Research map / Architectures and efficiency / Attention and transformers

Towards Universality: Studying Mechanistic Similarity Across Language Model Architectures

ICLR 2025 · Architectures and efficiency › Attention and transformers · approach: Attention head mechanisms · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Compares Pythia-160M Transformer and Mamba-130M via sparse autoencoder feature matching (max pairwise Pearson correlation) and circuit analysis of induction, finding high feature similarity and an "Off-by-One" motif where Mamba writes token info into SSM state at the next position.

Problem

Whether mechanistic findings (features and circuits) transfer across fundamentally different language model architectures, specifically Transformers versus Mamba (SSM) models, i.e. architectural universality.

Approach

Use sparse autoencoders trained on each model's residual stream to obtain architecture-agnostic feature bases, match features across Pythia and Mamba by max pairwise Pearson correlation, calibrate against neuron and seed-variant skylines, and then perform path-patching circuit analysis of induction in Mamba.

Key results

• Cross-architecture SAE MPPC averages 0.74 versus 0.76 for the Pythia model-seed skyline and 0.81 for the SAE-seed skyline; neuron baseline shows almost no similarity.
• More than 30% of Pythia SAE features find a Mamba match with MPPC above 0.95.
• Path patching shows Mamba's induction circuit relies on layer 17 SSM states, with key information entering at token B+1 via the local convolution (off-by-one motif).

Measured effects reported in the paper

Surprising findings

• Neuron-basis comparison between Pythia and Mamba shows almost no similarity, unlike prior same-architecture neuron universality findings.
• Matched features appear at roughly doubled depth (layer l in Pythia maps near layer 2l in Mamba).

Limitations

• SAEs capture only a subset of true features, and may capture combinations rather than the true features.
• Circuit universality analysis is qualitative and not automated or quantitatively measured.
• Feature similarity conclusions rest on several assumptions about SAE training and the MPPC metric, which is asymmetric.

Experimental setup

Datasets: The Pile OpenWebText

Benchmarks: OpenWebText feature activations Induction circuit path patching IOI task

Metrics: Max Pairwise Pearson Correlation (MPPC) logit difference monosemanticity score complexity score

Baselines: Neuron baseline Model seed variant skyline SAE seed variant skyline RWKV-169M

Builds on: Sparse autoencoders Pythia Mamba Path patching Universal neurons

Closest papers by meaning

• Transformers are SSMs: Generalized Models and Efficient Algorithms Through Structured State Space Duality (ICML 2024 · similarity 0.88)
• Transformer Key-Value Memories Are Nearly as Interpretable as Sparse Autoencoders (NeurIPS 2025 · similarity 0.87)
• Transcoders find interpretable LLM feature circuits (NeurIPS 2024 · similarity 0.87)
• Do Transformer Interpretability Methods Transfer to RNNs? (AAAI 2025 · similarity 0.87)
• SAEBench: A Comprehensive Benchmark for Sparse Autoencoders in Language Model Interpretability (ICML 2025 · similarity 0.87)
• Sparse Feature Circuits: Discovering and Editing Interpretable Causal Graphs in Language Models (ICLR 2025 · similarity 0.87)