Helge Spieker, Dennis Gross, Arnaud Gotliebcs.SE cs.AI
Test-set accuracy is silent on two issues that matter when deploying autoregressive neural sequence models: how much probability mass the system under test (SUT) places on constraint-violating alternatives that are reachable under sampling and what fraction of the input population satisfies a domain requirement. We answer both with probabilistic model checking. The pipeline extracts a discrete-time Markov chain (DTMC) from the SUT's token-by-token generation, verifies formal PCTL specifications with the PRISM model checker, and aggregates the per-input verdicts into a coverage curve over the input space. A soundness theorem establishes the DTMC as an under-approximation, so every verdict yields a certified interval on the SUT's true reachability probability. The coverage built from those verdicts is, therefore, conservative by construction. A counterexample-guided abstraction refinement (CEGAR) loop adaptively tightens the interval, and a maximum-likelihood algorithm extracts the most probable falsifying trace. Two case studies exercise the pipeline. On a GPT-2 computer-aided process-planning (CAPP) model with 100% test accuracy, the pipeline quantifies the probability mass greedy decoding hides, but that is reachable with sampling; and identifies the smallest training fraction at which an ordering requirement holds population-wide, neither of which test accuracy can report. We then verify the SMILES molecular generator with a 50x larger vocabulary. The only change is an external chemical-validity oracle, and the pipeline identifies the gap between structural completeness and chemical validity.
Small language models (sLLMs) are nowadays hosted on devices with limited memory and computational budget. In an autoregressive setup, inference is memory-bandwidth bound: uniform quantization is often detrimental to such models, since their architecture has limited redundancies and only a few layers are not very sensitive to lower precision. We propose a composite metric that combines two orthogonal criteria: information retention (measured in terms of a normalized SQNR-based coefficient) and throughput gains (modeled using a roofline-based latency analysis). By profiling Gemma 3 1B, we find that Feed-Forward Network blocks and the embedding matrix are the most promising targets for acceleration. For each candidate, we estimate a normalized quality score based on simulated quantization and a normalized speed score based on roofline modeling with no actual execution needed. We combine the two scores in a composite priority coefficient, allowing us to tune the trade-off between speed and quality as needed. Our metric is general and can be used to prioritize individual blocks, their projection sublayers, or transformer layers as a whole. We evaluate our approach on several model architectures, showing that our estimates have at around 4% prediction error for the accelerated speedup. We find that our method generally allocates more resources to the most expressive layers compared to evolutionary search, specialized accelerators, or Shapley-value-based approaches that require expensive approximate inference. Our analytical approach makes sLLM quantization a predictable engineering task.
Streaming autoregressive video models generate long videos chunk by chunk, using historical memory to maintain consistency. Existing methods typically expose subject and scene queries to history through similar policies. This stabilizes the subject, but can also lock backgrounds, viewpoints, and scene structure to previously generated states even when local motion continues. We call this failure memory-anchored scene under-progression; consistency and motion metrics alone can miss it. We introduce TetherMem, a training-free, query-aware spatiotemporal memory router for frozen video generators. TetherMem separates subject and scene queries and modulates historical access with region- and age-conditioned priors: subject queries retain identity-bearing history, while scene queries reduce reliance on subject history and stale backgrounds. Across 2,400 blinded pairwise judgments from 10 annotators, TetherMem achieves the highest estimated expected preference among eight streaming long-video baselines for overall quality (0.780) and scene progression (0.769). On complete 30-second videos, it sustains changes in background, viewpoint, and scene state while preserving subject recognizability and temporal continuity.
