arXiv:2508.05880v3 Announce Type: replace-cross Abstract: Understanding human emotions is central to user-facing AI applications, safety alignment, and the simulation of human behavior. As emotional...
arXiv:2610.27603v1 Announce Type: new Abstract: Empathetic interaction involves two distinguishable capacities: representing another person's emotional state (cognitive empathy) and generating an app...
arXiv:2606. 26987v1 Announce Type: cross Abstract: Recent work identified emotion vectors in Claude Sonnet 4.
The paper introduces Paraesthesia, a dynamic backdoor attack that uses emotionally styled inputs as triggers for large language models. By mapping target emotions into a valence–arousal space and rewriting a small subset of clean samples, the attack achieves over 98% success while minimally affecting clean performance. Experiments on four major LLMs show that the trigger cannot be fully explained by token-level cues and remains robust against several filtering and mitigation techniques.
Sofroniew et al. (2026) showed that emotion concepts in Claude Sonnet 4.5 are encoded as vectors whose geometry mirrors human affect psychology. This study replicates that finding using the base pretrained model google/gemma-2-27b, generating 205,200 Claude Sonnet 4.5 stories, extracting 171 emotion vectors, and recovering a similar affective circumplex with principal components explaining comparable variance. The analysis further identifies a sharp geometric seam at layers 22‑26, demonstrates that much of the geometry already exists in static token embeddings, and shows that the geometry predicts token‑level co‑activation with high correlation.
The paper demonstrates that a single internal direction in modern language models—called the valence axis (V-axis)—captures how positive or negative a sentence feels. By using only nine emotion category names and 50 short narrative paragraphs per emotion, the authors identify this axis via principal component analysis of frozen encoder embeddings, achieving 93% of supervised performance on SST‑2 and strong correlations with human valence ratings across images, audio, and brain recordings. The method transfers across modalities without target‑modality labels, but works only for continuous attributes and is specific to certain model families.
arXiv:2610.35411v1 Announce Type: new Abstract: We measure the emotion geometry of six LLMs spanning Google's Gemma 2 family. Using an extraction pipeline established by Sofroniew et al. (2026) on Cl...
arXiv:2603.04317v2 Announce Type: replace-cross Abstract: A growing literature shows that variables can be linearly decoded from the activations of large language models (LLMs). These range from prop...
The paper investigates why vision‑language models (VLMs) often miss hazards, attributing this to an overly conservative decision criterion rather than poor perception. By identifying and steering an affective ‘fear’ circuit within the model, the authors shift the decision threshold, dramatically increasing hazard recall from 27.0 % to 75.7 % without retraining. The study demonstrates that affect‑based inference-time manipulation can correct VLM decision biases in practical applications.
The study examines whether a decodable "empathy" direction can be used as a reliable causal lever in language models. Using EPITOME-derived facets of Recognition (cognitive) and Resonance (affective) across three instruction‑tuned LLMs, the authors find that while affective steering can partially raise affective scores, cognitive steering shows inconsistent or unmeasurable effects. The results highlight that decodability does not guarantee reliable control, especially for cognitive empathy, and that measurement sensitivity must be explicitly checked.
The study investigates how large language models (LLMs) assess psychological distress in online posts from six identity‑based communities. Through a perspectivist annotation task, 321 participants provided 9,587 judgments on 1,198 Reddit posts, revealing modest in‑group agreement (OR = 1.18) that varies across communities. When evaluated against these community‑specific labels, open‑weight LLMs consistently over‑estimate distress—achieving only 31–44% accuracy on posts perceived as none‑to‑mild—while newer models like GPT‑5 and Gemini 2.5 Pro show similar inflation, whereas Claude Opus 4 is more conservative. "whyItMatters":"The findings highlight that miscalibrated distress detection by LLMs can disproportionately impact the very communities they aim to serve, underscoring the need for equitable AI deployment in mental‑health contexts."
The paper investigates whether language‑grounded explanations improve trust in automated sheep pain detection from facial expressions. It finds that attention‑based explanations tied to the Sheep Pain Facial Expression Scale (SPFES) are largely ineffective, prompting the authors to replace the appearance bypass with a concept bottleneck that uses SPFES concept scores supervised by per‑region annotations. This new architecture slightly reduces performance but yields more semantically meaningful concepts, recovering minority pain states and revealing ear‑and‑eye severity orderings without explicit severity supervision, and highlights the pitfalls of pooled concept accuracy under clinical imbalance.