When correlation breaks an agent
Most RL learns from correlations — which states tend to precede high return. That is fragile, because a correlation can come from a confounder: a hidden cause that moves both the action and the outcome. Train a driving agent on human logs where the wipers are usually on when it rains, and it may learn 'wipers on → safe' — copying a marker of the cause rather than the cause itself. The instant the world changes, the spurious link snaps. This failure is most acute precisely in the offline and logged-data settings the frontier most wants to use.
Diagram of the agent–environment loop: the agent takes an action and receives a next state and reward.
Causal RL — interventions, not just observations
Causal RL equips agents with a model of cause and effect, not just statistics. Using structural causal models and the language of do-calculus, it distinguishes 'what tends to co-occur' from 'what happens if I intervene and do this.' A policy is itself an intervention, so this is the natural language for RL. The payoff lands on the very gaps from Guide 1: causal structure is the invariant that survives distribution shift, so it promises sturdier generalization, and reasoning about mechanisms — not just rewards — can sharply improve sample efficiency.
The heart of do-calculus: the effect of intervening to set x can differ from merely observing x — exactly the distinction a causal agent must respect.
Counterfactuals and credit
Causality also sharpens RL's oldest headache, credit assignment. A counterfactual question — 'would the outcome have been better had I acted differently, holding everything else fixed?' — is exactly the comparison an advantage tries to estimate, now on a principled footing. Counterfactual reasoning underlies cleaner off-policy evaluation and, in multi-agent settings, lets you isolate one agent's true contribution by imagining the world without its action.
The advantage function captures the counterfactual at the heart of credit assignment: how much better is action a than the state's baseline value?
The frontier, assembled
Step back and the pieces of this track form one picture of what a general agent would need. From Guide 1: cross the generalization and sample-efficiency gaps. From Guide 2: reuse knowledge — adapt fast, transfer, never forget, learn without rewards. From Guide 3: scale via the foundation-model recipe, with scaling laws and open-ended environments feeding endless diverse experience. From Guide 4: specify the right objective and stay safe. From this guide: ground it all in causal understanding. No single idea is sufficient; the frontier is their intersection.
An honest closing
These are open problems, not solved ones — which is the good news for you. RL today is roughly where supervised learning sat before its scaling era: powerful in narrow wins, searching for the recipe that makes it broadly reliable. If you want to contribute, pick one gap and make it measurable: build a benchmark that forces generalization, an environment generator that never repeats, a reward that resists hacking, or a causal method that survives a shift no correlational baseline can. The agents that finally learn the way we hoped will be built on exactly these questions.