Bebop: Improving Multi-Token Prediction Efficiency in Reinforcement Learning for Large Language Models
Researchers have introduced Bebop, a system that accelerates the rollout stage of reinforcement learning (RL) training for large language models by improving Multi-Token Prediction (MTP) acceptance rates. The work identifies that entropy fluctuations during RL training fundamentally degrade MTP performance, and proposes a novel end-to-end Total Variation (TV) loss combined with probabilistic rejection sampling to counteract this. The approach achieves up to 95% acceptance rates and 1.8x end-to-end training acceleration on Qwen3.5, Qwen3.6, and Qwen3.7 models, potentially reducing the cost and time of RL-based LLM post-training.
A team of researchers has published Bebop, a systematic study and practical framework for integrating Multi-Token Prediction (MTP) into large-scale reinforcement learning (RL) pipelines for large language models (LLMs). The paper identifies a core problem: MTP acceptance rates—critical for speculative decoding speedups—degrade during RL training due to entropy fluctuations, exhibiting a clear negative linear relationship with rising model entropy. To address this, the authors show that probabilistic rejection sampling outperforms greedy draft sampling in mitigating entropy-induced disturbances. They further argue that conventional MTP training objectives such as cross-entropy or KL divergence are suboptimal in this setting, and propose a novel end-to-end TV (Total Variation) loss that directly optimizes multi-step rejection sampling acceptance rates, yielding approximately 10 percentage point improvements and up to 95% acceptance rates. Experiments span mathematical reasoning, code generation, and agentic tasks, with results showing up to 25% extra inference throughput gains and up to 1.8x end-to-end acceleration in asynchronous RL training. Notably, the authors find that pre-RL MTP training with the proposed loss and sampling strategy eliminates the need for costly online MTP updates during RL, simplifying deployment. The work was validated on Qwen3.5, Qwen3.6, and Qwen3.7 model families.
What's missing
It is unclear whether the TV loss and rejection sampling approach generalizes beyond the Qwen model family to other LLM architectures. As a preprint, the work has not yet undergone formal peer review.
What different sources said
- arXiv cs.LGCenter
Breaking Entropy Bounds: Accelerating RL Training via MTP with Rejection Sampling
Related
Gut Bacteria Enzyme Found to Break Down Heat-Processed Food Compounds, Producing Novel Biogenic Amines
Researchers have discovered that an enzyme in common gut bacteria can degrade N-epsilon-carboxymethyllysine (CML), a compound formed during thermal food processing, producing previously unknown biogenic amines. The enzyme, ornithine decarboxylase SpeC from enterobacteria, acts on CML and related modified lysine derivatives through a low-level 'underground' catalytic activity. This finding suggests a previously unrecognized communication axis between thermally processed dietary compounds and gut microbial physiology, with potential implications for host health.
Full-Length Gene Sequencing Reveals Two Distinct Bacterial Communities in Black-Legged Ticks Expanding Into Canada
Researchers used Oxford Nanopore full-length 16S rRNA gene sequencing to characterize the microbiome of Ixodes scapularis black-legged ticks collected in Nova Scotia, Canada, distinguishing between tick-adapted bacteria and environmentally acquired bacteria. The study comes as I. scapularis — the primary vector of Lyme disease — is rapidly expanding northward into Canada due to climate change. The findings suggest that environmentally derived bacteria in tick microbiomes are not mere contamination, which has implications for how tick microbiome data is collected and interpreted across surveillance studies.
Study Identifies Metabolic Link Between Cell Envelope Stress and Biofilm Formation in Bacteria
Researchers have discovered that the metabolite acetyl-CoA directly inhibits enzymes that degrade the bacterial signaling molecule c-di-GMP, connecting cell envelope biosynthesis stress to biofilm formation in Pseudomonas aeruginosa. The study found that sub-inhibitory concentrations of antibiotics targeting early peptidoglycan biosynthesis — but not other antibiotic classes — elevate c-di-GMP levels by reducing phosphodiesterase activity, with acetyl-CoA competing for the enzyme active site. Because the relevant enzyme domain is broadly conserved across bacterial species, this checkpoint mechanism may be widespread and could have implications for understanding antibiotic-induced biofilm responses.