Researchers propose a convolutional neural network model of visual valence processing, combining a scene-encoding visual module with an emotional significance module, and test it on a novel Pavlovian learning paradigm. The model reproduces key observations from human associative learning studies, including association formation, generalization, and alignment of neural representations for conditioned and unconditioned stimuli. The work positions deep neural networks as viable computational models of affective learning, offering advantages over classical models like Rescorla-Wagner when applied to neural data.
OpenAI researchers discovered neurons in CLIP that respond to the same concept across literal, symbolic, and conceptual representations. This finding parallels multimodal neurons previously observed in biological brains and helps explain CLIP's ability to classify unusual visual renditions of concepts. The work is presented as a step toward understanding the associations and biases learned by CLIP and similar vision-language models.
A new arXiv paper investigates whether Transformer-based text and speech encoders (RoBERTa, wav2vec 2.0) recover the geometric structure of Russell's circumplex model of affect — a valence-arousal topology from psychology. Experiments on naturalistic datasets (MSP-Podcast) and LLM-generated stimuli show that multimodal fusion achieves perfect topological alignment with Russell's primary emotion ordering, and zero-shot generic text embeddings place fine-grained emotion terms near their human-mapped coordinates. The authors argue this structure is intrinsically encoded in the representations rather than being an artifact of labeling, bridging psychological theory and representation learning.
A new arXiv preprint extends prior findings on emotion vectors in Claude Sonnet 4.5 to two open-weight models, Apertus-8B-Instruct-2509 and Gemma-4-E4B-it, by extracting emotion contrast vectors across all layers. The authors recover valence geometry in both models (peak PC1-valence correlations of r=0.76 and r=0.83, near Claude's r=0.81) but find notable architectural differences: Gemma encodes valence strongly in early layers while Apertus shows the opposite pattern. Arousal encoding proves sensitive to the corpus used for extraction, suggesting uneven distribution of arousal-relevant cues across model-generated text.
OpenAI researchers trained a character-level language model on Amazon reviews to predict the next character and discovered it spontaneously learned a single neuron encoding sentiment with high accuracy. The system achieved state-of-the-art sentiment classification with minimal labeled data, demonstrating that unsupervised language modeling can yield interpretable, task-relevant representations. This was an early result connecting unsupervised pretraining to downstream NLP tasks.
This paper explores conditioning diffusion models on representations from pre-trained self-supervised models as an alternative to text prompts or semantic maps, which require large annotated datasets. The self-conditioning mechanism improves unconditional image generation quality and provides a controllable representation space. The authors identify directions of variation in this space and demonstrate smoothness and disentanglement properties, suggesting potential for fine-grained generative control without heavy annotation overhead.
Researchers introduce BabyCL, a continual learning framework that processes the SAYCam egocentric child video dataset in a single chronological pass rather than shuffled multi-epoch training, more closely mimicking how children actually encounter their environment. The system combines streaming visual representation learning with image-text contrastive objectives, a multi-stage temporal segmentation, and a dual replay buffer managing visual and multimodal histories. BabyCL outperforms streaming baselines on the SAYCam Labeled-S 4AFC benchmark under matched compute budgets, substantially closing the gap to offline training upper bounds. The work advances understanding of whether neural networks can acquire word-referent mappings under biologically plausible training conditions.
A new arXiv preprint proposes a neural memory architecture for real-time novel view synthesis from multi-view streaming video of dynamic scenes. The method decouples memory update frequency from memory application frequency, using periodic gradient-based updates with per-frame cross-view attention to handle deformations. Two mechanisms — an auxiliary Memory Loss and a Memory Caching strategy — prevent catastrophic forgetting over long contexts. The approach achieves state-of-the-art performance on dynamic human motion scenes with minute-scale online memorization at real-time speeds.
Researchers use Joint Energy-Based Models (JEMs) to isolate the effect of learning objective—independent of architecture, scale, and data—on human alignment in visual representations. By varying a single mixing coefficient between discriminative and generative training, they evaluate models across six human-alignment benchmarks and find that alignment peaks at intermediate points on the generative-discriminative continuum rather than at either extreme. The results suggest that hybrid objectives combining categorical structure from discriminative learning with input-structure sensitivity from generative learning yield the most human-like visual behavior. This challenges the framing of generative vs. discriminative as a binary choice for building human-aligned vision systems.