Autoregressive Model

Causal Transformers model sequences through an autoregressive factorization of the joint distribution, which enables efficient left-to-right decoding and conditional likelihood computation. However, they cannot tractably sample from or evaluate arbitrary conditionals -- e.g., a block of text conditioned on past and future tokens. Recent work aims to solve this problem through novel architectures, but they often lead to sub-optimal modeling of such conditionals and degraded generations. We propose Arbitrary Conditionals GPT (AC-GPT) which introduces a simple modification to standard causal Transformers to enable evaluating and sampling from arbitrary conditionals -- including past, future, and mixed contexts -- within a single forward pass. Unlike prior approaches, our method preserves the standard left-to-right ordering and next-token prediction objective essential for both strong performance and efficient training on natural language. Crucially, this compatibility allows existing LLMs to be fine-tuned for arbitrary conditioning. Our empirical results indicate that our method outperforms baselines on modeling arbitrary conditionals, without degrading standard left-to-right performance.

Open diffusion language models are marketed as parallel, non-autoregressive decoders, yet the order in which a shipped checkpoint actually commits its tokens is almost never measured. We instrument DiffusionGemma 26B, a masked discrete-diffusion mixture-of-experts model built on Gemma 4, hooking its sampler's accept step to record which canvas positions commit, when, and at what confidence. Across a 686-prompt, six-regime probe suite we find that its decoding is neither parallel nor block-autoregressive: it follows a partial left-to-right commit bias whose apparent strength depends almost entirely on the granularity at which you look. Order is weak token by token and strengthens smoothly as the analysis is coarsened, so the model's "block size" turns out to be an artifact of the measuring ruler rather than the architecture. The model commits in large simultaneous batches, leaving much of the within-batch order genuinely undefined rather than merely unobserved. The behaviour is regime-dependent: structured JSON is committed in essentially arbitrary order, and a position's commit confidence tracks correctness on mathematical reasoning but carries no signal on factual recall. Commitment is aggressive, finishing in a short late burst well inside the step budget, while task accuracy matches the model's autoregressive Gemma-4 sibling. Beyond these findings, our central contribution is methodological: measuring decoding order honestly demands handling trailing-EOS padding, within-regime confounding, commit non-monotonicity, block-size sensitivity, and large commit-batch ties, each of which can otherwise manufacture a decoding-order result that is not really there.

Real-time, full-duplex speech interaction is a key feature of next-generation spoken chatbots, allowing the model to listen and speak at the same time and to handle natural phenomena such as overlap, hesitation, and barge-in. Existing speech language models (SpeechLMs) such as LLaMA-Omni and GLM-4-Voice are still turn-based and rely on an external Voice Activity Detection (VAD) module to mark the end of the user's turn, which fundamentally limits their interactive ability. In this paper, we introduce BayLing-Duplex, a native full-duplex SpeechLM where a single autoregressive LLM decides when to listen, when to speak, and when to stop, with no auxiliary turn-taking module. The design adds only a few special tokens to the standard vocabulary, so it transfers across LLMs and reuses existing training and serving stacks with no architectural adaptation. Starting from the public GLM-4-Voice checkpoint and using only 400K full-duplex samples for fine-tuning followed by a lightweight DPO stage, BayLing-Duplex reaches 92% turn-taking success and 100% interruption success on InstructS2S-Eval, while improving the speech-response score from 2.17 to 3.39 over Moshi. BayLing-Duplex also matches or surpasses its turn-based counterpart on Llama Questions, Web Questions, and Alpaca-Eval, showing that simultaneous listen-and-speak modeling does not sacrifice response quality.

RL post-training has become increasingly pivotal for improving diffusion policies, but existing diffusion policy-gradient methods are often unstable and cannot achieve reliable policy improvement. We identify the cause as the double-drift phenomenon: optimizing a variational surrogate can let the ELBO separate from the true log-likelihood, which then makes the resulting proxy policy gradient misaligned with the true policy gradient of expected return. We propose \textbf{DiPOD}, a diffusion policy optimization framework that maintains tight-bound behavior throughout training by interleaving self-distillation with policy-improving gradient updates. This leads to a simple and practical algorithm: augmenting each diffusion policy-gradient update with an on-policy ELBO regularizer. Across diffusion language model post-training and continuous-control diffusion policies, DiPOD substantially stabilizes training and reaches higher rewards than previous methods.

We study generative modeling of Bach-style symbolic piano music using a shared MIDI corpus and three model families: autoregressive LSTMs with attention, latent-variable models including recurrent VAEs and vector-quantized VAEs, and generative adversarial networks. We compare their ability to model polyphonic note sequences, learn useful latent representations, and generate stylistically coherent compositions. Our experiments show that the autoregressive LSTM with attention produces the most musically coherent samples, while vector quantization helps mitigate posterior collapse and yields more structured outputs than conventional recurrent VAEs. The adversarial approach captures local pitch patterns but remains difficult to train and generalizes less reliably to Bach's style. These results highlight the relative strengths and failure modes of autoregressive, latent-variable, and adversarial approaches for symbolic music generation.

Speculative decoding speeds up LLM inference by using a draft model to generate tokens, with an acceptance-rejection scheme that ensures that the output matches the target distribution. Adapting this to continuous diffusions is difficult because speculative sampling requires drawing from a residual distribution. While straightforward in discrete spaces, efficiently sampling this residual in continuous space is non-trivial. Consequently, existing diffusion adaptations either use computationally inefficient sampling techniques or rely on an alternative scheme. In this work, we introduce a novel scheme that efficiently implements the original speculative sampling mechanism for diffusion models. Our approach offers a critical advantage over current methods: it enables us to adapt block verification from LLMs to diffusions -- which provably improves the acceptance rate of drafts. Furthermore, we formalize and analyze the Free Drafter, a heuristic self-speculative drafter for diffusions that requires no training. By enabling block verification, our Free Drafter yields up to a 6.3% speedup over existing speculative methods with no additional training and negligible overhead beyond the existing parallel verification pass.

Masked diffusion language models (MDLMs) such as LLaDA now rival autoregressive (AR) LLMs, but every existing knowledge-editing and unlearning method (ROME, MEMIT, etc.) targets AR transformers and either makes assumptions that fail under iterative denoising, or requires gradient updates whose backward-pass activations cost tens of GB of extra VRAM and which collapse MDLMs at standard learning rates. We introduce TimeROME-DLM, the first training-free, gradient-free, inference-time knowledge-editing framework for MDLMs. It couples two components: a Temporal Indirect Effect (TIE) causal-tracing protocol that identifies, for each fact, the coordinate whose intervention most strongly drives the object prediction at later denoising steps; and a closed-form, low-rank residual edit memory that aggregates subject keys and target deltas across all forget facts and applies a single ridge-regularised update at that coordinate at every diffusion forward, with sparsification to limit utility spillover. Backbone weights stay frozen; only three hyperparameters (alpha, lambda, q) are tuned on a small validation split. On TOFU forget01 with TOFU-finetuned LLaDA-8B-Base, TimeROME-DLM cuts forget-set log-probability by roughly 83 nats. The same configuration transfers to LLaDA-8B-Instruct, Dream-7B, MMaDA-8B, DiffuLLaMA-7B, and LLaDA-MoE-1.4B. It keeps retain-set log-probability nearly flat (within ~1 nat at the utility-safe operating point) across 50 sequentially inserted facts, delivers a four- to fourteen-fold wall-clock speedup with zero additional VRAM over the strongest converged training-time baseline, and scales sub-linearly to 400 facts. TimeROME-DLM closes the locate-then-edit gap between AR LLMs and MDLMs at a fraction of the computational cost.

