Research map / Large language models / Language, safety and interpretability

STDec: Spatio-Temporal Stability Guided Decoding for dLLMs

NeurIPS 2026 · Large language models › Language, safety and interpretability · approach: Parallel and early-exit decoding · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Training-free decoder for diffusion LLMs that replaces the global confidence threshold with token-adaptive thresholds from spatial (Gaussian-smoothed decoded neighborhood) and temporal (predicted-ID consistency) stability, yielding multi-x throughput gains at comparable scores.

Problem

Most dLLM decoders use a single global confidence threshold (top-k or fixed), ignoring local context from neighboring decoded tokens and cross-step consistency of predicted token IDs, which limits decoding speed.

Approach

Build a per-token threshold by Gaussian-smoothing a map that assigns low values to decoded tokens and high values to masked tokens (so masked tokens near decoded context get lower thresholds), then multiply by a relaxation factor for tokens whose top-1 ID has stayed unchanged for at least two consecutive steps and whose confidence was already above the low threshold.

Key results

• On MBPP with LLaDA-8B-Instruct, STDec achieves 14.17x speedup (TPS 80.47 vs 5.68) with score 38.40 vs 37.60.
• Averaged over five textual reasoning benchmarks with LLaDA, STDec gives 7.60x TPS speedup with average score 46.87 vs 46.59 for vanilla.
• On LaViDa-Reason multimodal math benchmarks, STDec averages 3.57x speedup (TPS 19.86 vs 5.57) with score 32.07 vs 31.75.

Surprising findings

• Mean smoothing (distance-agnostic) yields a notably larger score drop (HumanEval 40.85 vs 48.78 for Gaussian), showing distance-aware weighting matters.
• Temporal-aware decoding alone slightly raised GSM8K score (78.01 to 79.08) while accelerating 2.41x.

Limitations

• Gains are smaller for tasks generating few tokens, such as yes/no questions, as acknowledged by the authors.
• Hyperparameters (tau_high, tau_low, alpha, sigma) were tuned on HumanEval and checked on GSM8K; transfer to other settings is only partially shown.
• Several speedups are reported with some score drops (e.g., MathVista 46.20 vs 47.20 for STDec on LaViDa; grayed entries in tables).

Experimental setup

Datasets: HumanEval MBPP GPQA GSM8K MATH MathVerse MathVision MathVista GenEval DPGBench

Benchmarks: HumanEval MBPP GPQA GSM8K MATH MathVerse MathVision MathVista GenEval DPGBench

Metrics: TPS speedup Pass@1 accuracy

Baselines: Vanilla Dream Vanilla LLaDA Half-Step dKV-Cache Fast-dLLM (Parallel) LocalLeap Prefix-DLM

Builds on: LocalLeap Fast-dLLM dKV-Cache

Closest papers by meaning

• ST-Veto: Spatio-Temporal Token Veto for Diffusion MLLMs via Taylor Prediction and Visual Grounding (ICML 2026 · similarity 0.87)
• Free Draft-and-Verification: Toward Lossless Parallel Decoding for Diffusion Large Language Models (NeurIPS 2026 · similarity 0.87)
• Fast-dLLM++: Fr\'{e}chet Profile Decoding for Faster Diffusion LLM Inference (NeurIPS 2026 · similarity 0.87)
• PACE-dLLM: Elastic Block Decoding via Confidence Cliff Estimation for Diffusion Language Models (NeurIPS 2026 · similarity 0.86)
• AdaDecode: Accelerating LLM Decoding with Adaptive Layer Parallelism (ICML 2025 · similarity 0.86)
• DMax: Aggressive Parallel Decoding for dLLMs (NeurIPS 2026 · similarity 0.86)