SLICE transforms dense pretrained LLMs into interpretable mixture-of-experts architectures through continual pretraining, producing measurable interpretability gains at each training checkpoint. We're developing routing mechanisms and decomposition strategies that make monosemantic specialization an emergent property of scale rather than a constraint on it.
About the project
Modern interpretability research reverse-engineers functionality from tangled, polysemantic representations after training—an approach that doesn't scale to frontier models. We propose a different paradigm: restructure models into interpretable components during training.
SLICE (Specialized Learning with Interpretable Component Experts) converts dense MLPs into mixture-of-experts architectures where routing naturally encourages expert specialization. Like sparse autoencoders, we pursue unsupervised functional unit discovery, but instead of dictionary learning on activations, SLICE discovers functional units through routing, making the units actual computational pathways the model uses rather than learned reconstructions.
Theory of change
Current interpretability methods face a fundamental scaling problem: as models grow more capable and more dangerous to misuse, they become harder to interpret. Post-hoc techniques like circuit analysis require manual inspection; sparse autoencoders add computational overhead that grows with model size. Neither approach offers a path where interpretability improves as models scale. SLICE addresses this by making interpretability an emergent property of efficient architectures rather than a post-hoc tax on them. If successful, this enables:
Scalable oversight: Models with monosemantic expert pathways allow targeted monitoring; we can track which functional units activate for which inputs, providing transparency into model behavior without exhaustive manual analysis. Targeted intervention: Localized capabilities mean we can potentially ablate specific behaviors (e.g., harmful knowledge, deceptive reasoning) while preserving others. This directly supports machine unlearning and alignment fine-tuning.
Verification at deployment: For high-stakes deployments, SLICE-style models offer interpretability checkpoints, quantifiable metrics showing which experts handle which inputs. This provides evidence that models behave as intended. Future-proof methodology: By aligning interpretability with the bitter lesson (scaling compute beats encoding priors), SLICE remains relevant as frontier models grow. The hypothesis that larger models are easier to slice suggests interpretability becomes more tractable at scale, not less.
The core theory of change: if we can restructure models so that their internal computations route through semantically meaningful, inspectable pathways, we gain the transparency needed to safely deploy increasingly powerful systems.
Your role
Mentees, over the course of few weeks, will learn and be able to reason about the slice architecture and what it aims to do. Then they will be expected to autonomously explore the different configurations and training settings that optimize the interpretability + performance tradeoff.
Prerequisites
Required:
Highly proficient in Python and PyTorch Experience training a language model from scratch (not just fine-tuning; should understand the full training loop, loss computation, and optimizer dynamics) Strong ML engineering skills: comfortable debugging distributed training, managing GPU memory, and working with large codebases Familiarity with transformer architectures at the implementation level (attention, FFN layers, residual streams)
Strongly preferred:
Experience with mixture-of-experts architectures (routing mechanisms, load balancing, expert parallelism) Background in mechanistic interpretability (activation patching, circuit analysis, sparse autoencoders, causal abstraction) Understanding of scaling laws and compute-optimal training Hand-on experience with analyzing and "fixing" complex systems
Cherry on top:
Economics or information theory background (useful for thinking about resource allocation, sparsity, and efficiency tradeoffs) Jailbreaks and adversarial robustness Experience with LoRA or other parameter-efficient methods Prior exposure to interpretability research (MATS, graduate school or independent research)
Time commitment
30–40 hours
Location preference
Yes to UTC -8, but not a deciding factor
Application question(s)
Question 1 (200 words): When we slice a dense MLP along the input dimension to create experts, the resulting MoE is not mathematically equivalent to the original dense model. Consider a simple case: a dense MLP with weight matrices W_up (d_model → d_hidden) and W_down (d_hidden → d_model), converted into N experts by partitioning along d_hidden. Where exactly do the computations differ between the sliced MoE and the dense baseline? How would you quantify this difference, and what does it tell you about the reconstruction error introduced by slicing?
Question 2 (200 words): A core assumption behind SLICE is that neural networks encode more features than they have dimensions (superposition), which forces polysemanticity. This implies models are underparameterized relative to the features they represent. However, others argue LLMs are overparameterized—they have more parameters than necessary, enabling memorization and redundancy. Are LLMs underparameterized or overparameterized? How do we know? Does this depend on scale, and if so, how does that affect the SLICE hypothesis that larger models are easier to slice?
Question 3 (300 words): Read the MONET paper: https://arxiv.org/abs/2412.04139 Write an intuitive criticism of MONET's approach to achieving monosemantic experts. What assumptions does it make that might not hold? Where might the method struggle or fail to deliver on its interpretability promises?
About the mentors
Jaehyuk is interested in proactive interpretability, where interpretability in neural networks are induced, striking a balance of supervised and unsupervised feature learning.
Caleb Maresca