TalentRL learns a strategy layer above the skill library, and the strategy layer is what routes
AspireAI community · arXiv
Natural-language skills let LLM agents reuse procedures across tasks, but most systems pull individual skills from a flat SkillBank without modelling the high-level strategy that governs which to use. TalentRL adds that missing layer: coarse-grained talents represent recurring strategies for skill use while fine-grained skills are the executable procedures, and for each task the policy model selects a talent that guides execution and softly prioritises its associated skills during similarity-based retrieval. Because task outcomes depend jointly on talent selection, retrieved skill context and downstream execution, terminal rewards alone cannot isolate the routing decision - so TalentRL trains talent routing with reward contrasts between rollouts of the same task under matched interventions on the talent-conditioned context, and optimises downstream execution with group-relative advantages. Across seven open-domain and multi-hop question-answering benchmarks the authors report talent-aware supervised fine-tuning improves average success rate by 7.53%, 6.71% and 9.21% over the corresponding zero-shot baselines for Qwen2.5-7B, Qwen3-8B and Llama-3.1-8B respectively, that reinforcement learning improves those SFT checkpoints by a further 7.58% to 9.32%, and that TalentRL achieves the highest average success rate among all compared methods on all three backbones. On the agentic benchmarks ALFWorld and WebShop it reports the highest ALFWorld success rate (99.3%) and the highest WebShop score (90.7) among the compared methods. The limitation is that the gains are measured against a flat-skill baseline the authors define.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.