An AI-agent training method that turns a task's final checker into a guide for earlier decisions has reported large gains on two benchmark suites. VICT was tested on ALFWorld and WebShop, and with Qwen2.5-1.5B it exceeded GRPO by 18.2 points on ALFWorld average success and 24.9 points on WebShop strict success.
The question behind the work is whether a terminal verifier can do more than act as an outcome oracle. The authors ask whether it can trace training credit through a long sequence of actions. Their hypothesis is that verifier-backed credit can reward useful search or state-changing actions before an incorrect final commit without reinforcing the mistake.
Turning a final check into a trail of evidence
VICT exposes a verifier as executable or evidence-backed atoms, meaning smaller facts that can be checked or tied to evidence. It maps those atoms to observable evidence in an agent's trajectory and builds proof edges from actions to the facts they support. During training, it redistributes the advantage signal within each rollout group only through those verified links.
It keeps the original terminal reward and changes only the training-time advantage signal. When reliable proof is missing, it abstains. The design also requires no learned critic, process labels, branch rollouts or verifier access when the agent is running.
For ALFWorld and WebShop, the main setup used 16 task groups per update, with eight rollouts per group, across 300 updates. That made 128 rollouts per update and 38,400 training rollouts per seed. The evaluation sets contained 134 ALFWorld tasks and 500 WebShop tasks. The models were Qwen2.5-1.5B and Qwen2.5-7B Instruct, with three random seeds, a learning rate of 1 × 10−6, a KL coefficient of 0.01, and episode caps of 50 and 15 steps.
Success on ALFWorld and both normalized score and strict success on WebShop were the primary endpoints. A separate test reported pass@1 for Retail and Airline interactions in τ-bench.
The gains appeared during training as well as at the end
With Qwen2.5-7B, VICT reached 93.7 average success on ALFWorld and 83.6 strict success on WebShop. The reported gains over GRPO were 16.1 and 17.5 points. Every tested ablation or negative control lowered both primary metrics; removing the dependency core or proof edges cost 3.3 to 4.9 points compared with full VICT.
VICT also looked stronger across the training curve, not only at the final score. Validation AUC, which summarizes the validation-success curve over the same updates, was 78.9 plus or minus 1.4 on ALFWorld and 70.8 plus or minus 1.2 on WebShop. GRPO recorded 58.4 plus or minus 3.8 and 52.7 plus or minus 2.9. These are means and standard deviations over three seeds.
In the supplemental τ-bench test, conducted with Qwen3-8B under a separate simulated-user protocol, VICT reached pass@1 scores of 56.6 in Retail and 45.1 in Airline. Fission-GRPO scored 51.3 and 40.0. Because the backbone and protocol differed from the primary experiments, the authors treat this comparison as supplemental.
Checks support eligibility, not proof of cause
A hand audit examined 150 credited proof edges per domain. Agreement between two annotators was Cohen's kappa, an inter-annotator agreement statistic, of . In a fixed-continuation check, replacing 60 credited actions with no-ops reduced terminal verifier scores in 55 of 60 ALFWorld cases and 52 of 60 WebShop cases. The paper says this does not establish general causality.
Before policy updates, the eligibility-pass value was 100.0 in each listed domain. Proof coverage was 92.4 in ALFWorld, 88.7 in WebShop, 86.5 in τ-bench Retail and 84.9 in τ-bench Airline. Abstention was 8.1, 11.6, 14.3 and 15.8 in the same order. In other words, the method did not assign verifier-based credit to every trajectory.
A method tied to what the verifier can see
The method's scope follows the verifier. It requires executable or evidence-backed atoms and observable trajectory evidence, so the reported approach applies to tasks where those ingredients are available. When the evidence is not reliable enough, it abstains rather than force a credit assignment.
That machinery also takes time to build and run. Instrumentation audits reported 118 to 238 lines of code and 6.5 to 11.5 hours of engineering time, while training-time overhead on ALFWorld and WebShop ranged from 11.9% to 16.7%. The paper says the proof-edge audit does not establish general causality, so these figures do not show that any credited action was causally necessary.
The manuscript is an arXiv preprint, version 1, dated 28 August 2026. Its main evidence comes from ALFWorld and WebShop, with τ-bench reported separately under a different backbone and protocol.
Paper data and sources
Original title: VICT: Verifier-Instrumented Credit Tracing for Long-Horizon LLM Agent Reinforcement Learning
Authors: Pengcheng Li, Zhengyang Zhang, Dongxu Zhang et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-28
DOI: Not available
Original paper · Full text