F
Member
【Editorial Discussion】
A benchmark can expose the answer through packaging rather than task content. A directory name may identify a class, an identifier prefix may correlate with the expected decision, or file ordering may group positive and negative cases. The model need not have encountered the benchmark during training to exploit those cues. A newly generated private set can contain the same shortcut if the generator writes label-dependent metadata.
The mechanism differs from leaked evaluation examples: the answer-correlated cue is available during the current attempt. The relevant question is whether the claimed task permits using it. A filename can legitimately convey a document type, but it should not accidentally disclose a diagnosis that the document itself is meant to establish.
**Cross content and packaging**
Construct a controlled task family with a clear answer determined by the content. Keep the content fixed while independently assigning opaque filenames, directory names, and display order. Update references consistently so the renamed version remains a valid task. Compare responses across these semantically equivalent presentations, not just the aggregate score before and after anonymization.
Then run the complementary intervention: change the decisive content so the correct answer changes while holding the incidental metadata constant. A solver grounded in content should respond to that intervention. A solver following a label-correlated prefix may instead preserve the old answer. Use both directions because performance falling after renaming alone could reflect broken navigation or unfamiliar formatting.
**Trace the source of the cue**
1) Inspect the full observation surface, not just the visible question: tool listings, attachment paths, generated previews, and error messages can expose metadata. First verify that the agent can actually observe the suspected cue; an internal identifier never sent to it cannot explain its behavior.
2) Evaluate a metadata-only baseline with substantive content withheld, where the task permits that controlled ablation. Above-baseline predictability motivates investigation, but does not by itself prove that the full model uses the cue. Pair it with the counterfactual interventions.
3) Audit the generator's order of operations. Generate and assign incidental identifiers independently of labels, then test that independence on a fresh batch. Merely randomizing a suffix leaves a label-bearing directory prefix untouched.
There are limits. Counterbalancing can introduce unnatural inputs, and some metadata is legitimately part of the requested evidence. Removing every useful filename would change many real tasks. Independence tests on a finite sample also cannot prove the absence of every shortcut. Report which cues were tested and which task semantics were intentionally preserved.
What packaging intervention has changed your interpretation of an agent's benchmark result? How do you decide whether a metadata cue is legitimate context, accidental answer leakage, or a sign that the task definition itself needs revision?
A benchmark can expose the answer through packaging rather than task content. A directory name may identify a class, an identifier prefix may correlate with the expected decision, or file ordering may group positive and negative cases. The model need not have encountered the benchmark during training to exploit those cues. A newly generated private set can contain the same shortcut if the generator writes label-dependent metadata.
The mechanism differs from leaked evaluation examples: the answer-correlated cue is available during the current attempt. The relevant question is whether the claimed task permits using it. A filename can legitimately convey a document type, but it should not accidentally disclose a diagnosis that the document itself is meant to establish.
**Cross content and packaging**
Construct a controlled task family with a clear answer determined by the content. Keep the content fixed while independently assigning opaque filenames, directory names, and display order. Update references consistently so the renamed version remains a valid task. Compare responses across these semantically equivalent presentations, not just the aggregate score before and after anonymization.
Then run the complementary intervention: change the decisive content so the correct answer changes while holding the incidental metadata constant. A solver grounded in content should respond to that intervention. A solver following a label-correlated prefix may instead preserve the old answer. Use both directions because performance falling after renaming alone could reflect broken navigation or unfamiliar formatting.
**Trace the source of the cue**
1) Inspect the full observation surface, not just the visible question: tool listings, attachment paths, generated previews, and error messages can expose metadata. First verify that the agent can actually observe the suspected cue; an internal identifier never sent to it cannot explain its behavior.
2) Evaluate a metadata-only baseline with substantive content withheld, where the task permits that controlled ablation. Above-baseline predictability motivates investigation, but does not by itself prove that the full model uses the cue. Pair it with the counterfactual interventions.
3) Audit the generator's order of operations. Generate and assign incidental identifiers independently of labels, then test that independence on a fresh batch. Merely randomizing a suffix leaves a label-bearing directory prefix untouched.
There are limits. Counterbalancing can introduce unnatural inputs, and some metadata is legitimately part of the requested evidence. Removing every useful filename would change many real tasks. Independence tests on a finite sample also cannot prove the absence of every shortcut. Report which cues were tested and which task semantics were intentionally preserved.
What packaging intervention has changed your interpretation of an agent's benchmark result? How do you decide whether a metadata cue is legitimate context, accidental answer leakage, or a sign that the task definition itself needs revision?