Agent flip-flops usually occur near an ambiguous decision boundary, where company policies and available information do not clearly determine the right label.
2
Model disagreement is useful for active learning because it identifies the examples where human review can clarify labels or add missing information.
3
Semantic memory for domain policies and episodic memory for similar past cases can improve consistency without relying only on expensive fine-tuning.
Summary
Diane Lin explains why the same AI agent can give materially different answers for the same input. In sentiment analysis and cybersecurity alert triage, these flip-flops tend to happen near a decision boundary where the policy is unclear, important facts are missing, or human experts may reasonably disagree. Lin recommends treating disagreement as an active-learning signal. Running a model several times or comparing models surfaces cases for targeted human review, rather than requiring someone to inspect every output. Reviewers can clarify labels, add information, or express a company policy. Lin then proposes two forms of memory as a lighter alternative to immediate fine-tuning. Semantic memory stores explicit domain rules, while episodic memory retrieves similar cases and their previous decisions. In an experiment with 93 cybersecurity alerts run three times, episodic memory made about 15% of the alerts consistent, while 10% remained inconsistent for human review. The process also lets the agent adapt to customer preferences.
Lin distinguishes semantic inconsistency from harmless wording changes. The same model and input can produce materially different decisions across runs. In hotel sentiment analysis, a single evaluation run may miss this behavior, so repeated runs are needed to see the full pattern. The impact is sharper in cybersecurity. An agent may classify a failed Gmail login from a suspicious IP as benign in one run and suspicious in another. A suspicious alert may require action to stop an attack, while a benign alert can be ignored. Customers then have to ask which answer they should trust. Lin says this inconsistency can affect a vendor comparison because a consistently behaving product may beat one that flip-flops.
Disagreement clusters near an ambiguous decision boundary
The examples that change labels are concentrated in a gray zone near the decision boundary. A clearly positive hotel review stays positive across repeated runs, while a review containing both relatively positive and negative signals can flip. Lin says human experts may also disagree because the right label depends on the hotel's policy. In cybersecurity, repeated login attempts may indicate an attacker, but an enterprise may not want alerts when the attacker remains outside the environment. A successful password and MFA bypass changes the response because the attacker has entered. The label depends on the customer's notification preference and on information that separates those cases.
The model often exposes ambiguity instead of causing it
Lin argues that these cases are not usually an AI agent's fault. They are close to a boundary where human experts and traditional classifiers also struggle. The agent is pointing to ambiguity that already exists in the data, policy, or context. That changes the response: instead of only trying to make the model more deterministic, the team should find the uncertain cases and clarify what the business wants. Missing facts can matter as much as the label itself. For example, a cybersecurity alert may need to distinguish failed attempts from a successful login before the agent can consistently decide whether the event is benign or malicious.
Active learning focuses review on the examples with the most information
Active learning gives teams a way to inspect a large stream without labeling every item. An initial model predicts on unlabeled production data, then a selection strategy finds cases where the model is likely to learn from review. Traditional approaches can use uncertainty, such as probabilities close to 0.5, or query by committee, where several models or repeated runs disagree. Reviewers check the label and decide whether additional features or context are needed. The clarified labels and added information feed the next training cycle. This directs scarce human attention toward cases where it can change the model, instead of spending that attention on straightforward outputs.
Repeated disagreement is a better signal than the language model's confidence
For language models and agents, Lin recommends refining the selection strategy. She says uncertainty scores from language models are not reliable because a model can be confident while being wrong. Disagreement across repeated runs or different models gives a more useful signal that the verdict needs human guidance. Once those cases are collected, reviewers label them and provide feedback. Fine-tuning remains one possible response, but Lin describes it as expensive. Her proposed alternative is to augment the agent with semantic and episodic memory, which is lighter weight and easier to iterate.
Semantic memory turns company policy into an explicit rule
Semantic memory stores factual domain knowledge and business preferences that sharpen the decision boundary. For a hotel, a policy might say that a complaint about something outside the hotel's control should be classified as positive because the hotel cannot act on it. For a security customer, a password-spray event with repeated failed guesses and no successful login can be classified as benign, while a final successful login makes it malicious. Writing down this information separates cases that previously looked similar. It also gives human reviewers a clearer basis for applying labels consistently.
Episodic memory reuses decisions from similar past cases
Episodic memory applies earlier decisions without first requiring someone to distill their reasoning into a general rule. The agent retrieves a similar case, sees how it was labeled, and uses that decision as a reference. Lin says this reduces human intervention because the original review happened earlier. It is especially useful for recurring cybersecurity alerts and repeated false positives. Cases that match prior examples can be handled automatically. A new case with no matching past group still goes to human review. That review can later produce domain knowledge for semantic memory, so the two forms of memory work together.
The learning loop combines automatic reuse with targeted human review
Lin's workflow uses episodic memory first for recurring situations. Cases that remain unresolved, either because there is no useful past example or because the agent still disagrees, go to a human. The reviewer clarifies the label or adds the missing domain information, which can be stored in semantic memory for future decisions. This approach improves consistency while reducing quality-control work. Teams inspect the subset that active learning identifies as problematic rather than checking every agent output. The feedback also captures how a particular customer makes decisions, allowing the agent to adapt to that environment.
The experiment leaves a smaller set of difficult cases for people
Lin reports an experiment using 93 cybersecurity alerts, each run three times. Without the proposed solution, about a quarter of the alerts flip-flopped in their verdict. After applying the solution with episodic memory, about 15% became consistent, while 10% remained inconsistent. The remaining cases included alerts with no usable similar reference and cases where the agent still disagreed after consulting one. Lin's workflow sends those cases to human review, where additional knowledge can disambiguate them. The result is not that every ambiguity disappears automatically. Episodic memory handles recurring inconsistency, while people spend their time on the cases it cannot resolve.
"The disagreement from different runs or from different models actually give you a more reliable signal where the model actually not sure about its verdict."14:44
Who should watch
You are building an agent whose classifications change across repeated runs and need to explain that behavior to customers.
Your team has too many production outputs to review manually and needs a way to direct people toward ambiguous cases.
You want customer-specific policies and past decisions to affect an agent without making fine-tuning the first response.
