Agentic RL credit assignment improved by reusing evidence already in the rollout batch beat GiGPO by 3.4 and 5.9 points and GRPO by 16.4 and 8.3 points, with no extra rollouts
AspireAI community · arXiv
*In long-horizon agentic reinforcement learning, credit for a single action is estimated from one sampled continuation, so an action gets judged on whatever the model happened to do afterwards - and that is where local credit gets distorted.* ConCal is a recursive continuation-value estimator that matches visits of the same task by their incoming state-action-state transitions, combines the continuation evidence those visits share with each visit's own transition estimate, and propagates the calibrated values backward so earlier actions inherit evidence from later ones. With Qwen2.5-1.5B-Instruct it improves success rates over GiGPO by 3.39 points on ALFWorld and 5.86 points on WebShop, and over GRPO by 16.41 and 8.34 points, reaching 94.27 percent overall success on ALFWorld and a WebShop task score of 90.18 with 79.69 percent success, with no additional rollouts. The diagnostics explain why: shared continuation evidence is directly available for 55.4 percent of recorded actions on ALFWorld and 43.4 percent on WebShop, and recursive propagation extends that coverage to 92.5 and 67.5 percent, while centered-value error falls 34.3 and 16.8 percent. The limitation is the scale, a 1.5B model on two agent benchmarks, so whether the estimator still helps at frontier scale is open.
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.