Graph Memory Transformer (GMT)

Evelina Lamma; Niccol\`o Ferrari; Nicola Zanarini

arxiv: 2604.23862 · v3 · pith:HSPOK5VNnew · submitted 2026-04-26 · 💻 cs.LG · cs.AI· cs.CL

Graph Memory Transformer (GMT)

Nicola Zanarini , Niccol\`o Ferrari , Evelina Lamma This is my paper

Pith reviewed 2026-07-01 09:08 UTC · model grok-4.3

classification 💻 cs.LG cs.AIcs.CL

keywords Graph Memory Transformertransformer architectureFFN replacementmemory graphcentroid routinglanguage modelinginterpretabilitydecoder-only model

0 comments

The pith

A decoder-only transformer can replace its per-token FFN sublayers with a learned graph of 128 centroids linked by a directed transition matrix while training stably and exposing internal routing as inspectable quantities.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tests whether the feed-forward sublayer in a decoder-only transformer can be swapped for an explicit memory graph that routes representations across learned centroids instead of applying dense nonlinear transformations. The Graph Memory Transformer keeps causal attention unchanged but uses gravitational source routing and gated displacement readout to return source-to-target movements through a 128-by-128 transition matrix. The resulting 82.2-million-parameter model contains no dense FFN layers yet trains without instability and makes centroid usage, edge transitions, and movement vectors directly observable in the forward pass. It trails a 103-million-parameter dense baseline on validation loss and perplexity while remaining close on zero-shot benchmarks. The work positions the graph cell as a viable, structurally interpretable substitute for dense within-token computation.

Core claim

The FFN sublayer can be replaced by a memory cell that routes each token representation from an estimated source centroid toward a token-conditioned target centroid across a learned directed transition matrix, returning a gated displacement rather than a retrieved value; the resulting architecture trains stably as a fully decoder-only language model, exposes routing quantities for direct inspection, and achieves near-parity zero-shot behavior despite a remaining gap in validation loss relative to a dense baseline.

What carries the argument

Graph memory cell containing 128 centroids, a 128x128 directed transition matrix, gravitational source routing, token-conditioned target selection, and gated displacement readout, which substitutes for the usual FFN by computing per-token movement between memory states.

If this is right

Centroid usage statistics, transition-matrix structure, and source-to-target displacements become directly readable from the forward computation without auxiliary probes.
The model trains to convergence with zero dense FFN sublayers and 82.2 M total parameters.
Zero-shot benchmark scores remain close to those of the larger dense baseline despite the architectural substitution.
Further gains are expected from scaling the centroid bank or refining the routing kernels rather than from reintroducing dense layers.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The explicit transition matrix could be inspected post-training to recover reusable computational motifs that dense layers bury inside weight matrices.
Because movement rather than retrieval is returned, the same cell might be reused across multiple blocks without the parameter duplication typical of stacked FFNs.
If centroid count is increased while keeping the transition matrix sparse, the approach might close the remaining perplexity gap without restoring dense sublayers.

Load-bearing premise

Routing over a bank of 128 learned centroids connected by a directed transition matrix can supply the per-token nonlinear transformations normally performed by a dense FFN sublayer.

What would settle it

A controlled run in which the same architecture with the graph cell disabled (or replaced by a standard FFN of matched parameter count) produces validation loss indistinguishable from the reported 3.2903 baseline.

Figures

Figures reproduced from arXiv: 2604.23862 by Evelina Lamma, Niccol\`o Ferrari, Nicola Zanarini.

