Research on LLM-based agents has largely centered on cognitive competence, including reasoning, planning, and tool use. We explore a complementary dimension: the agent's capacity to exhibit state-dependent interactional modulation across extended interactions. We propose RollarTimerAgent, an architecture in which three independent LLM modules form a closed textual affective loop, acting as a semantic control system. A perceptual preprocessing module receives raw user input and produces an affectively annotated version that aims to retain most of the original semantic content while adding affective framing, attention tags, and intuitive associations. This annotated input, together with an internal affective signal from the preceding interaction, feeds into a deliberative core. An Affective Feedback Module then evaluates the completed exchange along five dimensions and injects its assessment back into the deliberative core for the next turn. In preliminary experiments spanning sustained adversarial interaction, jailbreak probing, deep technical dialogue, and strategic gameplay, we observe: (1) RollarTimerAgent does not improve reasoning performance and slightly degrades it on tasks requiring emotional stability; (2) however, it exhibits behavioral signatures not observed in our bare LLM baseline under the same protocol. These include stateful behavioral hysteresis—evidenced by accelerated selective engagement gating under sustained adversarial interaction (disengagement after 7–10 rounds vs. uniform response for the baseline), graded reduction of response verbosity, and explicit acknowledgment of architectural limitations. These findings suggest that structured textual affective feedback operates as a low-bandwidth control signal, altering long-horizon interactional patterns. We discuss limitations, including increased latency and backbone sensitivity.
Liuyan Lu (Fri,) studied this question.