TACT: Activation Steering to Reduce Overthinking and Overacting in Coding Agents
Researchers have unveiled TACT (Think-Act Calibration via activation Steering), a technique aimed at identifying and addressing agent drift in language model agents throughout extensive software engineering processes. Agent drift appears in two forms: overthinking, which involves excessive reasoning on familiar information, and overacting, characterized by making tool calls without considering new data. By examining hidden states in the residual stream, the researchers discovered that steps marked as overthinking, overacting, or calibrated align linearly along two drift axes (AUC ≈ 0.9). During testing, TACT maps each step's activation onto these axes, correcting drifted states to restore calibrated behavior. The research paper can be found on arXiv with ID 2605.05980.
Key facts
- TACT stands for Think-Act Calibration via activation Steering.
- Agent drift is defined as degradation over long trajectories in coding agents.
- Two failure modes are overthinking and overacting.
- Hidden states of trajectory steps separate linearly along two drift axes.
- AUC of the separation is approximately 0.9.
- TACT projects activations onto drift axes at test time.
- The method pulls drifted activations back toward calibrated behavior.
- The paper is on arXiv with ID 2605.05980.
Entities
Institutions
- arXiv