Read the Room Before Your Model Generates

Muyao Kong; Zhihao Lin

cs.AIOct 13, 2026

Read the Room Before Your Model Generates

Authors: Muyao Kong, Zhihao Lin

Organizations: New York University · Independent Researcher

Abstract

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, rather than 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. Scenario-Actor extracts scenario-relevant structures from Transformer hidden states and determines the temperature from scenario embedding, a structured geometric representation of the communicative setting. Empirical results show that on Llama-3.1-8B-Instruct, Scenario-Actor achieves strong goal-completion performance in interactive social scenarios, significant gains on open-ended instruction following, and competitive performance on multi-turn conversation compared to multiple baselines.

Figures & tables

Figure 1: AgentSense goal-outcome flip rate (top) and goal completion score (bottom) by changing temperature.

Figure 4: The $\operatorname{Scenarize}$ module.

Baseline	AgentSense	AlpacaEval	MT-Bench
Metric	Goal completion	LC-WR (%)	GPT-4 score
Scenario-Actor	0.8015	30.03	7.91
Greedy	0.7582	22.87	7.79
Fixed $T=0.5$	0.7540	23.63	7.79
Fixed $T=0.7$	0.7801	25.68	7.83
Fixed $T=1.0$	0.7692	23.42	7.61

Table 1: Benchmark results comparing Scenario-Actor with fixed-temperature and adaptive decoding baselines. Best results are highlighted .

Scenario Label	$\boldsymbol{\tau_{i}^{+}/\tau_{i}^{-}}$
rigorous	0.30
institutional	0.50
formal	0.55
emotional_support_neg	0.65
adversarial	0.75
emotional_support_pos	1.00

Table 2: Scenario labels and their endpoint temperatures.

Panel of Figure 5 — Figure 5: PCA (left) and UMAP (right) of scenario embedding geometry.

Validation	LLM-human	Human-human
Annotation $\kappa$	0.82 / 0.85	0.79
AE Agreement	86.00 / 86.00	80.00
AE judge $\kappa$	0.72 / 0.72	0.60
MT Pearson $r$	0.92 / 0.86	0.87
MT Spearman $\rho$	0.84 / 0.70	0.79
AS Agreement	80.00 / 72.00	76.00

Table 3: Agreement between LLM-based evaluators and human annotators. Values are rounded to two decimal places. Detailed validation metrics are reported in Appendix E .

Baseline	GSM8K
Metric	Accuracy (%)
Scenario-Actor	65.28
Greedy	77.18
Fixed $T=0.5$	69.52
Fixed $T=0.7$	69.07
Fixed $T=1.0$	62.40

Table 4: GSM8K results comparing Scenario-Actor with fixed-temperature and adaptive decoding baselines. The best result is highlighted .

Scenario	Count	Percentage
adversarial	350	26.5%
institutional	346	26.2%
brainstorming	332	25.2%
playful	79	6.0%
casual_conversation	67	5.1%
rigorous	66	5.0%

Table 5: GSM8K scenario count

Scenario Centroid	Cosine Similarity
adversarial	0.9765
institutional	0.9710
rigorous	0.9684
brainstorming	0.9510
playful	0.9499
casual_conversation	0.9463

Table 6: Mean cosine similarity between GSM8K and nine centroids, with gap of 0.0247 between the nearest and the second-nearest centroid per prompt

	Baseline
Benchmark	Scenario-Actor	Class-conditioned $T$
AgentSense	0.8015	0.7536
AlpacaEval	30.03	27.10
MT-Bench	7.91	7.64

Table 7: Comparison with the class-conditioned temperature baseline. Metrics are goal completion score, LC-WR (%), and GPT-4 score respectively. Best results are highlighted .

Metric	Layer 16	Layer 20	Layer 30
Recall@1	0.770	0.798	0.518
Eff. Rank	24.41	22.80	7.08
tanh Sat.	4.33%	15.42%	92.24%
Axis Corr.	0.60	0.80	0.94

Table 8: Layer-sweep diagnostics. Best results are highlighted

Geometry
Metric	$\boldsymbol{\beta=0.3}$	$\boldsymbol{\beta=1.0}$
Mean centroid cosine	0.9496	0.7103
Euclidean silhouette	0.1835	0.2668
Nearest-centroid agreement	0.6940	0.8877
Downstream
GSM8K Accuracy (%)	65.28	66.41

Table 9: Effect of the triplet-loss margin on scenario geometry and downstream performance. Full table can be found in Appendix G .

