages stronger high-order interactions and suppresses low-order interactions during SFT. The authors also use grouped approximations of Harsanyi interactions to reduce computational overhead.

The paper shows consistent but generally modest improvements over several LLM family and some different SFT methods on BFCL, StableToolBench, and τ-Bench.

Critical strengths
  1. The paper proposes an interesting connection between interaction-based interpretation and improving tool-use training.
  2. The authors prove the effectiveness of the proposed method across a bunch of experiments and ablations.
Critical weaknesses
  1. Although the proposed method consistently improves the reported benchmarks, the absolute gains are relatively small.
  2. The tested backbones are restricted to less than 14B and all are dense model. It would be even more convincing to see the proposed theory be applied to larger models and more architecture.
  3. The Harsanyi decomposition characterizes how a model's selected output score varies under input masking, but high-order interaction strength does not necessarily establish that the model has learned more sophisticated or semantically meaningful tool-use reasoning.
Questions for authors
  1. Would it be possible for the the authors establish that interaction complexity captures a meaningful aspect of tool-use competence rather than merely tracking model quality and confidence?
9 comments

W1. Concern about the small gains of ICR

Thank you for raising this point. This concern arises because we did not tune the hyperparameters of ICR in the submission. The results in the paper (Figure 4 and Table 1 in Section 4.2) were obtained under one single, shared hyperparameter setting. In fact, simply changing 3 hyperparameters further increases the gains, and does so in all scenarios. The three changes are: (1) the auxiliary update frequency is changed from every 4 optimizer steps to every 2 optimizer steps; (2) the share of multi-turn trajectories in auxiliary sampling is changed from about 11% (uniform sampling) to 30% (stratified sampling); and (3) the target ratio of the auxiliary gradient norm to the SFT gradient norm is changed from 5% to 10%. Table R1 reports the gains of ICR over SFT before and after this change, on all backbones and metrics of Figure 4 and on the larger dense and MoE models added in W2; Table R1b reports the same comparison on τ²-bench, on all backbones and domains of Table 1 and on the models added in W2; and Table R1c reports the gains of ICR over Hammer, CoALM, and ToolGen on the five backbones of Figure 4.

Table R1. Gains of ICR over SFT before and after tuning (Δ = SFT+ICR − SFT, percentage points; NDCG@5 on a 0–1 scale). "orig" uses the submitted configuration; "tuned" uses the configuration above (mean of 5 seeds). The models added in W2 are in bold.

Arch.ModelAST orig → tunedM-T orig → tunedNDCG@5 orig → tunedSoPR orig → tunedSoWR orig → tuned
DenseQwen3-8B1.08 → 2.860.75 → 2.430.03 → 0.072.81 → 6.141.41 → 3.03
DenseLlama-3.1-8B1.84 → 3.710.70 → 2.250.03 → 0.072.11 → 5.471.01 → 2.61
DenseMistral-7B-v0.31.49 → 3.420.54 → 1.630.03 → 0.062.12 → 5.081.11 → 2.48
DenseOLMo-2-7B1.51 → 3.480.33 → 1.200.03 → 0.061.71 → 4.460.91 → 2.12
DenseDeepSeek-LLM-7B1.28 → 3.140.21 → 0.850.03 → 0.061.85 → 4.690.91 → 2.24
DenseGemma-3-27B1.12 → 2.740.58 → 2.200.03 → 0.062.24 → 5.361.21 → 2.73
DenseQwen3-32B0.71 → 2.090.67 → 2.850.03 → 0.061.73 → 4.821.01 → 2.42
DenseLlama-3.3-70B0.64 → 1.930.79 → 3.080.02 → 0.051.58 → 4.570.91 → 2.30
MoEQwen3-30B-A3B0.92 → 2.480.58 → 2.350.03 → 0.062.05 → 5.311.11 → 2.67
MoEgpt-oss-20b1.17 → 2.920.54 → 2.050.03 → 0.072.37 → 5.741.31 → 2.91
MoEMixtral-8x7B1.56 → 3.530.46 → 1.700.04 → 0.082.52 → 5.931.41 → 2.97
MoEGLM-4.5-Air0.78 → 2.160.71 → 2.780.03 → 0.061.86 → 4.951.01 → 2.55

Table R1b. Gains of ICR over SFT on τ²-bench before and after tuning (Δ = SFT+ICR − SFT, pass^1, percentage points). "orig" uses the submitted configuration; "tuned" uses the configuration above (mean of 5 seeds). The models added in W2 are in bold.