Summaries of real-world events can become outdated as contexts evolve and new information arrives. A common response is to generate a new summary from the updated context, but full regeneration discards the previous draft, can obscure what changed, and may be unnecessary when only a few claims are unsupported. We study localized faithfulness repair: updating outdated spans in an existing summary while preserving supported content. We propose DETECT-REMASK-REPAIR, a diffusion-based framework that identifies, remasks, and repairs outdated regions with masked diffusion language models. To evaluate evolving-context summarization, we introduce StreamSum, a benchmark of synthetic event timelines. Experiments on DialogSum and StreamSum show that localized diffusion repair provides a controllable alternative to full rewriting: faithfulness-steered repair improves early drafts, one-step repair reduces repair cost to under half a second, with the framework enabling faithfulness-speed-preservation tradeoffs across datasets. We also find that the framework can provide a post-hoc correction step that improves faithfulness for autoregressive systems.

Auto-regressive models have emerged as powerful tools for sequential data, from language to video. Understanding how and why these models learn latent representations remains an open theoretical question. In this work, we demonstrate that when trained by empirical risk minimization on data from partially observed linear dynamical systems, two-layer linear auto-regressive models naturally learn to approximate Kalman filtering. In particular, we show that the learned hidden representation coincides, up to a similarity transformation, with the state estimates produced by the optimal (Kalman) filter, even though the model has no explicit knowledge of the underlying dynamics or state. The result follows from three main insights. First, we establish that the Kalman filter is well approximated by an auto-regressive model with bounded truncation error. Second, we show that despite non-convexity, the two-layer optimization landscape is benign, i.e., all stationary points are either strict saddles or global minima. Finally, as our main contributions, we provide finite-sample guarantees on prediction error, parameter estimation error, and latent state recovery. Numerical simulations support the theoretical results and demonstrate that the latent representations of auto-regressive models recover state estimates.

Reinforcement learning (RL) has become a key component in modern large language models, yet the rollout stage remains the key bottleneck in RL training pipelines. Although Multi-Token Prediction (MTP) offers a natural solution to accelerate rollouts through speculative decoding, many studies have observed that MTP acceptance rates degrade significantly during RL training, leading to limited speedup performance. To address this bottleneck, we present Bebop, a systematic study of MTP in LLM post-training, and offer practical recipes to integrate MTP into large-scale RL pipelines. First, we reveal that the MTP acceptance rate is fundamentally bounded by the fluctuation of model entropy, which demonstrates a clear negative linear relationship with the rise of entropy in the RL stage. Second, we show that probabilistic rejection sampling largely alleviates the disturbance introduced by entropy in RL compared to greedy draft sampling. We further identify that the conventional MTP training objectives (cross-entropy or KL) are suboptimal in such settings, and therefore we propose a novel end-to-end TV loss that directly optimizes multi-step rejection sampling acceptance rate, yielding ~10% acceptance rate improvements, achieving up to 95% acceptance rates and up to 25% extra inference throughput gains across mathematical reasoning, code generation, and agentic tasks. Third, we test various online MTP training strategies during RL and show that pre-RL MTP training with e2e TV loss and rejection sampling achieves a consistent acceptance rate and speedup throughout the entire RL, eliminating the need for costly online MTP updating. We provide extensive experiments and analysis that validate our findings. Experimental results show our method achieves up to 1.8x end-to-end acceleration in async RL training of Qwen3.5, Qwen3.6, and Qwen3.7 models.

Diffusion large language models (dLLMs) offer an efficient alternative to autoregressive models through parallel decoding, yet existing post-training methods largely rely on random masking strategies that overlook intrinsic token dependencies. In this work, we present an empirical analysis of attention in dLLMs and show that tokens attending more strongly to unmasked context exhibit greater generation stability and play a critical role in reasoning. Motivated by these findings, we propose AGDO, an attention-guided denoising and optimization framework that aligns both training and optimization with attention-derived dependencies. AGDO determines the denoising order based on attention structure and emphasizes attention-critical tokens during supervised fine-tuning and reinforcement learning. Experiments on mathematical and coding benchmarks demonstrate that AGDO consistently improves reasoning performance, outperforming state-of-the-art post-training methods for dLLMs.

Masked diffusion language models (dLLMs) have recently emerged as a competitive alternative to autoregressive language models, with the promise of faster inference via parallel token generation. A notable limitation of the masked formulation, however, is that once a token has been unmasked it can no longer be revised, leaving dLLMs vulnerable to early sampling mistakes. To address this, a growing body of work has sought to extend masked dLLMs with self-correcting (remasking) capabilities. One appealing subset of these methods does so in a training-free, post-hoc manner based on token confidences, with encouraging early reported results. In this work, we revisit the empirical evaluation of a representative post-hoc remasking method, WINO [Hong et al., 2026], and find that under standard decoding settings (shorter block lengths) it brings little-to-no benefit over confidence-based unmasking alone [Wu et al., 2025]. Extending the evaluation to non-greedy decoding, we find that while confidence-based remasking can mitigate errors introduced by increased stochasticity to some extent, it also exacerbates the diversity collapse previously reported for confidence-based unmasking. Overall, our results show that the benefits of post-hoc confidence-based remasking are highly setting-dependent, underscoring the need for a more comprehensive evaluation framework.

Can financial news reliably predict short-term stock movements? Despite advances in large language models, this question remains unresolved. We revisit this problem using a zero-shot natural language processing framework, investigating whether models can extract actionable signals from financial news without domain-specific training. We design a structured pipeline that combines zero-shot natural language inference with temporal aggregation, explicitly modelling recency and event-dependent impact horizons when integrating information across articles. To address the need for transparency in high-stakes settings, we introduce a multi-layered explainability framework that links predictions to token-level, article-level, and aggregate evidence, and produces grounded natural language rationales. Across multiple models and prediction horizons, we find that zero-shot approaches consistently fail to outperform simple baselines, with particularly weak performance on negative movements, suggesting deeper structural limitations in mapping news sentiment to short-term price dynamics. However, explainability signals reliably distinguish between trustworthy and unreliable predictions, offering practical value even when accuracy is limited. These findings highlight the limits of zero-shot financial NLP and motivate a shift toward decision-support systems that prioritise transparency and uncertainty awareness. Code: https://github.com/alimert05/zero-shot-stock-xai

Multi-turn image editing is essential for iterative design, yet current models often struggle with identity drift and error accumulation over successive steps. While existing research leverages video priors for consistency, their reliance on bidirectional attention is fundamentally misaligned with the causal, sequential nature of interactive editing. In this paper, we propose AnchorEdit, the first autoregressive (AR) diffusion-based framework designed specifically for high-resolution, long-term multi-turn editing. AnchorEdit bridges the gap between video priors and causal inference through a three-stage training curriculum: identity-preserving sing-turn pretraining, causal AR forcing fine-tuning with a novel self-rollout strategy to mitigate exposure bias, and consistency distillation for efficient 4-step generation. During inference, we introduce a memory mechanism to anchor the initial subject identity and ensure stable extrapolation across extended editing trajectories. To evaluate performance, we provide a new high-resolution multi-turn editing benchmark designed to stress-test long-horizon stability. Extensive experiments demonstrate that AnchorEdit achieves state-of-the-art results, maintaining exceptional subject fidelity and instruction following even over 10+ interaction rounds.

Large language models (LLMs) achieve remarkable performance across a wide range of tasks, but their autoregressive decoding process incurs substantial inference costs due to inherently sequential token generation. Speculative decoding addresses this bottleneck by employing a lightweight draft model to propose multiple future tokens that are subsequently verified in parallel by a larger target model. Recent work has demonstrated that diffusion language models are well suited for this setting, as they can generate entire blocks of draft tokens in parallel and thereby alleviate the sequential constraints of autoregressive drafting. A subtlety of this regime is that block-diffusion drafters generate tokens bidirectionally within a block, whereas verification is performed by an autoregressive target model that evaluates tokens in a strictly left-to-right manner, leaving a gap between the symmetric training-time objective and the asymmetric verification-time reward. In this work, we offer an empirical analysis of three training-time interventions that narrow this gap: token positional weighting, a first-error focal loss that targets the position that breaks the accepted prefix within each block, and a chain loss term that substitutes a differentiable surrogate for the expected accepted length. The three interventions act along orthogonal axes (position, block-conditional first error, joint prefix) and compose additively; they are likewise orthogonal to test-time alignment mechanisms such as multi-draft self-selection, with which they can in principle be combined. Across four target models and six reasoning, code, and dialogue benchmarks, the three interventions raise accepted draft length by 21-76% per benchmark over a position-uniform baseline, without adding additional forward passes and without changing the inference pipeline or the rejection-sampling exactness contract.

