REVIEW 60 cited by
DiffuCoder: Understanding and Improving Masked Diffusion Models for Code Generation
Not yet reviewed by Pith; the record is open.
This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.
SPECIMEN: schema-true, not a live event
T0 review · schema-true
One-sentence machine reading of the paper's core claim.
pith:XXXXXXXX · record.json · timestamp
DiffuCoder: Understanding and Improving Masked Diffusion Models for Code Generation
read the original abstract
Diffusion large language models (dLLMs) are compelling alternatives to autoregressive (AR) models because their denoising models operate over the entire sequence. The global planning and iterative refinement features of dLLMs are particularly useful for code generation. However, current training and inference mechanisms for dLLMs in coding are still under-explored. To demystify the decoding behavior of dLLMs and unlock their potential for coding, we systematically investigate their denoising processes and reinforcement learning (RL) methods. We train a 7B dLLM, \textbf{DiffuCoder}, on 130B tokens of code. Using this model as a testbed, we analyze its decoding behavior, revealing how it differs from that of AR models: (1) dLLMs can decide how causal their generation should be without relying on semi-AR decoding, and (2) increasing the sampling temperature diversifies not only token choices but also their generation order. This diversity creates a rich search space for RL rollouts. For RL training, to reduce the variance of token log-likelihood estimates and maintain training efficiency, we propose \textbf{coupled-GRPO}, a novel sampling scheme that constructs complementary mask noise for completions used in training. In our experiments, coupled-GRPO significantly improves DiffuCoder's performance on code generation benchmarks (+4.4\% on EvalPlus) and reduces reliance on AR bias during decoding. Our work provides deeper insight into the machinery of dLLM generation and offers an effective, diffusion-native RL training framework. https://github.com/apple/ml-diffucoder.
Forward citations
Cited by 60 Pith papers
-
DF3DV-1K: A Large-Scale Dataset and Benchmark for Distractor-Free Novel View Synthesis
DF3DV-1K supplies 1,048 scenes with clean and cluttered image pairs plus a challenging 41-scene subset to benchmark and improve distractor-free radiance field methods.
-
Learning from the Self-future: On-policy Self-distillation for dLLMs
d-OPSD reframes on-policy self-distillation for dLLMs via suffix conditioning from self-generated answers and step-level supervision, outperforming RLVR and SFT on reasoning benchmarks with ~10% of the optimization steps.
-
MaskForge: Structure-Aware Adaptive Attacks for Jailbreaking Diffusion Large Language Models
MaskForge reaches 79.3% average attack success rate on five dLLMs by adaptively searching and accumulating structural attack patterns with a UCB bandit, improving 17.6% over baselines and transferring to 88.2% on AdvBench.
-
SNLP: Layer-Parallel Inference via Structured Newton Corrections
SNLP enables layer-parallel Transformer inference by replacing sequential layer execution with structured Newton corrections and SNLP-aware training regularization, yielding up to 2.3x wall-clock speedup on 0.5B model...
-
Constrained Code Generation with Discrete Diffusion
Constrained Diffusion for Code (CDC) integrates constraint satisfaction into the reverse denoising process of discrete diffusion models via constraint-aware operators that use optimization and program analysis to stee...
-
Infinite Mask Diffusion for Few-Step Distillation
Infinite Mask Diffusion Models use stochastic infinite-state masks to overcome the factorization error lower bound in standard masked diffusion, achieving superior few-step performance on language tasks via distillation.
-
Masked Diffusion Language Models are Strong and Steerable Text-Based World Models for Agentic RL
Masked diffusion language models, not larger autoregressive LLMs, are the better building block for text-based world models in agentic RL, improving rollout fidelity, diversity, and downstream task success.
-
Simple Self-Conditioning Adaptation for Masked Diffusion Models
SCMDM is a post-training self-conditioning adaptation for masked diffusion models that reduces generative perplexity by nearly 50% on OWT and improves performance on images, molecules, and genomics.
-
Discrete Tilt Matching
DTM recasts dLLM fine-tuning as weighted cross-entropy matching of tilted local posteriors, with demonstrated gains on Sudoku and math tasks.
-
Discrete Tilt Matching
Discrete Tilt Matching recasts dLLM fine-tuning as state-level matching of tilted local unmasking posteriors, producing a stable weighted cross-entropy loss that improves Sudoku and Countdown performance when applied ...
