Abstract This paper investigates how large language models can be steered to act more strategically in text-based negotiation settings. Two prompt-based action space designs are compared, namely emotional tone prompts and explicit offer prompts, within a negotiation environment, and outcomes are compared in simulated dialogues. The results show that both approaches improve strategic outcomes compared to a baseline, with tone-based actions yielding higher agreement rates and offer-based actions providing more stable tradeoffs. These findings demonstrate how action space design influences agent behavior, providing insights for deployment of large language models in strategic negotiation scenarios to gain an advantage in, for example, online influence operations.

Abstract In adversarial contexts, success often hinges on understanding not just what the opponent knows, but what they believe and how they revise those beliefs. This study investigates how large language models can be made more resilient and strategically capable by modeling the opponent’s reasoning using Bayesian belief revision. By formalizing negotiations as Bayesian games of incomplete information, it is shown that models equipped with belief revision are better able to counter deceptive or willful-thinking adversaries. The findings underscore the role of second-order reasoning in adversarial settings, with implications for social manipulation in the context of, for example, online communication and intelligence gathering.

We consider the problem of how to move active sonars unpredictably in pursuit of a stealthy underwater vehicle. The search problem is formalized as an imperfect-information game played on a discretized nautical chart with fine-grained hydroacoustics. The game is solved approximately using public belief states and self-play following a game-theoretically sound approach. The solution method is shown empirically to approximate the Nash equilibrium in a restricted scenario small enough to be solvable with tabular methods from algorithmic game theory.

Recent experimental studies have explored how well adaptive honeypot allocation strategies defend against human adversaries. As the experimental subjects were drawn from an unknown, nondescript pool of subjects using Amazon Mechanical Turk, the relevance to defense against real-world adversaries is unclear. The present study reproduces the experiments with more relevant experimental subjects. The results suggest that the strategies considered are less effective against attackers from the current population. In particular, their ability to predict the next attack decreased steadily over time, that is, the human subjects from this population learned to attack less and less predictably.

Imitation learning is a promising approach for training autonomous vehicles, where a set of state-action pairs from human demonstrated driving is used as training data in a supervised learning manner. Dataset Aggregation (DAgger) is a common imitation learning algorithm, in which models are trained by iteratively collecting new data, aggregating it with old data, and retraining the model on the entire collected dataset. Data aggregation and retraining, however, lead to two main problems: (i) large memory consumption, and (ii) long training time. In this work, we present a fast and memory-efficient algorithm, called MarioDAgger, that improves DAgger by resolving the aforementioned problems. Unlike DAgger that requires a collection of old and new data to train the models, MarioDAgger uses only the new data and a few samples from the old data stored in a rehearsal buffer, which is updated iteratively using reservoir sampling. To prevent forgetting old knowledge, MarioDAgger uses a recent regularization tech-nique Elastic Weight Consolidation. We evaluate and compare MarioDAgger with SafeDAgger, a recent variant of DAgger that MarioDAgger builds upon, in the context of autonomous vehicles, and show that MarioDAgger achieves the same performance as SafeDAgger in half as many iterations, using significantly less memory space.

This paper presents our work on adapting and extending natural language interface (NLI) to database technology to support the specification and violation reporting of business rules. The resulting system allows non-technical users to author and manage a rulebook in controlled natural language - serving as a single point of definition that can be compiled into SQL to generate violation reports. To achieve this we represent business rules in tuple calculus, handle negation in our query re-writing algorithms and add support for natural language reflexives (e.g. 'its', 'themselves', etc.). Our results show a large class of business rules can be captured with these extensions. Although our approach is general, we present it applied to compliance checking of regulations over a materiel capability development information system at the Swedish Defence Materiel Administration. At EDBT we will also demonstrate this work over a more generic package delivery domain. While there has been recent effort in pursuing Semantics for Business Vocabulary and Business Rules (SBVR) in the semantic web and description logic communities, to our knowledge ours is the first attempt to provide this capability for ER-modeled relational databases.

We consider the complexity of the model checking problem for the logic of knowledge and past time in synchronous systems with perfect recall. Previously established bounds are k-exponential in the size of the system for specifications with k nested knowledge modalities. We show that the upper bound for positive (respectively, negative) specifications is polynomial (respectively, exponential) in the size of the system irrespective of the nesting depth.

The combination of first-order epistemic logic with formal cryptography offers a potentially powerful framework for security protocol verification. In this paper, cryptography is modelled using private constants and one-way computable operations, as in the applied Pi-calculus. To give the concept of knowledge a computational justification, we propose a generalized Kripke semantics that uses permutations on the underlying domain of cryptographic messages to reflect agents' limited resources. This interpretation links the logic tightly to static equivalence, another important concept of knowledge that has recently been examined in the security protocol literature, and for which there are strong computational soundness results. We exhibit an axiomatization which is sound and complete relative to the underlying theory of terms, and to an omega-rule for quantifiers. Besides standard axioms and rules, the axiomatization includes novel axioms for the interaction between knowledge and cryptography. As protocol examples we use mixes, a Crowds-style protocol, and electronic payments. Furthermore, we provide embedding results for BAN and SVO.