We introduce SPADE (SPlit And Delay Embeddings), an autoregressive transformer for sequences whose tokens carry multiple features. Rather than embedding these features jointly, SPADE embeds them independently. Delaying each feature stream relative to the previous one allows intra-token correlations to be learned by the standard self-attention mechanism. Applied to point-cloud calorimeter shower generation in the highly granular ILD detector, SPADE is competitive with the state of the art AllShowers model on photon showers, and substantially outperforms its VQ-VAE-based predecessor OmniJet-$α_C$. The mechanism is applicable to any generative task with multi-feature tokens, enabling LLM-style pretraining workflows for higher-dimensional data.

This paper introduces ARM, a discrete representation-based AutoRegressive Model that unifies image understanding, generation, and editing within a next-token prediction framework. ARM is built on three efforts: first, we train a discrete semantic visual tokenizer that maps images into compact token sequences. Our tokenizer is supervised with multiple objectives that jointly promote semantic discriminability, language alignment and faithful reconstruction, thereby supporting diverse tasks in a shared latent space. With this, we train a 7B autoregressive model over large-scale text and image token sequences, seamlessly developing vision-language perception and generation capabilities. Finally, to further improve preference-aligned behavior for text-to-image generation and instruction-guided editing, ARM applies reinforcement learning (RL) to optimize task-level objectives such as visual quality, instruction adherence, and edit consistency. Surprisingly, the results show that RL not only substantially improves performance on the target tasks (e.g., raising WISE overall from 0.50 to 0.56, GEdit-Bench-EN G_O from 5.75 to 6.68), but also induces cross-task synergy between text-to-image generation and editing. Collectively, these findings highlight autoregressive modeling, when paired with strong representations and preference optimization, as a scalable foundation for multimodal intelligence. Code: https://github.com/wdrink/ARM.

Large language model inference is bottlenecked by autoregressive decoding, where each token requires a full forward pass. Multi-token prediction (MTP) offers a promising acceleration path, but existing approaches suffer from a fundamental architectural flaw: the MTP head for the first token competes with the backbone's own language model (LM) head, leading to severe quality degradation when predictions are accepted. We identify this head-backbone competition as the root cause of repetitive and incoherent outputs in prior MTP-based acceleration methods. To address this, we propose Backbone-as-Architect, a design principle where the backbone LM head always generates the first token, and MTP heads are responsible only for subsequent tokens. Building on this principle, we introduce CLP (Collocation-Length Predictor), a lightweight span-level decision layer that predicts how many additional tokens can be safely accepted at each decoding step. CLP uses only a single linear layer (4.6K--7.7K parameters), replacing the over-engineered 1M-parameter gate networks used in prior work. Experiments on Qwen2.5 models (0.5B, 1.5B, 7B) show that CLP achieves 1.20x--1.29x speedup on 1.5B and 1.14x--1.20x on 7B, with zero quality degradation (repetition ratio < 0.02), while gate-based approaches fail to accelerate (1.07x) or produce severely degraded outputs (repetition ratio > 0.5%). We further demonstrate that shorter prediction horizons (k=2) recover 24% higher MTP head accuracy on large models, establishing a scaling-aware design principle. We identify MTP head prediction accuracy as the binding constraint on acceleration and establish a clear roadmap for future improvements.

Long-horizon autoregressive forecasting of oscillatory physical signals, such as seismograms, gravitational-wave strain, and similar wavefields is limited by error accumulation: as a causal model is fed its own outputs over hundreds of steps, small per-step errors compound into phase drift that pointwise metrics fail to detect. We ask when such rollout stays stable, using synthetic three-component seismograms as a physically structured testbed and the \textsc{SeismoGPT} autoregressive forecaster as the model under study. Through controlled, intra-architecture ablations evaluated on free-running rollout with paired significance tests, we isolate the contribution of each design choice. Multi-token prediction is the dominant stabilizer, accounting for almost the entire improvement over a single-token baseline ($+0.040$ median NCC); a horizon-embedding hybrid prediction head and a cross-horizon STFT-magnitude coherence loss each add a small but consistent further gain. Performance depends sharply on a context-ratio threshold near one, roughly the full P-S interval of observed signal, below which rollout generalization collapses. The dominant residual failure is a polarity inversion that a magnitude-based spectral loss cannot, by construction, penalize, identifying phase-aware objectives as the natural next step. We frame this as a controlled study of rollout stability on oscillatory wavefields, not a benchmark of forecasting architectures.

Masked diffusion language models can reduce inference steps by revealing multiple tokens per denoising iteration, but this parallelism is fragile: positions that are individually confident may be unsafe to commit together when their predictions are coupled. Existing training-free samplers such as Top-$k$, Fast-dLLM, and EB-Sampler mainly control how many tokens to reveal, while often ranking candidates by token-wise scores that ignore interactions within the selected set. We propose ADAS, a training-free reranking rule that leaves the base sampler's stopping rule unchanged and greedily discounts each token-wise confidence score according to its attention to already selected positions, weighted by their prediction uncertainty. Across LLaDA-8B-Base and Dream-7B-Base on the reasoning benchmarks GSM8K and MATH500 and the code benchmarks HumanEval and MBPP, plugging ADAS into all three samplers improves low-NFE performance at matched denoiser evaluations by $9.11$ and $10.46$ percentage points on average, respectively, with $3.1\%$ per-forward runtime overhead. Code is available at https://github.com/yusufsahin99/ADAS.

Autoregressive (AR) language modeling is the dominant paradigm for text generation, yet its sequential token-by-token decoding makes inference memory-bound and inefficient. Existing acceleration approaches, such as speculative decoding and diffusion language models, can yield speedups under certain conditions but do not directly address high-load batch serving--the scenario most critical for industrial-scale deployment. We introduce K-Forcing, a push-forward language modeling paradigm for joint next-k-token decoding. K-Forcing distills an existing AR model into a conditional push-forward mapping--one that transforms independent uniform noise variables into a joint sample of multiple future tokens in a single forward pass. This design preserves fixed-length outputs, reuses the AR teacher backbone, and remains compatible with standard AR serving infrastructure. We train this mapping via progressive self-forcing distillation, which gradually expands the prediction window while enabling the student to closely match the sequence distribution of the AR teacher. We evaluate K-Forcing on LM1B and OpenWebText using a standard causal Transformer backbone. When aggressively configured to generate k = 4 tokens per forward pass, K-Forcing delivers approximately 2.4-3.5x speedup across different batch sizes, while incurring modest quality degradation relative to its AR teacher. As inference increasingly dominates the lifetime compute cost of modern LLMs, K-Forcing offers a promising route toward accelerating AR generation under real-world high-load deployment.

Diffusion large language models (dLLMs) re-encode the entire prefix at every denoising step, causing recomputation that scales quadratically with context length and becomes prohibitive for long-context scenarios. We propose Prefilling-dLLM, a training-free prefill-decode disaggregation framework for dLLMs that partitions the prefix into N chunks, caches their KV representations once, and selects the top-K most relevant chunks with intra-chunk token sparsity for decoding, showing that sparse prefilling can outperform dense attention while reducing per-step complexity from quadratic in the full sequence length to quadratic only in the decode length. On LongBench and InfiniteBench, Prefilling-dLLM achieves state-of-the-art quality among dLLM acceleration methods, and an attention kernel that parallelizes decoding over the non-contiguously cached chunk KV yields 9.1--28.0x speedup at 8K--32K contexts. We further show that beginning-of-sequence tokens prepended to each chunk act as periodic attention anchors that eliminate the lost-in-the-middle phenomenon. Code is available at https://github.com/menik1126/Prefilling-dLLM.

