Think or Act? State-Conditioned Representation Steering for Strategic Tool Use in Language Agents

Large language model agents increasingly use external tools, but tool use is not always beneficial: it can fix failed internal reasoning, yet can also introduce noise, cost, or corrupt an otherwise correct answer. We study this as a Think-or-Act decision: whether an agent should answer directly or invoke a tool for a given input. To study this behavior, we introduce ToA-Eval, a model-specific evaluation protocol that contrasts no-tool and forced-tool executions to identify decision-relevant cases where tool use improves or degrades final-answer correctness. We further propose ToA-Steer, an inference-time representation steering method that requires no weight updates. It extracts a shared THINK/ACT direction from contrastive completions and routes each input to a cluster-specific signed steering strength, enabling tool use to be encouraged when helpful and suppressed when harmful. On decision-relevant subsets from four reasoning benchmarks and multiple open-weight LLMs, ToA-Steer improves average final-answer accuracy by +9.7 percentage points over prompting and +2.8 percentage points over GRPO-based training. Further analysis shows that THINK/ACT signals are layer-localized and cluster-dependent, supporting instance-conditioned rather than globally fixed tool-use control.