-
DMax: Aggressive Parallel Decoding for dLLMs
DMax uses On-Policy Uniform Training and Soft Parallel Decoding to enable aggressive parallelism in dLLMs, raising TPF on GSM8K from 2.04 to 5.47 and on MBPP from 2.71 to 5.86 while preserving accuracy.
-
A Comparative analysis of Layer-wise Representational Capacity in AR and Diffusion LLMs
Diffusion language models form more global representations with early-layer redundancy compared to autoregressive models, allowing layer skipping for up to 18.75% FLOP savings while maintaining over 90% performance.
-
DICE: Diffusion Large Language Models Excel at Generating CUDA Kernels
DICE, a family of diffusion LLMs trained with a bi-phase infilling-to-generation RL curriculum, achieves state-of-the-art CUDA kernel generation at data and token budgets far below prior specialized autoregressive models.
-
Not All Denoising Steps Are Equal: Model Scheduling for Faster Masked Diffusion Language Models
Early and late denoising steps in masked diffusion LMs are robust to smaller-model replacement, enabling 17% FLOPs reduction with modest generative quality loss.
-
The Flexibility Trap: Rethinking the Value of Arbitrary Order in Diffusion Language Models
Arbitrary-order generation lowers the reasoning Pass@k of diffusion LLMs; constraining RL to autoregressive order (JustGRPO) improves accuracy and preserves parallel decoding.
-
PartDiffuser: Part-wise 3D Mesh Generation via Discrete Diffusion
PartDiffuser is a semi-autoregressive discrete diffusion framework that generates high-fidelity 3D meshes from point clouds by combining inter-part autoregression with intra-part parallel diffusion using a part-aware ...
-
Coevolutionary Continuous Discrete Diffusion: Make Your Diffusion Language Model a Latent Reasoner
CCDD defines a joint multimodal diffusion on continuous representation space and discrete token space to combine expressivity with explicit token supervision for diffusion language models.
-
S2D2: Fast Decoding for Diffusion LLMs via Training-Free Self-Speculation
Training-free self-speculation reuses a block-diffusion model’s block-size-1 mode as a local AR verifier, improving accuracy–speed tradeoffs over confidence-threshold decoding.
-
Beyond On-Policy Exploration: Integrating External Policy Rollouts for Reinforcement Learning in Diffusion Language Models
ERILS mixes length-controlled external rollouts from a stronger policy with on-policy rollouts and processes their rewards separately, substantially improving dLLM reinforcement learning on reasoning tasks.
-
From Interface to Inference: Eliciting Any-Order Inference from Any-Order Models
Masked diffusion language models fail to exploit their any-order interface because of positional uncertainty; insertion-based (FlexMDM) and latent-segment (LatentMDM) variants recover distinct any-order inference beha...
-
Rethinking the Generation Order of Block Diffusion Language Models
BDLMs are training-biased toward left-to-right contexts, and PARD exploits that bias with leftmost-prefix parallel unmasking to improve quality–speed trade-offs over existing parallel DLM samplers.
-
Hierarchical Domain Generalization
Over infinite domains, hierarchy-uniform domain generalization is impossible for every nontrivial hypothesis class; a length-generalization bound is a property of the length hierarchy, not a hierarchy-free guarantee.
-
Nemotron-Labs-Diffusion: A Tri-Mode Language Model Unifying Autoregressive, Diffusion, and Self-Speculation Decoding
Joint AR–diffusion training yields one tri-mode LM that switches AR, diffusion, and self-speculation, beating open AR/diffusion models on accuracy and tokens-per-forward.
-
SLIM-RL: Risk-Budgeted Random-Masking RL for Diffusion LLMs Without Trajectory Slicing
SLIM-RL matches or exceeds TraceRL performance on MATH500, GSM8K, MBPP and HumanEval for diffusion LLMs by risk-budgeted random-masking RL without trajectory slicing.
-
Beyond Fully Random Masking: Attention-Guided Denoising and Optimization for Diffusion Language Models
AGDO improves dLLM reasoning performance by determining denoising order and emphasizing tokens based on attention-derived dependencies rather than random masking.
-
Back on Track: Aligning Rewards and States for Reasoning in Diffusion Large Language Models
PAPO improves reasoning performance in diffusion LLMs by converting sparse terminal rewards into dense step-wise credit and replaying real high-uncertainty trajectories, reporting gains up to 42.2% on Countdown.
-
NAVIRA: Decoupled Stochastic Remasking for Masked Diffusion Language Models
NAVIRA decouples quality scoring from regeneration via stochastic remasking in masked diffusion LMs, improving fluency and LLM-judge scores on a 170M model.