While speech quality is typically assessed on complete utterances, streaming and generative systems require incremental estimation from partial audio. Existing predictors assume full context, degrading on prefix-constrained inputs. Extending ARECHO, we propose ANCHOR, reformulating incremental assessment as a multi-resolution autoregressive task. It models chunk- and utterance-level quality within a single decoder using dual-resolution tokens and a resolution-aware hierarchy for coarse-to-fine refinement. Experiments show substantial robustness under partial input, including a 48% PLCMOS error reduction on 2-second prefixes. Convergence analysis reveals a 4-6 s effective perceptual context horizon. A stress test further isolates structured extrapolation biases under localized corruption. Results demonstrate that hierarchical supervision improves incremental prediction and elucidates how perceptual quality accumulates over time.

Insertion Language Models (ILMs) offer several advantages over left-to-right generation and mask-based generation. However, existing formulations of insertion-based generation have largely been ad-hoc. In this paper, we derive a diffusion-style denoising objective for ILMs from first principles by formulating the noising process as a continuous-time Markov chain on the space of variable-length sequences. We show that previous formulations of ILMs can be viewed as special cases of this denoising framework. Through empirical evaluation on a synthetic planning task, we show that the proposed approach retains the benefits of insertion-based generation over left-to-right generation and masked diffusion models. In language modeling, our diffusion-based approach is competitive with left-to-right generation and masked diffusion models, while offering additional flexibility in sampling compared to existing insertion language models.

Transitioning bidirectional video diffusion models into an autoregressive paradigm improves the interactivity of video world models, but existing causal pipelines need many stages (control fine-tuning, autoregressive training, causal initialization, few-step distillation) and still trail bidirectional models in quality due to error accumulation. Recent world models such as Yume-1.5 and Matrix-Game-3.0 instead adopt a bidirectional autoregressive approach, gaining fidelity and stable long-horizon rollout from self-correcting error propagation, yet open-source frameworks (e.g., minWM) support only causal models. We present BiWM, the first full-stack framework for interactive video world models under the bidirectional autoregressive paradigm, jointly optimizing generation quality and inference speed. From a pretrained video backbone, BiWM injects camera control by fine-tuning, then runs a few-step Distribution Matching Distillation (DMD) stage that turns the backbone into an action/camera-controllable world model: just two training stages instead of four in minWM, converging in a few hundred steps on 8xH200 GPUs. A single recipe spans Wan2.1-1.3B, Wan2.2-5B, HunyuanVideo-1.5-8B, and LTX-2.3-22B, and also supports secondary fine-tuning of existing bidirectional models. BiWM enables real-world camera control where minWM loses controllability, integrates pluggable history compression (FramePack-style and PackForcing-style) for long rollouts, and offers an optional NVFP4 4-bit training/inference pipeline. To counter DMD's mode-seeking degradation, we add GAN and mass-covering forward-KL objectives that preserve scene dynamics. We open-source BiWM for resource-constrained research and high-fidelity environment simulation.

Matrix-valued time series arise in a wide range of applications, such as spatio-temporal data from medical imaging and geophysics. Existing methods are mainly designed for static settings and lack adaptability to streaming and time-varying environments. Adaptive filtering techniques have also been largely limited to data with scalar or vector values, leaving adaptive forecasting for matrix-valued time series inadequately understood. To bridge these gaps, we develop an adaptive tensor regression framework that includes Matrix-on-Matrix (MoM) and Tensor-on-Matrix (ToM) formulations for streaming matrix-valued prediction. The two formulations differ in whether to directly model matrix-valued outputs or to exploit temporal structure via higher-order tensor representations. For the proposed tensor regression framework, we develop stochastic gradient descent (SGD) algorithms for online learning. We show that stacking multiple responses across time into higher-order tensors improves performance; in particular, the ToM achieves lower steady-state error and stronger denoising capability than MoM, motivating our focus on the ToM model. We further characterize the tracking behavior of SGD under time-varying dynamics. From a statistical perspective, we establish fixed-time recovery guarantees for ToM under general low-dimensional structures, including sparsity, low-rankness, and their joint sparselow-rank models.

We tightly characterize the VC dimension of depth-$L$ Transformers with a total of $W$ parameters, mapping an input sequence of length $T$ to a single output, establishing an upper bound of $O(L W \log (T W))$ and a nearly matching lower bound of $Ω(L W \log (T W / L))$. We further tightly characterize the sample complexity of chain-of-thought learning using such a Transformer, showing teacher forcing (i.e. selecting a predictor consistent with the entire chain-of-thought on training data) learns with sample complexity $O\left(L W \log \left(\left(T+T^{\prime}\right) W\right)\right)$ and that any learning rule that uses chain-of-thought data requires at least $Ω\left(L W \log \left(\left(T+T^{\prime}\right) W / L\right)\right)$ examples, where $T$ is the input length and $T^{\prime}$ is the number of autoregressive steps.

Foundation models offer a promising route to compress multi-modal physiological signals into compact representations of human health, with broad applications across sleep medicine, cardiology, neurology and other healthcare domains. Existing models have typically been trained with masked-reconstruction or contrastive objectives. However, masked reconstruction may be poorly suited to the stochastic nature of these signals, while contrastive approaches rely on positive-pair definitions despite the semantic invariances of physiological signals being poorly understood. In this work, we show that next-token prediction is a simple and scalable alternative. We develop Hypnos, a multi-modal sleep foundation model trained using eight different sensing modalities (e.g. EEG, ECG, respiratory signals) drawn from over 20,000 overnight polysomnography recordings. We tokenize each modality into streams of discrete tokens using residual vector quantization, then train a large auto-regressive RQ-Transformer to jointly predict the next token across all modalities in parallel. After training, Hypnos can be applied to continuous streams of sensor data from any subset of supported modalities, generating embeddings for downstream tasks. Across a range of benchmarks, Hypnos significantly outperforms existing foundation models. In sleep stage classification, we match the performance of strong supervised baselines on held-out test sets whilst using $100\times$ less labelled data. Hypnos even generalises to daytime physiology, surpassing a dedicated ECG foundation model at detecting atrial fibrillation. Our results demonstrate that next-token prediction is a strong self-supervised objective for representation learning from multi-modal physiological signals.

Continuous diffusion for categorical data is a framework belonging to the diffusion family and aiming at generating discrete data. The scientific interest to such models has been constantly increasing these days because researchers try to achieve a challenging goal of finding reasonable alternatives to autoregressive large language models. In this paper, we study the properties of the structure of the latent space corresponding to discrete tokens expressed in terms of Kullback-Leibler divergence on diffusion path measures and accuracy of the correct token prediction by the optimally trained diffusion model. We find that FSQ tokenization scheme has the latent space structure with the properties that make it best suited for continuous diffusion for categorical data as verified through rigorous theoretical analysis and numerical experiments. To validate our findings in real-life scenario, we train several text-to-speech diffusion models having speech tokens as intermediate acoustic features, and show that the one based on FSQ tokens indeed performs the best, and, moreover, it outperforms its strong LLM-based counterpart, at the same time being significantly smaller and faster.

Large models for time-series forecasting have been emerged as a promising paradigm for training models on heterogeneous collections of signals. These models typically rely on causal autoregressive architectures, where each observation is sequentially predicted from past. In practice, real-world time-series exhibit non-stationarities, which significantly influence predictive performance. To mitigate this, normalization is commonly employed. However, in efficient causal settings it might induce information leakage from future observations during training. Recent alternatives, including causal normalization and statistics computed from initial observations, have been proposed to address this issue, but their practical implications remain insufficiently understood. In this work, we evaluate normalization strategies for transformer-based large time-series models trained with patching and efficient causal strategy. We showcase that normalization choice significantly influences both training convergence and forecasting performance.

