A new training method for language-model agents recorded large gains on two long-horizon task benchmarks, according to a preprint. Called VICT, it was 18.2 points higher than outcome-level GRPO on ALFWorld average success and 24.9 points higher on WebShop strict success with Qwen2.5-1.5B. At the 7B scale, VICT reached 93.7 average success on ALFWorld and 83.6 strict success on WebShop, gains of 16.1 and 17.5 points over GRPO.
The result addresses a basic problem in long, multi-step agent training. A terminal verifier, essentially a program that checks whether the final task succeeded, can provide an outcome, but it does not automatically show which earlier actions deserve credit. The study asks whether that verifier can become a training-time credit tracer. VICT shifts credit assignment from rollout-side inference to verifier-side tracing, redistributing a group-relative training signal only along verifier-linked edges.
A sharper training signal
The method keeps the original terminal reward as the outcome anchor. It changes only the training-time advantage tensor, the internal signal used to weight actions, and abstains when no reliable verifier proof exists. In practical terms, the system can direct more credit toward actions connected to a valid proof instead of spreading the same outcome signal across an entire rollout.
VICT does not train a critic, require process labels, add branch rollouts, or expose verifier atoms at inference. Its primary runs used Qwen2.5-1.5B and Qwen2.5-7B Instruct models, with eight rollouts in each task group. The training caps were 50 steps for ALFWorld and 15 steps for WebShop.
For both primary benchmarks, the reported scale was 16 task groups per update, 128 rollouts per update and 300 updates. Evaluation covered 134 ALFWorld tasks and 500 WebShop tasks, with 38.4k training rollouts per seed. VICT results were reported as means and standard deviations over three random seeds, while the supplied analysis reports no inferential significance testing.
The benchmark gains
VICT also stayed ahead of the strongest fine-grained baseline in the listed comparisons. At the smaller model scale, its margins were 4.0 points on ALFWorld and 7.0 on WebShop. At the larger scale, the margins were 2.3 and 7.4 points. These are descriptive benchmark comparisons, not a claim that one training design is universally better.
Some results suggest the method can sit alongside rollout-side credit methods. Adding VICT to GiGPO produced reported gains of 3.8 points on ALFWorld and 11.9 on WebShop over GiGPO alone. The corresponding increases were 2.9 points for HCAPO and 9.0 for SALT in the listed comparisons. The combinations were 0.6 to 1.6 points above VICT alone, but the paper did not test every method-domain pair or a full factorial interaction.
A supplemental τ-bench test used the Fission-GRPO protocol. VICT reached 56.6 on Retail and 45.1 on Airline, compared with 51.3 and 40.0 for Fission-GRPO. The backbone, simulated-user setting and protocol differed from the primary experiments, so these figures are supporting evidence rather than a direct extension of the Qwen2.5 results.
What the tests do not show
The component checks pointed in the same direction, but they do not settle why the gains appeared. Replacing the verifier with outcome-only or randomized-edge controls lowered reported scores by 14.3 to 17.5 points. Dense, commit-only and temporal variants fell by 5.1 to 8.2 points, while removing the dependency core or proof edges lowered scores by 3.3 to 4.9 points. The paper says these controls do not isolate statistically independent causal effects.
Audits found that the tracing interface was useful but not complete. Proof coverage was 92.4% in ALFWorld and 88.7% in WebShop, while abstention was 8.1% and 11.6%. In the supplemental τ-bench Retail and Airline settings, coverage fell to 86.5% and 84.9%, and abstention rose to 14.3% and 15.8%. The system therefore sometimes lacked a reliable path for assigning action-level credit.
That boundary is central to interpreting the result. VICT depends on a terminal verifier with usable facts and action-to-proof links, and it is designed to abstain when those links are not reliable. The evidence does not show that credited actions were causally necessary, that the method will win on every task, or that its gains are universally additive with other methods. Because VICT changes the training-time advantage tensor, it also does not provide a policy-invariance guarantee.
For now, the strongest reading is narrow. On tasks with programmatic, auditable verifiers, VICT offers a way to make training credit more selective than an outcome-only update, with gains reported across the primary benchmark blocks. Whether that pattern holds for less instrumentable tasks, other model families and larger or more ambiguous proof structures remains open.
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