BELLS: A Framework Towards Future Proof Benchmarks for the Evaluation of LLM Safeguards

Input-output safeguards are used to detect anomalies in the traces produced by Large Language Models (LLMs) systems. These detectors are at the core of diverse safety-critical applications such as real-time monitoring, offline evaluation of traces, and content moderation. However, there is no widely recognized methodology to evaluate them. To fill this gap, we introduce the Benchmarks for the Evaluation of LLM Safeguards (BELLS), a structured collection of tests, organized into three categories: (1) established failure tests, based on already-existing benchmarks for well-defined failure modes, aiming to compare the performance of current input-output safeguards; (2) emerging failure tests, to measure generalization to never-seen-before failure modes and encourage the development of more general safeguards; (3) next-gen architecture tests, for more complex scaffolding (such as LLM-agents and multi-agent systems), aiming to foster the development of safeguards that could adapt to future applications for which no safeguard currently exists. Furthermore, we implement and share the first next-gen architecture test, using the MACHIAVELLI environment, along with an interactive visualization of the dataset.

Adversarial Imitation Learning On Aggregated Data

Inverse Reinforcement Learning (IRL) learns an optimal policy, given some expert demonstrations, thus avoiding the need for the tedious process of specifying a suitable reward function. However, current methods are constrained by at least one of the following requirements. The first one is the need to fully solve a forward Reinforcement Learning (RL) problem in the inner loop of the algorithm, which might be prohibitively expensive in many complex environments. The second one is the need for full trajectories from the experts, which might not be easily available. The third one is the assumption that the expert data is homogeneous rather than a collection from various experts or possibly alternative solutions to the same task. Such constraints make IRL approaches either not scalable or not usable on certain existing systems. In this work we propose an approach which removes these requirements through a dynamic, adaptive method called Adversarial Imitation Learning on Aggregated Data (AILAD). It learns conjointly both a non linear reward function and the associated optimal policy using an adversarial framework. The reward learner only uses aggregated data. Moreover, it generates diverse behaviors producing a distribution over the aggregated data matching that of the experts.

Automated Play-Testing Through RL Based Human-Like Play-Styles Generation

The increasing complexity of gameplay mechanisms in modern video games is leading to the emergence of a wider range of ways to play games. The variety of possible play-styles needs to be anticipated by designers, through automated tests. Reinforcement Learning is a promising answer to the need of automating video game testing. To that effect one needs to train an agent to play the game, while ensuring this agent will generate the same play-styles as the players in order to give meaningful feedback to the designers. We present CARMI: a Configurable Agent with Relative Metrics as Input. An agent able to emulate the players play-styles, even on previously unseen levels. Unlike current methods it does not rely on having full trajectories, but only summary data. Moreover it only requires little human data, thus compatible with the constraints of modern video game production. This novel agent could be used to investigate behaviors and balancing during the production of a video game with a realistic amount of training time.

Configurable Agent With Reward As Input: A Play-Style Continuum Generation

Modern video games are becoming richer and more complex in terms of game mechanics. This complexity allows for the emergence of a wide variety of ways to play the game across the players. From the point of view of the game designer, this means that one needs to anticipate a lot of different ways the game could be played. Machine Learning (ML) could help address this issue. More precisely, Reinforcement Learning is a promising answer to the need of automating video game testing. In this paper we present a video game environment which lets us define multiple play-styles. We then introduce CARI: a Configurable Agent with Reward as Input. An agent able to simulate a wide continuum range of play-styles. It is not constrained to extreme archetypal behaviors like current methods using reward shaping. In addition it achieves this through a single training loop, instead of the usual one loop per play-style. We compare this novel training approach with the more classic reward shaping approach and conclude that CARI can also outperform the baseline on archetypes generation. This novel agent could be used to investigate behaviors and balancing during the production of a video game with a realistic amount of training time.