**Figure 1.** Figure 1: Slot-routing flow at Block 00 view at source ↗

**Figure 2.** Figure 2: Slot-routing flow at Block 06. 35 view at source ↗

**Figure 3.** Figure 3: Slot-routing flow at Block 11. 36 view at source ↗

**Figure 4.** Figure 4: Topic-separated Block 11 routing flows for the narrative, political, view at source ↗

**Figure 5.** Figure 5: Slot-routing flow at Block 15. The sequence from Blocks 00, 06, view at source ↗

**Figure 6.** Figure 6: Active edge structure at Block 00. Darker cells indicate stronger view at source ↗

**Figure 7.** Figure 7: Active edge structure at Block 06 view at source ↗

**Figure 8.** Figure 8: Active edge structure at Block 11. 41 view at source ↗

**Figure 9.** Figure 9: Active edge structure at Block 15. These are the same representative view at source ↗

**Figure 10.** Figure 10: Slot-routing flow in block 13 for a political-text probe, illustrating view at source ↗

read the original abstract

We investigate whether the Feed-Forward Network (FFN) sublayer in a decoder-only transformer can be replaced by an explicit learned memory graph while preserving the surrounding autoregressive architecture. The proposed Graph Memory Transformer (GMT) keeps causal self-attention intact, but replaces the usual per-token FFN transformation with a memory cell that routes token representations over a learned bank of centroids connected by a learned directed transition matrix. In the base GMT v7 instantiation studied here, each of 16 transformer blocks contains 128 centroids, a 128 * 128 edge matrix, gravitational source routing, token-conditioned target selection, and a gated displacement readout. The cell therefore returns movement from an estimated source memory state toward a target memory state, rather than a retrieved value. The resulting model is a fully decoder-only language model with 82.2M trainable parameters and no dense FFN sublayers, compared with a 103.0M-parameter dense GPT-style baseline used in the evaluation. The base v7 model trains stably and exposes centroid usage, transition structure, and source-to-target movement as directly inspectable quantities of the forward computation. It remains behind the larger dense baseline in validation loss and perplexity (3.5995/36.58 vs. 3.2903/26.85), while showing close zero-shot benchmark behavior under the evaluated setting. These results are not intended as a state-of-the-art claim; they support the viability and structural interpretability of replacing dense within-token transformation with graph-mediated memory navigation. Broader scaling, optimized kernels, and more extensive benchmark evaluation are left for subsequent work.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

GMT replaces the FFN with a 128-centroid graph memory cell using gravitational routing and displacement readout, training stably at smaller size but trailing the dense baseline on loss and perplexity.

read the letter

The main thing to know is that this paper builds a decoder-only language model where each block's FFN is replaced by a graph memory cell with 128 centroids, a 128x128 directed transition matrix, gravitational source routing, token-conditioned targets, and gated displacement readout. The result is an 82M-parameter model with no dense sublayers that trains without instability and makes centroid usage plus source-to-target movement directly readable from the forward pass.

What is new is the specific combination of these routing and readout mechanics inside the per-token transformation while keeping causal attention untouched. Earlier memory work in transformers tends to add external stores or alter attention; here the within-block nonlinearity itself becomes navigation on a learned graph. The lower parameter count and the explicit inspectability are direct consequences of the design, and the zero-shot numbers landing close to the 103M baseline is a reasonable outcome for an initial instantiation.

The paper does well at showing that this cell can serve as a viable drop-in without breaking autoregressive training or requiring extra tricks for stability. The numbers are reported plainly, and the framing stays limited to viability plus structural interpretability rather than claiming equivalence.

The soft spots are the performance gap—validation loss 3.60 versus 3.29 and higher perplexity—and the lack of visible ablations on centroid count, transition matrix size, or routing variants. Without error bars or training curves it is hard to judge how robust the result is. The central assumption that graph navigation can handle the nonlinear work an FFN does is tested only by the fact that training succeeds, not by matching capacity.

This is for readers working on alternative transformer blocks or on interpretability through explicit structure. Someone looking for concrete proposals with baseline comparisons will find it useful even if the results are preliminary. It deserves a serious referee because the architecture is fully specified, the comparisons are external, and the claims are scoped appropriately.

I would send it for peer review. The idea is worth testing at scale and the current evidence is solid enough to justify referee time.

Referee Report

1 major / 2 minor

Summary. The paper proposes replacing the per-token FFN sublayers in a decoder-only transformer with an explicit Graph Memory cell (GMT v7) that routes representations over a bank of 128 learned centroids connected by a 128x128 directed transition matrix, using gravitational source routing, token-conditioned target selection, and gated displacement readout. The resulting 82.2M-parameter model preserves causal self-attention, trains stably, exposes centroid usage, transitions, and source-to-target movements as inspectable quantities, and achieves close zero-shot benchmark behavior to a 103M dense GPT-style baseline while reporting higher validation loss (3.5995 vs. 3.2903) and perplexity (36.58 vs. 26.85). The work frames these outcomes as evidence of viability and structural interpretability rather than performance equivalence.

Significance. If the empirical results hold under fuller verification, the work provides a concrete demonstration that dense FFN transformations can be substituted by graph-mediated memory navigation while retaining the autoregressive decoder structure and yielding directly interpretable internal quantities. This could support future research on transparent, memory-centric alternatives to standard transformer sublayers, particularly if scaling or kernel optimizations close the observed performance gap.

major comments (1)

Abstract: the claim that the model 'trains stably' rests on a single reported run without error bars, multiple seeds, or ablation on the routing components; this weakens the viability conclusion because the 0.3 loss gap to the baseline could reflect run-to-run variance rather than a reliable architectural property.

minor comments (2)

Abstract: the number of transformer blocks (stated as 16) and total layer count should be cross-referenced with the parameter count (82.2M) to allow direct comparison with the 103M baseline.
Abstract: the zero-shot benchmark results are described only qualitatively ('close behavior'); quantitative scores or a table would strengthen the interpretability claim.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for highlighting the limitations in our stability claim. We address the point directly below and propose a targeted revision.

read point-by-point responses

Referee: Abstract: the claim that the model 'trains stably' rests on a single reported run without error bars, multiple seeds, or ablation on the routing components; this weakens the viability conclusion because the 0.3 loss gap to the baseline could reflect run-to-run variance rather than a reliable architectural property.