Autoregressive models have emerged as an effective paradigm for point cloud generation. However, most existing approaches rely on heuristic tokenization strategies, such as spatial sorting or stochastic downsampling, which often disrupt intrinsic point cloud topology and weaken the structural coherence of the generated shapes. In this paper, we present PointRSP, an autoregressive framework that reformulates point cloud generation as a topology-preserving tessellation process via recursive spectral partitioning. Instead of constructing token sequences heuristically, we introduce a topology-aware partitioning autoencoder that decomposes an unstructured point cloud into a non-balanced binary tree through a hybrid recursive spectral partitioning strategy. This hierarchical representation provides a deterministic geometric blueprint that preserves topological relationships while capturing multiscale structural dependencies within a quantized latent space. To synthesize shapes in this space, we propose a dual-stream cascaded generator that jointly models structural evolution and feature synthesis. In addition, we design a geometry-calibrated positional encoding mechanism that anchors latent embeddings using multi-scale structural centers, which stabilizes cascaded generation during the early stages of structural formation. Extensive experiments show that PointRSP achieves state-of-the-art performance in generation quality and diversity, demonstrating strong generalization across complex 3D topologies.
Autoregressive rigging models such as UniRig and SkinTokens perform well on articulated characters, but their ability to generalize to plant structures remains largely unexplored, since plant topologies exhibit highly variable, non-canonical branching patterns that challenge learned skeletal priors. We evaluate these models for plant skeletal reconstruction using synthetic L-system-generated trees and real scanned data spanning monopodial, sympodial, whorled, and vine-like archetypes. Preliminary testing showed UniRig collapsing complex branching into near-linear chains, while SkinTokens preserved topology better but over-segmented branches and produced an unstable output space, so we focused on UniRig for its greater stability. Diagnosis traced the collapse to sampling-level suppression of branch tokens, and further analysis showed the frozen mesh encoder had limited sensitivity to structural variation, pointing to a geometric bottleneck in the tokenization pipeline rather than a purely learned bias. Building on these findings, we applied multi-round fine-tuning over multiple procedurally generated synthetic datasets. Across rounds, the model progressively recovered accurate branching topology and generalized beyond branch-only structures to plants with foliage, a harder case given the zero-thickness, mesh-normal-dependent geometry of leaves. The resulting model generalized well across diverse plant forms without leaf-specific architectural changes, indicating that targeted fine-tuning can substantially close the domain gap between character-rigging priors and plant skeletal structure. As such, our work points toward a viable path for automated plant rigging across both branch topology and foliage type, even those not considered in our findings.
We introduce Graded Large Language Models (GLLMs), an algebraic framework that equips the representation space of a transformer with a grading and propagates the induced weighted scalar action through embeddings, self-attention, and the training objective. The construction extends the theory of graded neural networks and graded transformers to autoregressive language models while preserving expressive power, asymptotic computational complexity, and inference cost. The governing geometric picture is that of geometric invariant theory. The benefit of a grading is expressed by a Kempf--Ness functional on the grading torus; the grades that improve upon the uniform architecture form an open convex cone whose membership is decided by a Hilbert--Mumford-type criterion pairing a grade direction against two measurable profiles of the target and the data; the optimal grades are the coincidence point of two moment maps, given in closed form; and the ordinary transformer appears as a semistable isotropic point on the boundary of the cone: one member of a larger graded family rather than a distinguished optimum. Separately, for level-stratified targets we prove a minimax separation between the graded prior and its absence: over all estimators the risks of the graded and uniform target classes separate throughout an explicit window of sample sizes, by a factor that decays exponentially in the number of levels under geometric stratification. Both profiles are estimable offline, so the optimal grades solve a convex program certified before training begins. Because the grading is absorbed into the learned parameters after training, every GLLM compiles to a standard transformer of identical architecture and inference complexity.
