RISED is a method for training LLM agents across multiple interactive environments, improving generalist agents by considering relationships between environments and using rubrics to guide learning. It outperforms other methods in multi-environment RL and can characterize behavioral changes.

A new dataset (AED) for identifying errors and failures in AI agents is introduced, with 50,228 error-diagnosis pairs from various environments and policy models. The dataset aims to support cross-setting failure analysis and re-diagnosis without repeating the original rollout. The authors present a five-stage pipeline for collecting natural failures, generating diagnoses, and checking proposed corrections against recorded evidence.

Researchers explored character training as a method to instill risk aversion in AI agents to prevent misaligned behavior. They trained models using constant absolute risk aversion (CARA) and found that character-trained models performed competitively with baselines and better out-of-distribution. Token budget and model choice are key factors in instilling risk aversion.

Researchers introduced SkillGym, a pipeline to train skill-use agents with verifiable environments. The pipeline crawls skills from the internet, builds difficulty-controlled tasks, and collects verified trajectories for supervised finetuning. This improves LLM performance on skill-use benchmarks.

This paper explores the effect of prompt choice on transfer in on-policy distillation, a technique for training AI agents. The study finds that a few well-chosen prompts can be as effective as a large pool of prompts, but the effectiveness depends on the teacher-student pair and target capability. This research has implications for the development of AI agents and the choice of prompts for on-policy distillation.

The paper presents UpliftMem, a method for learning memory retrieval for large language model (LLM) agents. UpliftMem learns which memory sets improve execution without relying on costly outcome feedback, instead using set-level execution uplift relative to the same executor without memory. This approach is evaluated on three benchmarks (ALFWorld, WebShop, and BigCodeBench) and outperforms other baselines.

This paper studies the use of intuitive prompting to improve the fidelity of language models (LLMs) simulating social media reactions. The authors found that instructing LLMs to respond intuitively and immediately resulted in higher fidelity and better performance on unfamiliar content, suggesting potential applications for general-purpose simulated users.