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

Fast and Fluent Diffusion Language Models via Convolutional Decoding and Rejective Fine-tuning

NeurIPS 2025 · 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.

Proposes Conv, a normalization-based convolutional decoding that narrows the diffusion LM decoding window without hard blocks, plus R2FT, a rule-based rejective fine-tuning that penalizes repetitive corruptions, improving open-ended generation of masked diffusion LMs at few steps.

Problem

Masked diffusion LMs suffer a long decoding-window problem where tokens far from the context become repetitive or generic, and semi-autoregressive block decoding that mitigates it degrades quality at small step counts (time-interval expansion) and loses bidirectionality.

Approach

Convolutional decoding rescales per-position token probabilities by a tanh of the number of unmasked neighbors within a kernel, softly narrowing the decoding window instead of hard blocking; R2FT adds a DPO/SimPO-style negative loss where rejected samples are rule-based repetition corruptions of SFT data, reducing the prior and repetition preference.

Key results

• Small 180M MDLM with R2FT and top-k + Glob + Conv reaches 45.92 AlpacaEval win rate (46.92 LC) at L=1024, S=128, versus 23.31 for categorical decoding after R2FT and 14.54 for LLaDA decoding without R2FT.
• On LLaDA-8B-Base with LoRA SFT plus R2FT, top-k + Glob + Conv reaches 55.65 AlpacaEval win rate (55.96 LC), above LLaDA-8B-Instruct's 36.52.
• On GSM8K with LLaDA-8B-Instruct, Conv raises accuracy from 29.49 (LLaDA decoding) to 39.34 at L=512, S=64.

Measured effects reported in the paper

Surprising findings

• R2FT lowers performance under categorical decoding, indicating it works synergistically with top-k decoding.
• LLaDA-style semi-AR decoding with small step counts yields mostly broken sentences, while R2FT rescues it.

Limitations

• Bidirectional generation is not evaluated on downstream benchmarks; only conceptual/proof-of-concept sampling patterns are shown.
• EOS-fill is designed for unidirectional generation and is not applicable to bidirectional settings.
• Evaluation relies on GPT-4-turbo G-eval and small-scale experiments with limited model sizes.

Experimental setup

Datasets: Alpaca instruction dataset OpenWebText GSM8K MT-Bench Mintaka (Wiki entity prompts)

Benchmarks: AlpacaEval MT-Bench Arena-Hard-Auto GSM8K

Metrics: length-controlled win rate win rate GSM8K accuracy

Baselines: SEDD MDLM categorical decoding LLaDA decoding (semi-AR) LLaDA-8B-Instruct vicgalle/gpt2-alpaca

Builds on: MDLM DPO SimPO LLaDA Block diffusion

Closest papers by meaning

• Scaling Diffusion Language Models via Adaptation from Autoregressive Models (ICLR 2025 · similarity 0.93)
• From Next-Token to Next-Block: A Principled Adaptation Path for Diffusion LLMs (NeurIPS 2026 · similarity 0.92)
• Simple and Effective Masked Diffusion Language Models (NeurIPS 2024 · similarity 0.92)
• Breaking AR’s Sampling Bottleneck: Provable Acceleration via Diffusion Language Models (NeurIPS 2025 · similarity 0.92)
• Beyond Autoregression: Fast LLMs via Self-Distillation Through Time (ICLR 2025 · similarity 0.92)
• dKV-Cache: The Cache for Diffusion Language Models (NeurIPS 2025 · similarity 0.92)