Some effects on generalization gradients of tandem schedules.

Halpern et al. (1966) trained pigeons to peck a key for food on two color schedules.

One color always paid off. The other color never did.

The team then tested how the birds spread their pecks across new colors that looked like the trained ones.

When the birds stopped pecking during the no-pay color, their peak of responding slid away from that color on the test.

If both colors kept the birds pecking at the same speed, the peak stayed put—even when one color still paid more.

The take-home: final response rate, not reinforcement rate, steers the peak.

Reynolds (1968) from the same lab soon echoed the idea: squash the response to one stimulus and the gradient peak shifts.

Dukhayyil et al. (1973) later pushed back. They ran more birds and found no tidy link between last-minute response rate and peak shift. The clash warns us the rule may be fragile.

Rogers-Warren et al. (1976) looked like another contradiction: when they equalized reinforcement across colors, behavioral contrast vanished. Yet M et al. saw contrast whenever response rates differed. Same pigeons, same keys—only the method changed. The two studies sit side-by-side, not face-to-face.

When you set up a discrimination program, watch the learner’s response rate in each stimulus, not just how often you deliver reinforcement. If one stimulus shuts the client down, expect the peak of correct responses to drift away from that stimulus. Use that drift as a live signal: the client is suppressing, so adjust prompting or reinforcement before errors lock in.