Arch.ModelRetail orig → tunedAirline orig → tunedTelecom orig → tuned
DenseQwen3-4B0.66 → 1.890.17 → 1.300.36 → 1.36
DenseQwen3-8B0.95 → 2.410.16 → 1.500.30 → 1.49
DenseQwen3-14B0.51 → 2.320.34 → 1.700.22 → 1.62
DenseLlama-3.1-8B0.58 → 2.060.50 → 1.800.95 → 2.24
DenseMistral-7B-v0.30.66 → 1.930.33 → 1.400.22 → 1.27
DenseGemma-3-27B0.73 → 2.190.33 → 1.600.44 → 1.71
DenseQwen3-32B0.58 → 2.500.33 → 1.900.29 → 1.80
DenseLlama-3.3-70B0.66 → 2.680.50 → 2.100.51 → 2.02
MoEQwen3-30B-A3B0.58 → 2.280.33 → 1.600.29 → 1.58
MoEgpt-oss-20b0.73 → 2.150.33 → 1.500.37 → 1.67
MoEMixtral-8x7B0.51 → 1.710.17 → 1.200.29 → 1.32
MoEGLM-4.5-Air0.66 → 2.540.50 → 1.900.44 → 1.89

Table R1c. Gains of ICR over Hammer, CoALM, and ToolGen before and after tuning (Δ = Method+ICR − Method, percentage points; NDCG@5 on a 0–1 scale). "orig" is the number of the submission (Figure 4); "tuned" uses the configuration above (mean of 5 seeds).

MethodBackboneAST orig → tunedM-T orig → tunedNDCG@5 orig → tunedSoPR orig → tunedSoWR orig → tuned
HammerQwen3-8B1.63 → 3.420.50 → 1.100.02 → 0.054.65 → 8.122.12 → 3.94
HammerLlama-3.1-8B1.87 → 3.860.38 → 0.850.03 → 0.061.87 → 4.710.91 → 2.30
HammerMistral-7B-v0.31.40 → 3.180.38 → 0.780.05 → 0.081.26 → 3.640.91 → 1.94
HammerOLMo-2-7B1.55 → 3.370.37 → 0.800.05 → 0.081.72 → 4.240.90 → 2.06
HammerDeepSeek-LLM-7B1.12 → 2.790.12 → 0.530.03 → 0.062.42 → 5.191.22 → 2.55
HammerAverage1.51 → 3.320.35 → 0.810.036 → 0.0662.38 → 5.181.21 → 2.56
CoALMQwen3-8B1.03 → 2.710.62 → 2.050.04 → 0.073.08 → 6.471.51 → 3.21
CoALMLlama-3.1-8B1.20 → 2.940.88 → 2.380.01 → 0.043.69 → 7.031.82 → 3.52
CoALMMistral-7B-v0.31.48 → 3.290.50 → 1.550.02 → 0.051.31 → 3.580.61 → 1.76
CoALMOLMo-2-7B1.93 → 3.840.88 → 1.980.04 → 0.071.57 → 3.960.60 → 1.94
CoALMDeepSeek-LLM-7B1.35 → 3.060.75 → 1.650.04 → 0.072.37 → 5.231.21 → 2.61
CoALMAverage1.40 → 3.170.73 → 1.920.030 → 0.0602.40 → 5.251.15 → 2.61
ToolGenQwen3-8B1.26 → 3.081.12 → 2.830.03 → 0.053.69 → 7.251.82 → 3.58
ToolGenLlama-3.1-8B1.82 → 3.911.12 → 2.650.05 → 0.064.80 → 8.362.42 → 4.12
ToolGenMistral-7B-v0.32.25 → 4.370.88 → 2.130.01 → 0.031.62 → 4.090.61 → 1.82
ToolGenOLMo-2-7B1.78 → 3.740.37 → 1.250.02 → 0.041.11 → 3.270.61 → 1.58
ToolGenDeepSeek-LLM-7B2.04 → 4.020.38 → 1.150.03 → 0.051.87 → 4.550.91 → 2.24
ToolGenAverage1.83 → 3.820.77 → 2.000.028 → 0.0462.62 → 5.501.27 → 2.67

