From Tokens to Steps: Verification-Aware Speculative Decoding for Efficient Multi-Step Reasoning

• URL: http://arxiv.org/abs/2604.15244v1
• Date: Thu, 16 Apr 2026 17:20:13 GMT
• Title: From Tokens to Steps: Verification-Aware Speculative Decoding for Efficient Multi-Step Reasoning
• Abstract summary: SpecGuard is a speculative decoding framework that performs step-level verification using only model-internal signals.<n> Experiments show that SpecGuard improves accuracy by 3.6% while reducing latency by 11%, outperforming both SD and reward-guided SD.
• Score: 18.37034672013596
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Speculative decoding (SD) accelerates large language model inference by allowing a lightweight draft model to propose outputs that a stronger target model verifies. However, its token-centric nature allows erroneous steps to propagate. Prior approaches mitigate this using external reward models, but incur additional latency, computational overhead, and limit generalizability. We propose SpecGuard, a verification-aware speculative decoding framework that performs step-level verification using only model-internal signals. At each step, SpecGuard samples multiple draft candidates and selects the most consistent step, which is then validated using an ensemble of two lightweight model-internal signals: (i) an attention-based grounding score that measures attribution to the input and previously accepted steps, and (ii) a log-probability-based score that captures token-level confidence. These signals jointly determine whether a step is accepted or recomputed using the target, allocating compute selectively. Experiments across a range of reasoning benchmarks show that SpecGuard improves accuracy by 3.6% while reducing latency by ~11%, outperforming both SD and reward-guided SD.

