A new training trick lets AI agents grade their own work mid-task, without hiring another AI to watch.
Researchers studying how to train large language models on multi-step tasks found a workaround for one of reinforcement learning's annoying problems: knowing which step in a long chain of actions actually helped. The standard method, Group Relative Policy Optimization (GRPO), only scores a task after it's done, so a model that nails steps one through nine and botches step ten gets the same verdict as one that failed from the start. Teams have tried patching this with full task replays, outside AI judges checking each step, or reward checks that need the correct answer already in hand - all of which are slow, expensive, or impractical outside of controlled benchmarks. The new paper proposes PAIR (Prefix-Aware Internal Reward), a model that reads an LLM's own internal hidden states to guess whether each step is on track, without any external help.
It matters because PAIR fixes a specific failure mode other papers glossed over: once an early step in a chain goes wrong, a hidden-state probe tends to just agree with the model's own confused narrative, measuring consistency with a bad prefix rather than grounded correctness. PAIR adds a second, attention-based check that resists that kind of contamination, and the paper reports the best accuracy on exactly those messy trajectories while running cheap enough to score every step during training, not just the final output.
That is a real gap, not just a tweak to an existing benchmark - but the result lives on one paper's test set so far, and a claim of negligible inference cost is the kind of thing that gets more expensive the moment someone else tries to run it at scale.