arXiv subjects

Benjamin Shih

Publications and source records attributed to Benjamin Shih.

3 recordsLinked to original sources

arXiv2026

How Do Language Models Choose Between Context and Memory?

When contextual information conflicts with the knowledge stored in model parameters, activation directions can be used to decode and steer which source the model follows. However, steering along a direction does not establish causality: whether the unedited model would naturally use that direction or whether the direction is reusable across tasks. We test these distinctions through counterfactual experiments in unambiguous settings. First, we estimate authority directions from agreement prompts, in which the context and parametric knowledge support the same answer. We then interchange naturally occurring coordinates along these directions between matched prompts that direct the model to prioritize either the supplied context or its parametric knowledge. Across Qwen, Llama, and OLMo models, this intervention reproduces 30-68% of the authority-induced shift in source choice, whereas matched controls reproduce almost none. To test cross-task reuse, we learn authority directions on two tasks separately and see that cross-task transferability closes only 9% of the authority gap while the local direction learned on the given task closes 57%. These results distinguish authority representation, causal use, and cross-task causal reuse, and suggest that authority computations may be task-dependent, rather than reusable across tasks.

cs.LG↗

arXiv2026

When Does Activation Steering Change What a Model Computes From?

Activation steering can reliably change an agent's output by modifying its internal activations. Yet arriving at the same answer need not involve the same computation: behavioral equivalence does not imply mechanistic equivalence. We test whether an activation edit changes the state used in subsequent computation or biases that computation toward the desired output. In a controlled state-tracking task, trace supervision creates an editable register: changing its internal value makes the model apply the next operation to the edited state. We next ask whether widely used mean activation steering admits a similar interpretation: does steering recreate the internal configuration the model naturally uses when performing a task? If so, transplanting that activation should be effective, and steering should remain effective at the scale of the natural source-to-target activation change. Neither prediction holds in the two selected model-task settings. Replacing the activation at one layer produces less than 2% of the target-answer margin gain from patching through all remaining layers, while natural-scale steering is similarly ineffective. The steering vectors have norms 77 and 26 times the median natural change in Qwen and Llama, producing 54% and 82% of the reference effect. Thus, a successful steering intervention need not reproduce the natural target activation at the intervention layer. More generally, an intervention should be interpreted as changing the computational state only when a later computation uses the edited value according to the semantics of that state.

cs.LG↗

arXiv2024

Transformers as Neural Operators for Solutions of Differential Equations with Finite Regularity

Neural operator learning models have emerged as very effective surrogates in data-driven methods for partial differential equations (PDEs) across different applications from computational science and engineering. Such operator learning models not only predict particular instances of a physical or biological system in real-time but also forecast classes of solutions corresponding to a distribution of initial and boundary conditions or forcing terms. % DeepONet is the first neural operator model and has been tested extensively for a broad class of solutions, including Riemann problems. Transformers have not been used in that capacity, and specifically, they have not been tested for solutions of PDEs with low regularity. % In this work, we first establish the theoretical groundwork that transformers possess the universal approximation property as operator learning models. We then apply transformers to forecast solutions of diverse dynamical systems with solutions of finite regularity for a plurality of initial conditions and forcing terms. In particular, we consider three examples: the Izhikevich neuron model, the tempered fractional-order Leaky Integrate-and-Fire (LIF) model, and the one-dimensional Euler equation Riemann problem. For the latter problem, we also compare with variants of DeepONet, and we find that transformers outperform DeepONet in accuracy but they are computationally more expensive.

cs.LG↗