Below you can see a list of past talks at the Logic Seminar in this season. For forthcoming talks, see the main page.
Thursday, June 11, 2026
Grigore Roșu (Pi Squared Labs & University of Illinois Urbana-Champaign)
FastSet: Verified Settlement for AI-Native Work
Abstract:
Fast is a verifiable settlement infrastructure for AI-native work: humans, agents, services, and merchants coordinating work and settling outcomes, including payments, programmatically. This talk presents recent progress on Fast at Pi Squared Labs, starting with the FastSet weak consensus protocol that powers the Fast network. Unlike traditional systems that serialize transactions into one global order, FastSet processes independent claims in parallel, avoiding unnecessary consensus bottlenecks while preserving precise settlement semantics.
A central theme of the talk is verification. Fast does not treat correctness as an afterthought or a layer of audits around an implementation. The protocol is designed so that critical behavior can be specified, executed, and checked with formal methods. This gives stronger guarantees about what the system is allowed to accept, not just whether the network accepted it. The talk will also connect the protocol to real products: fast.xyz as the user and developer entry point, app.fast.xyz as an app for holding user and agent accounts and signing transactions, and shop.fast.xyz as a working example of agent-first commerce, where AI agents can discover merchants, prepare purchases, and participate in checkout flows.
Thursday, June 4, 2026
Cristian Kevorchian (University of Bucharest)
SQAN: A Neuro-Symbolic Architecture for Normative Reasoning
Abstract:
Henry Kautz’s 2020 taxonomy of neuro-symbolic systems spans six levels of integration, from neural networks with symbolic interfaces (Type 1) to systems where symbolic structure emerges from neural substrate (Type 6). The research frontier sits at Type 5, exemplified by Logic Tensor Networks. We present SQAN, a neuro-symbolic architecture for normative reasoning situated squarely at Type 5, with three layers: multilingual BERT embeddings of ontological entities on the cosine hypersphere; a Vietoris–Rips complex whose Atkin Q-connective components partition the ontology; and a Logic Tensor Network quantifying deontic axioms over those components. One observation gestures toward Type 6: on the Universal Declaration of Human Rights ontology, the Q-connective partition recovers 30 components with 100% ontological type purity — symbolic structure aligning with the neural substrate without type supervision. We demonstrate viability across the pipeline (1.89× semantic separation over typing baselines; LTN predicates at F1=0.80 with a 60% gain from Q-modulation) and introduce Derivability, a non-vacuous satisfaction metric that correctly rejects inferences outside the source ontology — illustrated via the Kevorkian assisted-dying case. SQAN offers a viable, interpretable, epistemically disciplined route to normative reasoning at the differentiable-logic frontier.
Thursday, May 28, 2026
Bogdan Ichim (University of Bucharest & IMAR) and Ștefania-Alexandra Cristea (University of Bucharest)
An Introduction to Voting Theory
Abstract:
We make a brief introduction to Voting Theory, including the Condorcet and Borda paradoxes, as well as the Condorcet efficiency of a voting system. We also present some computational results obtained under the assumption of the Impartial Anonymous Culture.
Finally, we examine the potential of natural language processing techniques for estimating the electorate’s intentions and forecasting final election outcomes.
Thursday, May 21, 2026
Radu Negulescu (The Informational Buildup Foundation)
Durable Alignment via Orthogonal Memory in Large Language Models
Abstract:
Introducing an architecture for update-stable alignment in LLMs, based on the Informational Buildup Framework (IBF), that encodes behavioral corrections in a gradient-isolated orthogonal memory space with no update path from the base model.
Alignment here denotes persistence of a specified target behavior on a delimited input region, not its specification.
The design avoids parameter entanglement and retrieval inconsistency by placing constraints outside the model’s weights, governed by dynamics that reinforce internally consistent corrections and dissolve contradicted ones.
On a frozen Mistral-7B, the system installs novel facts absent from the base model, achieves accuracy on counterfactual override while preserving unchanged facts, enables targeted geometric unlearning with near-zero collateral drift on matched controls, and maintains bounded amplitudes under continuous adversarial injection. Corrections remain effective after LoRA and instruction tuning, demonstrating durable alignment under model evolution.
