J. Clayton Kerce

J. Clayton Kerce

Transformers resist surgical control. Ablating an attention head identified as critical for capitalization produces minimal behavioral change because distributed redundancy compensates for damage. This Hydra effect renders interpretability illusory: we may identify components through correlation, but cannot predict or control their causal role. We demonstrate that architectural interventions can expose hidden modularity. Our approach combines dual-stream processing separating token and contextual representations, per-layer supervision providing independent gradient signal at each depth, and gated attention regularizing toward discrete activation patterns. When trained with per-layer supervision, models produce ablation effects 5 to 23 times larger than architecturally identical controls trained with standard objectives. This enables 4 times greater control leverage on targeted behaviors: scaling identified attention heads produces smooth, predictable changes in model output. The key finding is architectural. Without per-layer supervision, ablation damage concentrates near zero with low variance (Winograd standard deviation 0.63%). With per-layer supervision, effects spread widely (standard deviation 6.32%), revealing which predictions depend on which circuits. The larger variance is not measurement noise but the signature of unmasked modularity. We validate our approach through three components: engineered features that capture computational dynamics rather than vocabulary structure (validated by near-zero correlation with raw activation clustering), an architecture providing positive control for modularity, and causal experiments demonstrating functional reorganization where different tasks route through different attention heads. This es tablishes a methodology for transforming interpretability from passive observation to active control.

J. Clayton Kerce, Alexis Fox

Standard transformers entangle all computation in a single residual stream, obscuring which components perform which functions. We introduce the Dual-Stream Transformer, which decomposes the residual stream into two functionally distinct components: a token stream updated by attention and a context stream updated by feed-forward networks. Information flow between attention heads is controlled through a hierarchy of mixing strategies, from fully independent (maximum interpretability) to dense (standard transformer behavior). This design exposes a tunable tradeoff between interpretability and performance. We measure this tradeoff on language modeling tasks at 29M parameters. Fully independent head mixing increases validation loss by 8\% relative to dense baselines. The recommended Kronecker mixing strategy, which permits scalar communication between heads while preserving within-head structure, costs only 2.5\%. All configurations maintain functional generation under attention amplification (scaling logits by factors up to 16 at inference time), with degradation ranging from 16\% to 27\%. This robustness suggests the architectures learn discrete algorithms that operate independently of soft probabilistic mixing. The architecture provides a foundation for interpretable language models where internal structure is exposed by design. \footnote{This work was partially supported by DARPA Contract HR001125C0302.}

Muralikrishnna G. Sethuraman, Ali Payani, Faramarz Fekri et al.

Visual question answering (VQA) has been gaining a lot of traction in the machine learning community in the recent years due to the challenges posed in understanding information coming from multiple modalities (i.e., images, language). In VQA, a series of questions are posed based on a set of images and the task at hand is to arrive at the answer. To achieve this, we take a symbolic reasoning based approach using the framework of formal logic. The image and the questions are converted into symbolic representations on which explicit reasoning is performed. We propose a formal logic framework where (i) images are converted to logical background facts with the help of scene graphs, (ii) the questions are translated to first-order predicate logic clauses using a transformer based deep learning model, and (iii) perform satisfiability checks, by using the background knowledge and the grounding of predicate clauses, to obtain the answer. Our proposed method is highly interpretable and each step in the pipeline can be easily analyzed by a human. We validate our approach on the CLEVR and the GQA dataset. We achieve near perfect accuracy of 99.6% on the CLEVR dataset comparable to the state of art models, showcasing that formal logic is a viable tool to tackle visual question answering. Our model is also data efficient, achieving 99.1% accuracy on CLEVR dataset when trained on just 10% of the training data.