-
dMoE: dLLMs with Learnable Block Experts
dMoE aggregates token expert distributions to block level in dLLMs, cutting unique experts from 69.5 to 14.6, memory by 76-80%, and latency by 1.14-1.66x while retaining 99.11% performance.
-
GDSD: Reinforcement Learning as Guided Denoiser Self-Distillation for Diffusion Language Models
GDSD reduces RL for dLLMs to likelihood-free self-distillation via a normalization-free logit-matching objective, outperforming ELBO methods with more stable training on LLaDA-8B and Dream-7B.
-
The Confidence Shortcut: A Reasoning Failure Mode of Masked Diffusion Models
Confidence-based decoding and training in masked diffusion models shortcut long-range dependencies in reasoning, producing errors on complex inputs that random masking avoids.
-
SNLP: Layer-Parallel Inference via Structured Newton Corrections
SNLP achieves up to 2.58x wall-clock speedup on 0.5B Transformers via architecture-specific Newton corrections (IDN/HCN) that enable layer-parallel inference while preserving perplexity in milder settings.
-
SNLP: Layer-Parallel Inference via Structured Newton Corrections
Structured Newton Layer Parallelism (IDN/HCN) replaces sequential Transformer layers with cheap parallel Newton-style updates, trading finite-iteration bias for up to 2.58× inference speedup.
-
Pushing Biomolecular Utility-Diversity Frontiers with Supergroup Relative Policy Optimization
SGRPO is a GRPO-style framework that constructs set-level diversity rewards via supergroup sampling and leave-one-out redistribution to expand the utility-diversity Pareto frontier in biomolecular design tasks.
-
Pushing Biomolecular Utility-Diversity Frontiers with Supergroup Relative Policy Optimization
SGRPO expands the utility-diversity Pareto frontier in biomolecular design by using supergroup sampling and leave-one-out diversity rewards combined with utility signals.
-
Simple Self-Conditioning Adaptation for Masked Diffusion Models
SCMDM adapts trained masked diffusion models to condition denoising steps on their own prior clean predictions, cutting generative perplexity nearly in half on open-web text while improving discretized image, molecule...
-
Stability-Weighted Decoding for Diffusion Language Models
Stability-Weighted Decoding improves diffusion LLM accuracy by modulating token scores with temporal stability from KL divergence between prediction steps.
-
DF3DV-1K: A Large-Scale Dataset and Benchmark for Distractor-Free Novel View Synthesis
DF3DV-1K supplies 1,048 real scenes with clean/cluttered image pairs and a 41-scene hard subset to benchmark and improve distractor-free radiance-field methods.
-
Dataset-Level Metrics Attenuate Non-Determinism: A Fine-Grained Non-Determinism Evaluation in Diffusion Language Models
Dataset-level metrics in diffusion language models mask substantial sample-level non-determinism that varies with model and system factors, which a new Factor Variance Attribution metric can decompose.
-
LightningRL: Breaking the Accuracy-Parallelism Trade-off of Block-wise dLLMs via Reinforcement Learning
A GRPO-based post-training method for block-wise diffusion LLMs improves average tokens-per-forward from ~3 to 7.3 while preserving accuracy.
-
Diffusion-State Policy Optimization for Masked Diffusion Language Models
DiSPO is a plug-in credit-assignment method for masked diffusion LMs that optimizes intermediate filling decisions via branched completions from rollout-cached logits.
-
Diffusion-State Policy Optimization for Masked Diffusion Language Models
DiSPO optimizes intermediate decisions in masked diffusion LMs by branching at selected masked states, resampling tokens, scoring completions, and updating only new tokens using a derived policy-gradient estimator tha...
-
UnMaskFork: Test-Time Scaling for Masked Diffusion via Deterministic Action Branching
UnMaskFork applies MCTS with deterministic cross-model unmasking actions and node caching to outperform existing test-time scaling baselines for masked diffusion language models on code and math tasks.
-
Efficient-DLM: From Autoregressive to Diffusion Language Models, and Beyond in Speed
Efficient-DLM converts AR models to dLMs via block-wise causal attention and position-dependent masking, yielding higher accuracy and 2.7-4.5x throughput than Dream 7B and Qwen3 4B.
-
Saber: An Efficient Sampling with Adaptive Acceleration and Backtracking Enhanced Remasking for Diffusion Language Model
Saber improves both speed and accuracy of diffusion language models on code generation by dynamically adjusting unmasking steps and reverting low-confidence tokens via backtracking.