Long-horizon weather forecasting is a fundamental challenge in atmospheric science, for which autoregressive Deep Learning Weather Prediction (DLWP) has emerged as the primary paradigm. Although the autoregressive pipeline is highly scalable and flexible, its prediction errors grow rapidly over long forecasting horizons. In this work, we study this error growth phenomenon from both theoretical and empirical perspectives. Our analysis reveals that the growth is driven by a feedback loop between output errors and input distribution shifts. Specifically, the autoregressive process amplifies small initial output errors, which progressively corrupt subsequent input distributions, echoing the butterfly effect in atmospheric science and ultimately deteriorating forecasting accuracy over longer horizons. Furthermore, we show that this distributional shift originates at the earliest stage of inference, with out-of-distribution signatures detectable as early as the first autoregressive step. To mitigate this issue, we propose \textbf{Self-Output Fine-Tuning (SOFT)}, a plug-and-play strategy that leverages the model's own one-step predictions to calibrate the biased input distribution encountered at the first step. Extensive experiments demonstrate that, despite its simplicity, SOFT achieves state-of-the-art performance on long-horizon forecasting tasks and substantially reduces both prediction errors and distributional discrepancy. The success of SOFT highlights the importance of reexamining the fundamental pipeline of deep learning weather prediction, representing a critical pipeline advance for atmospheric science.
Autoregressive (AR) streaming models have emerged as a powerful paradigm for long video generation. However, the linearly growing Key-Value (KV) cache poses a significant bottleneck, leading to memory overload and degraded inference throughput. A common compression method is to drop redundant KV tokens, which often breaks long-range dependencies, resulting in temporal flickering and identity loss. In this paper, we propose Instance-Specific Parametric Absorption (ISPA), a novel framework that shifts the KV cache compression from discarding to distilling. The core idea is to transit a subset of layers from Full-Attention (F-Layers) to memory-efficient Local-Attention (L-Layers) by "absorbing" historical context into the model's weights. Specifically, during a brief warmup phase, ISPA monitors the output discrepancy between global and local attention. At the transition point, we solve a closed-form least-squares problem to compute an instance-specific weight modulation that compensates for the missing history. Experiments across architectures (1.3B to 14B) demonstrate that ISPA can remove up to 50\% of the KV cache with near-lossless visual quality. We hope this perspective encourages future work to explore parametric memory consolidation beyond external token-level cache management for streaming generative models.
Matthias Brändel, Stephan Köhler, Oliver Rheinbachcs.LG
We study how the next-token prediction of an autoregressive Transformer language model changes under small perturbations of earlier input token embeddings. Motivated by operator learning and iterative solvers for differential equations, we investigate how the influence of one token on another decays with distance in a trained model. In multilevel methods for differential equations, such as domain decomposition, multigrid, and multilevel preconditioning, one often exploits a separation between strong local interactions and weaker but essential global interactions. The latter correspond to the long tail of the Green's function and are typically handled by a coarse-level operator. Inspired by this perspective, we compute an empirical, distance-resolved gradient profile of token dependencies using autograd. Experiments on trained Pythia models and Qwen2.5-0.5B show that, over the measured distance range, the median Jacobian sensitivity is much better described by a power-law-type decay than by an exponential alternative: the diagonal-normalized profile is well described by $$\overline G(r) \approx γ+β(r+1)^{-p}$$ with exponents $p \approx 0.7$--$0.9$ (typically $0.8$--$0.9$). This behavior appears on coherent text from Gutenberg and WikiText-103. Token-shuffling experiments show that the power-law profile persists even when syntax and prediction quality collapse, whereas randomly initialized models do not exhibit it. The slowly decaying long-range sensitivity thus appears to be a learned property of trained autoregressive Transformer operators. These findings suggest that hierarchical or coarse-level mechanisms in language models may be able to exploit the long-tailed sensitivity profiles.
The AI community has framed the relationship between large language models (LLMs) and world models as a dichotomy: LLMs predict tokens; world models simulate reality. Yann LeCun argues in 2022 that reaching general intelligence requires abandoning autoregressive token prediction in favour of latent-space architectures. This framing is unnecessarily binary. Two claims will be defended. First, LLMs are a degenerate special case of world models: the state space is the set of all token sequences, the only action is appending one token, and world models are therefore a strict generalisation of LLMs, not a replacement. Second, there is a natural continuous spectrum from NTP to JEPA, with multi-token prediction, future-summary prediction, and next-latent prediction as intermediate stations already populated by current research. Moving along this spectrum relaxes the LLM constraints one by one. It also progressively surrenders the two practical advantages that make LLMs trainable at scale: internet-scale self-supervised data, and a transformer architecture co-designed for discrete token prediction. Both are examined as open research questions: the data question (the cliff from self-supervised text to instrumented action-labelled environments) and the architecture question (whether the transformer generalises to continuous-state prediction, or whether a new primitive is needed).