Authors: We agree the single-run observation is a genuine limitation that weakens the strength of the 'trains stably' phrasing. The reported run converged without divergence, NaNs, or exploding gradients, but we lack multiple seeds or variance estimates. We will revise the abstract to replace 'trains stably' with 'trains without divergence in the reported run' and add a short limitations paragraph noting the absence of multi-seed statistics and routing ablations. The 0.3 loss gap is presented as an observed difference rather than a claim of equivalence; we will not reinterpret it as architectural superiority. revision: partial

Circularity Check

0 steps flagged

No significant circularity

full rationale

The GMT paper is an empirical architecture proposal that replaces FFN sublayers with a graph memory cell (128 centroids, 128x128 transitions, gravitational routing, gated readout) while keeping causal attention. It reports forward-pass quantities, stable training, and benchmark numbers against a dense baseline without any derivation chain, uniqueness theorem, or self-citation that reduces a claimed result to its own inputs by construction. No equations are presented as first-principles predictions; the work is limited to viability and interpretability of the stated cell, making the derivation self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 1 invented entities

The central claim rests on the untested premise that a graph memory cell can substitute for dense FFN transformations; the design introduces several new learned structures whose sufficiency is assumed rather than derived.

free parameters (2)

number of centroids per block
Fixed at 128 in base v7; chosen as part of the instantiation rather than derived.
transition matrix size
128x128 edge matrix per block; design choice not justified from first principles.

axioms (1)

domain assumption A learned directed graph over centroids can perform the nonlinear per-token mapping previously done by an FFN.
Invoked when the memory cell is substituted for the FFN sublayer.

invented entities (1)

gravitational source routing and token-conditioned target selection no independent evidence
purpose: To determine source and target centroids for each token
New routing mechanism introduced in the memory cell; no independent evidence outside this design.

pith-pipeline@v0.9.1-grok · 5823 in / 1451 out tokens · 37580 ms · 2026-07-01T09:08:29.786278+00:00 · methodology

Review history (2 revisions) →

discussion (0)

Reference graph

Works this paper leans on

46 extracted references

[1]

Gomez,  Lukasz Kaiser, and Illia Polosukhin

Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez,  Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In�� , volume 30, 2017

2017
[2]

Transformer feed-forward layers are key-value memories

Mor Geva, Roei Schuster, Jonathan Berant, and Omer Levy. Transformer feed-forward layers are key-value memories. In�� , pages 5484–5495, 2021

2021
[3]

Lo- cating and editing factual associations in GPT

Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. Lo- cating and editing factual associations in GPT. In�� , volume 35, pages 17359–17372, 2022

2022
[4]

A practical review of mechanistic interpretability for transformer-based language models.�� , 2024

Daking Rai, Yilun Zhou, Shi Feng, Abulhair Saparov, and Ziyu Yao. A practical review of mechanistic interpretability for transformer-based language models.�� , 2024

2024
[5]

Memory networks

Jason Weston, Sumit Chopra, and Antoine Bordes. Memory networks. In�� , 2015

2015
[6]

End-to-end mem- ory networks

Sainbayar Sukhbaatar, Jason Weston, and Rob Fergus. End-to-end mem- ory networks. In�� , volume 28, 2015

2015
[7]

Large memory layers with product keys

Guillaume Lample, Alexandre Sablayrolles, Marc’Aurelio Ranzato, Lu- dovic Denoyer, and Herv´ e J´ egou. Large memory layers with product keys. In�� , volume 32, 2019

2019
[8]

Rabe, DeLesley Hutchins, and Christian Szegedy

Yuhuai Wu, Markus N. Rabe, DeLesley Hutchins, and Christian Szegedy. Memorizing transformers. In�� , 2022

2022
[9]

Memory layers at scale

Vincent-Pierre Berges, Barlas Oguz, Daniel Haziza, Wen-Tau Yih, Luke Zettlemoyer, and Gargi Ghosh. Memory layers at scale. In�� , volume 267 of�� , pages 3831–3842, 2025

2025
[10]

MemoryLLM: Plug-n- play interpretable feed-forward memory for transformers.�� , 2026

