AutoHebb: Self-Modifying Neural Networks that Zero-Shot Infer and Continually Learn

Daniel Levine, David van Dijk, Kenneth KayarXiv 2026 · arXiv

TL;DR

AutoHebb is a novel neural network architecture that achieves superior zero-shot inference and continual learning by combining meta-learning with autonomous, neuromodulated Hebbian plasticity. Unlike standard models that rely on external gradients, AutoHebb generates its own learning signals, allowing it to "discover" efficient learning algorithms that generalize far beyond their training data.

What Problem Does This Paper Address?

Artificial intelligence still struggles with two hallmarks of biological intelligence: Zero-Shot Inference (ZI)—the ability to reason about novel combinations of familiar elements (e.g., "if A > B and B > C, then A > C")—and Continual Learning (CL)—the ability to learn new information over time without forgetting the old.

Current approaches often fall into two traps. Either they rely on external components (like researcher-calculated gradients or memory buffers) that don't exist in the brain, or they use architectures so complex that they are impossible to analyze mathematically. The authors of AutoHebb sought a "fully neural" model that is biologically grounded, autonomously controlled, and simple enough to be reverse-engineered into human-understandable algorithms.

Methods

The AutoHebb architecture consists of two feedforward subnetworks:

  1. The Main Network: A standard multi-layer perceptron (MLP) containing both fixed weights and plastic weights.
  2. The Control Network: An auxiliary network that monitors the main network's activity and environmental feedback to produce a scalar neuromodulator signal ($m$).
Figure 1: AutoHebb architecture. Network weights (W) are either fixed (black) or plastic (red). The architecture consists of a "main" network (here an MLP with hidden layers h) and an auxiliary "control" network (here an MLP with hidden layers c) that receives two inputs: (1) internally generated activity from the main network (here from last hidden layer hlast), and (2) external input reflecting environmental (external) feedback given an output (y) generated by the network. The control network generates a scalar-valued neuromodulatory signal (m, red circle) that gates changes to the plastic weights within network (here Wn∗). See Appendix E for further details; additional network variants in Figure 16.

Learning happens via a 3-factor Hebbian rule: updates to plastic weights are determined by the product of pre-synaptic activity, post-synaptic activity, and the global neuromodulator signal. Crucially, the network is trained using meta-learning (an "outer loop") to find the optimal fixed weights. These fixed weights act as "inductive biases" that enable the "inner loop" (the Hebbian updates) to learn specific tasks efficiently during the network's "lifetime."

Figure 2: Meta-learning in AutoHebb. In Auto- Hebb networks, fixed (non-plastic) network weights were identified through meta-learning optimization, which involves two levels of adaptation: an "inner" and "outer" loop. In the inner loop, corresponding to learning, (3-factor Hebbian) plastic updates to the network occur across an episode, i.e. a sequence of tri- als (where each trial involves input, output, control, and update steps; see main text). At the start of each episode, fixed weights are frozen and plastic weights are set to 0. In the outer loop, corresponding to meta- learning (or meta-training), the network’s fixed weights are updated via optimization based on performance within the episode and/or testing following the episode.

Datasets and Benchmarks

The authors evaluated AutoHebb on several cognitively demanding tasks:

  • Transitive Inference (TI): A ranking task where the model learns adjacent pairs (A>B, B>C) and must infer non-adjacent relationships (A>C).
  • Transitive Equivalence (TE): A grouping task where the model must infer that if A=B and B=C, then A=C.
  • Continual Learning Variants: "Linking" tasks (TI-L and TE-L) where the model must connect two separate lists or groups based on a single new piece of information.
  • Multi-task ZI: Learning TI followed by TE (or vice versa) without being told which task is currently active.

Key Insights and Ideas

  • Autonomous Control is Key: By generating its own neuromodulatory signal, AutoHebb can self-regulate. For example, it can "shut off" learning once a task is mastered, preventing "overexposure" from corrupting its internal representations.
  • Emergent "Length Generalization": Even when meta-trained on only 8 items, AutoHebb successfully generalized its ranking abilities to 20+ items. This suggests the meta-learning process discovered a truly abstract algorithm for ordering, rather than just memorizing patterns.
  • Algorithmic Tractability: Because the architecture is feedforward and modular, the authors successfully reverse-engineered the network's behavior. They found it "discovered" a scalar rank representation for items, using subtraction to make decisions.
  • The Essential Trinity: The paper demonstrates that meta-learning, Hebbian plasticity, and autonomous neuromodulation are jointly essential. Removing any one component caused performance to collapse on at least one of the benchmark tasks.

Key Results

AutoHebb consistently outperformed popular meta-learning models like MAML, Reptile, and ANML.

  • Zero-Shot Accuracy: On the TI task, AutoHebb achieved ~98% accuracy compared to ~70-74% for MAML and Reptile. On the TE task, most baselines performed at chance (~50%), while AutoHebb reached ~94%.
  • Continual Learning: AutoHebb was the only model that could successfully perform the "Linking" tasks and the Multi-task ZI benchmark without explicit task cues.
  • Robustness: While other models' performance decayed when shown the same pair too many times (overexposure), AutoHebb remained stable by down-regulating its own plasticity.
Multi-task performance

Limitations and Future Work

While powerful, the current study focused on relatively low-dimensional, symbolic inputs. The authors acknowledge that scaling AutoHebb to high-dimensional sensory data (like raw images or natural language) is a major next step. They also suggest that AutoHebb modules could be integrated into larger architectures, like Transformers, to provide them with more robust online learning and "test-time" adaptation capabilities.

Summary updated

Abstract

Humans and animals have a unique capacity to generalize and learn in novel circumstances, including "0-shot" inference (ZI) and continual learning (CL) -- two powerful cognitive abilities. Despite this importance, efforts to model ZI and CL are seriously limited in several ways. First, existing models typically lack components critical to ZI and CL in the brain, namely, strong inductive biases (for ZI) and autonomous control of learning (for CL). Second, models can be less tractable to analysis if model components are overly complex (lacking simplicity), or too task-restricted if model components are overly bespoke (lacking expressivity). Here we introduce autonomous Hebbian networks ("AutoHebb"), a class of online learning feedforward neural networks that realize strong inductive biases and autonomously controlled learning, and are both simple and expressive. Unlike standard models that learn via externally controlled optimization (e.g. gradient descent), AutoHebb networks combine external feedback with internally generated activity to produce a global learning signal ('neuromodulator') fed back to the network. The neuromodulator in turn gates a biologically grounded local learning rule (3-factor Hebbian rule) that updates networks' plastic weights. Using meta-learning to search for fixed (non-plastic) weights, we find that AutoHebb networks (1) succeed on ZI tasks, (2) show emergent (untrained) ZI/CL abilities, (3) admit reverse-engineering into explicit algorithms, and (4) outperform modern learning models on ZI/CL tasks. Lastly, performance comparisons to related models implied an essential role for three model components: meta-learning, Hebbian plasticity, and autonomous control of learning. These findings indicate that AutoHebb may be useful in hypothesis generation, analysis, and discovery of biologically grounded forms of learning and generalization.
BibTeX
@misc{levine2026autohebb,
  author = {Daniel Levine and David van Dijk and Kenneth Kay},
  title = {{AutoHebb: Self-Modifying Neural Networks that Zero-Shot Infer and Continually Learn}},
  year = {2026},
  eprint = {2610.40757},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2610.40757},
}
Figure