Method	Latency (ms/episode)	Peak memory (GiB)		Throughput
Method	Latency (ms/episode)	Allocated	Reserved	Output (tokens/s)	Request (episodes/s)
Scenario-Actor	97,000	18.59	18.71	21.59	0.01031
Greedy	87,410	18.36	18.48	21.84	0.01144
Fixed $T=0.5$	88,110	18.38	18.50	21.57	0.01135
Nucleus ( $p=0.95$ )	88,370	18.38	18.50	21.49	0.01132
Random- $T$	87,340	18.38	18.50	21.59	0.01145

Table 10: End-to-end efficiency over 245 AgentSense templates in 15-turn closed-loop interactions.

Benchmark	Comparison	Statistical significance
Benchmark	Comparison	Difference	95% CI	$p$ -value
AgentSense ( $\Delta$ Goal completion)	Scenario-Actor vs Greedy	$+4.33$ pp.	$[+1.55,+7.15]$	$2.44\times 10^{-3}$
	Scenario-Actor vs Fixed $T=0.5$	$+4.75$ pp.	$[+1.82,+7.76]$	$1.72\times 10^{-3}$
	Scenario-Actor vs Fixed $T=0.7$	$+2.14$ pp.	$[-0.54,+4.89]$	$1.22\times 10^{-1}$
	Scenario-Actor vs Fixed $T=1.0$	$+3.23$ pp.	$[+0.41,+6.11]$	$2.63\times 10^{-2}$
	Scenario-Actor vs Fixed $T=1.5$	$+10.26$ pp.	$[+6.97,+13.53]$	$8.73\times 10^{-10}$

Table 13: Statistical significance testing results on AgentSense, AlpacaEval, MT-Bench, and GSM8K. Differences are computed as Scenario-Actor minus the corresponding baseline.

Table 14: Scenario labels and their corresponding communicative descriptions.

Figure 6: Source composition by scenario.

Geometry metric	$\boldsymbol{\beta=0.3}$	$\boldsymbol{\beta=1.0}$	$\boldsymbol{\Delta}$
Mean pairwise centroid cosine	0.9496	0.7103	$-0.2393$
Centered centroid cosine	0.5258	0.5057	$-0.0201$
Centroid effective rank at 95%	5	6	$+1$
Euclidean silhouette	0.1835	0.2668	$+0.0833$
Nearest-centroid agreement	0.6940	0.8877	$+0.1937$
tanh saturation ( $>0.95$ )	0.0938	0.0347	$-0.0591$

Table 18: Scenario geometry under two triplet-loss margins.

Classic AI Scaffolding for LLM Social Agents
Sep 1, 2026Anatole GershmanCognitive Architectures for AI Agents Human Behavior Simulation
Carnegie Mellon University (retired)
Principled Under Pressure: Post-Training Decides Whether LLMs Act on Their Own Moral Judgment
Oct 6, 2026Orion Reblitz-RichardsonMoral Reasoning in Language Models Language Model Post-Training
Distiller Labs
Steer Like the LLM: Activation Steering that Mimics Prompting
May 5, 2026Geert Heyman, Frederik VandeputteLanguage Model Steering LLM Prompting
Nokia Bell Labs, Belgium.

Sep 1, 2026J/K move · Enter open · S save

Classic AI Scaffolding for LLM Social Agents

Anatole Gershman

Carnegie Mellon University (retired)

Cognitive Architectures for AI Agents Human Behavior Simulation LLM Agents Large Language Model-Based Social Simulation

Open paper AskPDF arXiv

Large language models can produce locally plausible social turns, but fluent next-turn generation is not enough for social simulation. Human encounters such as restaurant lunches and hotel check-ins are bounded social episodes with roles, scripts, material state, obligations, commitments, timing, and closure conditions. We present EpisodeSim, a hybrid LLM-agent architecture that represents classic-AI structures as natural-language control state interpreted by LLM calls. A World Master maintains shared reality, constructs scenes, adjudicates proposed actions, tracks effects and obligations, and controls closure. Experiments with small qualitative ablations on two held-out settings support a design claim: LLM fluency supplies local texture, but coherent social simulation benefits from persistent classic-AI-style scaffolding that organizes behavior over time.

Abstract

Explore similar work

Classic AI Scaffolding for LLM Social Agents

Principled Under Pressure: Post-Training Decides Whether LLMs Act on Their Own Moral Judgment

Steer Like the LLM: Activation Steering that Mimics Prompting