In the attention economy, sensational content exposes consumers to excessive emotional stimulation, hindering calm decision-making. This study proposes Multi-Agent LLM-based Emotional deToxification (MALLET), a multi-agent information sanitization system consisting of four agents: Emotion Analysis, Emotion Adjustment, Balance Monitoring, and Personal Guide. The Emotion Analysis Agent quantifies stimulus intensity using a 6-emotion BERT classifier, and the Emotion Adjustment Agent rewrites texts into two presentation modes, BALANCED (neutralized text) and COOL (neutralized text + supplementary text), using an LLM. The Balance Monitoring Agent aggregates weekly information consumption patterns and generates personalized advice, while the Personal Guide Agent recommends a presentation mode according to consumer sensitivity. Experiments on 800 AG News articles demonstrated significant stimulus score reduction (up to 19.3%) and improved emotion balance while maintaining semantic preservation. Near-zero correlation between stimulus reduction and semantic preservation confirmed that the two are independently controllable. Category-level analysis revealed substantial reduction (17.8-33.8%) in Sports, Business, and Sci/Tech, whereas the effect was limited in the World category, where facts themselves are inherently high-stimulus. The proposed system provides a framework for supporting calm information reception of consumers without restricting access to the original text.
Figures & tables
Fig. 1: Overview of the MALLET System.
Mode
Sˉ
HIR
EBI
FRE
RAW (original)
0.482
0.446
0.316
39.91
BALANCED (neutralized)
0.419
0.354
0.338
30.23
COOL (neutralized + suppl.)
0.389
0.330
0.346
34.54
TABLE I: Primary Metrics across Presentation Modes.
Metric
Comparison
t
padj
d
S
RAW vs BAL
5.93
1.3×10−8
0.210
S
RAW vs COOL
7.97
2.1×10−14
0.282
EBI
RAW vs BAL
− 4.26
2.3×10−5
− 0.151
EBI
RAW vs COOL
− 5.67
3.9×10−8
− 0.201
FRE
RAW vs BAL
17.23
5.2×10−56
0.609
FRE
RAW vs COOL
9.60
4.9×10−20
0.339
TABLE II: Results of Paired t -Tests (Holm Correction, n=800 ).
Category
SˉRAW
SˉBAL
SˉCOOL
Δ BAL
Δ COOL
World
0.659
0.660
0.645
− 0.1%
2.2%
Sports
0.393
0.312
0.260
20.6%
33.8%
Business
0.500
0.396
0.350
20.8%
30.0%
Sci/Tech
0.374
0.308
0.299
17.8%
20.0%
TABLE III: Mean Stimulus Score by Category.
Fig. 2: Mean Probability Distribution of the Six Emotions across the Three Conditions.
Verification method
BALANCED
COOL
SBERT similarity (mean)
0.846
0.827
NLI entailment rate
95.1%
27.8%
NLI neutral rate
3.0%
68.6%
NLI contradiction rate
1.9%
3.6%
TABLE IV: Results of Semantic Preservation Verification.
Fig. 3: Scatter Plot of Stimulus Reduction Magnitude and SBERT Similarity.
Consumer ID
Sˉ
HIR
EBI
anger
0
0.540
0.53
0.615
0.398
1
0.503
0.46
0.641
0.322
2
0.537
0.50
0.609
0.364
3
0.459
0.44
0.616
0.313
4
0.457
0.41
0.603
0.328
TABLE V: Weekly Summary of Pseudo-Consumers.
ID
Advice
0
Your dominant emotion this week is anger at 39.8%, combined with a high HIR of 0.53, indicating frequent engagement with conflict-driven news. Try diversifying your news consumption toward science, technology, or positive stories. Additionally, switching your display mode to COOL or BALANCED can help moderate emotional impact.
3
Your dominant emotion this week is joy (0.404), with moderate anger at 0.313 and HIR of 0.44. To maintain a positive balance while reducing anger, try incorporating more science and technology stories. Since your HIR is below 0.5, you can continue with your current display mode but consider mixing in more balanced topics.
TABLE VI: Examples of Personalized Advice.
ID
Rec.
Rationale
P1
COOL
Gentle and context-rich rewriting reduces anxiety and prevents being overwhelmed.
P2
BAL
Compatibility between factual accuracy and moderate emotional tone suits a balanced consumer.
P3
RAW
Preserves the boldness and dramatic expression of the original while maintaining accuracy, matching the stimulus-seeking preference.
TABLE VII: Mode Recommendation Results Based on Consumer Personas.
Text detoxification requires balancing toxicity removal, semantic preservation, and fluency. Existing methods typically optimize an unconstrained scalar objective---enabling safety scores to mask semantic loss---or rely on open-ended iterative debate prone to semantic drift and high computational cost. We propose AMEND, an inference-time framework that decouples detoxification into role-conditioned proposal generation, isolated multidimensional criticism, and constraint-first selection. AMEND evaluates diverse candidates with role-blind critics and enforces strict semantic and fluency feasibility before ranking by utility. To maximize efficiency, AMEND Lite restricts subsequent search to at most one targeted repair round only when residual toxic spans remain. On ParaDetox, one-round selection over diverse roles improves the joint evaluation metric by +0.0508 over the best single role. On a frozen confirmation test set, AMEND Lite achieves a joint score of 0.6597±0.0062, outperforming unconstrained multi-round loops by +0.0489 (p<0.05) while reducing inference requests by 28.4%.
Tianyu Wang
School of Computer Science and Technology, Soochow University
Despite their wide-ranging benefits, LLM-based agents deployed in open environments can be exploited to produce manipulative material. In this study, we task LLMs with propaganda objectives and analyze their outputs using two domain-specific models: one that classifies text as propaganda or non-propaganda, and another that detects rhetorical techniques of propaganda (e.g., loaded language, appeals to fear, flag-waving, name-calling). Our findings show that, when prompted, LLMs exhibit propagandistic behaviors and use a variety of rhetorical techniques in doing so. We also explore mitigation via Supervised Fine-Tuning (SFT), Direct Preference Optimization (DPO), and ORPO (Odds Ratio Preference Optimization). We find that fine-tuning significantly reduces their tendency to generate such content, with ORPO proving most effective.
Julia Jose, Ritik Roongta, Rachel Greenstadt
Department of Computer Science and Engineering, New York University, New York, NY, USA
Emotional expression can influence the safety decisions of large language models (LLMs), offering a potential avenue for improving safety alignment. Existing studies have mainly focused on how emotional expressions facilitate attacks under harmful requests, while overlooking their effects on benign requests. We find that emotional expression can also systematically increase refusal tendencies on benign requests, leading to unnecessary over-refusal. Based on this observation, we propose emotion-guided refusal subspace steering (EmoRSS), an activation-steering method that mitigates emotion-induced over-refusal while preserving refusal behaviour on harmful requests. Specifically, we first identify a refusal-sensitive layer using layer-wise linear probes and construct a refusal subspace from sparse autoencoder (SAE) features aligned with the probe direction. Next, we use paired regular and emotional requests with the same queries to estimate the mean activation shift in the features defining the refusal subspace. Finally, we decode this shift into an activation intervention vector and apply it in the reverse refusal direction during inference, without updating the backbone parameters. Experiments on two LLMs show that, when requests contain emotional expressions, our method achieves a more favourable trade-off between refusing harmful requests and answering benign ones than prior over-refusal mitigation baselines, while better preserving general task performance.
Shuyi Miao, Yaojin Ma, Chenhang Cui +5
Beihang University · University of Science and Technology Beijing · National University of Singapore +2