# VCBM focuses RL training on decisions tied to rewards

> A University of Manchester dissertation tests a method for assigning more credit to consequential actions, with promising results in two different settings.

Oossa · 2026-10-20 · https://oossa.com/en/vcbm-focuses-rl-training-on-decisions-tied-to-rewards

A University of Manchester dissertation introduces Visual Causal-chain Bookmarking (VCBM), a method for helping reinforcement-learning agents learn which actions led to delayed rewards. In a watch-repair simulation, it reached 95.07% correct decisions in an average of 1,532.3 episodes across 20 runs, compared with 3,327.1 episodes for the RUDDER baseline. In AppWorld, VCBM-blended training raised task success from 0.354 to 0.458, but that comparison used only one run per method, so it is not a statistical confirmation.

## The facts

- The watch-repair experiment used 20 seeds; VCBM matched the known causal structure in all of them.
- The AppWorld comparison used Qwen2.5-7B-Instruct and one seed per method.

## Why it matters

For researchers training agents on tasks with delayed rewards, VCBM offers a tested way to emphasize consequential decisions, though the AppWorld result needs more runs before its reliability is clear.

## Sources & references

1. [Visual Causal-chain Bookmarking: Retrieval-Weighted Credit Assignment for Reinforcement Learning from Delayed Reward](https://arxiv.org/abs/2610.24032) – arXiv cs.LG, 2026-10-20

Last updated: 2026-10-20