Each section below summarizes one paper's reported method, experiments, and limits from its full review. The studies use different tasks and metrics; their results are not directly comparable.
arXiv:2610.41177
TraceSpec: Understanding and Improving Speculative Decoding for LLM Agents
TraceSpec studies speculative decoding for LLM agents at the level of individual trajectory decisions rather than aggregate task averages. It introduces TraceSpec-Traj, a replay corpus with 10,000 training states and 2,500 test states from web research and software engineering, and combines dynamic execution with fixed-prefix replay to expose differences across context length, workload, execution phase, and task progress. Its adaptive training method supervises intermediate assistant turns and reallocates a fixed prediction budget using state-conditioned, prefix-aware acceptance gaps.
How it works
TraceSpec-Traj is constructed from DeepResearch-traj, Nebius SWE-agent trajectories, and R2E-Gym development and testing-agent collections. It contains 10,000 training states and 2,500 test states, equally divided between web research and software engineering, with disjoint source tasks across splits. Each state records actual tokenized prefix length, workload, one of four phases—planning, action, reflection, or answer—preceding tool interactions, and trajectory position. The evaluation has two complementary modes.
What the paper evaluates
Dynamic evaluation uses 100 tasks from BrowseComp-Plus and 100 from SWE-bench Verified. Replay evaluates the 2,500 TraceSpec-Traj test states. The comparison includes DFlash, EAGLE-3, DSpark, LongSpark, DLite, and TraceSpec with Qwen3-4B, Qwen3-8B, and Qwen3-14B targets at temperatures 0 and 1, producing 108 configurations. DFlash, DSpark, and TraceSpec use seven-token draft blocks; dynamic episodes allow 50 action steps and 32,768 context tokens, while replay allows 36,864 context tokens and both modes allow up to 4,096 new tokens per call. The reported metrics are accepted length, verification yield, and generation throughput.
Limit to keep in view
The evaluation uses 100 tasks from each of BrowseComp-Plus and SWE-bench Verified, so the reported serving behavior may not cover other agent environments, tool APIs, or interaction styles.
arXiv:2611.07783
PBPO: Prefix-Block Policy Optimization for Diffusion Language Models
PBPO adapts reinforcement-learning policy optimization to masked diffusion language models by preserving block-level conditioning while applying importance weighting and clipping independently to each token. It constructs a prefix-block state that reveals completed response blocks and masks the current and future blocks, approximates block likelihood with one shared forward pass, and diversifies repeated updates through prompt and prefix masking. Across LLaDA-8B-Instruct and LLaDA-1.5, six reasoning, planning, and coding benchmarks, and generation lengths of 128, 256, and 512, PBPO reports competitive results and more stable training than the direct block-level objective, although it still relies on a surrogate likelihood and analysis assumptions.
How it works
For a response divided into blocks of length B_l, the prefix-block state z_i^{<b} reveals blocks before b and masks every position from block b onward. The dLLM predicts all masked positions from this shared state, so a single forward pass supplies the token probabilities needed for the current block. The direct block surrogate multiplies those probabilities and forms a block ratio across all tokens. PBPO changes only the weighting granularity.
What the paper evaluates
The evaluation covers six benchmarks. GSM8K and MATH500 measure mathematical reasoning with final-answer matching; Countdown measures valid target-reaching expressions; Sudoku measures complete-puzzle correctness; HumanEval and sanitized MBPP use execution-based pass@1. Training uses the GSM8K training split, MATH excluding MATH500, three-number TinyZero Countdown problems, a 4 × 4 Sudoku split, and KodCode-Light-RL-10K, with held-out evaluation sets described in the main body. PBPO is initialized from LLaDA-8B-Instruct and LLaDA-1.5 and compared with d1, wd1, GDPO, and DMPO.
