Same Signal, Opposite Meaning: Direction-Informed Adaptive Learning for LLM Agents

Researchers demonstrate that adaptive compute gates for LLM agents produce unstable and reversible signals across different environments and models, where the same confidence metric predicts both beneficial and harmful outcomes. They propose DIAL, a learned gating mechanism trained through counterfactual exploration, which outperforms fixed-direction baselines by accounting for task-specific utility directions.

This research addresses a critical failure mode in adaptive inference systems for large language models. Current methods rely on static assumptions about when additional computation helps—typically using confidence or uncertainty as gates. The work reveals that these assumptions break down systematically: a low-confidence signal might indicate states where extra computation helps in one environment but harms performance in another, even with identical tasks and different model backbones.

The distinction between compute need and compute suitability represents an important conceptual advance. High uncertainty can signal either decision-difficult states where reasoning benefits from expanded search, or intervention-unsuitable states where the model lacks sufficient context for productive rollout exploration. This bidirectional relationship explains why naive gating mechanisms can actively degrade performance by selecting wrong states for computation allocation.

DIAL addresses this by treating gate direction as a learnable parameter per environment-backbone pair, discovered through signal-agnostic counterfactual exploration. Rather than imposing fixed utility directions, the system learns which states genuinely benefit from extra compute in its specific configuration. Across six environments and three backbone models, this approach achieves superior success-cost trade-offs.

For the AI research community, this work highlights the brittleness of seemingly robust adaptive inference heuristics and validates the value of environment-specific calibration. The findings suggest that simple scaling of compute at test-time requires more sophisticated gating than confidence thresholds. Future systems deploying adaptive compute in production environments should account for backbone-specific and task-specific utility reversals rather than assuming universal signal directions.

• →Fixed-direction gating signals for LLM adaptive compute show reversed utility across environments, sometimes selecting harmful rather than beneficial computation states
• →DIAL learns direction-specific gates per environment-backbone pair through counterfactual exploration, outperforming assumptions of static utility directions
• →High uncertainty signals are ambiguous—indicating either decision-difficult states where reasoning helps or intervention-unsuitable states where rollouts fail
• →Naive adaptive compute mechanisms can actively harm performance by precisely selecting states unsuitable for extra computation
• →Environment and model architecture substantially affect whether gating signals correlate with compute utility, requiring calibration rather than universal heuristics