Thursday, May 14, 2026
Răzvan Diaconescu (IMAR) and Ionuț Țuțu (IMAR)
Concept creation and abstraction
Abstract:
AI is facing a tough wall regarding concept creation and abstraction. We will discuss how these two processes are fundamentally important capabilities/functions of the human mind, and try to understand why AI has not been able to touch any of them. Then, we will present our middle-way approach to the problem that avoids both nihilism (i.e., concept creation and abstraction are fundamentally beyond any AI scope) and essentialism (i.e., AI will eventually, and quite soon, surpass by far any form of human intelligence). This is based on mathematical modelling of concept creation and abstraction, following work from neuroscience and algebraic semiotics, with the ultimate goal of implementing such mathematical models as part of the IBF initiative. Although our preliminary approach is inherently very technical, based on the advanced mathematics of category theory and institution theory, we will keep our discourse accessible to a wider audience by deliberately limiting the mathematical technicalities.
Thursday, May 7, 2026
Cristian Bereanu (University of Bucharest & IMAR)
On the gradient method
Abstract:
In this talk I will present the gradient descent method (Cauchy) and Nesterov’s accelerated gradient descent (1983). After that I will talk about a second-order ordinary differential equation which is the limit of Nesterov’s accelerated gradient method (Su, Boyd, Candes, J. Machine Learning Research, 2016). If time permits I will present “A variational perspective on accelerated method in optimization” by Wibisono, Wilson, Jordan (PNAS, 2016).
Thursday, April 30, 2026
Ioana Leuștean (University of Bucharest) and Bogdan Macovei (University of Bucharest)
Łukasiewicz logic and neural networks
Abstract:
While Multi-Layer Perceptrons (MLPs) are powerful classifiers, their “black-box” nature remains a barrier to interpretability. We propose a novel methodology for extracting symbolic explanations from neural networks using Łukasiewicz logic. Our presentation is divided into two parts: first, we explore the algebraic structures of Łukasiewicz logic and its enhancement with rational coefficients, demonstrating how specific MLP architectures correspond naturally to these logical formulas. Second, we show that network outputs can be decomposed into Disjunctive Normal Form (DNF), providing clear insights into decision boundaries and feature constraints. We validate this approach on both simple and multiclass problems, maintaining high performance while ensuring symbolic clarity. A key contribution is our algorithm for transforming arbitrary formulas into DNF via piecewise affine decomposition.
Thursday, April 23, 2026
Ciprian Păduraru (University of Bucharest)
Reinforcement Learning: Foundations, Algorithms, and LLM Applications II
Abstract:
This two-part seminar introduces the main ideas and methods of reinforcement learning, from foundational concepts to modern applications. The first part covers the basic framework of reinforcement learning, with emphasis on Deep Q-Networks (DQN) and an introduction to policy gradients. The second part focuses on more advanced algorithms such as Proximal Policy Optimization (PPO) and explores the growing role of reinforcement learning in the training and alignment of large language models.
Thursday, April 2, 2026
Radu Negulescu (The Informational Buildup Foundation)
The Informational Buildup Framework: A Computable Theory of Coherence-Preserving Dynamics
Abstract:
Mainstream AI treats information as something to store, transmit, and optimize. This talk starts from a different premise: information is not data, but the achievement of structural alignment between a system’s internal configuration and the structure of its environment.
From this premise, we introduce the Informational Buildup Framework (IBF), and here develop a computable formulation of its core dynamics through two coupled equations. The first is a Law of Motion that drives the system toward greater coherence. The second is the Modification Dynamics, which continuously reshape the coherence landscape in response to local discrepancy signals. In this framework, memory, agency, intelligence, and self-correction are not separate modules, but they emerge from the dynamics themselves.
To test IBF, we instantiate it in a controlled non-stationary environment and in chess, evaluated independently with Stockfish. Despite storing no raw training data, the framework retains knowledge better than replay-based baselines and yields significant behavioral gains from reward-free dynamics alone.
The framework predicts an emergent agency channel, confirmed empirically, that determines what the system learns. Ablation studies reveal a cascade of capacities: memory forms first, agency guides exploration, agency enables self-correction, and curated memory supports intelligence. When agency is removed, the cascade fails and the system forgets.