Diffusion Language Models (DLMs) enable parallel text generation by iteratively denoising a full sequence, offering attractive flexibility compared to auto-regressive (AR) decoding. However, existing methods fail to fully capture token relationships, leading to a performance gap relative to AR baselines, especially as the degree of parallelism increases. In this paper, we give a systematic analysis of the gap, identifying three key factors: (i) model capacity, (ii) dependency, and (iii) invariance. To address these issues, we first propose an invariant energy (Inv-E) together with an effective sampling-based estimator to handle the invariance issue. By further combining with the independent energy (Ind-E), we obtain a unified energy (Uni-E), that accounts for all these factors. Uni-E enjoys a unique advantage: it can be computed exactly without sampling-based partition estimation. Besides, Uni-E is model agnostic and can therefore be scaled to models of arbitrary size. We further prove that Uni-E can correct the distribution shift caused by dependency and invariance. Extensive experiments across Diffusion Language Models (DLMs) and Diffusion Large Language Models (DLLMs) demonstrate the effectiveness of the proposed Uni-E.

Autoregressive (AR) models have demonstrated strong potential in visual generation, offering superior performance with simple architectures and optimization objectives. However, existing methods are typically limited to single-modality conditions, e.g., text, restricting their applicability in real-world scenarios that demand image synthesis from diverse controls. In this work, we present OmniGen-AR, a unified autoregressive framework for Any-to-Image generation. By discretizing various visual conditions through a shared visual tokenizer and text prompts with a text tokenizer, OmniGen-AR supports a broad spectrum of conditional inputs within a single model, including text (text-to-image generation), spatial signals (segmentation-to-image and depth-to-image), and visual context (image editing, frame prediction, and text-to-video generation). To mitigate the risk of information leakage from condition tokens to content tokens, we introduce Disentangled Causal Attention (DCA), which separates the full-sequence causal mask into condition causal attention and content causal attention. It serves as a training-time regularizer without affecting the standard next-token prediction during inference. With this design, OmniGen-AR achieves new state-of-the-art or at least competitive results across a range of benchmark, e.g., 0.63 on GenEval and 80.02 on VBench, demonstrating its effectiveness in flexible and high-fidelity visual generation.

Recent progress in speech dialogue systems requires Text-to-Speech (TTS) models to be faster and more responsive. Modern speech dialogue systems impose two primary requirements on TTS models: low latency and support for streaming inputs and outputs. However, most existing single-codebook LLM-based TTS methods rely on multi-stage pipelines that lack native streaming capabilities. These systems typically suffer from high end-to-end latency due to slow autoregressive prediction and multi-step flow matching. To address these limitations, we propose FlashTTS, an open-source and low-latency streaming TTS framework. FlashTTS introduces a lagged multi-track architecture that natively processes streaming text and speech inputs, thereby eliminating the need for sentence-level buffering. To accelerate acoustic generation, we integrate parallel Multi-Token Prediction (MTP) with an X-pred mean flow matching decoder. This configuration achieves high-fidelity token-to-mel generation in exactly two function evaluations (2-NFE). By jointly optimizing input processing and decoding efficiency, FlashTTS offers a practical foundation for real-time speech dialogue systems. Experiments show that FlashTTS substantially reduces First-Packet Latency to 325ms compared to robust streaming baselines, all while preserving strong zero-shot voice cloning and cross-lingual intelligibility. Speech samples are available. The model code and checkpoints will be released as open source.

Deep reinforcement learning (DRL) frameworks for portfolio optimization have shown promise for their ability to learn allocation rules dynamically from market data. However, these models fail to account for fat-tailed returns, which characterize actual market behavior with more frequent extreme events. Furthermore, historical data is treated homogeneously, without accounting for temporal importance, leading models to fail during regime changes. We propose a new BAVAR-BLED algorithm that combines methods derived from Bayesian-Averaging Vector Autoregressive (BAVAR) and the Black-Litterman model using Elliptical Distributions (BLED) within a TD3 architecture. BAVAR captures a set of vector autoregressive representations that consider multi-scale temporal features, enabling adaptive allocation decisions based on regime-aware estimates of return expectations and dispersion matrices. These estimates serve as prior inputs to BLED, a model that uses Student's t-distributions, allowing for more realistic fat tail return estimates. The BAVAR-BLED algorithm uses transformer networks for view construction and CNNs for risk-aversion estimates, which modify dynamic allocation decisions based on market conditions. An evaluation of 29 Dow Jones Industrial Average constituents over a decade-long market period shows that BAVAR-BLED significantly outperforms state-of-the-art methods, achieving Sharpe and Sortino ratios of 1.72 and 2.70, respectively, and total returns of 57.26%.

Codec-based autoregressive (AR) speech language models have achieved strong text-to-speech (TTS) quality by modeling speech as sequences of discrete audio tokens with large pretrained backbones. However, this token-level formulation creates a structural efficiency bottleneck: speech-token sequences are much longer than text sequences, requiring the AR backbone to perform causal computation at every token position and maintain a KV cache that grows with the sequence length. We introduce TLDR, a patch-based autoregressive framework that accelerates codec-based AR-TTS by shifting the causal modeling from token-level speech sequences to patch-level sequences. TLDR groups consecutive codec tokens into compact latent patches using a lightweight compressor, models the resulting shorter patch sequence with a frozen pretrained AR-TTS backbone adapted by LoRA, and reconstructs fine-grained speech tokens within each patch using a speaker-conditioned extractor. With a patch size of 4, TLDR achieves a 1.8x inference speedup over the baseline AR-TTS model and reduces global KV-cache memory by up to 75%. Experimental results indicate that patch-level global causal modeling can be a practical way to reduce the inference cost of pretrained codec-based AR-TTS systems without replacing the existing modules.

Vision-language models (VLMs) with varying performance and resource requirements are widely deployed, making it difficult for users to select the most appropriate one among numerous VLM candidates. Existing work reveals the performance paradox phenomenon in language models and focuses on routing methods to solve it. However, developing a router for VLM selection is still a critical yet challenging problem, which primarily faces: 1) lack of specialized data, 2) ineffective feature representation, and 3) rigid model space and costly adaptation. In this paper, we construct a multimodal dataset for VLM selection, containing the outputs of seven mainstream VLMs on 32,626 unique image-text queries. We then propose ARMS, a router for VLM selection. ARMS enhances input signals with VLM profiles, employs a simple but effective architecture to improve representations of queries and VLM capabilities. To improve ARMS' adaptation to new VLMs, we propose two extension training strategies: incremental training and independent training. Experimental results on both in-distribution and out-of-distribution test sets demonstrate the effectiveness of ARMS. In particular, using our training strategy, ARMs (only 800M in size) can adapt to a broader VLM space and defeat commercial models like GPT-4o that are hundreds of times larger in scale. Our code, models, and datasets are available in the anonymous repository.

We introduce Contrast Sensitive Flow (CSFlow), a weighting scheme that connects the human eye's Contrast Sensitivity Function (CSF) to the iterative denoising steps of flow matching. Because real-world images concentrate signal at low spatial frequencies, these components reach high signal-to-noise ratio earlier during continuous diffusion than high-frequency components. When generating images with diffusion or flow matching models, this induces a soft autoregressive structure in Fourier space, where coarse image content stabilizes before fine detail. Meanwhile, the human visual system is unequally sensitive to spatial frequencies: very low and very high frequencies require significantly higher contrast to be perceived. We for the first time merge these observations through two contributions: (1) a metric that estimates which frequencies are generated at each reverse flow interval and (2) timestep weights obtained by aligning the frequencies generated at each noise level with human contrast sensitivity. We validate our contributions experimentally showing that these weights can improve generative performance by lowering FID by 4.7%, increasing Inception Score by 2.2% and improving GenEval scores by 2.5% using inference-only timestep modification or short fine-tuning. Qualitatively, we find that our CSFlow weights lead to better visual realism and less cartoonish appearance of generated images.

