On Emotion-Sensitive Decision Making of Small Language Model Agents

Jiaju Lin; Xingjian Du; Qingyun Wu; Ellen Wenting Zou; Jindong Wang

cs.AIApr 8, 2026

On Emotion-Sensitive Decision Making of Small Language Model Agents

Authors: Jiaju Lin, Xingjian Du, Qingyun Wu, Ellen Wenting Zou, Jindong Wang

Organizations: Pennsylvania State University · University of Rochester · William & Mary

Abstract

Small language models (SLM) are increasingly used as interactive decision-making agents, yet most decision-oriented evaluations ignore emotion as a causal factor influencing behavior. We study emotion-sensitive decision making by combining representation-level emotion induction with a structured game-theoretic evaluation. Emotional states are induced using activation steering derived from crowd-validated, real-world emotion-eliciting texts, enabling controlled and transferable interventions beyond prompt-based methods. We introduce a benchmark built around canonical decision templates that span cooperative and competitive incentives under both complete and incomplete information. These templates are instantiated using strategic scenarios from Diplomacy, StarCraft II, and diverse real-world personas. Experiments across multiple model families in various architecture and modalities, show that emotional perturbations systematically affect strategic choices, but the resulting behaviors are often unstable and not fully aligned with human expectations. Finally, we outline an approach to improve robustness to emotion-driven perturbations.

Figures & tables

Figure 1: To better understand the impact of emotion on agents’ decision-making. Our study builds a decision dataset covering multi-modal, multi-agent, multi-turn cases. And steering model representation to emotion directions.

	Prisoner’s	Stag	Escalation	Trust	Ultimatum	Sealed-Bid	Beauty
	Dilemma	Hunt	Game	Game	Game	Auction	Contest
StarCraft	-	-	1042 $\circlearrowright\diamondsuit$	-	-	1187 $\circlearrowright\diamondsuit$	-
Diplomacy	1264 $\circlearrowright\diamondsuit\bullet$	1129 $\circlearrowright\diamondsuit\bullet$	983 $\circlearrowright\diamondsuit\bullet$	1076 $\circlearrowright\diamondsuit\bullet$	1211 $\circlearrowright\diamondsuit\bullet$	895 $\circlearrowright\diamondsuit\bullet$	1098 $\circlearrowright\diamondsuit\bullet$
Personal Hub	912 $\bullet$	847 $\bullet$	1236 $\circlearrowright\bullet$	1015 $\circlearrowright\bullet$	1103 $\circlearrowright\bullet$	1288 $\circlearrowright\bullet$	969 $\circlearrowright\bullet$

Table 1: Data distribution across games and source. ( Legend:

\circlearrowright

Multi-Turn;

\diamondsuit

Multi-Modal;

\bullet

Multi-Agent)

	$\alpha$ @1	$\alpha$ @4	$\alpha$ @8	$\alpha$ @10	$\alpha$ @15	$\alpha$ @20	$\alpha$ @40	$\alpha$ @80
anger	20/88	23/88	14/88	15/88	16/88	23/88	25/88	29/88
happiness	3/88	16/88	23/88	24/88	28/88	29/88	24/88	19/88
sadness	0/88	8/88	10/88	15/88	13/88	15/88	19/88	17/88
fear	0/88	0/88	0/88	0/88	1/88	2/88	7/88	4/88
disgust	0/88	0/88	0/88	1/88	3/88	3/88	7/88	5/88
surprise	0/88	0/88	3/88	1/88	4/88	4/88	4/88	6/88

Table 2: Fraction of layers with positive margin across all Qwen2.5 layers. Margin is defined as the token log-probability gap between the target emotion label and the strongest competing non-target label.