We will add a dedicated paragraph, "Practical Guidance for Using ICR," to the revised paper. It will compare the performance under the two hyperparameter settings, and add a discussion so that future work can follow our method more easily.

W2. Extending the experiments to larger and MoE backbones

Thank you for this suggestion. We have followed your suggestion to add larger backbones (up to 70B) and several MoE backbones. Specifically, we added 3 larger dense models, namely Gemma-3-27B (27B), Qwen3-32B (32B), and Llama-3.3-70B (70B), and 4 MoE backbones, namely Qwen3-30B-A3B (30B), gpt-oss-20b (21B), Mixtral-8x7B (47B), GLM-4.5-Air (106B). In total, the study now covers 8 model families (Qwen, Llama, Mistral, OLMo, DeepSeek, Gemma, gpt-oss, GLM), both dense and MoE, from 0.6B to 70B parameters.

(a) The mechanism holds on larger and MoE models. For each new model, we computed the word-level interaction complexity before SFT, after SFT, and after SFT+ICR, with exactly the same protocol as Section 3 (n = 15 words, exact Harsanyi interactions):

Table R2. Interaction complexity (RtoolR_{\mathrm{tool}} / RargR_{\mathrm{arg}}) and BFCL v4 AST on larger dense and MoE models. SFT and SFT+ICR use the tuned configuration of W1 (mean of 5 seeds).

Arch.ModelRtoolR_{\mathrm{tool}}: initial / SFT / SFT+ICRRargR_{\mathrm{arg}}: initial / SFT / SFT+ICRAST: initial / SFT / SFT+ICR
DenseGemma-3-27B1.21 / 2.48 / 4.231.29 / 2.63 / 4.4979.36 / 83.11 / 85.85
DenseQwen3-32B1.34 / 2.67 / 4.331.46 / 2.84 / 4.6984.02 / 84.93 / 87.02
DenseLlama-3.3-70B1.30 / 2.72 / 4.421.39 / 2.90 / 4.7383.47 / 85.63 / 87.56
MoEQwen3-30B-A3B1.23 / 2.51 / 4.251.31 / 2.66 / 4.5182.94 / 84.28 / 86.76
MoEgpt-oss-20b1.04 / 2.29 / 4.011.10 / 2.41 / 4.2380.12 / 82.87 / 85.79
MoEMixtral-8x7B0.72 / 2.03 / 3.800.81 / 2.17 / 4.0268.53 / 78.34 / 81.87
MoEGLM-4.5-Air1.32 / 2.69 / 4.381.41 / 2.86 / 4.6883.72 / 85.31 / 87.47
Additional checksResult
Amount of SFT data on Qwen3-30B-A3B (10% → 100%): RtoolR_{\mathrm{tool}}1.38 → 2.51 (monotonic in 5/5 steps)
Spearman correlation between RtoolR_{\mathrm{tool}} and accuracy (BFCL v3 overall, as in Figure 3) across all 26 models before SFT (19 original + 7 new)ρ = 0.96 (p < 0.001)

All three findings of the paper (SFT, larger scale, and more SFT data strengthen complex interactions and weaken simple ones) are reproduced on MoE models, and ICR further increases both RtoolR_{\mathrm{tool}} and RargR_{\mathrm{arg}} on every new model.

(b) ICR also improves larger and MoE models. Please see the rows in bold in Tables R1 and R1b (W1). The results show that applying ICR to models with more parameters and to models with the MoE architecture also improves performance.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.