- 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
- 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.
- 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
- 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.
- 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
(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 over orders 1–5 / 11–15. The last column counts the runs in which both and are higher for SFT+ICR than for SFT.
| Arch. | Model | : SFT → SFT+ICR | : SFT → SFT+ICR | Low: SFT → SFT+ICR | High: SFT → SFT+ICR | Runs with ↑ |
|---|---|---|---|---|---|---|
| Dense | Qwen3-4B | 2.21 → 4.12 | 2.34 → 4.38 | 0.92 → 0.66 | 2.03 → 2.72 | 5/5 |
| Dense | Qwen3-8B | 2.38 → 4.18 | 2.52 → 4.44 | 0.90 → 0.68 | 2.14 → 2.84 | 5/5 |
| Dense | Qwen3-14B | 2.55 → 4.27 | 2.70 → 4.55 | 0.87 → 0.67 | 2.22 → 2.86 | 5/5 |
| Dense | Llama-3.1-8B | 2.26 → 4.17 | 2.39 → 4.41 | 0.92 → 0.66 | 2.08 → 2.75 | 5/5 |
| Dense | Mistral-7B-v0.3 | 1.98 → 3.74 | 2.11 → 3.95 | 0.96 → 0.70 | 1.90 → 2.62 | 5/5 |
| Dense | Gemma-3-27B | 2.48 → 4.23 | 2.63 → 4.49 | 0.88 → 0.66 | 2.18 → 2.79 | 5/5 |
| Dense | Qwen3-32B | 2.67 → 4.33 | 2.84 → 4.69 | 0.85 → 0.66 | 2.27 → 2.86 | 5/5 |
| Dense | Llama-3.3-70B | 2.72 → 4.42 | 2.90 → 4.73 | 0.83 → 0.66 | 2.26 → 2.92 | 5/5 |
| MoE | Qwen3-30B-A3B | 2.51 → 4.25 | 2.66 → 4.51 | 0.87 → 0.67 | 2.18 → 2.85 | 5/5 |
| MoE | gpt-oss-20b | 2.29 → 4.01 | 2.41 → 4.23 | 0.91 → 0.69 | 2.08 → 2.77 | 5/5 |
| MoE | Mixtral-8x7B | 2.03 → 3.80 | 2.17 → 4.02 | 0.95 → 0.69 | 1.93 → 2.62 | 5/5 |
| MoE | GLM-4.5-Air | 2.69 → 4.38 | 2.86 → 4.68 | 0.84 → 0.65 | 2.26 → 2.85 | 5/5 |
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 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 interaction | Qwen3-8B | Llama-3.1-8B | Llama-3.3-70B | Qwen3-30B-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. | Model | Retail orig → tuned | Airline orig → tuned | Telecom orig → tuned |
|---|---|---|---|---|
| Dense | Qwen3-4B | 0.66 → 1.89 | 0.17 → 1.30 | 0.36 → 1.36 |
| Dense | Qwen3-8B | 0.95 → 2.41 | 0.16 → 1.50 | 0.30 → 1.49 |
| Dense | Qwen3-14B | 0.51 → 2.32 | 0.34 → 1.70 | 0.22 → 1.62 |
| Dense | Llama-3.1-8B | 0.58 → 2.06 | 0.50 → 1.80 | 0.95 → 2.24 |
| Dense | Mistral-7B-v0.3 | 0.66 → 1.93 | 0.33 → 1.40 | 0.22 → 1.27 |
| Dense | Gemma-3-27B | 0.73 → 2.19 | 0.33 → 1.60 | 0.44 → 1.71 |
| Dense | Qwen3-32B | 0.58 → 2.50 | 0.33 → 1.90 | 0.29 → 1.80 |
| Dense | Llama-3.3-70B | 0.66 → 2.68 | 0.50 → 2.10 | 0.51 → 2.02 |
| MoE | Qwen3-30B-A3B | 0.58 → 2.28 | 0.33 → 1.60 | 0.29 → 1.58 |
| MoE | gpt-oss-20b | 0.73 → 2.15 | 0.33 → 1.50 | 0.37 → 1.67 |
| MoE | Mixtral-8x7B | 0.51 → 1.71 | 0.17 → 1.20 | 0.29 → 1.32 |
| MoE | GLM-4.5-Air | 0.66 → 2.54 | 0.50 → 1.90 | 0.44 → 1.89 |
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 S: instead of optimizing the interactions over all 2n subsets S⊆N, it only samples the subsets S of specific orders, namely those made of whole groups of words. For example, when n=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 n are low, and orders above two thirds of n 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 215 to 25. 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 Rtool / word-level Rarg / gain in BFCL v4 AST over SFT (points). b: 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.
Under all five groupings and on all four models, the word-level Rtool and Rarg 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: