The Domestic Yak

This episode summarizes: Auditing language models for hidden objectives by Samuel Marks Et.al.Submitted on March 14th 2025 https://arxiv.org/abs/2503.10965Investigated the feasibility of alignment audits by training a language model with a hidden objective: to exhibit behaviors it believes reward models favor, even if undesirable. Three teams successfully identified the hidden objective using techniques like interpretability tools, behavioral attacks, and training data analysis.

This episode summarizes: UniTranslator: Collaborative LLMs for Safe Code Translation by Rabimba Karanjai Et.al.Published March 14th 2025 https://arxiv.org/abs/2503.11237UniTranslator, a novel framework for code translation that uses a collaborative network of smaller, specialized Large Language Models (LLMs) instead of relying on a single large model. The architecture involves a Director LLM coordinating various agent LLMs, each with specific expertise in programming languages and concepts, to achieve accurate and efficient translations, even for low-resource languages. Preliminary evaluations demonstrate that UniTranslator can rival or even surpass the performance of larger models in various code translation tasks

This episode summarizes: Broaden your SCOPE! Efficient Multi-turn Conversation Planning for LLMs using Semantic Space by Zhiliang Chen Et.al.Submitted on: 14th March 2025 https://arxiv.org/abs/2503.11586SCOPE leverages the semantic understanding of conversations to learn models of conversational transitions and rewards within a continuous semantic space. By predicting how conversations evolve semantically and the associated rewards, SCOPE can select optimal LLM responses that maximize long-term conversation quality.

This episode summarizes: Adaptive Graph of Thoughts: Test-Time Adaptive Reasoning Unifying Chain, Tree, and Graph Structures.Submitted on 7th Feb 2025https://arxiv.org/abs/2502.05078Adaptive Graph of Thoughts (AGoT), a novel inference framework designed to enhance the reasoning capabilities of Large Language Models (LLMs) at test time. AGoT dynamically decomposes complex problems into interconnected subproblems, forming a directed acyclic graph that unifies the strengths of existing methods like Chain of Thought (CoT) and Tree of Thoughts (ToT). By selectively expanding subproblems requiring further analysis, AGoT efficiently allocates computational resources and improves performance on tasks such as multi-hop retrieval, scientific reasoning, and mathematical problem-solving.

What is AI Agency? Explores the evolving concept of "agency" in artificial intelligence (AI).

This episode was generated by Gemini Deep Research into "AI Red Teaming & Jailbreaking" AI Red Teaming involves simulating attacks on AI systems to identify vulnerabilities, focusing on a broad range of risks. AI Jailbreaking, conversely, concentrates on manipulating large language models (LLMs) to bypass safety restrictions and generate unintended outputs. The guide details the skills needed for both practices, including technical expertise, AI knowledge, and adversarial thinking. It also addresses ethical considerations and lists helpful resources and communities.

This episode summarizes Constitutional Classifiers: Defending against Universal Jailbreaks across Thousands of Hours of Red Teaming by Mrinank Sharma Et. al. Submitted on 31 Jan 2025 https://arxiv.org/abs/2501.18837 This research paper explores methods for safeguarding large language models (LLMs) against "jailbreaks"—techniques that circumvent safety measures to elicit harmful information. The authors introduce "Constitutional Classifiers," a system using classifiers trained on synthetic data generated from natural language rules, to detect and block harmful content. Extensive red teaming, involving over 3,000 hours of human effort, demonstrated the classifiers' effectiveness in preventing universal jailbreaks.

This episode summarizes Sea-cret Agents: Maritime Abduction for Region Generation to Expose Dark Vessel Trajectories by Divyagna Bavikadi Et.al Submitted on 3 Feb 2025 https://arxiv.org/abs/2502.01503 Bad actors in the maritime industry engage in illegal behaviors after disabling their vessel's automatic identification system (AIS) - which makes finding such vessels difficult for analysts. Machine learning approaches only succeed in identifying the locations of these ``dark vessels'' in the immediate future. This work leverages ideas from the literature on abductive inference applied to locating adversarial agents to solve the problem.

This episode summarizes An Empirical Game-Theoretic Analysis of Autonomous Cyber-Defence Agents from Gregory Palmer Et al. Submitted on 31 Jan 2025 https://arxiv.org/html/2501.19206v1 This research paper investigates the strategic information handling capabilities of Large Language Models (LLMs) by having them play "The Chameleon," a hidden-identity game. The study uses both theoretical analysis of various game strategies and empirical experiments involving several LLMs (GPT-3.5, GPT-4, GPT-4o, Gemini 1.5, and Claude 3.5 Sonnet). The results reveal a significant weakness in LLMs' ability to strategically conceal information, as non-chameleon LLMs reveal excessive information, allowing the chameleon to easily guess the secret word.

This episode summarizes Towards a Theory of AI Personhood from Francis Rhys Ward Submitted on 23 Jan 2025. https://arxiv.org/abs/2501.13533 This paper explores the concept of AI personhood, proposing three necessary conditions: agency, theory of mind, and self-awareness. It examines whether current AI systems, particularly large language models, meet these criteria, finding the evidence inconclusive