Gaussian-corrupted sentence embeddings have no direct linguistic interpretation, yet continuous diffusion language models can generate fluent text from them. We study this puzzle through Embedded Language Flows (ELF) and identify a decoder-basin mechanism: our evidence suggests that denoising becomes reliable when trajectories reach regions where the native decoder can read stable tokens. We introduce a diagnostic protocol for denoisability, semantic recoverability, order sensitivity, decoder compatibility, and trajectory reliability. It exposes failures hidden by scalar metrics: low mean-squared error can discard linguistic content, low perplexity can reflect low-entropy collapse, and clean latent reconstruction can coexist with a narrow decoder basin. A decoder-margin bound explains why token recovery depends on margin and local decoder sensitivity, not latent error alone. Auditing public ELF checkpoints reveals an interface phase diagram: early predictions are weakly readable, mid-trajectory disagreement marks a competition region, and late predictions enter a high-margin decoder basin. Once inside, token realization is surprisingly simple on generated ELF states: frozen T5 (Text-to-Text Transfer Transformer) token-embedding lookup recovers $93$--$96\%$ of native decoder decisions, and a single linear readout reaches $97.9\%$ agreement at 32k samples, leaving an $\approx1.1$--$1.2$ perplexity gap in a structured residual tail. Under conservative held-out gates, a margin rule exits roughly $17$--$28\%$ earlier in denoising steps under an explicit diagnostic monitor. Boundary checks on LangFlow, BitstreamDiffusion, and the Continuous Latent Diffusion Language Model (Cola-DLM) show that the same interface questions remain meaningful when the state object and decoder change. Continuous and latent diffusion language models should therefore be evaluated as representation-decoder systems.

Reinforcement learning (RL) holds immense promise for enhancing the reasoning capabilities of diffusion large language models (dLLMs). However, progress is fundamentally constrained by a dual misalignment between authentic generation trajectory and the gradient update process: (i) Process-reward misalignment. Sparse, terminal rewards are indiscriminately assigned to all intermediate steps of the generation process, failing to provide discriminative credit assignment. (ii) State-trajectory misalignment. Policy updates are often diverted toward artificial, out-of-trajectory states, squandering gradients on less informative samples. To address these limitations, we introduce Process Aligned Policy Optimization (PAPO), a novel framework that holistically aligns the RL update with the dLLM's generative trajectory via Step-Aware Process Rewards (SPR) that transform sparse terminal rewards into dense, step-wise credit, and Entropy-Guided Historical Re-enactment (EHR) that replays authentic trajectories at high-uncertainty steps. Extensive experiments on four benchmarks demonstrate that PAPO significantly outperforms baselines, achieving gains of up to 4.5% on GSM8K, 4.8% on MATH500, 42.2% on Countdown and 16.1% on Sudoku.

Diffusion and continuous flow-based language models have emerged as the leading non-autoregressive alternatives to language modeling. Progress in both paradigms is overwhelmingly tracked by generative perplexity (gen-PPL): the per-token negative log-likelihood of samples under a frozen autoregressive (AR) scorer such as gpt2-large, typically paired with an empirical-entropy guardrail to rule out low-entropy collapse. We argue that this metric is unsound. By construction, gen-PPL measures only predictability under the scoring AR, not grammaticality or semantic coherence -- and the set of predictable but still low-quality sequences is combinatorially large. To make this concrete, we construct a suite of zero-parameter, deliberately naive samplers that achieve state-of-the-art gen-PPL on LM1B and OpenWebText at non-degenerate entropy, surpassing recently published diffusion and continuous-flow models while producing text that is incoherent by construction. We recommend evaluation suites that directly quantify the distributional divergence between generated and reference text, and use such a suite to re-benchmark recent non-autoregressive models, recovering a more faithful picture of the current state of the art.

Block-wise semi-autoregressive decoding is the standard inference paradigm for diffusion large language models (DLMs), but it imposes a strict dependency between blocks: the next block cannot begin until the current block is fully decoded or its denoising budget is exhausted. We observe that once a block exposes a reliable delimiter boundary or stable semantic prefix, continuation generation need not wait for every residual token to be resolved. We propose AsyncLane, a training-free decoding scheduler that decouples refinement from advancement. AsyncLane forks a generate lane at observed delimiter boundaries into a refine lane and a continuation generate lane: the prefix remains editable, while the continuation advances before prefix refinement finishes. The resulting lane tree records decoding dependencies and output order, while execution proceeds over the active lane set. To make this asynchronous schedule efficient under bidirectional attention, AsyncLane combines shared-prefix lane batching, lookahead draft reuse, cascading termination, and compact cache refresh with refresh-logit reuse, preventing model-call cost from scaling directly with the number of lanes. AsyncLane is a drop-in replacement for block-wise DLM samplers and requires no retraining. Experiments on mathematical reasoning and code generation show that AsyncLane consistently improves throughput while maintaining competitive quality. Across LLaDA and Dream backbones, AsyncLane achieves the highest TPS in all evaluated benchmark-length settings; relative to the fastest competing baseline, it reaches peak speedups of 2.95x on LLaDA and 3.04x on Dream, with especially large gains under longer generation budgets.

We extend activation steering to diffusion language models (DLMs) and study a novel problem that arose due to the inference mechanism of DLMs: Modifying a text in-place to manifest a different concept. We propose TimpaTeks, an automatic in-place text modification mechanism using DLMs. Experiments on IMDB movie reviews (sentiment) and a synthetic Cats and Dogs Dataset (arbitrary, more unconventional concept steering) show that TimpaTeks provides a feasible novel mechanism to steer diffusion language model outputs in-place. TimpaTeks enables in-place modification while simultaneously lowers sentence perplexity and retaining the original sentence structre without the need of instruction tuned models. TimpaTeks is also computationally cheaper than prompt-based DLM steering, as it performs denoising in-place rather than constructing an additional prompt-conditioned output sequence.

Diffusion language models generate text through iterative denoising, offering a powerful alternative to autoregressive generation. However, discrete language spaces lack a natural neighborhood structure for defining effective perturbations, so some artificial corruption schemes are proposed in the forward process. Such prescribed forward processes often produce states that are mathematically convenient but misaligned with drafts and errors encountered during generation, resulting in degraded sample quality. To address this limitation, we propose FReDA, a forward-free diffusion language model that eliminates the need for a hand-designed forward process. We formulate diffusion language modeling as recursive distribution refinement, in which model-generated drafts serve as implicit intermediate states, and the learned refinement model progressively moves the draft distribution toward the target distribution. Concretely, FReDA refines drafts by proposing candidate draft sequences and either directly performing self-refinement or selecting among parallel candidates via best-of-N refinement. With this design, FReDA is neighborhood-agnostic, model-complexity-aware, and compatible with flexible refinement parameterizations. Extensive evaluations in the sub-8B regime show that FReDA-4B outperforms larger diffusion base models on reasoning and coding benchmarks, achieving absolute gains of up to 15%, while reaching a 1.5-1.8x average speedup over diffusion baselines and scaling effectively with additional refinement computation.

Designing OLED molecules with targeted optical properties remains challenging due to the scarcity of high-quality data and the limited reliability of conditional control in generative models across chemical motifs. Here, we benchmark a token-conditioned autoregressive language model for OLED molecular generation in a realistic low-data regime. A GPT2 model is pretrained on large chemical corpora, augmented with discrete property tokens, and fine-tuned using multi-task optimisation. Conditioning targets vertical absorption energy and oscillator strength, with the HOMO-LUMO gap included as an auxiliary electronic descriptor. Generated molecules are evaluated at the TDDFT level to assess distributional fidelity and controllability. The generated library reproduces the dominant optical-property support of the training distribution while shifting towards lower molecular weight and fewer heavy atoms. Token-level control is consistently directional across conditioning bins, but is not fully orthogonal and exhibits local calibration irregularities. A chemotype-resolved analysis further shows that controllability depends strongly on local electronic environments: moderately conjugated aromatic-carbon motifs are associated with improved joint target satisfaction, whereas electron-withdrawing motifs, particularly aryl nitriles, show systematic red-shifting and reduced controllability. These results establish a quantitative benchmark for conditional OLED molecular generation and show that model reliability must be assessed in chemically meaningful subspaces rather than from aggregate property distributions alone.

