Embedded Files
Objective of the Workshop
Objective of the Workshop
Are you surprised about the ability of LLMs to perform complex reasoning tasks? Do you have a particular keenness for cracking the code of their inner workings? If the answer is yes, that's the case you are working in mechanistic interpretability or in neuro-symbolic reasoning of LLMs. Then, this is the workshop for you! Observing properties and understanding the inner workings are two sides of the same coin.
Another way is possible (e.g., Position Paper: MeMo: Towards Language Models with Associative Memory Mechanisms - ACL Anthology)!
The workshop aims to accept focussed contributions on:
Mechanistic Interpretability by-design
Neuro-symbolic approaches by-design
Work in progress and blue sky ideas are in the spirit of the workshop.
Physical Venue & Registration
Physical Venue & Registration
Venue: Room "Archimede", Edificio Didattica Ingegneria, Departement of Enterprise Engineering, Univeristy of Rome Tor Vergata, Viale del Politecnico, 1, Roma, Italy
Registration (both Pysical and Virtual): The registration is mandatory but free of charge. Register here.
Program
Program
26/6/2026 - Hybrid Workshop (Physical and Virtual Formats) - Timetable in CEST
11.00 - 11.30 (in-person) Fabio Massimo Zanzotto Opening remarks and MeMo
11.30 - 13.00 First Session
11.30 - 11.50 (in-person) Francesca Franzon, Matéo Mahaut Developmental Pathways of Context Use in Language Models. Repetition as a Primitive Processing Step
11.50 - 12.10 (in-person) Giovanni Trappolini, Andrea Santilli, Franco Maria Nardini , Fabrizio Silvestri Toward Geometric Memory in LLMs: Measuring Structural Inheritance in Knowledge Editing
12.10 - 12.25 (online) Muhammad Huzaifa Imran Approximate Symbolic Recovery of Transformer Attention Module Computations via KAN Surrogates
12.25 - 12.40 (online) Mehran Haddadi Comparing Traceability and Controllability of MeMo Representations against Transformers
12.40 - 13.00 (in-person) Giacomo De Luca, Matteo Rinaldi What Fires Is Not What Steers: Activation-Based and Intervention-Based Labels Disagree on the Same SAE Features
13.00 - 14.00 Light Lunch
14.00 - 14.30 Invited Talk: (in-person) Leonardo Ranaldi Evolving Agents via Quasi-Symbolic Memory: Bridging Neural and Symbolic Reasoning By-Design
14.30 - 16.00 Second Session
14.30 - 14.45 (online) Raju Kumar Yadav Routable Composition: Per-Position Weights as a Mechanistic Probe and Symbolic Interface for MeMo's Associative Memory Layers
14.45 - 15.00 (online) Lakshmi Chakradhar Vijayarao AIG: Towards a Mechanistic Science of Retrieval Integrity in Associative Memory Systems
15.00 - 15.15 (online) Fatemeh Shahrokhshahi Towards Logical Reasoning by Design: Encoding Inference Rules in MeMo’s Associative Memory
15.15 - 15.30 (online) Karthik Sairam Neuro-Symbolic MeMo: Injecting Structured Semantic Representations into Associative Memory Layers for Mechanistic Interpretability
15.30 - 15.45 (online) Doruk Benli DualMemo: From Episodic Memorization to Semantic Generalization in MeMo Language Models
15.45 - 16.00 Abdulhady Abas Transparent by Design: A Neuro-Symbolic Approach to Mechanistic Interpretability via Associative Memory
16.00 - 16.15 Coffee Break
16.15 - 17.40 Third Session
16.15 - 16.35 (online) Tiago Fernandes Tavares Mechanistic Interpretability of Sequence Models via Representational Alignment with Hidden Markov Models: A Case Study on Beatles Chords
16.35 - 16.50 (online) Jesús Jorge Armenta Segura Preliminary Evidence of Representation Changes Induced by Associative Memorization on Transformer-based Language Models, Using MeMo Models
16.50 - 17.15 (online) Faisal Ahamed Khan Federated Associative Assimilation: Decentralized Knowledge Acquisition for Modular Language Models
17.15 - 17.30 (online) Peiran Li Towards Version-aware Operations and Transaction Memories for Multi-layer MeMo
17.30 - 17.45 (online) Ritu Bains From Memorisation to Reasoning: Compositional Inference in Correlation Matrix Memory Language Models
17.45 - 18.00 (online) Steven Au Exploring MeMo as an Associative Memory Layer for Longitudinal Health Conversation
18.00 - 18.30 Open discussion and Closing Session
Program Committee
Program Committee
Organizer: Fabio Massimo Zanzotto, Human-centric ART, University of Rome Tor Vergata
PC Members:
Leonardo Ranaldi, Human-centric ART, University of Rome Tor Vergata & University of Edimburgh
Elena Sofia Ruzzetti, Human-centric ART, University of Rome Tor Vergata
Roberto Basili, SAG, University of Rome Tor Vergata
Danillo Croce, SAG, University of Rome Tor Vergata
Cristina Giannone, Almawave S.p.A
Raniero Romagnoli, Almawave S.p.A
Federico Ranaldi, Human-centric ART, University of Rome Tor Vergata
Contacts: fabio.massimo.zanzotto@uniroma2.it
Call for Papers
Call for Papers
Are you surprised about the ability of LLMs to perform complex reasoning tasks? Do you have a particular keenness for cracking the code of their inner workings? If the answer is yes, that's the case you are working in mechanistic interpretability or in neuro-symbolic reasoning of LLMs. Then, this is the workshop for you! Observing properties and understanding the inner workings are two sides of the same coin.
Another way is possible (e.g., Position Paper: MeMo: Towards Language Models with Associative Memory Mechanisms - ACL Anthology)!
The workshop aims to accept focussed contributions on:
Mechanistic Interpretability by-design
Neuro-symbolic approaches by-design
Work in progress and blue sky ideas are in the spirit of the workshop.
Your contribution must be between four pages and six pages, excluding references, and should be prepared using the Overleaf Submission Template. The proceedings will be submitted to CEUR-WS.org for online publication. Please submit your contribution via EasyChair at the designated submission link. The reviewing process is single-blind.
Submission Deadline: 31/5/2026 12:00 PM AoE
Acceptance Notification Date: 14/6/2026
Camera ready: 21/6/2026
Workshop Date: 26/6/2026
Page updated
Google Sites
Report abuse