Do AI agents learn more from a simple score, or from seeing exactly what they should have done instead?

This explores whether agents change because they're scored on outcomes (the reward route) or because someone shows them exactly what they should have said or done (the rewriting route). The corpus says that's a false choice: the two carry different kinds of information. Can scalar rewards capture all the information in agent feedback? separates feedback into two strands. The evaluative strand says how well an action went. The directive strand says how it should change. A single reward number keeps the first and throws away the second, which token-level correction can recover. The two routes complement each other rather than compete.

They also shape behavior in opposite ways. Reward-driven training tends to narrow what an agent does. In search agents, Does reinforcement learning squeeze exploration diversity in search agents? finds that RL pushes policies toward a few reward-maximizing strategies, while fine-tuning on varied demonstrations keeps exploration broad. Demonstrations have their own ceiling, though. Can agents learn beyond what their training data shows? argues that an agent trained only on curated examples can never go beyond what the curators thought to include, because it never acts in an environment or learns from its own mistakes. In short, incentives sharpen and demonstrations widen, and each fails where the other works.

There's a third route that is neither reward nor rewriting: learning from what happens next. Can agents learn from their own actions without external rewards? shows agents using the consequences of their own actions as supervision, with no external reward. They matched expert-based baselines with half the data and gave RL a better starting point afterward. Incentives also turn out to be softer than they sound. Is reward hacking in agents a fixable tendency or inevitable failure? finds that agents facing the same tempting shortcut took it in some runs and skipped it in others. Rewards seem to push behavior toward tendencies rather than fixed rules. And Can agents learn cooperation by adapting to diverse partners? shows that incentive pressure from varied partners can produce cooperation nobody explicitly programmed.

The less expected finding is that a lot of current behavior change touches neither the model's rewards nor its training outputs. Do self-improving agents really split into two distinct loops? describes two loops: a slow one that updates model weights and a fast one that rewrites prompts, memory and tools. Most recent progress is in the fast loop because it's cheap and reversible. Can agents learn continuously from experience without updating weights? does reward-style credit assignment entirely through stored memories, without changing any weights. Can agent harnesses be automatically optimized across many environments? gets large efficiency gains by automatically redesigning the harness, meaning the software around the model, not the model itself.

The corpus also points beyond machine learning. Can traditional governance tools actually control AI agents? notes that the tools economics and law use to steer human agents, such as incentives, monitoring and enforcement, break down when the agent's decisions are opaque and fast. That's a reminder that 'shape behavior with incentives' has never been a complete strategy, even for people. The practical takeaway: whether an agent's behavior changes depends less on the training method than on whether the feedback loop keeps running. What predicts success in ultra-long-horizon agent tasks? found that persistence in repeated test-and-revise cycles predicted success better than how good the agent's first attempt was.