A preprint reports that IAPO scored 42.18% on the held-out τ²-Bench domain-macro pass1 test at the 8B scale, compared with 29.61% for matched GRPO, a difference of 12.57 percentage points. The comparison comes from computational tests of Qwen3-4B and Qwen3-8B across three service-agent benchmarks, with BFCL-v4 Multi-Turn used to assess function-calling retention.
IAPO is built around a typed influence-dependency graph for each completed interaction. The graph tracks observed support-use and failed-use dependencies among trainable actions, then routes the original trajectory-level advantage—the feedback signal used to update the policy—toward those actions. The method leaves the reward, rollout process, group normalization and clipped-loss implementation unchanged.
How the comparison was set up
The test was designed as a same-protocol comparison. IAPO and GRPO used the same training protocol; their stated difference was the map that assigns a trajectory’s advantage to individual action tokens. Training used the airline, retail and telecom portions of τ²-Bench, a 178-task training split.
The Qwen3-4B and Qwen3-8B policies were trained with thinking disabled on eight NVIDIA H20 GPUs, using 16 samples per prompt and a maximum of 30 turns. Dependency annotation used Qwen3-32B at temperature 0.
The gap appeared across several benchmarks
The pattern extended beyond the headline τ²-Bench comparison. IAPO’s score was 4.61 percentage points higher on UserBench and 4.41 points higher on AgentChangeBench than GRPO’s. On BFCL-MT, however, the two results were close: 39.36% for IAPO and 39.04% for GRPO.
Within τ²-Bench, the gap varied by domain: 6.94 percentage points on airline, 12.57 on retail and 18.19 on telecom. Telecom therefore had the largest reported difference.
On the listed out-of-domain checks, IAPO was higher on every split named in the analysis. The three UserBench travel splits showed differences of 4.97, 5.09 and 3.45 percentage points; AgentChangeBench showed differences of 4.44 points in banking and 4.40 in education.
Function-calling performance was also compared with the untrained base models. At 8B, BFCL-MT was 39.36% for IAPO and 39.38% for the base; at 4B, it was 31.89% and 31.38%, respectively. The supplied analysis describes the 8B result as essentially unchanged and the 4B result as higher than the base.
Where the method changed the feedback
A secondary measure was macro tool-call redundancy. At 4B, it was 29.74 for IAPO versus 46.00 for GRPO.
The routing map did not distribute feedback uniformly. Positive routing weights had a standard deviation of 0.222 and ranged from 0.637 to 1.505. In the audit, 86.5% of positive-branch steps moved more than 10% from uniform credit, compared with 27.1% of negative-branch steps.
An audit across all 178 training tasks and 2,349 annotated action nodes found a 100% parse rate. The reported agreement measures included a mean edge-typed F1 of 0.87, source Jaccard of 0.84 and pairwise Cohen’s kappa of 0.83; routing correlations ranged from 0.73 to 0.77 for the positive branch and 0.96 to 1.00 for the negative branch.
What the test cannot settle
Those figures indicate consistency in parsing and routing within the audit, but they are not validation against an independently established gold standard. They therefore do not show that the extracted edges are ground-truth causal dependencies.
The comparison also has statistical limits. It was a computational model-training experiment rather than a randomized experiment, and the preprint reports no confidence intervals, p-values or formal significance tests. The main trained-method results are given as means with standard deviations over three seeds, with each seed represented by the mean of its final three checkpoints.
The evidence is confined to Qwen3-4B and Qwen3-8B, training on one 178-task τ²-Bench split and the listed benchmark evaluations. These are model-training results, not evidence from human participants or real service outcomes.
The routing design was also tested with a different clipping setting. The preprint describes the default c=0.25 as having the strongest joint in-domain and transfer profile; with c=0.5, the reported scores were 42.29% on τ²-Bench, 17.78% on UserBench and 25.67% on AgentChangeBench.
Taken together, the results are a report of higher benchmark scores for IAPO than matched GRPO in the tested Qwen3 setup, not a causal demonstration that the training method produced those differences. The evidence does not cover performance in other models, settings or real-world services.
The work is labeled a preprint, and its front matter lists WeChat, Tencent Inc. as the affiliation. No funding source is reported in the supplied text.
Paper data and sources
Original title: IAPO: Influence-Aware Policy Optimization for Credit Assignment in Multi-Turn Service Agents
Authors: Bo Ren, Yirong Mao, Yi Yang, Wenhui Que
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-25
DOI: Not available
Original paper · Full text