Diffusion language models (DLMs) offer substantial speed advantages through parallel decoding, but the lack of token dependencies limits generation quality compared to autoregressive (AR) models. Recent progress attempts to bridge the gap via importance sampling, with DLM being the proposal and AR being the target. However, due to the huge gap between their distributions, the sampling requires a large number of particles and is thus expensive to compute. In this paper, we introduce PoE-Bridge, a novel decoding framework that drastically improves generation speed and accuracy by introducing an intermediate distribution to bridge the gap. The distribution is constructed as a Product-of-Experts (PoE) of the DLM proposal and the AR target. With the intermediate distribution, we first use the DLM to draft multiple continuations in parallel, then apply rejection sampling to verify the drafted tokens and move the resulting candidates toward the PoE. We then use importance sampling to further correct the PoE-aligned candidates toward the AR target. We further propose several improved techniques, including mixed-temperature sampling for enhanced diversity and elastic rejection windows for reducing wasted verification. Empirically, PoE-Bridge achieves significantly improved accuracy with $5\times$ speedup over the standard DLM decoding approach, and recovers at least 95% of the target AR model's performance, efficiently advancing most of the quality gap on challenging mathematical reasoning and coding tasks. Our code is available at https://github.com/juntongshi48/poe-bridge.

We study the transformation of autoregressive models (ARLMs) into diffusion language models (DLMs). Rather than pretraining from scratch, prior work replaces the causal attention in ARLMs with bidirectional attention and then trains the resulting model using a DLM objective. However, these approaches incur two distribution shifts. First, transitioning from a next-token prediction objective to a DLM objective can discard knowledge acquired by the ARLM during training. Second, standard DLMs suffer from a train-inference mismatch, as the training loss is defined on randomly masked sequences rather than the trajectories encountered at inference produced by confidence-based decoding. To address both challenges, we introduce an On-Policy Diffusion Language Model (OPDLM) in which On-Policy Distillation (OPD) is employed for ARLM-to-DLM transformation. Specifically, OPDLM is trained via self-OPD, where the student, an ARLM with bidirectional attention, generates its own trajectories, and the teacher, the original frozen ARLM, distills its knowledge by providing target logits on these trajectories. By training directly in an on-policy manner, OPDLM eliminates the train-inference mismatch in DLMs, while distillation from the original model enhances knowledge retention from the ARLM. Empirical results demonstrate that OPDLM requires 15x to 7,000x fewer training tokens with strong performance across a wide variety of tasks. OPDLM avoids the prohibitive cost of DLM pretraining and positions DLM transformation as a form of ARLM post-training.

Discrete diffusion language models generate text by iteratively denoising an entire response in parallel. At each step, they predict tentative tokens for every masked position, committing the confident predictions to the output and discarding the unconfident ones. We show that the discarded tokens are in fact a useful lookahead signal for retrieval-augmented generation: even low-confidence tokens often surface salient entities early in the denoising trajectory, enabling retrieval of stronger evidence before the output is finalized. We exploit this through Self-Augmenting Retrieval for Diffusion Language Models (SARDI), a dynamic RAG framework that uses these lookahead tokens to guide retrieval during denoising. SARDI is training-free, retriever-agnostic, and applicable to any reasoning-capable discrete diffusion language model. Across five multi-hop QA benchmarks, SARDI outperforms current training-free diffusion and autoregressive retrieval baselines at up to $8\times$ higher throughput.

Large language models often improve reasoning by generating explicit chain-of-thought (CoT), demonstrating the importance of intermediate computation. However, textual CoT forces this computation through a discrete, serial, and communication-oriented token stream: each reasoning step must be verbalized before the model can proceed, even when the underlying update is semantic, uncertain, or only partially formed. Latent reasoning offers a higher-bandwidth alternative by performing intermediate computation in compact continuous states before committing to text. Yet existing latent-reasoning methods often sacrifice key advantages that make CoT effective in autoregressive language models, including native left-to-right generation, probabilistic sampling, compatibility with KV-cache decoding, and tractable likelihood estimation. We propose NF-CoT, a latent reasoning framework that preserves these advantages by modeling continuous thoughts with normalizing flows. NF-CoT instantiates a TARFlow-style normalizing flow inside the LLM backbone, defining a tractable probability model over compact continuous thoughts distilled from explicit CoT. Continuous-thought positions are generated by an NF head, while text positions are generated by the standard LM head within the same causal stream. This design provides exact likelihoods for latent thoughts, enables probabilistic left-to-right decoding with the original KV cache, and supports direct policy-gradient optimization in the latent reasoning space. On code-generation benchmarks, NF-CoT improves pass rates over explicit-CoT and prior latent-reasoning baselines while substantially reducing intermediate-reasoning cost.

Machine unlearning aims to remove targeted knowledge from a trained model while preserving its general capabilities. For autoregressive language models, not all tokens in a forget sample are equally relevant to forgetting. Existing approaches either ignore this heterogeneity or rely on auxiliary models, heuristics, or external annotations to estimate each token's relevance for forgetting. We instead characterize it through the interaction with the retain objective: a token is forget-specific to the extent that minimizing the forget loss on that token does not conflict with retain optimality. We formalize this perspective as a joint optimization problem over the model parameters and the token weights and show that, under a natural separation condition, the resulting objective recovers the oracle forget-specific token support. Motivated by this formulation, we introduce Alternating Token-Weighted Unlearning (ATWU), a lightweight framework that jointly learns token forget-specificity and model parameters during unlearning using a simple linear scorer over the hidden states, without external token level supervision. Across TOFU and RWKU, ATWU achieves state of the art forget-retain trade-offs, outperforming sample-level methods, probability-based token weighting heuristics, and auxiliary-model-based approaches. Moreover, the learned scores align substantially better with ground truth forget-specific spans, indicating that ATWU identifies semantically meaningful token level forgetting signals. Overall, our results suggest that retain conflict provides an effective criterion for identifying what language models should forget, enabling unsupervised learning of token level forget-specificity directly from model representations with minimal computational overhead.

Lossless pixel-level image transmission is a fundamental regime beyond semantic communications, because exact recovery requires both accurate symbol probability modeling and reliable delivery over noisy channels. This paper proposes DDM-SSCC, a discrete-diffusion-model-based separate source-channel coding framework for lossless image transmission. Different from raster-order autoregressive coding, the proposed source codec adapts a diffusion language model to pixel-token restoration and performs synchronized reverse arithmetic coding under bidirectional attention, allowing multiple masked tokens to be coded within one reverse denoising step. This progressive restoration process also yields a more favorable source representation for noisy transmission, since newly restored tokens can serve as bidirectional context in subsequent denoising steps. To bridge the gap between generation-oriented masked denoising and lossless arithmetic coding, we further introduce a Halton-guided denoising order, a mask-ratio-aware cosine schedule, and a lightweight temperature calibration module. These designs respectively improve spatial coverage, adapt the denoising pace to context reliability, and calibrate the probability tables used by arithmetic coding. Experiments on CIFAR10, DIV2K-LR-X4, and Kodak over additive white Gaussian noise and Rayleigh fading channels show that DDM-SSCC achieves better exact-recovery performance than representative lossless and semantic communication baselines, while ablation studies verify the effectiveness of the proposed denoising order, schedule, and calibration modules.

We propose a statistical-field framework for text generated by large language models (LLMs), treating token embeddings as continuous spin variables on a one-dimensional chain. Defining a susceptibility from the connected two-point correlator and an order parameter from the ensemble-averaged embedding field, we vary the \texttt{softmax} temperature $T$ and observe a sharp susceptibility peak near a characteristic $T_c$ with power-law-like scaling, a concurrent rapid change in the order parameter, and a collapse onto a single semantic direction below $T_c$. The intrinsic dimension estimated by the two nearest neighbor (TwoNN) method independently corroborates these findings, reaching a minimum near $T_c$. Results are robust across model scales (Qwen3: 0.6B--32B) and prompt categories. While the phenomenology closely resembles a continuous phase transition, the non-equilibrium nature of autoregressive generation warrants further investigation. Our framework provides quantitative tools for probing the collective statistical structure of LLM outputs and suggests connections between decoding strategies and critical phenomena.