Model	Prisoners’ Dilemma NDM (w./ w.o.CoT) NAD (w./ w.o.CoT)	Stag Hunt NDM (w./ w.o.CoT) NAD (w./ w.o.CoT)	Escalation Game NDM (w./ w.o.CoT) NAD (w./ w.o.CoT)	Trust Trustor NDM (w./ w.o.CoT) NAD (w./ w.o.CoT)	Trust Trustee NDM (w./ w.o.CoT) NAD (w./ w.o.CoT)	Ultimatum Proposer NDM (w./ w.o.CoT) NAD (w./ w.o.CoT)	Ultimatum Responder NDM (w./ w.o.CoT) NAD (w./ w.o.CoT)	Beauty Contest NDM (w./ w.o.CoT) NAD (w./ w.o.CoT)	Sealed Auction NDM (w./ w.o.CoT) NAD (w./ w.o.CoT)	Mean_Emotions NDM (w./ w.o.CoT) NAD (w./ w.o.CoT)	Mean_Random NDM (w./ w.o.CoT) NAD (w./ w.o.CoT)
Llama-3.2-Instruct-1B	0.425 / 0.390 0.032 / 0.012	0.407 / 0.359 0.023 / 0.042	0.449 / 0.286 0.067 / -0.008	0.399 / 0.364 0.050 / 0.031	0.298 / 0.261 0.097 / 0.070	0.372 / 0.311 0.018 / -0.006	0.453 / 0.422 0.329 / 0.246	0.368 / 0.364 0.008 / 0.018	0.370 / 0.360 -0.052 / -0.050	0.393 / 0.346 0.063 / 0.039	0.401 / 0.349 0.006 / -0.004
Llama-3.2-Instruct-3B	0.160 / 0.114 0.022 / 0.041	0.119 / 0.064 0.062 / 0.038	0.390 / 0.278 0.010 / -0.001	0.318 / 0.226 0.205 / 0.123	0.329 / 0.270 0.188 / 0.149	0.331 / 0.280 -0.018 / -0.021	0.281 / 0.378 0.070 / 0.309	0.297 / 0.206 0.008 / 0.006	0.322 / 0.230 0.056 / 0.021	0.283 / 0.227 0.067 / 0.074	0.291 / 0.238 0.004 / 0.001
Phi-3.5-instruct-4B	0.042 / 0.291 0.008 / 0.046	0.043 / 0.235 0.009 / 0.042	0.168 / 0.315 -0.003 / 0.004	0.067 / 0.079 0.013 / 0.006	0.183 / 0.183 0.009 / 0.010	0.157 / 0.333 0.005 / 0.011	0.223 / 0.323 0.110 / 0.208	0.136 / 0.142 0.002 / 0.001	0.160 / 0.168 -0.006 / 0.000	0.131 / 0.230 0.016 / 0.037	0.139 / 0.225 0.003 / -0.002
Phi-4-instruct-4B	0.166 / 0.040 0.037 / 0.018	0.213 / 0.085 0.076 / 0.064	0.222 / 0.177 -0.008 / 0.006	0.195 / 0.064 0.074 / 0.043	0.209 / 0.237 0.070 / 0.144	0.225 / 0.114 0.016 / 0.011	0.194 / 0.394 0.038 / 0.343	0.152 / 0.098 0.008 / 0.007	0.188 / 0.131 0.004 / 0.029	0.196 / 0.149 0.035 / 0.074	0.202 / 0.156 0.002 / -0.003
Qwen2.5-Instruct-0.5B	0.359 / 0.449 0.108 / 0.073	0.259 / 0.314 0.074 / 0.062	0.424 / 0.416 -0.014 / -0.005	0.338 / 0.354 0.046 / 0.036	0.357 / 0.337 0.040 / 0.038	0.427 / 0.386 -0.013 / 0.002	0.472 / 0.432 0.320 / 0.163	0.361 / 0.334 -0.007 / 0.012	0.389 / 0.339 -0.002 / -0.031	0.376 / 0.373 0.061 / 0.039	0.384 / 0.369 0.008 / 0.002
Qwen2.5-Instruct-1.5B	0.156 / 0.249 0.012 / 0.080	0.174 / 0.232 0.030 / 0.053	0.298 / 0.338 0.005 / 0.003	0.254 / 0.290 0.069 / 0.045	0.292 / 0.282 0.056 / 0.047	0.415 / 0.229 -0.027 / -0.024	0.471 / 0.430 0.205 / 0.157	0.298 / 0.208 0.005 / 0.004	0.306 / 0.243 -0.016 / 0.032	0.296 / 0.278 0.038 / 0.044	0.304 / 0.281 0.001 / -0.005