Controlled ablations confirm the predicted dependencies between mechanisms and capacities: when a mechanism is removed, the corresponding capability degrades in the expected way.
Thursday, March 26, 2026
Ciprian Păduraru (University of Bucharest)
Reinforcement Learning: Foundations, Algorithms, and LLM Applications I
Abstract:
This two-part seminar introduces the main ideas and methods of reinforcement learning, from foundational concepts to modern applications. The first part covers the basic framework of reinforcement learning, with emphasis on Deep Q-Networks (DQN) and an introduction to policy gradients. The second part focuses on more advanced algorithms such as Proximal Policy Optimization (PPO) and explores the growing role of reinforcement learning in the training and alignment of large language models.
Thursday, March 19, 2026
Gheorghe Ștefănescu (University of Bucharest)
Adaptive virtual organisms: A compositional model for complex hardware-software binding
Abstract:
The relation between a structure and the function it runs is of interest in many fields, including computer science, biology (organ vs. function) and psychology (body vs. mind). Our paper addresses this question with reference to computer science recent hardware and software advances, particularly in areas as Robotics, Self-Adaptive Systems, IoT, CPS, AI-Hardware, etc.
At the modelling, conceptual level our main contribution is the introduction of the concept of “virtual organism” (VO), to populate the intermediary level between reconfigurable hardware agents and intelligent, adaptive software agents. A virtual organism has a structure, resembling the hardware capabilities, and it runs low-level functions, implementing the software requirements. The model is compositional in space (allowing the virtual organisms to aggregate into larger organisms) and in time (allowing the virtual organisms to get composed functionalities).
The virtual organisms studied here are in 2D (two dimensions) and their structures are described by 2D patterns (adding time, we get a 3D model). By reconfiguration an organism may change its structure to another structure in the same 2D pattern. We illustrate the VO concept with a few increasingly more complex VO’s dealing with flow management or a publisher-subscriber mechanism for handling services. We implemented a simulator for a VO, collecting flow over a tree-structure (TC-VO), and the quantitative results show reconfigurable structures are better suited than fixed structures in dynamically changing environments. Finally, we briefly show how Agapia – a structured parallel, interactive programming language where dataflow and control flow structures can be freely mixed – may be used for getting quick implementations for VO’s simulation.
Bibliography:
C. I. Păduraru. G. Ștefănescu, Adaptive Virtual Organisms: A Compositional Model for Complex Hardware-software Binding, Fundamenta Informaticae 173(2-3) (2020), 139-176, doi:10.3233/FI-2020-1919.
Thursday, March 12, 2026
Gheorghe Ștefănescu (University of Bucharest)
Mixed control-flow and data-flow computing models
Abstract:
We start with a brief presentation of the evolution of formalisms for understating brain neuronal activity and the connections with computing machines, particularly with an emphasis on the works of Boole (1854), McCulloch and Pitts (1943), Kleene (1951) et al. Then we go to the technical part of the talk.
Symmetric monoidal category with feedback (a.k.a. traced monoidal category), introduced by Ștefănescu in 1986, is a versatile algebraic structure. Initially, an additive interpretation of the monoidal operation was considered, aiming to model control structures like flowchart schemes or finite automata. Later on, with a multiplicative interpretation of the monoidal operation, the structure was successfully used to capture parallel computing models, in particular data-flow networks.
Classical sequential programs have limited parallelism, while parallel data-flow networks have limited control. Hence, it is desirable to have a setting freely mixing control and parallelism. We propose symmetric semiringal categories (having two monoidal operations: addition and multiplication) with feedback as an algebraic candidate to study this complicated setting. Finally, we present a few partial results (the construction of the free mixalgebra and the related free distributed category; an example showing the formal extraction of parallel components of a cyclic sequential program).
Thursday, March 5, 2026
Marius Popescu (University of Bucharest)
NeurASP, a differentiable extension of the Answer Set Programming
Abstract:
We present NeurASP, a differentiable extension of the Answer Set Programming. We describe the basic concepts and syntax of ASP, the associated differentiable extension including semantics, inference, and learning procedures.