Ajay Jaiswal, Lauren Hannah, Han-Byul Kim, Duc Hoang, Arnav Kundu, Mehrdad Farajtabar, and Minsik Cho. MemoryLLM: Plug-n- play interpretable feed-forward memory for transformers.�� , 2026. 58

2026
[11]

Zoom in: An introduction to circuits.��, 2020

Chris Olah, Nick Cammarata, Ludwig Schubert, Gabriel Goh, Michael Petrov, and Shan Carter. Zoom in: An introduction to circuits.��, 2020

2020
[12]

Thread: Circuits.��, 2020

Nick Cammarata, Shan Carter, Gabriel Goh, Chris Olah, Michael Petrov, Ludwig Schubert, Chelsea Voss, Ben Egan, and Swee Kiat Lim. Thread: Circuits.��, 2020

2020
[13]

Prototype transformer: Towards language model architectures interpretable by design.�� , 2026

Yordan Yordanov, Matteo Forasassi, Bayar Menzat, Ruizhi Wang, Chang Qi, Markus Kaltenberger, Amine M’Charrak, Tommaso Salvatori, and Thomas Lukasiewicz. Prototype transformer: Towards language model architectures interpretable by design.�� , 2026

2026
[14]

A comprehensive study of knowledge editing for large language models.�� , 2024

Ningyu Zhang, Yunzhi Yao, Bozhong Tian, Peng Wang, Shumin Deng, Mengru Wang, Zekun Xi, Shengyu Mao, et al. A comprehensive study of knowledge editing for large language models.�� , 2024

2024
[15]

Neural turing machines

Alex Graves, Greg Wayne, and Ivo Danihelka. Neural turing machines. �� , 2014

2014
[16]

Hybrid computing using a neu- ral network with dynamic external memory.��, 538(7626):471–476, 2016

Alex Graves, Greg Wayne, Malcolm Reynolds, Tim Harley, Ivo Dani- helka, Agnieszka Grabska-Barwi´ nska, Sergio G´ omez Colmenarejo, Ed- ward Grefenstette, Tiago Ramalho, John Agapiou, Adri` a Puigdom` enech Badia, Karl Moritz Hermann, Yori Zwols, Georg Ostrovski, Adam Cain, Helen King, Christopher Summerfield, Phil Blunsom, Koray Kavukcuoglu, and Demis Has...

2016
[17]

Hopfield networks is all you need

Hubert Ramsauer, Bernhard Sch¨ afl, Johannes Lehner, Philipp Seidl, Michael Widrich, Lukas Gruber, Markus Holzleitner, Thomas Adler, David Kreil, Michael K Kopp, G¨ unter Klambauer, Johannes Brand- stetter, and Sepp Hochreiter. Hopfield networks is all you need. In �� , 2021

2021
[18]

Generalization through memorization: Nearest neighbor lan- guage models

Urvashi Khandelwal, Omer Levy, Dan Jurafsky, Luke Zettlemoyer, and Mike Lewis. Generalization through memorization: Nearest neighbor lan- guage models. In�� , 2020. 59

2020
[19]

Jacobs, Michael I

Robert A. Jacobs, Michael I. Jordan, Steven J. Nowlan, and Geoffrey E. Hinton. Adaptive mixtures of local experts.�� , 3(1):79– 87, 1991

1991
[20]

Outrageously large neural networks: The sparsely-gated mixture-of-experts layer

Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. In�� , 2017

2017
[21]

Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity

William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. �� , 23(120):1–39, 2022

2022
[22]

MemoryFormer: Minimize transformer computation by removing fully-connected layers.�� , 2024

Ning Ding, Yehui Tang, Haochen Qin, Zhenli Zhou, Chao Xu, Lin Li, Kai Han, Heng Liao, and Yunhe Wang. MemoryFormer: Minimize transformer computation by removing fully-connected layers.�� , 2024

2024
[23]

Graph transformers: A survey

Ahsan Shehzad, Feng Xia, Shagufta Abid, Ciyuan Peng, Shuo Yu, Dongyu Zhang, and Karin Verspoor. Graph transformers: A survey. �� , pages 1–20, 2026

2026
[24]

Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D

Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D. Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Lit...

