John Winnicki

John Winnicki, Abeynaya Gnanasekaran, Eric Darve

Sparse autoencoders (SAEs) extract millions of interpretable features from a language model, but flat feature inventories aren't very useful on their own. Domain concepts get mixed with generic and weakly grounded features, while related ideas are scattered across many units, and there's no way to understand relationships between features. We address this by first constructing a strict domain-specific concept universe from a large SAE inventory using contrastive activations and a multi-stage filtering process. Next, we build two aligned graph views on the filtered set: a co-occurrence graph for corpus-level conceptual structure, organized at multiple levels of granularity, and a transcoder-based mechanism graph that links source-layer and target-layer features through sparse latent pathways. Automated edge labeling then turns these graph views into readable knowledge graphs rather than unlabeled layouts. In a case study on a biology textbook, these graphs recover coherent chapter and subchapter-level structure, reveal concepts that bridge neighboring topics, and transform messy sentence-level activity containing thousands of features into compact, readable views that illustrate the model's local activity. Taken together, this reframes a flat SAE inventory as an internal knowledge graph that converts feature-level interpretability into a global map of model knowledge and enables audits of reasoning faithfulness.

Benjamin Shih, John Winnicki, Eric Darve

A central hope behind process supervision is that models can expose intermediate variables that matter for their later behavior. For this to help with alignment, a scratchpad must be tied to the computation: when the model writes a state, later steps should compute from that state. To test this requirement, we use a controlled state-tracking task with a known update rule, comparing models trained to report only the final state with models trained to write intermediate states before giving the final answer. At evaluation, we edit the internal representation of one written state while leaving the visible scratchpad text fixed. Because the transition rule is known, the edit has a single correct downstream consequence. In Qwen2.5-Coder-7B, the state-writing model predicts the next phase bit implied by the edited state on 80% and 91% of held-out examples across the two task variants, while pretrained and final-answer-only controls remain near baseline. Additional controls rule out generic next-token steering and copying another continuation: the prediction depends on both the edited state and the current move. The same causal-use pattern replicates across model families. Together, these results suggest a sharper goal for scratchpad oversight: not just to make intermediate reasoning legible, but to train written states that the model uses as part of its computation.

Elliot L. Epstein, John Winnicki, Thanawat Sornwanee et al.

Large language models (LLMs) excel at numerical estimation but struggle to correctly quantify uncertainty. We study how well LLMs construct confidence intervals around their own answers and find that they are systematically overconfident. To evaluate this behavior, we introduce FermiEval, a benchmark of Fermi-style estimation questions with a rigorous scoring rule for confidence interval coverage and sharpness. Across several modern models, nominal 99\% intervals cover the true answer only 65\% of the time on average. With a conformal prediction based approach that adjusts the intervals, we obtain accurate 99\% observed coverage, and the Winkler interval score decreases by 54\%. We also propose direct log-probability elicitation and quantile adjustment methods, which further reduce overconfidence at high confidence levels. Finally, we develop a perception-tunnel theory explaining why LLMs exhibit overconfidence: when reasoning under uncertainty, they act as if sampling from a truncated region of their inferred distribution, neglecting its tails.