Memory formation is fundamental to intelligence, yet whether deep neural networks preserve identifiable memory traces analogous to biological memory units remains an open question. This work introduces a geometric framework to identify such "AI engrams" by formalizing the neuroscientific criteria of specificity, reactivation, sufficiency, and necessity into a constrained inverse problem. We derive a closed-form estimator that isolates individual memory traces from globally entangled parameters, and show that this biologically-derived solution corresponds to a natural gradient update on the parameter manifold. AI engrams enable surgical manipulation of learned knowledge: any subset of memories can be composed or erased through linear arithmetic, without iterative optimization. Experiments ranging from simple MLPs to LLMs demonstrate the causal validity and substantial scalability of AI engrams. Together, these results bridge theories of biological memory and artificial representation learning and offer geometric insight into how deep networks simultaneously support functional specificity within distributed storage.
While language models remain frozen at their training state, the world evolves continuously. Knowledge editing has emerged as a key alternative to full retraining, but its deployment is bottlenecked by the erosion of core capabilities: mathematical and programmatic reasoning collapse while encyclopedic recall remains intact. We trace this asymmetric degradation to a distributional mismatch. Covariance-based editors preserve only the subspaces spanned by their reference corpus, but fail to capture the operative distribution shaped by post-training such as SFT and DPO. Static external corpora, including Wikipedia and even the original pretraining mixture, cannot recover this shifted manifold. We propose Moir, which estimates the preservation covariance $C$ directly from the model itself by sampling from its own decoding distribution. Seeding generation with a single random vocabulary token bypasses the instruction-following templates that otherwise dominate sampled outputs, exposing the broader subspaces the model has internalized. Moir requires no external data and serves as a drop-in component for any covariance-based editor, a practical advantage given that the pre- and post-training corpora of most modern LLMs are not publicly accessible. Across OLMo-2, Llama-3.1, and Qwen-3 (7-8B), under both MEMIT and AlphaEdit and in batch and sequential regimes, Moir consistently extends preservation in the most vulnerable domains, most strikingly on Qwen3-8B after 20,000 AlphaEdit batch edits, it retains 79.9% GSM8K accuracy compared to 10.9% with the Wikipedia baseline. These results suggest that aligning the preservation distribution with the model's operative distribution is a key factor in non-destructive editing, and that the model itself may be the most accessible source of that distribution for deployed systems.
Large language models often encounter conflicts between in-context knowledge retrieved at inference time and parametric knowledge acquired during pretraining. Models that accept external knowledge uncritically are vulnerable to misinformation, whereas models that adhere rigidly to parametric knowledge fail to benefit from retrieval. Despite the widespread adoption of retrieval-augmented generation, we still lack a systematic understanding of what shapes knowledge-arbitration strategies during training. This gap risks producing pretrained models with undesirable arbitration behaviors and, consequently, wasting substantial computational resources after the pretraining budget has already been spent. To address this problem, we present the first controlled study of how training conditions influence models' use of in-context and parametric knowledge, and how they arbitrate between them. We train transformer-based language models on a synthetic biographies corpus while systematically controlling various conditions. Our experiments reveal that intra-document repetition of facts fosters the development of both parametric and in-context capabilities. Moreover, training on a corpus that contains inconsistent information or distributional skew encourages models to develop robust strategies for leveraging parametric and in-context knowledge. Rather than viewing these non-ideal properties as artifacts to remove, our results indicate that they are important for learning robust arbitration. These insights offer concrete, empirical guidance for pretraining models that harmoniously integrate parametric and in-context knowledge.
Large language models (LLMs) exhibit impressive in-context learning (ICL) abilities, enabling them to solve wide range of tasks via textual prompts alone. As these capabilities advance, the range of applicable domains continues to expand significantly. However, identifying the precise mechanisms or internal structures within LLMs that allow successful ICL across diverse, distinct classes of tasks remains elusive. Physics-based tasks offer a promising testbed for probing this challenge. Unlike synthetic sequences such as basic arithmetic or symbolic equations, physical systems provide experimentally controllable, real-world data based on structured dynamics grounded in fundamental principles. This makes them particularly suitable for studying the emergent reasoning behaviors of LLMs in a realistic yet tractable setting. Here, we mechanistically investigate the ICL ability of LLMs, especially focusing on their ability to reason about physics. Using a dynamics forecasting task in physical systems as a proxy, we evaluate whether LLMs can learn physics in context. We first show that the performance of dynamics forecasting in context improves with longer input contexts. To uncover how such capability emerges in LLMs, we analyze the model's residual stream activations using sparse autoencoders (SAEs). Our experiments reveal that the features captured by SAEs correlate with key physical variables, such as energy. These findings demonstrate that meaningful physical concepts are encoded within LLMs during in-context learning. In sum, our work provides a novel case study that broadens our understanding of how LLMs learn in context.
Quantifying entropy production (EP) is essential to understand stochastic systems at mesoscopic scales, such as living organisms or biological assemblies. However, without tracking the relevant variables, it is challenging to figure out where and to what extent EP occurs from recorded time-series image data from experiments. Here, applying a convolutional neural network (CNN), a powerful tool for image processing, we develop an estimation method for EP through an unsupervised learning algorithm that calculates only from movies. Together with an attention map of the CNN's last layer, our method can not only quantify stochastic EP but also produce the spatiotemporal pattern of the EP (dissipation map). We show that our method accurately measures the EP and creates a dissipation map in two nonequilibrium systems, the bead-spring model and a network of elastic filaments. We further confirm high performance even with noisy, low spatial resolution data, and partially observed situations. Our method will provide a practical way to obtain dissipation maps and ultimately contribute to uncovering the nonequilibrium nature of complex systems.
A collective flashing ratchet transports Brownian particles using a spatially periodic, asymmetric, and time-dependent on-off switchable potential. The net current of the particles in this system can be substantially increased by feedback control based on the particle positions. Several feedback policies for maximizing the current have been proposed, but optimal policies have not been found for a moderate number of particles. Here, we use deep reinforcement learning (RL) to find optimal policies, with results showing that policies built with a suitable neural network architecture outperform the previous policies. Moreover, even in a time-delayed feedback situation where the on-off switching of the potential is delayed, we demonstrate that the policies provided by deep RL provide higher currents than the previous strategies.
Dong-Kyum Kim, Youngkyoung Bae, Sangyun Lee et al.
This Letter presents a neural estimator for entropy production, or NEEP, that estimates entropy production (EP) from trajectories of relevant variables without detailed information on the system dynamics. For steady state, we rigorously prove that the estimator, which can be built up from different choices of deep neural networks, provides stochastic EP by optimizing the objective function proposed here. We verify the NEEP with the stochastic processes of the bead-spring and discrete flashing ratchet models, and also demonstrate that our method is applicable to high-dimensional data and can provide coarse-grained EP for Markov systems with unobservable states.