thinker:sai-sumedh-r-hindupurSai Sumedh R. Hindupur
Co-first author; prior work on SAE assumptions/concept geometry.
Authored papers (1)
Sparse autoencoders carry a hidden i.i.d.-over-time prior that conflicts with how language model activations behave, and replacing it with a temporal prior recovers structure SAEs miss. Rewriting the SAE objective as MAP estimation shows the sparsity penalties of ReLU, TopK, BatchTopK and JumpReLU SAEs are additive over tokens (Prop. 4.1). They therefore assume concepts are independent across time and that sparsity is time-invariant. Gemma-2-2B and Llama-3.1-8B activations contradict this. Intrinsic dimension, measured with a Gram-matrix U-statistic, grows with sequence position, while a permuted stationary surrogate stays flat. Up to 80% of a token's variance is explained by a 500-token context. Temporal Feature Analysis (TFA) splits each activation into a predictable component, projected onto past context through attention, and a sparse novel residual assumed i.i.d. At Gemma-2-2B layer 12 it matches SAEs on NMSE (0.139 vs 0.139 for BatchTopK on web text). Its predictive code recovers event boundaries in 50 GPT-5 stories, keeps garden-path parses stable, and recovers the graph geometry in the Park et al. in-context graph task. It supports counterfactual edits on 4/9 nodes, where no SAE succeeds. On Ultrachat it also lowers tortuosity from about 25–30 to 2.0. Interpretability tools need inductive biases matched to the data's temporal generative structure, and features are better treated as evolving low-dimensional manifolds than as isolated directions.
More papers — OpenAlex / S2
Affiliations (1)
- SEAS, Harvard University(institute)
Co-authors (12)
- Ba, Demba2 shared
- Bigelow, Eric J.2 shared
- Costa, Valerie2 shared
- Fel, Thomas2 shared
- Hindupur, Sai Sumedh R.2 shared
- Lin, Johnny2 shared
- Lubana, Ekdeep Singh2 shared
- Mueller, Aaron2 shared
- Murthy, Sonia Krishna2 shared
- Patel, Oam2 shared
- Rager, Can2 shared
- Tuckute, Greta2 shared
Their work is cited by (5)
- Do Sparse Autoencoders Capture Concept Manifolds?1× refs
- Manifold Steering Reveals the Shared Geometry of Neural Network Representation and Behavior1× refs
- Structuring Sparsity: Block-Sparse Featurizers Capture Visual Concept Manifolds1× refs
- Turn-Averaged SAEs for Feature Discovery and Long-Context Attribution1× refs
- Arithmetic in the Wild: Llama uses Base-10 Addition to Reason About Cyclic Concepts1× refs
Other inbound relations (1)
Recent mentions (1)
- papers-typedsingh-2025-priors-time.md