TL;DR
When LLM agents fail to act, is it because they don't understand the situation or because they are simply "too shy"? This paper introduces FLIPACT, a diagnostic tool that uses Signal Detection Theory to separate a model's sensitivity (its ability to recognize when an action is needed) from its criterion (its internal threshold for acting). The authors demonstrate that while sensitivity scales with model size, the criterion is a tunable "personality trait" that can be shifted via activation steering and distillation without compromising the model's underlying knowledge.
What Problem Does This Paper Address?
Large Language Model (LLM) agents must decide if they should act before they decide how. Current benchmarks usually report a single accuracy score for this decision. However, accuracy conflates two distinct types of failure. As shown in Figure 1a, two models can have nearly identical accuracy while exhibiting opposite behaviors: one might be "conservative" (refusing to call a tool even when necessary), while the other is "permissive" (calling tools when they aren't needed).
Standard Signal Detection Theory (SDT) separates these as sensitivity ($d'$) and criterion ($c$). The challenge is that existing benchmarks use binary labels that don't provide a shared "evidence scale," making it impossible to compare criteria across different tasks or models.
Methods
The authors developed FLIPACT, a diagnostic built on qualitative process theory—essentially, how tools change physical states (e.g., a humidifier adds moisture to a dry room).
Key methodological innovations include:
- The 2x2 Quad Design: Questions are grouped into "quads" where the required knowledge is held fixed while the physical state flips (e.g., testing both a humidifier and a dehumidifier in both wet and dry rooms). This ensures the model must actually read the state, not just memorize tool definitions.
- Rule-Based Evidence ($\Delta U$): Instead of human-authored labels, a rule computes an evidence score ($\Delta U$) representing how much an action helps. This fixes the "Yes/No" boundary at zero for every model.
- Fitting Response Curves: By plotting a model's "Yes-rate" against the evidence score, the authors fit a logistic curve. The slope ($a$) represents sensitivity, and the crossing point ($\tau$) represents the criterion.
- Interventions: The authors used activation steering to find a "criterion direction" in the model's residual stream. They then used distillation to bake this steered behavior into the model's weights, effectively moving the decision threshold without changing the model's sensitivity.
Datasets
- FLIPACT: The primary new benchmark, consisting of 1,880 questions across six physical axes (thermal transfer, moisture, mechanical cohesion, etc.).
- External Benchmarks: To prove their findings generalize, the authors tested on When2Call, BFCL (Berkeley Function Calling Leaderboard), QuestBench, When2Tool, ACPBench, and CaT-Bench.
Key Insights and Ideas
- Sensitivity is a Capability, Criterion is a Setting: Sensitivity ($a$) consistently rises with model size within a family. In contrast, the criterion ($\tau$) is not determined by size; it is a stable, per-model setting. For example, two different releases of Mistral-Small-24B sit on opposite sides of the zero-evidence boundary.
- Models are "Step Functions": LLMs tend to respond to the sign of evidence rather than its magnitude. As shown in Figure 4, Yes-rates jump sharply at the zero-boundary but remain flat even as the evidence becomes much stronger.
- Default Conservatism: When "Yes" means "act," most open models are naturally conservative, often due to a linguistic preference for answering "No."
- Independence of Thresholds: You can successfully move a model's criterion (making it more or less likely to act) without changing its sensitivity. This shift even transfers to benchmarks that don't use a Yes/No format.
Key Results
- Validation: FLIPACT sensitivity ($a$) correlates strongly with external $d'$ scores on three out of four major when-to-act benchmarks.
- Successful Steering: Activation steering moved the criterion toward acting on 11 out of 16 models tested.
- Distillation Transfer: Distilling a model from its own steered copy successfully moved the criterion in the weights. On external benchmarks, accuracy rose whenever the shift moved the model away from its initial bias (e.g., making a conservative model more permissive on a task that requires frequent tool use).
- Preserved Knowledge: During criterion distillation, the "criterion-free AUC" (a measure of pure sensitivity) remained unchanged, proving the model didn't lose its "knowledge" while its "willingness to act" was adjusted.
Limitations and Future Work
- Domain Specificity: FLIPACT currently focuses on basic physical knowledge and a single question template. Whether these findings hold for more abstract reasoning or social common sense remains to be seen.
- Transfer Limits: While the criterion shift generalizes well, supervised training on labels (sensitivity learning) did not transfer stably to external benchmarks, suggesting that "learning to act" is more task-specific than "adjusting the threshold."
- Future Directions: The authors suggest expanding the build to multi-step agent scenarios where a model's tendencies can be tracked across a sequence of interdependent decisions.
Summary updated