Can a causal graph explain why a reinforcement learning agent acted? ── A preprint proposing CausalXRL, which learns the causal structure behind an agent's decisions and returns explanations at more than one level

1. What problem does the paper address?

The problem is that a reinforcement learning agent can score well while people remain unable to inspect the reasons for its decisions. Reinforcement learning trains behaviour from trial and reward, and it is widely used to train autonomous agents. The authors note that this success often comes at the cost of interpretability, which lowers trust and makes it harder to debug and improve agent behaviour^[1].

Being unable to inspect a decision is a separate problem from performance. An agent with good scores may still be relying on cues nobody intended it to use. If the reasons are hidden, there is also no clear place to start when performance drops. The paper is therefore best read as work on making learned behaviour readable, not as work on raising scores. The source is a preprint on arXiv, so everything below should be read as the authors' claims and as results within the authors' own evaluation.

2. What does it propose?

The core idea is to represent the agent's decision-making as a causal graph and to use that structure as the explanation. A causal graph draws the factors involved in states and actions as nodes, with arrows showing which factor influences which. The authors claim that reasoning over this graph yields explanations that are transparent, structured and organised at multiple levels^[1].

The multi-level aspect appears to be what sets the proposal apart. The abstract says that, in a real-time strategy game, participants found the method useful for inspecting high-level strategies^[1]. The aim seems to be explaining not only why a single move was taken but also which plan that move belongs to. How the graph is learned, how many levels there are and in what form the explanation is shown to people are not described in the abstract.

3. What does it show, within the abstract?

The abstract reports case studies and a two-part evaluation^[1].

• A quantitative analysis, in benchmark reinforcement learning environments, of explanation fidelity and of how efficiently the causal structure is learned. The authors state that CausalXRL gave faithful explanations while learning causal structures efficiently.
• A qualitative expert study in MicroRTS, a real-time strategy benchmark. Participants reportedly found CausalXRL useful for inspecting high-level strategies.

The abstract does not say which environments were used, what the method was compared against, or how faithful the explanations were in numbers. The number of experts and the study procedure are not given either. What can be confirmed from the abstract is that the authors measured fidelity and learning efficiency and report favourable results. MicroRTS is published as a two-player research game that is deliberately simpler than full commercial strategy games^[5]. An evaluation in that setting does not automatically carry over to explanations in more complex environments.

4. A worked example

Consider a quality reviewer inspecting a support system that uses reinforcement learning to order work items. This is an illustrative assumption, not an experiment from the paper, and it does not imply that reinforcement learning is commonly used for such ordering in pharmaceutical operations.

5. What is not new, and where are the limits?

Using causal models to explain reinforcement learning is not new in itself. Madumal and colleagues earlier described learning a structural causal model during training and generating explanations from counterfactual analysis, with a user study built on StarCraft II^[4]. Where CausalXRL stands depends on how it improves on such work in multi-level explanation and in the efficiency of structure learning. The abstract does not provide that comparison.

• Fidelity depends on how it is measured. The abstract says the explanations were faithful but does not name the metric or the baseline. Some metrics can reward explanations that merely sound plausible.
• The expert study is qualitative. Participants finding the tool useful is not the same as showing that explanations helped them find faults faster or make fewer wrong judgements.
• The environments are limited. The results come from benchmark environments and MicroRTS. Whether the structure can be learned as well in real work settings with many more state factors is not shown.
• On review status. The comments field on the arXiv page carries the authors' note that the paper was accepted to ICML 2026^[1]. This article, however, relies on the abstract of the preprint posted on arXiv, and what was checked during review cannot be determined from it.

6. Why is this topic clustering now?

In the selection record, this paper heads a cluster of size 3. The cluster also contains a paper that reframes continual reinforcement learning as an integration problem^[2] and a paper that uses hierarchical hints as rewards for mathematical reasoning^[3]. Only the cluster signal fired, the one driven by researchers writing papers. The paper had 0 upvotes and 0 GitHub stars at selection time, so neither readers' votes nor implementers were following it yet.

The keywords shared by the three papers include reinforcement, reward and reasoning. One reading is that, in the same week, several papers revisited reinforcement learning not just as a tool for raising scores but in terms of how rewards are given, how learning continues and how learned decisions can be read. The cluster, though, is built mechanically from vocabulary overlap, and the three papers do not answer the same question. It cannot be concluded from this cluster that explainability is becoming a stronger research trend.

7. Where does this connect to pharmaceutical and regulatory work?

The connection is the expectation that a learned system's decisions can be checked by people after the fact. When computerised systems are used in pharmaceutical operations, organisations have long been expected to explain and verify what the system did and how. A system that learns its policy from trial and error, as reinforcement learning does, is harder to explain than one built from written rules. Presenting decisions as causal structure can be read as an attempt to address that difficulty.

The paper also suggests concrete questions for adoption. Does the causal direction shown in an explanation make sense to domain experts? How will anyone test that the explanation matches the actual decision? And can it be shown from records, not just from opinion, that reviewers who read the explanations actually caught errors? Within the abstract, the third question has not yet been addressed.

The paper does not evaluate pharmaceutical or medical tasks, and it does not consider regulatory requirements. What it supports is a way of framing how explanation mechanisms should be assessed. It does not support the conclusion that adding causal-graph explanations makes decisions safer.

The question is not whether an explanation is attached, but whether it matches the actual decision and whether it helped someone find an error.