2020
[25]

Using the output embedding to improve language models

Ofir Press and Lior Wolf. Using the output embedding to improve language models. In�� , pages 157–163, 2017

2017
[26]

On layer normalization in the transformer architecture

Ruibin Xiong, Yunchang Yang, Di He, Kai Zheng, Shuxin Zheng, Chen Xing, Huishuai Zhang, Yanyan Lan, Liwei Wang, and Tie-Yan Liu. On layer normalization in the transformer architecture. In�� , volume 119 of �� , 2020. 60

2020
[27]

Neural discrete representation learning

Aaron van den Oord, Oriol Vinyals, and Koray Kavukcuoglu. Neural discrete representation learning. In�� , volume 30, 2017

2017
[28]

OpenWebText Corpus

Aaron Gokaslan and Vanya Cohen. OpenWebText Corpus. �� , 2019

2019
[29]

Language models are unsupervised multitask learners

Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsupervised multitask learners. Technical report, OpenAI, 2019

2019
[30]

Think you have solved question answering? try ARC, the AI2 reasoning challenge.�� , 2018

Peter Clark, Isaac Cowhey, Oren Etzioni, Tushar Khot, Ashish Sabhar- wal, Carissa Schoenick, and Oyvind Tafjord. Think you have solved question answering? try ARC, the AI2 reasoning challenge.�� , 2018

2018
[31]

The language model evaluation harness, July 2024

Leo Gao, Jonathan Tow, Baber Abbasi, Stella Biderman, Sid Black, Anthony DiPofi, Charles Foster, Laurence Golding, Jeffrey Hsu, Alain Le Noac’h, Haonan Li, Kyle McDonell, Niklas Muennighoff, Chris Ociepa, Jason Phang, Laria Reynolds, Hailey Schoelkopf, Aviya Skowron, Lin- tang Sutawika, Eric Tang, Anish Thite, Ben Wang, Kevin Wang, and Andy Zou. The langu...

2024
[32]

HellaSwag: Can a machine really finish your sentence? In�� , pages 4791–4800, 2019

Rowan Zellers, Ari Holtzman, Yonatan Bisk, Ali Farhadi, and Yejin Choi. HellaSwag: Can a machine really finish your sentence? In�� , pages 4791–4800, 2019

2019
[33]

PIQA: Reasoning about physical commonsense in natural lan- guage

Yonatan Bisk, Rowan Zellers, Ronan Le Bras, Jianfeng Gao, and Yejin Choi. PIQA: Reasoning about physical commonsense in natural lan- guage. In�� , volume 34, pages 7432–7439, 2020

2020
[34]

WinoGrande: An adversarial winograd schema challenge at scale

Keisuke Sakaguchi, Ronan Le Bras, Chandra Bhagavatula, and Yejin Choi. WinoGrande: An adversarial winograd schema challenge at scale. In�� , vol- ume 34, pages 8732–8740, 2020

2020
[35]

Generating diverse high-fidelity images with VQ-V AE-2

Ali Razavi, Aaron van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with VQ-V AE-2. In�� , volume 32, pages 14837–14847, 2019

2019
[36]

Taming transformers for high-resolution image synthesis

Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In�� 61 �� , pages 12873–12883, 2021

2021
[37]

How contextual are contextualized word represen- tations? comparing the geometry of BERT, ELMo, and GPT-2 embed- dings

Kawin Ethayarajh. How contextual are contextualized word represen- tations? comparing the geometry of BERT, ELMo, and GPT-2 embed- dings. In�� , pages 55–65, 2019

2019
[38]

Interpretability in the wild: A circuit for indirect object identification in GPT-2 small.�� , 2022

Kevin Wang, Alexandre Variengien, Arthur Conmy, Buck Shlegeris, and Jacob Steinhardt. Interpretability in the wild: A circuit for indirect object identification in GPT-2 small.�� , 2022

2022
[39]

Mavor-Parker, Aengus Lynch, Stefan Heimersheim, and Adri` a Garriga-Alonso

Arthur Conmy, Augustine N. Mavor-Parker, Aengus Lynch, Stefan Heimersheim, and Adri` a Garriga-Alonso. Towards automated circuit discovery for mechanistic interpretability. In�� , volume 36, 2023

2023
[40]

Sparse autoencoders find highly interpretable features in language models.�� , 2023

Hoagy Cunningham, Aidan Ewart, Logan Riggs, Robert Huben, and Lee Sharkey. Sparse autoencoders find highly interpretable features in language models.�� , 2023

2023
[41]

Exploring activation pat- terns of parameters in language models.�� , 2024

Yudong Wang, Damai Dai, and Zhifang Sui. Exploring activation pat- terns of parameters in language models.�� , 2024

2024
[42]

Neuron-guided interpretation of code LLMs: Where, why, and how?�� , 2025

Zhe Yin, Xiaodong Gu, and Beijun Shen. Neuron-guided interpretation of code LLMs: Where, why, and how?�� , 2025

2025
[43]

Memorizing normality to detect anomaly: Memory-augmented deep autoencoder for unsupervised anomaly detection

