Gaia Molinaro

Jérémy Perez, Grgur Kovač, Corentin Léger et al.

As large language models (LLMs) start interacting with each other and generating an increasing amount of text online, it becomes crucial to better understand how information is transformed as it passes from one LLM to the next. While significant research has examined individual LLM behaviors, existing studies have largely overlooked the collective behaviors and information distortions arising from iterated LLM interactions. Small biases, negligible at the single output level, risk being amplified in iterated interactions, potentially leading the content to evolve towards attractor states. In a series of telephone game experiments, we apply a transmission chain design borrowed from the human cultural evolution literature: LLM agents iteratively receive, produce, and transmit texts from the previous to the next agent in the chain. By tracking the evolution of text toxicity, positivity, difficulty, and length across transmission chains, we uncover the existence of biases and attractors, and study their dependence on the initial text, the instructions, language model, and model size. For instance, we find that more open-ended instructions lead to stronger attraction effects compared to more constrained tasks. We also find that different text properties display different sensitivity to attraction effects, with toxicity leading to stronger attractors than length. These findings highlight the importance of accounting for multi-step transmission dynamics and represent a first step towards a more comprehensive understanding of LLM cultural dynamics.

Julien Pourcel, Cédric Colas, Gaia Molinaro et al.

The ability to invent novel and interesting problems is a remarkable feature of human intelligence that drives innovation, art, and science. We propose a method that aims to automate this process by harnessing the power of state-of-the-art generative models to produce a diversity of challenging yet solvable problems, here in the context of Python programming puzzles. Inspired by the intrinsically motivated literature, Autotelic CodE Search (ACES) jointly optimizes for the diversity and difficulty of generated problems. We represent problems in a space of LLM-generated semantic descriptors describing the programming skills required to solve them (e.g. string manipulation, dynamic programming, etc.) and measure their difficulty empirically as a linearly decreasing function of the success rate of Llama-3-70B, a state-of-the-art LLM problem solver. ACES iteratively prompts a large language model to generate difficult problems achieving a diversity of target semantic descriptors (goal-directed exploration) using previously generated problems as in-context examples. ACES generates problems that are more diverse and more challenging than problems produced by baseline methods and three times more challenging than problems found in existing Python programming benchmarks on average across 11 state-of-the-art code LLMs.

Yuzheng Xu, Annya Dahmani, Matthew D. Blanchard et al.

Human-AI complementarity, the idea that combining human and AI judgments can outperform either alone, offers a promising pathway toward robust oversight of advanced AI systems. However, whether human-AI complementarity can be achieved on realistic tasks remains an open question. We investigate this through two approaches: hybridization and two AI assistance methods (top-2 assistance and subtask delegation), evaluated on a multi-domain dataset of 1,886 samples spanning knowledge, factuality, long-context reasoning, and deception detection. We find only modest complementarity gains. Baseline hybridization yields just +0.4 percentage points (pp) over AI alone (69.3\% vs 68.9\%), limited both by a small complementarity region (only 8.9\% of items where AI errs but humans do not) and the inability of confidence-based routing to identify it, since the model's confidence is similarly distributed across correct and incorrect predictions. Applied when AI has low confidence, top-2 assistance increases human accuracy from 28.4\% to 38.3\%, surpassing AI alone (37.7\%) -- but primarily because humans adopt correct AI suggestions, not because they successfully override AI errors. These findings suggest that the primary bottleneck is not human task accuracy per se, but the ability to route decisions to humans when it matters and to design assistance methods that enable humans to catch AI mistakes. Our quantitative and qualitative analyses pinpoint where and why each method succeeds or fails, offering concrete targets for future work. We will release our dataset and code upon request to support progress toward more effective human-AI collaboration for AI oversight.

Gaia Molinaro, Dave August, Danielle Perszyk et al.

As large language models (LLMs) get integrated into human decision-making, they are increasingly choosing goals autonomously rather than only completing human-defined ones, assuming they will reflect human preferences. However, human-LLM similarity in goal selection remains largely untested. We directly assess the validity of LLMs as proxies for human goal selection in a controlled, open-ended learning task borrowed from cognitive science. Across four state-of-the-art models (GPT-5, Gemini 2.5 Pro, Claude Sonnet 4.5, and Centaur), we find substantial divergence from human behavior. While people gradually explore and learn to achieve goals with diversity across individuals, most models exploit a single identified solution (reward hacking) or show surprisingly low performance, with distinct patterns across models and little variability across instances of the same model. Even Centaur, explicitly trained to emulate humans in experimental settings, poorly captures people's goal selection. Chain-of-thought reasoning and persona steering provide limited improvements. These findings highlight the uniqueness of human goal selection, cautioning against replacing it with current models in applications such as personal assistance, scientific discovery, and policy research.

Gaia Molinaro, Anne G. E. Collins

Reinforcement learning agents learn from rewards, but humans can uniquely assign value to novel, abstract outcomes in a goal-dependent manner. However, this flexibility is cognitively costly, making learning less efficient. Here, we propose that goal-dependent learning is initially supported by a capacity-limited working memory system. With consistent experience, learners create a "compressed" reward function (a simplified rule defining the goal) which is then transferred to long-term memory and applied automatically upon receiving feedback. This process frees up working memory resources, boosting learning efficiency. We test this theory across six experiments. Consistent with our predictions, our findings demonstrate that learning is parametrically impaired by the size of the goal space, but improves when the goal space structure allows for compression. We also find faster reward processing to correlate with better learning performance, supporting the idea that as goal valuation becomes more automatic, more resources are available for learning. We leverage computational modeling to support this interpretation. Our work suggests that efficient goal-directed learning relies on compressing complex goal information into a stable reward function, shedding light on the cognitive mechanisms of human motivation. These findings generate new insights into the neuroscience of intrinsic motivation and could help improve behavioral techniques that support people in achieving their goals.