NgramRL is a new method for accelerating reinforcement‑learning post‑training of large language models by using a per‑prompt n‑gram drafter built from recent rollout history, eliminating the need for auxiliary models or persistent GPU state. It incorporates an adaptive fan‑out prefetcher that predicts verification outcomes to hide CPU cache lookup and scheduling delays, and a causal budget allocator that focuses prefetches on high‑probability outcomes. Experiments on two model families and three reasoning/structured‑generation workloads show that NgramRL achieves 3.60–7.77 token savings per target‑model forward in the warm‑cache regime, up to 1.99× faster than SuffixDecoding and 7.0× faster than prompt‑lookup decoding, with the prefetcher improving throughput by 1.38× on Spider.
arXiv:2606. 18967v1 Announce Type: new Abstract: Reinforcement learning (RL) has become a representative post-training paradigm for LLMs, enabling strong reasoning and agentic capabilities.
The paper introduces AdaptiveSpec, a training‑free speculative decoding method that simultaneously adapts the per‑step verification rule and the draft‑tree shape using signals generated during decoding. It replaces the fixed token‑match rule with a margin‑based threshold and adjusts tree depth, width, and node count based on draft confidence and recent acceptance history, allowing the total draft count to vary. Experiments on SGLang show up to 56% throughput gains over EAGLE‑3 while maintaining 93% of lossless task accuracy on GSM8K, MATH‑500, and HumanEval across three models.
The paper introduces APEX, a learned controller that improves speculative decoding for large language models by dynamically selecting the best speculation strategy and adjusting draft depth during generation. APEX uses request‑level routing among EAGLE‑3, n‑gram, and draft‑model speculation, and block‑level depth adaptation based on causal decoding signals and verifier feedback. Integrated into vLLM and tested with Qwen3‑8B, APEX achieves up to 5.24× speedup over autoregressive decoding while reducing wasted tokens by 41% compared to fixed‑depth n‑gram speculation.
The paper introduces the Expected Decoding Rounds (EDR) objective, a theoretical framework that models speculative decoding as a Markov reward process and directly optimizes the number of decoding rounds. EDR weights local rejection costs by state occupancies, eliminating auxiliary hyperparameters and enabling unbiased stochastic optimization via temporal‑difference gradients. Applying EDR to fine‑tune state‑of‑the‑art draft models (DSpark and DFly) improves mean accepted length and outperforms prior training objectives across nine benchmarks in math reasoning, code generation, and chat.
The paper introduces DCOCO (DFlash Contextual Corrector), a lightweight mechanism that restores intra-block dependencies in speculative decoding by evaluating multiple candidate tokens for contextual compatibility across adjacent draft positions. DCOCO maintains parallel drafting while enabling contextually informed token selection at low additional cost. Across multiple verifier backbones and benchmarks, DCOCO achieves the highest average throughput, delivering speedups of up to 6.6× over vanilla decoding and outperforming competing methods by 6–23% under concurrent serving loads.
arXiv:2608. 10362v1 Announce Type: cross Abstract: Speculative decoding accelerates autoregressive large language model (LLM) inference by using a lightweight draft model to speculate multiple tokens, reducing expensive target model decoding steps.
The paper introduces Self‑Sparse Speculative Rollouts (S$^3$R), a method that uses sparse attention to propose tokens and then verifies them with full attention, thereby reducing KV‑cache traffic while preserving the dense‑policy distribution for reinforcement learning rollouts. S$^3$R eliminates the need for a separate drafter by sharing policy parameters between drafting and verification, and a KV‑aware scheduler limits cache thrashing to maintain parallelism. Experiments on long‑context mathematical reasoning with the Qwen3 model show that S$^3$R achieves 1.35–1.39× higher end‑to‑end training throughput with comparable task performance.
arXiv:2608. 04962v1 Announce Type: new Abstract: Reinforcement learning (RL) post-training improves the reasoning capabilities of large language models, but autoregressive rollout generation remains a major efficiency bottleneck.
arXiv:2606. 01019v1 Announce Type: cross Abstract: Large Language Model (LLM) generation remains expensive because autoregressive decoding calls the model once for each new token.
arXiv:2609.24197v1 Announce Type: new Abstract: Speculative decoding losslessly accelerates large language model inference by having a lightweight draft model predict future tokens for verification b...
Draft-OPD introduces an on‑policy distillation method for speculative draft models, addressing the mismatch between supervised fine‑tuning and inference by letting the target model supervise the drafter on draft‑induced states. The approach uses target‑assisted rollouts for stable continuations and replays drafting from error positions exposed during verification, enabling the drafter to learn from both accepted and rejected proposals. Experiments demonstrate that Draft‑OPD achieves more than five‑fold lossless acceleration across diverse tasks, outperforming prior draft models such as EAGLE‑3 and DFlash by 23 % and 13 % respectively.
The paper introduces DLoop, a looped speculative decoding technique that adaptively performs multiple drafting stages before verification, allowing a draft model to continue generating tokens while confident. By verifying all accumulated draft tokens together and training the draft model to handle its own hidden states for unverified tokens, DLoop reduces the number of target‑model forward passes needed. Experiments across several speculative decoding methods show wall‑clock speedups of 5–41 % without sacrificing lossless decoding.