NgramRL drafts RL rollouts from the model's own previous answers and committed up to 7.77 tokens per forward pass
AspireAI community · arXiv (Nilesh Prasad Pandey)
RL post-training for reasoning models is dominated by autoregressive rollout generation, which the paper states can exceed 70 percent of training time. Existing speculative-decoding methods fit this setting poorly: model-based drafters consume scarce GPU memory and drift out of alignment as the policy evolves, while model-free methods do not exploit repetition across training epochs. NgramRL's observation is that recent rollouts of the same prompt are themselves a high-quality, self-updating draft source. It builds a per-prompt n-gram drafter from bounded recent rollout history, with no auxiliary model, no drafter training and no persistent GPU-resident drafter state, while target-model verification preserves the rollout distribution. An adaptive fan-out prefetcher predicts likely verification outcomes to hide CPU cache lookup and scheduling. Across two model families and three reasoning and structured-generation workloads, NgramRL committed 3.60 to 7.77 tokens per target-model forward pass in the warm-cache regime, up to 1.99x more than SuffixDecoding and 7.0x more than prompt-lookup decoding. On Spider, the prefetcher covered 96.6 percent of Qwen3-8B outcomes at a candidate budget of 16 and improved rollout throughput by 1.38x over synchronous NgramRL, approaching the 1.40x overlap ceiling. The honest caveat: the speedups are measured in the warm-cache regime, so the gains depend on reusing rollouts for the same prompt rather than on general-purpose decoding.
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.