Quick Takeaways
-
Reward labels aren’t guaranteed to teach: The test hid a routing rule that only past success/failure labels could reveal.
-
Models reacted very differently: GPT-6 Sol reached 94.4% accuracy with true labels but just 7.8% with flipped labels; Haiku improved modestly, while Qwen showed little clear learning from true labels.
-
Context matters: These results differ from a paper finding little effect from reward labels across other tasks—and come from one deliberately designed routing task.
-
Test your own agent: Flip its memory labels and see whether performance changes. If it doesn’t, the agent may not be using the rewards.
False rewards changed the route
An AI agent can review its past attempts, along with scores, before making its next choice. The idea sounds simple: learn from success and avoid failure. But does the model actually use those scores?
In a small test, each model chose among four delivery routes. A hidden rule linked each made-up region to one route. Only the reward labels revealed that rule. Then I flipped the labels, marking wrong answers as right and right answers as wrong.
GPT-6 Sol reacted sharply. In later trials, it answered correctly 94.4% of the time with accurate labels, but only 7.8% with flipped labels. It followed the bad advice. Claude Haiku 5.5 improved modestly with accurate labels, while Qwen3 4B showed no clear benefit from them in the main comparison.
One test does not settle the debate
The results suggest that reward labels can guide some models. They also show that models differ. In a fixed-memory test, GPT-6 Sol and Haiku followed flipped labels. Qwen often copied routes it had seen, even when the labels warned against them.
A separate study found that changing or removing reward labels had little effect across several models and tasks. That finding does not directly conflict with this test. Here, the hidden rule made labels the only useful clue. Other tasks may let models draw on knowledge they already have.
The test also had limits. It used one simple routing task and three models. One model’s early results exposed a bias in the region names, and the researcher later widened the confidence ranges after correcting the analysis. Those details matter: a striking result still needs careful checks.
Test your agent before relying on scores
For teams building AI agents, the practical lesson is clear: do not assume a model learns from its memory, and do not assume it ignores reward labels. Test the behavior.
A low-cost check is to compare performance with accurate labels against performance with flipped labels. If the results barely change, the scores may not guide the model. Teams can also compare labeled memories with memories that show only successful attempts.
This experiment cost about $2.55 in API fees and used 0.65 Colab compute units. That makes a small test accessible. Still, teams should repeat it on their own tasks. A useful reward system for one model may confuse another.
Continue Your Tech Journey
Learn how the Internet of Things (IoT) is transforming everyday life.
Discover archived knowledge and digital history on the Internet Archive.
AITechV1
