Large language models (LLMs) increasingly operate as components inside larger software systems. In that setting, correctness is not all that matters: operators also need to see the evidence behind an answer, to know when the model is unsure, and to keep latency and token cost predictable. We ask whether a small, transparent control layer can steer how a fixed base model responds, without retraining it or wrapping it in a heavyweight agent. Inspired by active inference, the selector reads a compact task state and picks one of three response styles (direct, grounded, or bounded) by scoring each against an explicit objective over helpfulness, truthfulness, groundedness, calibration, hallucination risk, and verbosity. Every decision it makes is logged. On 2,000 questions spanning HotpotQA, MuSiQue, StrategyQA, and TruthfulQA, the controller makes response style a settable property: support exposure and uncertainty handling rise substantially, and the controlled system is preferred in most pairwise comparisons on evidence-rich tasks, while it loses ground where a concise commitment is rewarded. Task accuracy decreases in aggregate (from 0.318 to 0.233); this is the cost of the shift, not its objective, and the gains are strongly dataset-dependent rather than a uniform improvement in dependable behavior. We quantify the token and latency cost and trace the effect dataset by dataset. In a supplementary study with thinking mode enabled, the control-layer effect replicates on a second model family (DeepSeek-V4-Pro), and the state-conditioned selector outperforms a static always-grounded prompt, which collapses on low-evidence tasks; the policy choice is also robust to the scoring weights. The study concerns controllability rather than task performance: response style can be steered and audited at low cost, and the conditions under which this helps or hurts are reported per dataset.
更多
原文链接
分享
引用