Masked diffusion language models generate text by iteratively unmasking many tokens in parallel, but this speed comes with a correction problem: tokens generated in the same step are predicted from marginal distributions, and early local dependency errors can later contaminate the context. PRISM addresses this by learning token-level quality scores and remasking unreliable tokens, but its inference rule is coupled: the same forward pass both detects low-quality tokens and computes logits for their replacements, so the erroneous tokens still condition regeneration. We propose NAVIRA, an inference-time decoding policy that separates these two operations and samples remasking positions stochastically. A first forward pass scores tokens; selected tokens are masked; a second forward pass regenerates from the cleaned context. Temperature-controlled remasking reduces repeated correction of the same positions and balances fluency against diversity. In controlled experiments with a 170M masked diffusion language model, decoupling improves fluency, while scheduled stochastic remasking preserves entropy and achieves stronger LLM-judge scores under larger forward-pass budgets. These results show that remasking policy, not only the learned quality signal, is central to reliable masked-diffusion text generation.

Generating executable tool plans requires selecting appropriate subsets from tool libraries, a combinatorial search problem with an exponentially large solution space. However, we identify a critical misalignment in predominant approaches: standard autoregressive (AR) decoding suffers from early commitment, where initial token choices rigidly constrain the search trajectory. A controlled study shows that masked denoising raises Pass@10 solution coverage from 0.320 to 0.943 over AR sampling under matched compute. Motivated by this, we propose DiG-Plan, a framework that decouples combinatorial exploration from structural refinement. DiG-Plan employs a diffusion-based proposer to generate diverse tool sets via iterative refinement, followed by an AR refiner for dependency prediction. On TaskBench, DiG-Plan improves over AR baselines by a 10% relative margin, with the largest gains on complex compositional tasks; API-Bank results show that the propose-refine-select design remains effective across domains. Code is available at https://github.com/puddingyeah/DiG-Plan.

Parametric retrieval augmentation encodes document information into lightweight, document-specific modules such as LoRA adapters, reducing the need to include all evidence as input context. However, it remains unclear how this parameter-side memory interacts with context-side memory stored in the KV cache. We study this interaction in document-level question answering by progressively evicting document key-value states and measuring when a document LoRA contributes beyond the retained context. We find that document LoRA adds little when the KV cache is largely intact, but becomes increasingly useful under aggressive compression, recovering 13-21 ROUGE-L points when no document context remains. The gain is largest when the base model encodes the document, and the adapter is applied only during answer generation, suggesting that document LoRA is better understood as decoding-time parametric memory than as a document encoder. Finally, QA-style supervision produces substantially stronger adapters than raw-context next-token-prediction. These results position document LoRA as a complementary memory channel whose value emerges precisely when context-side evidence is scarce.

The Deep Graph Generation's panorama spans two extremes: one-shot and sequential models. The former generates nodes and edges jointly, while the latter samples them autoregressively. Each method performs better in different graph domains depending on size and topology, but neither is applicable to all graph categories. For instance, one-shot methods struggle with generating large graphs, while sequential methods underperform on smaller graphs. A possible way to overcome these limitations is to flexibly combine the two methods in a unique system. In this work, we propose the FLAGG (Flexible Autoregressive Graph Generation) framework, which sequentially generates portions of graphs with one-shot models. FLAGG can apply any one-shot model to make it autoregressive, allowing flexibility in choosing the sequential policy. This policy is specified through a stochastic node removal process, which an Insertion Model learns to reverse. We evaluate FLAGG with the DiGress one-shot model on several data sets of different graph sizes and domains. We show that the approach outperforms both one-shot and autoregressive baselines in terms of sampling quality.

Diffusion large language models (DLLMs) enable non-autoregressive generation by iteratively denoising corrupted token sequences with bidirectional context. Despite their ability to update multiple positions in parallel, inference remains costly due to the many denoising steps required for high-quality generation. We propose SAID, a Scaffold-Aware Iterative Decoding framework that accelerates DLLMs by reallocating computation across tokens. SAID first spends denoising computation on scaffold tokens to establish the coarse semantic structure, and then completes predictable detail tokens with fewer steps. We further adapt SAID to block-wise diffusion decoding and introduce Confidence-Hierarchical Layered Generation (CHLG), which assigns additional steps only to low-confidence tokens. Experiments on LLaDA-8B and LLaDA 1.5 across math, coding, and knowledge benchmarks show that SAID significantly accelerates DLLM inference with a maximum speedup of 9.1x while maintaining competitive performance. Our code is publicly available: https://github.com/TH-AI-Lab-PKU/SAID.

Diffusion language models (DLMs) generate text through iterative denoising, and blockwise decoding improves their practicality by committing tokens in local blocks. However, existing blockwise methods typically rely on fixed block sizes or delimiter-based runtime signals, which do not necessarily align with semantic boundaries. In this paper, we propose SemBlock, a semantic-boundary-driven dynamic block decoding framework for diffusion LLMs. SemBlock formulates dynamic block construction as semantic boundary prediction and trains lightweight predictors on frozen LLaDA hidden states. To provide supervision, we construct SemBound, a semantic-boundary dataset that derives boundary labels from discourse units, reasoning steps, and implementation spans across natural language, math, and code tasks. During inference, SemBlock uses predicted boundary probabilities to select the ending position of each dynamic block. Experiments on GSM8K, IFEval, MATH, and HumanEval show that SemBlock consistently improves over fixed-block decoding and AdaBlock. Our code is publicly available: https://github.com/TH-AI-Lab-PKU/SemBlock.

Diffusion large language models (DLLMs) have recently emerged as a promising alternative to autoregressive LLMs by generating text through iterative masked denoising with bidirectional context. However, their large model sizes and iterative denoising process introduce substantial memory and computational overhead, motivating post-training quantization for efficient deployment. In this paper, we identify two key challenges for low-bit DLLM quantization: state-dependent activation disparity and temporal error accumulation. Masked and unmasked tokens exhibit different activation distributions within each denoising step, while quantization errors can accumulate across steps during iterative decoding. To address these challenges, we propose STaR-Quant, a state-time consistent PTQ framework for DLLMs. STaR-Quant introduces State-Guided Activation Transformation (SGAT) to assign masked and unmasked tokens to different activation transformation spaces with a unified static weight-side transformation. It further introduces Temporal Attention Compensation (TAC) to correct the quantized attention representation via a lightweight block-diagonal affine mapping. Experiments on representative DLLMs demonstrate that STaR-Quant consistently improves low-bit weight-activation quantization over strong PTQ baselines, while delivering up to 1.69x speedup and 3.14x memory saving over FP16 deployment.

Large Language Models (LLMs) are increasingly deployed across diverse applications, raising critical questions for governance, accountability, and data provenance. Understanding which training data most influenced a model's output remains a fundamental open problem. We address this challenge through training data attribution (TDA) for auto-regressive LLMs by expanding upon the inverse formulation: How would training data be affected if the model had seen the generated output during training? Our method perturbs the base model using bidirectional gradient optimization (gradient ascent and descent) on a generated text sample and measures the resulting change in loss across training samples. Our framework supports attribution at arbitrary data granularity, enabling both factual and stylistic attribution. We evaluate our method against baselines on pretrained models with known datasets, and show that it outperforms previous work on influence metrics, thereby enhancing model interpretability, an essential requirement for accountable AI systems.

Diffusion large language models (dLLMs) offer bidirectional attention and parallel generation, enabling them to exploit global context and naturally support format-constrained tasks like parseable JSON or reasoning templates. While straightforward fixed anchors can enforce such constraints, they often impose rigid spans, leading to truncated reasoning or redundant content. To overcome this, we propose Dynamic Infilling Anchors (DIA), a training-free method that dynamically estimates end-anchor positions to adjust generation length before iterative infilling. This flexible mechanism ensures structural correctness and semantic coherence, avoiding the inefficiencies of fixed-span methods. Experiments on reasoning benchmarks demonstrate that DIA substantially improves format compliance and answer accuracy, achieving significant zero-shot gains on GSM8K and MATH. These results establish DIA as a robust pathway toward reliable, structure-aware generation.