Tag
A preprint proposing Verification Learning Theory (VLT), a formal process model of how AI-assisted claim checking affects retention of independent judgment, predicting that perceived assistant fallibility can improve delayed unaided performance through evidence checking.
This paper compares human and large language model linguistic productivity in extending novel verbs from active to passive contexts, finding that humans generalize based on abstract syntactic form even when no meaning can be assigned, while LLMs succeed only when a meaning similar to an existing verb exists.
The paper trains contrastive vision-language models on 2,441 hours of egocentric video from the BabyView corpus to test cross-situational word learning, finding that internet-pretrained visual representations far outperform those learned from developmental data, and that while models develop human-like semantic structure, their learning trajectories diverge from children's.
作者分享通过大声朗读自己的思考过程（类似链式思维）来提升解谜和问题解决效率的经验，并联系到 LLM 的链式思维机制与橡皮鸭调试法的认知原理。
A study in Proceedings of the Royal Society B found that 70 domestic dogs, tracked via AI head-gaze analysis, looked at the rounded shape when hearing 'Bouba' and the spiky shape when hearing 'Kiki', suggesting the sound-symbolic bouba/kiki effect may extend beyond humans.
该论文提出一个形式化的嵌套模块化层级（单元级、正交、线性），发现尽管单个神经元常常承担多种功能，神经网络在子空间层面仍表现出清晰的功能模块化，并在视觉、语言模型和多任务RNN中通过选择性消融得到验证，且这些子空间与脑数据有差异化对应。
Steinert-Threlkeld and Szymanik argue that machine learning models can serve as 'cognitive thermometers' for measuring semantic complexity, showing that learning-based measures converge with and sometimes better explain semantic universals than traditional logical complexity approaches.
A Verge commentary argues that the computational model of mind, which AI industry leaders like Demis Hassabis and Elon Musk rely on, is a warped and limited view of cognition, and that AI products function like 'cognitive junk food' undermining human thought.
UC Berkeley researchers propose a hybrid system treating cognitive algorithm discovery as a program refinement problem, where LLM agents revise human-authored probabilistic programs based on behavioral data mismatches, consistently improving model fit in a problem-solving paradigm.
This FAST workshop paper argues that LLM-based agents should treat evidence preparation—acquiring, transforming, or probing evidence before deciding—as a first-class design capability called epistemic action, enabled by 'epistemic scaffolding' interfaces and tools.
A user questions the functional differences between an artificial fly brain and a real fly brain, arguing that if the end results are the same, the distinction in mental processes may not matter.
The article explores why Lisp's parenthesis-heavy syntax is perceived as difficult to read, drawing on cognitive science principles like the Stroop effect to explain how syntax design influences human perception.
The paper proposes a unified computational theory integrating concepts and chunks from cognitive psychology, introduces the Trellis implementation, and demonstrates its ability to learn context-free grammars through experiments.
The article cautions against overusing metaphors in scientific explanations, using the brain-computer analogy as an example to show how such abstractions can conceal information and may be historically limited.
This paper surveys the linear representation hypothesis across AI and related fields, analyzing inconsistencies and proposing a more rigorous formalization to make it a falsifiable scientific claim.
This paper introduces WROP, a dataset for training object permanence in world models, and evaluates 14 video models, releasing PWM-WROP, a 16B world model that ranks among top continuation models.
The author describes memorizing the first 100 digits of pi using memory techniques like the method of loci and PAO systems, turning memorization into an engineering problem.
CogGym is a scalable framework for comparing human and AI cognition using cognitive experiments, revealing that larger language models better mimic human reasoning but still lag behind formal benchmarks.
A tweet recommending 15 books that explore how the world works across politics, wealth, human nature, cognition, and AI, stressing that the real information gap is in not reading these accessible books.
This paper examines the syntax and semantics of goal representations, analyzing their compositionality and role in rational behavior for cognitive science and AI.