Human decision-makers terminate evidence accumulation using flexible decision rules

This study demonstrates that human decision-makers employ flexible, adaptive rules to terminate evidence accumulation in response to varying task conditions, thereby optimizing the trade-off between speed and accuracy to maximize reward rates in a manner consistent with rationally bounded normative principles.

The Gist

The Big Idea: The Brain's "Stop Button"

Imagine you are trying to decide which way to turn while driving. You see a sign, but it's a bit blurry. Do you turn immediately (fast but risky), or do you wait a few more seconds to get a clearer view (slow but safe)?

In psychology, this is called evidence accumulation. Your brain gathers clues until it feels confident enough to make a choice. The moment you decide to stop gathering clues and commit to a choice is governed by a "decision rule."

For decades, scientists thought our brains used a fixed rule: "I will wait until I have gathered X amount of evidence, no matter what." It's like setting a timer on your phone: "I will wait exactly 10 seconds, then I must decide."

This paper argues that our brains are much smarter and more flexible than that. We don't just use a fixed timer. We have a smart, adjustable stop button that changes based on the situation, the cost of being wrong, and how clear the information is.

---

The Experiment: The "Pigeon Game"

To test this, the researchers created a video game called the "Pigeon Task."

• The Setup: A pigeon starts in the middle of a screen and walks randomly toward either a left or right pile of seeds.
• The Goal: You have to press a key to guess which pile the pigeon will eventually reach.
• The Catch: You have a limited budget of "steps" (time).
  • If you guess too fast, you might be wrong (low accuracy).
  • If you wait too long, you run out of steps (low speed).
  • If you are wrong, you lose coins or steps.

Because the pigeon's path was shown on the screen, the researchers could see exactly how much "evidence" (how far the pigeon had walked) you had when you decided to stop. This allowed them to see your "decision rule" in real-time.

---

Three Key Discoveries

The researchers tested three different scenarios to see how people adjusted their "stop buttons."

1. Changing the Stakes (The "Cost of Being Wrong" Test)

The Scenario:

• Round A: If you are wrong, you lose nothing.
• Round B: If you are wrong, you lose a lot of money.
• Round C: If you are wrong, you lose a lot of time.

The Result:

• In Round A (no penalty), people were reckless. They stopped very early, guessing quickly because being wrong didn't hurt.
• In Round B (money penalty), people became cautious. They waited longer, gathering more evidence to be sure, because the cost of a mistake was high.
• In Round C (time penalty), people didn't change much. Since losing time was the penalty, waiting longer didn't help much; they just kept their usual pace.

The Analogy:
Think of this like crossing a street.

• If you are in a parking lot (no penalty), you might dart across without looking.
• If you are in front of a speeding truck (high penalty), you wait until you are 100% sure it's safe.
• If you are just trying to save 5 seconds of your life (time penalty), you might not change your behavior much because the risk/reward balance is different.

Takeaway: We adjust our patience based on how much we have to lose.

2. Changing the Clarity (The "Foggy vs. Clear Day" Test)

The Scenario:

• Block 1: The pigeon's path is very clear (High Signal).
• Block 2: The pigeon's path is very foggy and hard to see (Low Signal).
• The Twist: Sometimes the fog changed between blocks (predictable). Sometimes the fog changed randomly on every single step (unpredictable).

The Result:

• Predictable Fog: When people knew a block would be foggy, they adjusted. They waited longer and gathered more evidence because the clues were weak. When the air was clear, they stopped sooner.
• Unpredictable Fog: When the fog changed randomly every second, people did not adjust. They used the same "stop button" for both clear and foggy moments.

The Analogy:
Imagine you are a detective.

• If you know you are entering a dark room (predictable fog), you bring a flashlight and look around carefully before making a guess.
• If you are in a room where the lights flicker randomly (unpredictable fog), you can't adjust your strategy for every flicker. You just use your best average guess and hope for the best.

Takeaway: We can adapt to predictable changes in difficulty, but we struggle to adapt to random, moment-to-moment chaos.

3. Changing Mid-Stream (The "Plot Twist" Test)

The Scenario:
The pigeon starts walking in the fog (low clarity), but halfway through the walk, the fog suddenly lifts (high clarity). Or vice versa: it starts clear, then gets foggy.

