- Rating
- 1
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper argues that multi-turn tool-using agents should be judged on efficiency as well as correctness: even successful trajectories often contain redundant or corrective tool calls that inflate interaction cost. It proposes an iterative DPO pipeline that samples several successful trajectories per query, ranks them by tool-call overhead, and turns best-vs-worst differences into step-level preference pairs via two rules (detour detection and first-divergence under parallel tool calls), keeping only pairs the current greedy policy fails. When cross-trajectory contrasts flatten out, an intra-trajectory pruning stage removes tool calls that the BFCL checker confirms are unnecessary and runs a final DPO stage on cleaned prefixes. On BFCL v3 multi-turn subsets, an 8B model trained this way surpasses a much larger prior best model while sharply cutting excess tool calls; trained only on the Base subset, the policy transfers to the harder subsets with fewer calls, tokens, and inference steps.
Critical strengths
- Reducing extra tool calls in agent trajectory is genuine problem, especially in improving the latency concerns for agentic AI applications. This paper addresses exactly that.
- The small model, once trained, seems to significantly reduce the number of tool calls on BFCL v3.
- Upon training, accuracy on some harder sets also improves in addition to efficiency.
Critical weaknesses
- The evaluation scope is limited to BFCL v3. In order for the technique to be considered generalizable, empirical analysis on multiple benchmarks is required.
- The baseline (untrained model) is a weaker choice. The paper mentions a few other efficiency-focused methods that should be considered and compared as baselines.
- Per-component isolation ablations are missing.
- The pipeline seems to be applied differently for different backbones. Authors should provide analysis on how tedious hand-tuning of the technique is.
- The method is heavily dependent on the BFCL checker. But in many cases where there are a very high number of tools, there can be many solution paths possible; such an oracle checker might not be directly applicable in such cases.
- Error bars and CIs are not reported. This makes it hard to gauge where the improvement comes from.
Questions for authors
- Does the technique generalize to other environments like TauBench or ToolSandbox?
- Is there a variant of the technique for other environments where oracle checker might not be available?
- How sensitive is the excess-call reduction to the choice of reference/ground-truth trajectory used to define "excess"?
To compare our method with other baselines on matched data, We split the 200 base samples in BFCL multi-turn into training and evaluation by stratified sampling over the involved tool classes, yielding 150 training samples and 50 evaluation samples, with no overlap. The miss_func, miss_param, and long_context subsets share the same ids as the base samples. Once an id is held out for evaluation, its counterparts in these augmented subsets are also excluded from training and trajectory mining. Evaluation is conducted on the 50 base samples together with their three augmented subsets (200 instances in total), while trajectory mining uses only the 150 base samples.
We compare our method with two iterative baselines, trajectory-level SFT and trajectory-level DPO. In each iteration, we draw K=6 samples at temperature T=0.7. Trajectory-level SFT fine-tunes on the sampled trajectory with the fewest tool calls. Trajectory-level DPO contrasts the trajectory with the most tool calls against the one with the fewest. For both baselines, user messages and tool results are masked, so the loss is computed only on assistant outputs. Each method is trained for 4 iterations. All methods are evaluated on the same held-out set of 200 instances. The table below reports accuracy. Due to the time limit, the held-out evaluation is reported for BitAgent, the backbone on which the paper’s main gains were claimed. The same split for xLAM, and a regeneration check of pruned trajectories, will be included in the revision.
The following table reports the inference cost on the same test set. Inference steps and tokens are measured during evaluation. Excess tools are the generated tool calls beyond the ground-truth calls.
τ-bench is difficult for small open-source models. On retail, xLAM reaches 65/115, above 50%, but remains below mainstream closed-source models. On airline, xLAM-2-8b-fc-r reaches only 35.2, too low for efficiency optimization to be meaningful, so we focus on τ-bench retail. Under the time limit we ran one iteration. Accuracy rises from 65/115 to 70/115. The gain is on the 76 task ids that were not used to build training pairs, from 32/76 to 40/76. On the 30 unseen tasks solved by both the original and the updated model, the total trajectory length falls from 950 to 936. More results will be put in the revision soon.
The validity check used for pruning is the official BFCL multi-turn checker, the tool the benchmark itself uses to decide whether an answer is correct. It is bound to the scenario of each task. For a given example, the checker instantiates the simulated tool backends from that example's initial state and involved tool classes, executes both the proposed calls and the ground-truth calls, and applies two tests after every turn. The resulting backend state must match the ground-truth state, and every ground-truth tool return for the current turn must already be covered by the returns obtained so far, including returns from earlier turns. A call is removed only when both tests still pass. A limitation of our method is that we do not provide this evaluation ourselves. Deciding whether a backend state is sufficient to solve the task is a criterion that should be supplied by the benchmark authors, so the pruning stage applies to benchmarks that provide such a checker.
Excess is the model's tool-call count minus the call count of the benchmark's annotated ground-truth trajectory, on the same samples. It measures how many extra calls remain relative to that reference, not the target of training. During training we rank sampled trajectories against each other by their own tool-call counts. On a fixed set of samples, the ground-truth count cancels out, so the reported reduction in excess equals the reduction in the model's own calls. A longer or shorter reference trajectory would shift the absolute excess of every method by the same amount, and would not change the reduction.