2026-10-14 — views
Scenario-Actor: Adaptive Temperature Selection for LLMs
Read this because Moves beyond fixed heuristics by using geometric representations of communicative settings to improve instruction following and social scenario performance.
New framework adapts LLM decoding temperature based on inferred situational characteristics from hidden states.
Rethinking Temperature Selection in LLMs
Large language models (LLMs) generally rely on sampling temperature to regulate the amount of randomness in their outputs. Mainstream methods for selecting temperature typically involve fixed, heuristic global settings and context or token-level calibration. These approaches often lack a geometrically interpretable representation of the communicative scenario. In this work, we introduce a simple and interpretable temperature selection framework, the Scenario-Actor, which adapts decoding to the inferred situational characteristics.
How Scenario-Actor Works
The Scenario-Actor framework operates by extracting scenario-relevant structures from Transformer hidden states. It determines the temperature from a scenario embedding, which serves as a structured geometric representation of the communicative setting. This method allows the model to adjust its randomness based on the specific context of the interaction rather than relying on static parameters.
Empirical Results
Empirical results demonstrate the effectiveness of this approach. On Llama-3.1-8B-Instruct, Scenario-Actor achieves strong goal-completion performance in interactive social scenarios. The framework also shows significant gains on open-ended instruction following tasks. Furthermore, it delivers competitive performance on multi-turn conversation compared to multiple baselines. These findings suggest that adapting decoding based on inferred situational characteristics can enhance model behavior in complex interactions.
Implications for Model Behavior
By moving away from fixed heuristics, the Scenario-Actor framework offers a more nuanced way to control model output. The use of scenario embeddings provides a geometric interpretation of the communicative setting, allowing for more dynamic adjustments. This approach highlights the potential for leveraging hidden state structures to improve the alignment of LLM outputs with specific situational requirements.
Practitioner note
Developers should consider integrating scenario-aware temperature selection into their decoding pipelines to improve performance in interactive and open-ended tasks. Monitor how geometric representations of communicative settings affect goal completion and instruction following in your specific use cases.