The Result:
People noticed the change!

• If the clues got better (fog lifted), they waited a bit longer to gather the new, high-quality evidence.
• If the clues got worse (fog rolled in), they stopped gathering evidence sooner because the new clues weren't worth the extra time.

The Analogy:
Imagine you are cooking a stew.

• You start tasting it, but the flavors are muted (foggy). You decide to wait.
• Suddenly, you add a powerful spice (clarity improves). You realize, "Wow, now I can taste it perfectly!" so you wait a little longer to get the perfect flavor.
• Or, you add too much salt (clarity gets worse). You realize, "This is ruined," so you stop tasting and just serve it (or throw it out) immediately.

Takeaway: We can change our minds during a single decision if the situation changes.

---

The "Good Enough" Conclusion

The most important finding isn't just that we are flexible, but how flexible we are.

The researchers found that people didn't always make the mathematically perfect decision to get the maximum possible reward. Instead, they aimed for "Good Enough."

The Analogy:
Imagine you are trying to find the highest point on a hill to get the best view.

• The Perfect Optimizer: Would climb every single inch of the hill to find the absolute peak, even if it takes hours.
• The Human: Sees a nice spot with a great view, says, "That's good enough," and sits down. They don't waste energy climbing the last few inches because the view doesn't get much better.

Why? Because our brains are "bounded." We have limited energy and time. If the reward for being slightly more accurate is tiny, our brains say, "Not worth the extra effort." We settle for a "satisficing" (satisfying + sufficing) strategy.

Summary

This paper shows that human decision-making is not a rigid robot process. We are like smart drivers:

1. We slow down when the road is dangerous (high cost of error).
2. We speed up when the road is clear (high evidence quality).
3. We react to sudden changes in the road (mid-decision adjustments).
4. But, we only do the math if it's worth the effort. If the view is "good enough," we stop looking for the perfect spot.

We are flexible, adaptive, and surprisingly rational, but we are also practical and lazy when the extra effort doesn't pay off.

Technical Summary

1. Problem Statement

Decision-making models, particularly sequential sampling models like the Drift-Diffusion Model (DDM), traditionally assume that decisions are terminated when accumulated evidence reaches a fixed, predefined threshold (decision bound). This bound balances the trade-off between speed and accuracy. While "fixed" bounds are optimal for static environments, and "collapsing" bounds are optimal for specific dynamic conditions, existing research has two major limitations:

1. Trial-Averaging Bias: Most models infer decision rules by fitting data across many trials, obscuring the rules governing individual decisions.
2. Static Assumptions: Existing models often assume decision-relevant quantities (like evidence quality or reward structure) remain constant within a trial. It is unclear how humans adapt decision rules when these quantities change dynamically within a single decision process.

The core question is: Do human decision-makers use flexible, adaptive rules to terminate evidence accumulation in response to changing task conditions (rewards, costs, and evidence quality) across and within trials?

2. Methodology

The authors developed a novel experimental paradigm called the "Pigeon Task" to directly observe decision variables and termination rules on a trial-by-trial basis.

• Task Design:

  • A visual stimulus (a pigeon) performs a biased random walk from a central starting point toward one of two seed piles (left or right).
  • The pigeon's position represents the decision variable.
  • Participants press a key at any time to stop the walk and commit to a choice.
  • Feedback: Continuous feedback is provided on total coins earned and steps taken.
  • Constraints: Each block is limited to 600 total pigeon steps. Decisions consume steps (time cost); errors incur penalties (coins lost or steps deducted).
  • Advantage: Unlike standard tasks where the decision variable is latent, here it is explicitly visible, allowing direct measurement of the decision bound (pigeon position at the moment of choice).

• Experimental Cohorts:
  Three cohorts of 60 participants each performed variants of the task manipulating different variables:

  • Cohort 1 (Reward/Cost Structure): Tested three blocks with different reward/cost structures (e.g., Block 1: No error penalty; Block 2: High coin penalty for errors; Block 3: High step penalty for errors).
  • Cohort 2 (Across-Trial SNR Variability): Tested blocks where Signal-to-Noise Ratio (SNR) was either fixed per block or varied randomly between trials.
  • Cohort 3 (Within-Trial SNR Variability): Tested blocks where SNR changed predictably within a single trial (e.g., Low $\to$ High or High $\to$ Low) at a specific, cued time point.

