The paper examines the potential of internal model representations for gauging confidence in agentic systems. Agentic workflows present complex failure modes involving planning, tool invocation, and dynamic environment interactions, distinguishing them from traditional machine learning systems. The research focuses on whether changes in residual-stream representations across interaction trajectories – captured by the Latent Trajectory Dynamics (LTD) method – can predict eventual task success. Simultaneously, the Action Representation Probe (ARP) method predicts success from representations formed at action decisions. Across three interactive benchmarks (Bash, SQL, Python) and three model families (Qwen14B, Qwen7B, DeepSeek6.7B), LTD and ARP consistently outperformed surface level generation and sequence-based calibration baselines. The methods provide a zero-overhead reliability monitor without prompt alterations or multi-sample rollouts. This allows for continuous monitoring of agent confidence during operation.
Source: https://arxiv.org/abs/2609.09448