Read the Room Before Your Model Generates
Published in AACL-IJCNLP 2026 (Main Conference), 2026
Authors: Muyao Kong*, Zhihao Lin* (*equal contribution)
arXiv: 2610.18820
Most LLM deployments pick a sampling temperature with a fixed global setting or token-level calibration, neither of which reflects the communicative situation a prompt sets up. We propose Scenario-Actor, a scenario-adaptive decoding framework that extracts scenario-relevant structure from Transformer hidden states, learns a scenario embedding, and selects the sampling temperature for each prompt from that embedding.
On Llama-3.1-8B-Instruct, Scenario-Actor achieves strong goal completion in interactive social scenarios, improves open-ended instruction following, and stays competitive on multi-turn conversation against several decoding baselines.
Recommended citation: Muyao Kong*, Zhihao Lin*. (2026). "Read the Room Before Your Model Generates." AACL-IJCNLP 2026 (Main Conference). arXiv:2610.18820.
Download Paper