Related papers

• Verification-Aware Training for Speculative Decoding [57.84976863792784]
  We introduce Verification-Aware Training (VAT), a plug-in framework that simulates verification at every training step and turns the resulting accept and reject patterns into supervision.<n>VAT modifies only the training objective, so it can be layered on top of existing methods without changing the draft architecture, the target model, or the inference procedure.<n>VAT improves average acceptance length by up to 11.4% and wall-clock speedup by up to 8.7%, with consistent gains across math, code, and chat benchmarks.
  arXiv  Detail & Related papers  (2026-08-31T01:42:10Z)
• Self-Improving In-Context Learning [3.9202238580555417]
  We propose to improve in-context learning by optimizing the embeddings of a few prompt test at time.<n>We formalize this signal as a bounded, self-supervised calibration procedure.<n>Across a comprehensive suite of ICL tasks, the proposed calibration improves or matches upon the base model and consistently outperforms classification-specific baselines on most tasks.
  arXiv  Detail & Related papers  (2026-05-22T03:01:34Z)
• Calibrated Speculative Decoding: Frequency-Guided Candidate Selection for Efficient Inference [27.59556627479635]
  Calibrated Speculative Decoding recovers valid tokens discarded by standard verification.<n>Online Correction Memory aggregates historical rejections to propose recurring divergence patterns as rescue candidates.<n>Semantic Consistency Gating verifies candidate admissibility using probability ratios instead of exact token matching.
  arXiv  Detail & Related papers  (2026-04-15T09:01:54Z)
• Towards Verifiable AI with Lightweight Cryptographic Proofs of Inference [3.3323431541048385]
  We present a verification framework and protocol that replaces full proofs with a lightweight, sampling-based approach.<n>We formalize the conditions under which trace separation between functionally dissimilar models can be leveraged to argue the security of verifiable inference protocols.<n>Our approach reduces proving times by several orders of magnitude compared to state-of-the-art cryptographic proof systems.
  arXiv  Detail & Related papers  (2026-03-19T15:24:27Z)
• Training-free Dropout Sampling for Semantic Token Acceptance in Speculative Decoding [13.249778063956917]
  Speculative decoding accelerates large language model inference by proposing tokens with a lightweight draft model.<n>This work introduces DropMatch, a novel approach that matches draft tokens to the predictive distribution of the target model.<n> Experiments across multiple benchmarks demonstrate that our approach increases acceptance length while maintaining competitive task performance.
  arXiv  Detail & Related papers  (2026-02-11T04:53:33Z)
• TriSpec: Ternary Speculative Decoding via Lightweight Proxy Verification [63.65902785448346]
  Speculative decoding offers significant speed-ups through its lightweight drafting and parallel verification mechanism.<n>We propose TriSpec, a novel ternary SD framework that introduces a lightweight proxy to significantly reduce computational cost.<n>Experiments on the Qwen3 and DeepSeek-R1-Distill-Qwen/LLaMA families show that TriSpec achieves up to 35% speedup over standard SD.
  arXiv  Detail & Related papers  (2026-01-30T17:04:18Z)
• AdaSPEC: Selective Knowledge Distillation for Efficient Speculative Decoders [36.345954548346235]
  Speculative Decoding (SD) accelerates large language model inference by employing a small draft model to generate predictions.<n> Knowledge Distillation (KD) aims to minimize the KL divergence between the draft and target models across all tokens.<n>We propose AdaSPEC, a novel method that incorporates selective token filtering into the KD process.
  arXiv  Detail & Related papers  (2025-10-22T17:13:00Z)
• LaSeR: Reinforcement Learning with Last-Token Self-Rewarding [54.72617309922891]
  Reinforcement Learning with Verifiable Rewards (RLVR) has emerged as a core paradigm for enhancing the reasoning capabilities of Large Language Models (LLMs)<n>Previous practice requires the LLM to sequentially generate solutions and self-verifications using two separate prompt templates, which significantly reduces efficiency.<n>We propose LaSeR (Reinforcement Learning with Last-Token Self-Rewarding), an algorithm that simply augments the original RLVR loss with a MSE loss.
  arXiv  Detail & Related papers  (2025-10-16T17:55:11Z)
• Every Step Counts: Decoding Trajectories as Authorship Fingerprints of dLLMs [63.82840470917859]
  We show that the decoding mechanism of dLLMs can be used as a powerful tool for model attribution.<n>We propose a novel information extraction scheme called the Directed Decoding Map (DDM), which captures structural relationships between decoding steps and better reveals model-specific behaviors.
  arXiv  Detail & Related papers  (2025-10-02T06:25:10Z)
• Towards Better Code Generation: Adaptive Decoding with Uncertainty Guidance [42.737012213197865]
  AdaDec is an adaptive decoding framework that employs a lookahead-based, uncertainty-aware pause-and-rerank mechanism.<n>AdaDec achieves up to 20.9% absolute gains in Pass@1 accuracy compared with greedy decoding.<n>By applying reranking only when necessary, AdaDec reduces computational overhead and latency, enhancing efficiency alongside reliability.
  arXiv  Detail & Related papers  (2025-06-10T16:49:46Z)
• Refine, Discriminate and Align: Stealing Encoders via Sample-Wise Prototypes and Multi-Relational Extraction [57.16121098944589]
  RDA is a pioneering approach designed to address two primary deficiencies prevalent in previous endeavors aiming at stealing pre-trained encoders. It is accomplished via a sample-wise prototype, which consolidates the target encoder's representations for a given sample's various perspectives. For more potent efficacy, we develop a multi-relational extraction loss that trains the surrogate encoder to Discriminate mismatched embedding-prototype pairs.
  arXiv  Detail & Related papers  (2023-12-01T15:03:29Z)
• KECOR: Kernel Coding Rate Maximization for Active 3D Object Detection [48.66703222700795]
  We resort to a novel kernel strategy to identify the most informative point clouds to acquire labels. To accommodate both one-stage (i.e., SECOND) and two-stage detectors, we incorporate the classification entropy tangent and well trade-off between detection performance and the total number of bounding boxes selected for annotation. Our results show that approximately 44% box-level annotation costs and 26% computational time are reduced compared to the state-of-the-art method.
  arXiv  Detail & Related papers  (2023-07-16T04:27:03Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.