Limit to keep in view
PBPO does not evaluate the exact diffusion response likelihood; it relies on a factorized one-step surrogate conditioned on a partially masked state.
arXiv:2611.08362
SpecCTMC: Exact Speculative Simulation of Diffusion Language Model Continuous-Time Markov Chains
SpecCTMC is a training-free speculative sampler for masked diffusion language models that accelerates exact continuous-time Markov chain (CTMC) generation without a separate draft model. It samples the CTMC’s event times in advance, drafts successive position-token updates using earlier model predictions, evaluates the tentative intermediate states in batches, and verifies proposals sequentially with residual correction at the first rejection. Under the paper’s stated conditions, this preserves the joint distribution of event times, token positions, token values, and resulting states produced by DirectCTMC.
How it works
The state x is a token sequence, and A(x) is the set of positions still containing the mask token; m(x) counts them. An event mark z=(d,v) selects a masked position d and a non-mask replacement token v. The target generator assigns rate pθ(v|x,d)/t to that mark. Because the token probabilities at each masked position are normalized, the total exit intensity is m(x)/t, and the event mark distribution is uniform over masked positions with token values drawn from the model.
What the paper evaluates
The authors evaluate LLaDA-8B-Instruct and Dream-v0-Instruct-7B on GSM8K, MATH, MBPP, HumanEval, and IFEval. They report accuracy, wall-clock time, and sequential target-model evaluation reduction relative to DirectCTMC. Generation lengths are 32, 64, 128, 256, and 512 tokens at sampling temperature one; speculation depth is selected separately per condition from 2 through 8 using warmed timing on five fixed questions. The paper reports results averaged across fixed evaluation examples and sampling seeds, and states that wall-clock measurements generally use one NVIDIA B200 GPU.
Limit to keep in view
Exactness depends on the paper’s stated conditions, including valid normalized target transition rows and correct acceptance and residual correction; implementations should verify these assumptions.
arXiv:2611.05894
RECAP: Context-Aware Prediction Recycling with Request-Local and Persistent Memories for Multimodal Speculative Decoding
RECAP is a training-free speculative decoding method for vision-language models that reuses target-model predictions as draft proposals. It addresses a weakness in token-only recycling: predictions stored by the last token can conflate different textual prefixes and images. RECAP keeps one-, two-, and three-token prediction memories for the current request, alongside a persistent token-level memory, then fuses available rows using specificity priors and entropy-based confidence. Best-first search builds a candidate tree, suffix retrieval adds observed continuations, and a single packed target pass verifies proposals while preserving greedy decoding.
How it works
At each evaluated position, the target model produces logits and RECAP retains a top-K candidate set with a normalized sparse distribution. It indexes that row by the final one, two, or three history tokens. Longer keys distinguish more prefixes, while shorter keys remain available when longer matches are absent. The two longer-context maps reset at each request, making the request's image and prompt implicit in their contents; the token-level map persists across requests.
What the paper evaluates
Evaluation uses Qwen2.5-VL-7B-Instruct and LLaVA-1.6 Vicuna-7B on an NVIDIA A100-SXM4-80GB GPU. Decoding is greedy, batch size one, with at most 200 new tokens. The seven single-turn benchmarks are MMT-Bench, SEED-Bench, ScienceQA, OCRBench, ChartQA, MathVista, and TextVQA, with 1,000 inputs per benchmark. MMSpec contributes 600 inputs across six categories and includes 223 generation turns from multi-turn conversations, for 723 turns total. Persistent memory is carried across each benchmark's inputs after warmup and cleared at benchmark boundaries; contextual maps reset per request.
Limit to keep in view
The evaluated setting is greedy generation at batch size one on two 7B vision-language models with image-based inputs; larger models, concurrent serving, and video are not evaluated.
Read the full reviews
- TraceSpec: Understanding and Improving Speculative Decoding for LLM Agents · arXiv:2610.41177
- PBPO: Prefix-Block Policy Optimization for Diffusion Language Models · arXiv:2611.07783
- SpecCTMC: Exact Speculative Simulation of Diffusion Language Model Continuous-Time Markov Chains · arXiv:2611.08362
- RECAP: Context-Aware Prediction Recycling with Request-Local and Persistent Memories for Multimodal Speculative Decoding · arXiv:2611.05894
Source note: findings and limitations above come from the linked full paper reviews. Follow each original arXiv paper for the authors' complete methods and results.