Dong Gong, Lingqiao Liu, Vuong Le, Budhaditya Saha, Moussa Reda Mansour, Svetha Venkatesh, and Anton van den Hengel. Memorizing normality to detect anomaly: Memory-augmented deep autoencoder for unsupervised anomaly detection. In�� , pages 1705–1714, 2019

2019
[44]

PaDiM: A patch distribution modeling framework for anomaly detection and localization

Thomas Defard, Aleksandr Setkov, Angelique Loesch, and Romaric Au- digier. PaDiM: A patch distribution modeling framework for anomaly detection and localization. In�� , pages 475–489, 2021

2021
[45]

Towards total recall in industrial 62 anomaly detection

Karsten Roth, Latha Pemula, Joaquin Zepeda, Bernhard Sch¨ olkopf, Thomas Brox, and Peter Gehler. Towards total recall in industrial 62 anomaly detection. In�� , pages 14318–14328, 2022

2022
[46]

MVTec AD: A comprehensive real-world dataset for unsupervised anomaly detection

Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Ste- ger. MVTec AD: A comprehensive real-world dataset for unsupervised anomaly detection. In�� , pages 9592–9600, 2019. 63 � �� This appendix records implementation-fidelity details neede...

2019

[1] [1]

Gomez,  Lukasz Kaiser, and Illia Polosukhin

Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez,  Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In�� , volume 30, 2017

2017

[2] [2]

Transformer feed-forward layers are key-value memories

Mor Geva, Roei Schuster, Jonathan Berant, and Omer Levy. Transformer feed-forward layers are key-value memories. In�� , pages 5484–5495, 2021

2021

[3] [3]

Lo- cating and editing factual associations in GPT

Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. Lo- cating and editing factual associations in GPT. In�� , volume 35, pages 17359–17372, 2022

2022

[4] [4]

A practical review of mechanistic interpretability for transformer-based language models.�� , 2024

Daking Rai, Yilun Zhou, Shi Feng, Abulhair Saparov, and Ziyu Yao. A practical review of mechanistic interpretability for transformer-based language models.�� , 2024

2024

[5] [5]

Memory networks

Jason Weston, Sumit Chopra, and Antoine Bordes. Memory networks. In�� , 2015

2015

[6] [6]

End-to-end mem- ory networks

Sainbayar Sukhbaatar, Jason Weston, and Rob Fergus. End-to-end mem- ory networks. In�� , volume 28, 2015

2015

[7] [7]

Large memory layers with product keys

Guillaume Lample, Alexandre Sablayrolles, Marc’Aurelio Ranzato, Lu- dovic Denoyer, and Herv´ e J´ egou. Large memory layers with product keys. In�� , volume 32, 2019

2019

[8] [8]

Rabe, DeLesley Hutchins, and Christian Szegedy

Yuhuai Wu, Markus N. Rabe, DeLesley Hutchins, and Christian Szegedy. Memorizing transformers. In�� , 2022

2022

[9] [9]

Memory layers at scale

Vincent-Pierre Berges, Barlas Oguz, Daniel Haziza, Wen-Tau Yih, Luke Zettlemoyer, and Gargi Ghosh. Memory layers at scale. In�� , volume 267 of�� , pages 3831–3842, 2025

2025

[10] [10]

MemoryLLM: Plug-n- play interpretable feed-forward memory for transformers.�� , 2026

Ajay Jaiswal, Lauren Hannah, Han-Byul Kim, Duc Hoang, Arnav Kundu, Mehrdad Farajtabar, and Minsik Cho. MemoryLLM: Plug-n- play interpretable feed-forward memory for transformers.�� , 2026. 58

2026

[11] [11]

Zoom in: An introduction to circuits.��, 2020

Chris Olah, Nick Cammarata, Ludwig Schubert, Gabriel Goh, Michael Petrov, and Shan Carter. Zoom in: An introduction to circuits.��, 2020

2020

[12] [12]

Thread: Circuits.��, 2020

Nick Cammarata, Shan Carter, Gabriel Goh, Chris Olah, Michael Petrov, Ludwig Schubert, Chelsea Voss, Ben Egan, and Swee Kiat Lim. Thread: Circuits.��, 2020

2020

[13] [13]

Prototype transformer: Towards language model architectures interpretable by design.�� , 2026

Yordan Yordanov, Matteo Forasassi, Bayar Menzat, Ruizhi Wang, Chang Qi, Markus Kaltenberger, Amine M’Charrak, Tommaso Salvatori, and Thomas Lukasiewicz. Prototype transformer: Towards language model architectures interpretable by design.�� , 2026

2026

[14] [14]

A comprehensive study of knowledge editing for large language models.�� , 2024

Ningyu Zhang, Yunzhi Yao, Bozhong Tian, Peng Wang, Shumin Deng, Mengru Wang, Zekun Xi, Shengyu Mao, et al. A comprehensive study of knowledge editing for large language models.�� , 2024

