arXiv:2602.05183v2 Announce Type: replace-cross Abstract: Large language models (LLMs) are increasingly trained in complex Reinforcement Learning, multi-agent environments, making it difficult to understand how behavior changes over training. Sparse Autoencoders (SAEs) have recently shown to be useful for data-centric interpretability. In this work, we analyze large-scale reinforcement learning training runs from the sophisticated environment of Full-Press Diplomacy by applying pretrained SAEs, alongside LLM-summarizer methods. We introduce Meta-Autointerp, a method for grouping SAE features into interpretable hypotheses about training dynamics. We discover fine-grained behaviors including role-playing patterns, degenerate outputs, language switching, alongside high-level strategic behaviors and environment-specific bugs. Through automated evaluation, we validate that 90% of discovered SAE Meta-Features are significant, and find a surprising reward hacking behavior. However, through two user studies, we find that even subjectively interesting and seemingly helpful SAE features may be worse than useless to humans, along with most LLM generated hypotheses. However, a subset of SAE-derived hypotheses are predictively useful for downstream tasks. We further provide validation by augmenting an untrained agent's system prompt, improving the score by +14.2%. Overall, we show that SAEs and LLM-summarizer provide complementary views into agent behavior, and together our framework forms a practical starting point for future data-centric interpretability work on ensuring trustworthy LLM behavior throughout training.

Background: Mobile phone ecological momentary assessment (EMA) methods are a well-established measure of eating and drinking behaviors, but compliance can be poor. Micro-EMA (μEMA), which collects information with a single tap response to brief questions on smartwatches, offers a novel application that may improve response rates. To our knowledge, there is no data evaluating μEMA to measure eating habits in children or in low-to-middle-income countries. Objective: In this study, we investigated the feasibility of micro-EMA to measure eating patterns in Malaysian children and adolescents. Methods: We invited 100 children and adolescents aged 7-18 years in Segamat, Malaysia, to participate in 2021-2022. Smartwatches were distributed to 83 children and adolescents who agreed to participate. Participants were asked to wear the smartwatch for 8 days and respond to 12 prompts per day, hourly, from 9AM to 8PM, asking for information on their meals, snacks, and drinks consumed. A questionnaire captured their experiences using the smartwatch and μEMA interface. Response rate (proportion of prompts responded to) assessed participants’ adherence. We explored associations between response rate with time of day, across days, age, and sex using multilevel binomial logistic regression modeling. Results: Eighty-two participants provided usable smartwatch data. The median number (IQR) of meals, drinks, and snacks per day was 2 (2-4), 3 (1-5), and 1 (0-2), respectively, on the first day of the study. The median response rate across the study was 68% (IQR 50-83). The response rate decreased across study days from 74% (68-78) on Day 1 to 40% (30-50) on Day 7 (odds ratio [OR] per study day 0.73, 95% CI 0.64-0.83). Response rate was lowest at the start of the day and highest between the hours of 12 PM and 2 PM. Female participants responded to more prompts than male participants (OR 1.72, 95% CI 1.03-2.86). There was no evidence of differential response by age (OR 0.73, 95% CI 0.41-1.28). Most participants (65%) rated their experience using the smartwatch positively, with 33% saying they were happy to participate in future studies using the smartwatch. For children that did not wear the smartwatch for the full study duration (n=22), discomfort was the most common complaint (41%). Conclusions: In this study of the feasibility of μEMA on smartwatches to measure eating in Malaysian children, we found the method was acceptable. However, response rates declined across study days, resulting in substantial missingness. Future studies (eg, through focus groups) should explore approaches to improving response to event prompts, trial alternative devices to increase children’s comfort, and evaluate revised protocols for reporting of intake events.