Research

Overview

My research aims to understand how large language models internally organize computation,
and why modular reasoning sometimes emerges, sometimes fails to emerge, and sometimes
collapses into low-rank shared subspaces. I work at the intersection of:

  • Mechanistic interpretability
  • Representation geometry & modularity
  • Optimization dynamics (gradient flow, efficiency bias)
  • Activation-based evaluation of long-form reasoning and creativity

I am particularly interested in developing a theory of modularity that connects
optimization behavior, representation geometry, and task-level functional structure.

🔹 1. Demand-Driven Modularity in Transformers

with Prof. Robert Dick, University of Michigan (2025– )

Modularity is often assumed to “naturally emerge” in Transformers, but my multi-task
experiments reveal a different picture: modularity only appears when functional conflict
is sufficiently strong, not when input distributions differ.

Key findings

  • Distribution shift alone does not produce modules.
    Tasks such as Movie vs. Food sentiment share a dominant low-rank space.

  • Functional conflict induces clear representational separation.
    Math vs. Sentiment or Logic vs. Emotion leads to disentangled subspaces.

  • Transformers exhibit an efficiency bias.
    They reuse shared features aggressively and only separate when forced.

Methods

  • Layer-wise representation geometry (SVD / CKA / PCA curvature)
  • ΔW/W gradient-flow tracking across fine-tuning
  • Task-probing via L0/L12 linear classifiers
  • Causal ablations on singular directions vs. neuron subsets

Contribution

This work provides a unified optimization-based explanation for why modularity
emerges selectively, challenging the assumption that modularity is an inherent property
of LLMs rather than an optimization outcome.

Status

Preparing a manuscript targeting ICML 2026.

🔹 2. Low-Rank Collapse and Representation Bottlenecks

Across diverse task pairs, I consistently observe that the final layers—especially Layer 12
in DeBERTa-v3-base—collapse into a dominant low-rank subspace that captures >95% of task
variance. This low-rank bottleneck often prevents the formation of distinct experts.

Insights

  • A few singular directions encode nearly all task-specific behavior.
  • Large groups of FFN neurons are functionally redundant.
  • Removing top singular directions significantly degrades task accuracy,
    while ablating many neurons barely affects performance.

These findings support the hypothesis that LLMs prefer shared, compressed manifolds
unless conflict forces expansion into modular subspaces.

🔹 3. Early-Layer Gradient Starvation Revisited

Existing interpretations attribute early-layer stagnation to vanishing gradients.
My analysis shows the opposite: early layers achieve feature sufficiency early in training.

Evidence

  • L0 probe accuracy reaches 1.0, indicating “perfect” token-level feature extraction.
  • ΔW/W updates become tiny not due to gradient loss, but because further feature changes
    do not improve the objective    .
  • Late layers absorb coherent, task-differentiating gradients.

This reframes “gradient starvation” as a functional, not optimization-failure phenomenon.

🔹 4. Activation Subspaces as Evaluation Rubrics

with Prof. Yixin Cao, Fudan University (2024– )

I explore whether interpretable activation subspaces can serve as computational rubrics
for evaluating reasoning, process correctness, and creativity in generative models.

Goals

  • Identify latent activation directions that correspond to reasoning quality
  • Build activation-based rubrics that evaluate process, not just final answer
  • Validate causal link via activation steering and ablation

This line of work connects interpretability with LLM evaluation, moving beyond token-level
metrics toward “reasoning-aware” assessment.

Selected Slides & Writing Samples

  • đź”— Research Slides (Nov 2025) — [link coming soon]
  • đź”— Technical Report Draft — [link coming soon]
  • đź”— ICML 2026 Manuscript (in preparation) — [abstract coming soon]

Future directions

I plan to pursue a PhD on:

  • The causal structure of reasoning in LLMs
  • When and how modular reasoning can be reliably induced
  • Activation geometry as a foundation for interpretable evaluation
  • Designing objectives that promote functional specialization
  • Understanding optimization biases that govern emergent structures

My long-term goal is to build a principled framework for modular, interpretable,
and controllable reasoning systems.