2024

[15] [15]

Neural turing machines

Alex Graves, Greg Wayne, and Ivo Danihelka. Neural turing machines. �� , 2014

2014

[16] [16]

Hybrid computing using a neu- ral network with dynamic external memory.��, 538(7626):471–476, 2016

Alex Graves, Greg Wayne, Malcolm Reynolds, Tim Harley, Ivo Dani- helka, Agnieszka Grabska-Barwi´ nska, Sergio G´ omez Colmenarejo, Ed- ward Grefenstette, Tiago Ramalho, John Agapiou, Adri` a Puigdom` enech Badia, Karl Moritz Hermann, Yori Zwols, Georg Ostrovski, Adam Cain, Helen King, Christopher Summerfield, Phil Blunsom, Koray Kavukcuoglu, and Demis Has...

2016

[17] [17]

Hopfield networks is all you need

Hubert Ramsauer, Bernhard Sch¨ afl, Johannes Lehner, Philipp Seidl, Michael Widrich, Lukas Gruber, Markus Holzleitner, Thomas Adler, David Kreil, Michael K Kopp, G¨ unter Klambauer, Johannes Brand- stetter, and Sepp Hochreiter. Hopfield networks is all you need. In �� , 2021

2021

[18] [18]

Generalization through memorization: Nearest neighbor lan- guage models

Urvashi Khandelwal, Omer Levy, Dan Jurafsky, Luke Zettlemoyer, and Mike Lewis. Generalization through memorization: Nearest neighbor lan- guage models. In�� , 2020. 59

2020

[19] [19]

Jacobs, Michael I

Robert A. Jacobs, Michael I. Jordan, Steven J. Nowlan, and Geoffrey E. Hinton. Adaptive mixtures of local experts.�� , 3(1):79– 87, 1991

1991

[20] [20]

Outrageously large neural networks: The sparsely-gated mixture-of-experts layer

Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. In�� , 2017

2017

[21] [21]

Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity

William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. �� , 23(120):1–39, 2022

2022

[22] [22]

MemoryFormer: Minimize transformer computation by removing fully-connected layers.�� , 2024

Ning Ding, Yehui Tang, Haochen Qin, Zhenli Zhou, Chao Xu, Lin Li, Kai Han, Heng Liao, and Yunhe Wang. MemoryFormer: Minimize transformer computation by removing fully-connected layers.�� , 2024

2024

[23] [23]

Graph transformers: A survey

Ahsan Shehzad, Feng Xia, Shagufta Abid, Ciyuan Peng, Shuo Yu, Dongyu Zhang, and Karin Verspoor. Graph transformers: A survey. �� , pages 1–20, 2026

2026

[24] [24]

Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D

Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D. Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Lit...

2020

[25] [25]

Using the output embedding to improve language models

Ofir Press and Lior Wolf. Using the output embedding to improve language models. In�� , pages 157–163, 2017

2017

[26] [26]

On layer normalization in the transformer architecture

Ruibin Xiong, Yunchang Yang, Di He, Kai Zheng, Shuxin Zheng, Chen Xing, Huishuai Zhang, Yanyan Lan, Liwei Wang, and Tie-Yan Liu. On layer normalization in the transformer architecture. In�� , volume 119 of �� , 2020. 60

2020

[27] [27]

Neural discrete representation learning

Aaron van den Oord, Oriol Vinyals, and Koray Kavukcuoglu. Neural discrete representation learning. In�� , volume 30, 2017

2017

[28] [28]

OpenWebText Corpus

Aaron Gokaslan and Vanya Cohen. OpenWebText Corpus. �� , 2019

2019

[29] [29]

Language models are unsupervised multitask learners

Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsupervised multitask learners. Technical report, OpenAI, 2019

2019

[30] [30]

Think you have solved question answering? try ARC, the AI2 reasoning challenge.�� , 2018

Peter Clark, Isaac Cowhey, Oren Etzioni, Tushar Khot, Ashish Sabhar- wal, Carissa Schoenick, and Oyvind Tafjord. Think you have solved question answering? try ARC, the AI2 reasoning challenge.�� , 2018

2018

[31] [31]

The language model evaluation harness, July 2024

Leo Gao, Jonathan Tow, Baber Abbasi, Stella Biderman, Sid Black, Anthony DiPofi, Charles Foster, Laurence Golding, Jeffrey Hsu, Alain Le Noac’h, Haonan Li, Kyle McDonell, Niklas Muennighoff, Chris Ociepa, Jason Phang, Laria Reynolds, Hailey Schoelkopf, Aviya Skowron, Lin- tang Sutawika, Eric Tang, Anish Thite, Ben Wang, Kevin Wang, and Andy Zou. The langu...