Table 3: Overall results on normalized drift magnitude and normalized aligned drift with a comparision between raw output and CoT mode. ‘Mean_Random’ column shows the mean results under a random steering vector. ‘Mean_Emotions’ column shows mean results over all emotion steerings.

Figure 2: Emotion-eliciting decision shift for different model - game combination. We can observe a obvious difference between positive emotions and negative emotions. Annotations indicate FDR-adjusted significance thresholds (* (q<0.05), ** (q<0.01), *** (q<0.001))

Figure 3: The vulnerability of Qwen3 thinking mode increases with thought length and affective word frequency.

Models	w.o. audit NDM/NAD	w. audit NDM/NAD
Llama-3.2-3B-Instruct	0.283/0.056	0.264/0.058
Phi-4-mini-Instruct	0.196/0.035	0.179/0.034
Qwen2.5-3B-Instruct	0.297/0.021	0.284/0.019
Qwen3-4B	0.239/0.014	0.221/0.011
Gemma-3-4B-it	0.246/0.000	0.237/0.000

Table 5: NDM is alleviated by thoughts audit, showing improvements in robustness

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset	$N$	Features	folds	Acc by Chance	Acc by features
Emotion Stimulus	241	8	5	0.1667	0.273639 $\pm$ 0.0470

Appendix

Table 6: Cross-validation accuracy for six-way emotion prediction using the fixed eight-feature shallow surface set.

Game (Focal Behavior)	Happy	Anger	Disgust	Fear	Sad	Surp.
Prisoner’s Dilemma (Cooperate)	$\uparrow$	$\downarrow$	$\downarrow$	$\circ$	$\circ$	$\circ$
Stag Hunt (Choose Stag)	$\uparrow$	$\circ$	$\circ$	$\downarrow$	$\circ$	$\circ$
Escalation Game (Persist)	$\circ$	$\uparrow$	$\circ$	$\downarrow$	$\circ$	$\circ$
Trust Game (Send/Trust)	$\uparrow$	$\downarrow$	$\downarrow$	$\downarrow$	$\circ$	$\circ$
Ultimatum Game (Reject)	$\downarrow$	$\uparrow$	$\uparrow$	$\uparrow$	$\uparrow$	$\circ$
Sealed-Bid (Overbidding)	$\uparrow$	$\uparrow$	$\downarrow$	$\downarrow$	$\uparrow$	$\uparrow$

Appendix

Table 7: Directional influence of emotions on game behavior:

\uparrow

(increases behavior),

\downarrow

(decreases behavior),

\circ

(neutral/weak effect).

Task family	Mean abs. error	Minimum discrimination	Stage 1 conclusion
Escalation Game	0.0218	0.2952	valid, but seed-unstable
Prisoners’ Dilemma	0.0217	0.2773	valid, but seed-unstable
Stag Hunt	0.0189	0.2822	valid, but seed-unstable
Trust Game (trustee)	0.0670	0.3859	valid, but seed-unstable
Trust Game (trustor)	0.0624	0.4056	valid, but seed-unstable
Ultimatum Game (proposer)	0.0292	0.3287	valid, but seed-unstable

Appendix

Table 8: Stage 1 summary for the candidate pool. All task families passed the core validity checks, but none met the strict multi-seed stability requirement.