• Data Analysis:

  • Non-Decision Time (NDT): Estimated by finding the delay that maximizes the congruence between the pigeon's position and the participant's choice.
  • Decision Time (DT): Response Time (RT) minus NDT.
  • Decision Bound: Defined as the absolute distance of the pigeon from the center at the DT.
  • Bias Correction: A simulation-based linear regression was used to correct for estimation biases at short decision times.
  • Normative Modeling: Simulations were run to calculate the optimal reward-rate functions for various bound strategies under the specific task conditions.

3. Key Contributions

1. Direct Measurement of Individual Bounds: The study moves beyond trial-averaged inference, providing empirical data on the decision bounds used for individual decisions.
2. Evidence of Flexibility: It demonstrates that humans do not rely on a single static rule but employ flexible strategies that adapt to the temporal structure of evidence and rewards.
3. Distinction Between Predictability and Adaptation: The study isolates how predictability (block-wise vs. trial-wise vs. within-trial) dictates the ability to adapt decision bounds.
4. Satisficing vs. Maximizing: It provides evidence that humans aim for "good enough" (satisficing) reward rates rather than strict mathematical optimization, particularly when reward-rate functions are shallow.

4. Key Results

A. Baseline Behavior

• Participants generally used bounds consistent with accumulate-to-bound models.
• Bounds were relatively constant over time (with a slight collapse) and varied noisily around a mean.
• There was a strong positive correlation between average decision time and average bound height across participants, confirming the speed-accuracy trade-off.

B. Adaptation to Reward/Cost Structures (Across Blocks)

• Participants adjusted bounds systematically based on the cost of errors.
  • Low Cost (No penalty): Used lower bounds (faster, guessing strategy).
  • High Cost (Coin penalty): Used higher bounds (slower, accuracy-focused).
  • High Cost (Step penalty): Showed minimal adjustment, as the step penalty discouraged guessing regardless of bound height.
• Conclusion: Humans adapt to block-level changes in reward structure to maximize reward rate, though they tend to "satisfice" rather than strictly optimize.

C. Adaptation to Evidence Quality (SNR)

• Predictable Across-Block Changes: When SNR changed between blocks (e.g., Low SNR block vs. High SNR block), participants adapted by using higher bounds for Low SNR and lower bounds for High SNR. This aligns with normative theory: low-quality evidence requires a higher threshold to accumulate sufficient weight for an accurate decision.
• Unpredictable Across-Trial Changes: When SNR varied randomly between trials within a block, participants did not adjust bounds. They used a single, average bound (closer to the High SNR regime) for both conditions. This suggests humans do not infer trial-specific SNR to adjust bounds online in unpredictable environments.
• Predictable Within-Trial Changes: When SNR changed predictably within a trial (e.g., Low $\to$ High):
  • Participants adjusted bounds after the changepoint.
  • Low $\to$ High: Bounds increased after the change (waiting for better evidence).
  • High $\to$ Low: Bounds decreased after the change (avoiding wasting time on degrading evidence).
  • The direction of adjustment correlated with the gradient of the reward-rate function, indicating a rational, bounded implementation of normative principles.

5. Significance and Implications

• Neural Mechanisms: The findings suggest that neural correlates of decision bounds (observed in areas like LIP, FEF, and SC) are not fixed thresholds but flexible control variables. The brain likely implements dynamic adjustments based on the predictability of the environment and the specific goals (reward rate) of the decision-maker.
• Theoretical Refinement: The study challenges the universality of fixed-bound models. It supports a "bounded rationality" framework where humans use adaptive rules that are "good enough" (satisficing) rather than perfectly optimal, especially when the cost of fine-tuning outweighs the marginal gain in reward.
• Predictability is Key: The ability to adapt decision rules depends heavily on the predictability of the change. Humans adapt well to block-level and within-trial predictable changes but fail to adapt to trial-by-trial unpredictable changes, likely due to computational constraints or the lack of a reliable inference mechanism for trial-specific SNR.

In summary, this paper provides robust empirical evidence that human decision termination is a highly flexible process, dynamically tuned to the statistical structure of rewards and evidence quality, provided those changes are predictable.