2024

[32] [32]

HellaSwag: Can a machine really finish your sentence? In�� , pages 4791–4800, 2019

Rowan Zellers, Ari Holtzman, Yonatan Bisk, Ali Farhadi, and Yejin Choi. HellaSwag: Can a machine really finish your sentence? In�� , pages 4791–4800, 2019

2019

[33] [33]

PIQA: Reasoning about physical commonsense in natural lan- guage

Yonatan Bisk, Rowan Zellers, Ronan Le Bras, Jianfeng Gao, and Yejin Choi. PIQA: Reasoning about physical commonsense in natural lan- guage. In�� , volume 34, pages 7432–7439, 2020

2020

[34] [34]

WinoGrande: An adversarial winograd schema challenge at scale

Keisuke Sakaguchi, Ronan Le Bras, Chandra Bhagavatula, and Yejin Choi. WinoGrande: An adversarial winograd schema challenge at scale. In�� , vol- ume 34, pages 8732–8740, 2020

2020

[35] [35]

Generating diverse high-fidelity images with VQ-V AE-2

Ali Razavi, Aaron van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with VQ-V AE-2. In�� , volume 32, pages 14837–14847, 2019

2019

[36] [36]

Taming transformers for high-resolution image synthesis

Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In�� 61 �� , pages 12873–12883, 2021

2021

[37] [37]

How contextual are contextualized word represen- tations? comparing the geometry of BERT, ELMo, and GPT-2 embed- dings

Kawin Ethayarajh. How contextual are contextualized word represen- tations? comparing the geometry of BERT, ELMo, and GPT-2 embed- dings. In�� , pages 55–65, 2019

2019

[38] [38]

Interpretability in the wild: A circuit for indirect object identification in GPT-2 small.�� , 2022

Kevin Wang, Alexandre Variengien, Arthur Conmy, Buck Shlegeris, and Jacob Steinhardt. Interpretability in the wild: A circuit for indirect object identification in GPT-2 small.�� , 2022

2022

[39] [39]

Mavor-Parker, Aengus Lynch, Stefan Heimersheim, and Adri` a Garriga-Alonso

Arthur Conmy, Augustine N. Mavor-Parker, Aengus Lynch, Stefan Heimersheim, and Adri` a Garriga-Alonso. Towards automated circuit discovery for mechanistic interpretability. In�� , volume 36, 2023

2023

[40] [40]

Sparse autoencoders find highly interpretable features in language models.�� , 2023

Hoagy Cunningham, Aidan Ewart, Logan Riggs, Robert Huben, and Lee Sharkey. Sparse autoencoders find highly interpretable features in language models.�� , 2023

2023

[41] [41]

Exploring activation pat- terns of parameters in language models.�� , 2024

Yudong Wang, Damai Dai, and Zhifang Sui. Exploring activation pat- terns of parameters in language models.�� , 2024

2024

[42] [42]

Neuron-guided interpretation of code LLMs: Where, why, and how?�� , 2025

Zhe Yin, Xiaodong Gu, and Beijun Shen. Neuron-guided interpretation of code LLMs: Where, why, and how?�� , 2025

2025

[43] [43]

Memorizing normality to detect anomaly: Memory-augmented deep autoencoder for unsupervised anomaly detection

Dong Gong, Lingqiao Liu, Vuong Le, Budhaditya Saha, Moussa Reda Mansour, Svetha Venkatesh, and Anton van den Hengel. Memorizing normality to detect anomaly: Memory-augmented deep autoencoder for unsupervised anomaly detection. In�� , pages 1705–1714, 2019

2019

[44] [44]

PaDiM: A patch distribution modeling framework for anomaly detection and localization

Thomas Defard, Aleksandr Setkov, Angelique Loesch, and Romaric Au- digier. PaDiM: A patch distribution modeling framework for anomaly detection and localization. In�� , pages 475–489, 2021

2021

[45] [45]

Towards total recall in industrial 62 anomaly detection

Karsten Roth, Latha Pemula, Joaquin Zepeda, Bernhard Sch¨ olkopf, Thomas Brox, and Peter Gehler. Towards total recall in industrial 62 anomaly detection. In�� , pages 14318–14328, 2022

2022

[46] [46]

MVTec AD: A comprehensive real-world dataset for unsupervised anomaly detection

Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Ste- ger. MVTec AD: A comprehensive real-world dataset for unsupervised anomaly detection. In�� , pages 9592–9600, 2019. 63 � �� This appendix records implementation-fidelity details neede...

2019