-
Enhancing Reasoning for Diffusion LLMs via Distribution Matching Policy Optimization
DMPO improves diffusion LLM reasoning by matching the model's distribution to the reward-tilted target via importance-weighted denoising cross-entropy plus weight-baseline subtraction, beating GRPO-style baselines on ...
-
ParallelBench: Understanding the Trade-offs of Parallel Decoding in Diffusion LLMs
Parallel decoding in diffusion LLMs degrades quality on token-dependent tasks, and existing strategies cannot adapt parallelism to preserve accuracy, as shown by the new 17-task ParallelBench benchmark.
-
Fine-Tuning Masked Diffusion for Provable Self-Correction
PRISM fine-tunes any masked diffusion model with a binary-cross-entropy loss so its new head provably estimates per-token quality p(x_i=y_i|y⊕m_i) and can remask low-quality tokens at inference.
-
FS-DFM: Fast and Accurate Long Text Generation with Few-Step Diffusion Language Models
FS-DFM enables 1024-token generation at perplexity parity with 1024-step baselines using only 8 steps via explicit step-budget training, reliable updates, and teacher guidance.
-
Inpainting-Guided Policy Optimization for Diffusion Large Language Models
IGPO injects partial ground-truth reasoning hints into masked diffusion LLM generation during RL sampling, restoring non-zero advantages and improving math benchmarks, but as written the RL training set overlaps the e...
-
Revolutionizing Reinforcement Learning Framework for Diffusion Large Language Models
TraceRL aligns the RL objective with the DLM's actual step-by-step decoding, producing TraDo-4B/8B models that beat autoregressive baselines on math reasoning.
-
Layer-Parallel Inference Reduces Encrypted Nonlinear Depth in Transformers
SNLP reduces encrypted Transformer nonlinear depth from L sequential stages to (L−N)+K, cutting symbolic bootstraps ~2.65× with lower error amplification than sequential inference.
-
TACG: Trajectory-Aware Commit Gating for Diffusion Language Model Decoding
Trajectory-aware commit gating (TILG + History Gate + capped extra promotion) improves or preserves DLLM accuracy while reducing steps and raising tokens-per-forward without retraining.
-
CoGate: Confidence-Gated Co-Decoding for Secure Code Generation
Gating a security expert by its own confidence (max-prob or entropy) fixes inverted acceptance in co-decoding and improves joint functional-security code generation, especially on OOD CWEs.
-
Layer-Parallel Inference Reduces Encrypted Nonlinear Depth in Transformers
SNLP reduces symbolic FHE bootstraps from 53 to 20 on a 0.5B model with +1.2% PPL degradation and lower polynomial-error amplification than sequential inference.
-
Dynamic Infilling Anchors for Format-Constrained Generation in Diffusion Large Language Models
DIA is a training-free method that dynamically adjusts anchor positions in diffusion LLMs to improve format compliance and accuracy on reasoning benchmarks like GSM8K and MATH.
-
WaveFilter: Enhancing the Long-Context Capability of Diffusion LLMs via Wavelet-Guided KV Cache Filtering
WaveFilter applies wavelet decomposition to filter critical tokens for sparse KV caching, improving long-context performance of diffusion LLMs as a plug-and-play addition to existing methods.
-
Reinforcement Learning from Denoising Feedback
RLDF is a new RL paradigm for diffusion language models that optimizes toward clipped clean states with weighted timestep sampling and reports substantial gains on reasoning benchmarks for LLaDA and Dream.
-
DACA-GRPO: Denoising-Aware Credit Assignment for Reinforcement Learning in Diffusion Language Models
DACA-GRPO adds denoising-aware credit assignment and bias-reduced likelihood estimation to GRPO, delivering consistent gains up to 36.3pp on math, code, constraint, and schema benchmarks for diffusion LLMs.
-
DMax: Aggressive Parallel Decoding for dLLMs
DMax enables faster parallel decoding in diffusion language models by using on-policy training to recover from errors and soft embedding interpolations for iterative revision, boosting tokens per forward pass roughly ...
-
A Comparative analysis of Layer-wise Representational Capacity in AR and Diffusion LLMs
Diffusion-trained LLaDA retains >90% accuracy when six of 32 layers are skipped, while autoregressive Qwen2.5 collapses after two layers, but the comparison is confounded by instruction-tuning status.
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.