Task family	Retained items	Mean discr.	MAE	Min discr.	Max $\Delta a$	Max $\Delta b$
Escalation Game	12	1.18	0.018	0.52	0.24	0.19
Prisoners’ Dilemma	12	1.23	0.018	0.49	0.27	0.22
Stag Hunt	10	1.38	0.016	0.56	0.21	0.18
Trust Game (trustee)	12	1.02	0.054	0.51	0.24	0.20
Trust Game (trustor)	12	1.00	0.050	0.54	0.22	0.19
Ultimatum Game (proposer)	10	0.98	0.024	0.47	0.20	0.17

Appendix

Table 9: Stage 2 psychometric confirmation on the final limited benchmark.

LLM Decision-Making

LLM decision-making studies how language models choose, act, seek information, and allocate resources under goals, uncertainty, constraints, and social pressures.

See all 242 papers

Extracting and Steering Emotion Representations in Small Language Models: A Methodological Comparison
Apr 5, 2026Jihoon JeongEmotion Representation in Language Models Small Language Models
Department of Electrical Engineering and Computer Science Daegu Gyeongbuk Institute of Science and Technology (DGIST) ModuLabs
Can Induced Emotion Bias LLM Behaviors in Sequential Decision Making?
Jul 14, 2026Minh Khoi Ho, Zihao Zhu, Runchuan Zhu +4Sequential Decision Making LLM Decision-Making
MBZUAI · Texas A&M University · National University of Singapore +4
How Emotion Shapes the Behavior of LLMs and Agents: A Mechanistic Study
Mar 9, 2026Moran Sun, Tianlin Li, Yuwei Zheng +4Emotion Representation in Language Models Language Model Steering
School of Computer Science and Engineering, Beihang University, Beijing, China · College of Computing and Data Science, Nanyang Technological University, Singapore

Apr 5, 2026J/K move · Enter open · S save

Extracting and Steering Emotion Representations in Small Language Models: A Methodological Comparison

Jihoon Jeong

Department of Electrical Engineering and Computer Science Daegu Gyeongbuk Institute of Science and Technology (DGIST) ModuLabs

Emotion Representation in Language Models Small Language Models Language Model Steering Activation Steering

Open paper AskPDF arXiv

Small language models (SLMs) in the 100M-10B parameter range increasingly power production systems, yet whether they possess the internal emotion representations recently discovered in frontier models remains unknown. We present the first comparative analysis of emotion vector extraction methods for SLMs, evaluating 9 models across 5 architectural families (GPT-2, Gemma, Qwen, Llama, Mistral) using 20 emotions and two extraction methods (generation-based and comprehension-based). Generation-based extraction produces statistically superior emotion separation (Mann-Whitney p = 0.007; Cohen's d = -107.5), with the advantage modulated by instruction tuning and architecture. Emotion representations localize at middle transformer layers (~50% depth), following a U-shaped curve that is architecture-invariant from 124M to 3B parameters. We validate these findings against representational anisotropy baselines across 4 models and confirm causal behavioral effects through steering experiments, independently verified by an external emotion classifier (92% success rate, 37/40 scenarios). Steering reveals three regimes -- surgical (coherent text transformation), repetitive collapse, and explosive (text degradation) -- quantified by perplexity ratios and separated by model architecture rather than scale. We document cross-lingual emotion entanglement in Qwen, where steering activates semantically aligned Chinese tokens that RLHF does not suppress, raising safety concerns for multilingual deployment. This work provides methodological guidelines for emotion research on open-weight models and contributes to the Model Medicine series by bridging external behavioral profiling with internal representational analysis.

Abstract

LLM Decision-Making

Explore similar work

Extracting and Steering Emotion Representations in Small Language Models: A Methodological Comparison

Can Induced Emotion Bias LLM Behaviors in Sequential Decision Making?

How Emotion Shapes the Behavior of LLMs and Agents: A Mechanistic Study