Pixel-space continuous-token autoregressive (AR) generation directly models images as sequences of raw pixel patches, avoiding discrete tokenization or a separately pretrained tokenizer. However, it faces coupled challenges: high-dimensional patch generation causes large single-step errors, and teacher-forced training creates a train--inference gap that makes these errors accumulate across AR steps. Existing fixes such as $x$-prediction and input noise injection only partially mitigate these issues. Exact rollout training better matches inference-time conditions, but is impractical due to prohibitively slow sequential sampling. We propose \emph{Parallel Rollout Approximation} (PRA), a scalable framework that addresses both challenges jointly. PRA generates low-dimensional intermediate states instead of high-dimensional pixel patches, then maps them back to pixel-space tokens with a pixel decoder, preserving a pixel-in, pixel-out AR interface. It also constructs inference-like pixel inputs through the same intermediate-state-to-pixel path used at inference, independently across positions, approximating the pixel-feedback interface encountered during inference-time rollout while retaining parallel teacher-forced training. On class-conditional ImageNet-1K generation at $256\times256$ resolution, PRA-S with 135M parameters achieves an FID of 2.58, surpassing the previous billion-scale pixel-space AR result of 3.60. Scaling to PRA-L with 511M parameters further improves FID to 1.94, establishing a new state of the art among pixel-space AR models. Beyond generation, PRA achieves higher ImageNet classification probing accuracy than other AR and diffusion baselines, suggesting its potential for unified pixel-space image generation and understanding.
Speculative decoding (SD) has emerged as a key solution to accelerate the inference of autoregressive models. However, in the field of image generation, it faces the challenge of low acceptance rates, and directly relaxing its criteria leads to degradation in image quality. In this paper, we propose a novel content-aware speculative decoding algorithm, termed CSD, which integrates an entropy-based probability relaxation mechanism with an optimal resampling strategy to enhance the inference efficiency for autoregressive image generation. By leveraging the informational uncertainty inherent in different regions of an image, CSD dynamically adjusts the acceptance probability of candidate tokens, increasing the acceptance rate in low-detail areas to accelerate generation. Moreover, a distribution alignment filter is introduced to ensure the output distribution to be aligned with the target model, which significantly improves the generative quality. Experiments conducted on Lumina-mGPT and Janus-Pro demonstrate that the superiority of the proposed CSD. Our source code is available at https://github.com/aderfebr/CSD.
Xingyu Su, Jacob Helwig, Shubham Parashar +6cs.CL cs.AI
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.
Thomas T. Zhang, Alok Shah, Yifei Zhang +3cs.LG cs.AI eess.SY
Many modern applications of deep learning involve training a neural network via a one-step prediction loss (e.g., $L^2$ regression, cross-entropy), but deploy the network by rolling out along its own predictions. Key examples include autoregressive language modeling, flow-based generative modeling, and robot policy learning. It is well-documented that these settings induce a phenomenon we call test-time feedback (TTF): the mismatch between the training/validation loss and downstream metrics of interest, such as task success rate and generation quality, which grows with task length. While data curation, architecture, and objective design have been proposed to combat train-test shift in TTF settings, this paper proposes optimization as a new design axis to mitigate error accumulation. Specifically, we introduce a new optimization paradigm called double-preconditioning (DoPr) uniquely tailored to the challenges of TTF. DoPr combines gradient-wise preconditioning, as in Adam and Muon, with activation-wise preconditioning (AP), such as in KFAC. We show that the addition of AP yields a drop-in intervention for increasing downstream model performance across a range of TTF settings. Interestingly, these gains in test-time performance do not consistently accompany improvements in validation loss, opening new questions about how to properly evaluate models trained with one-step supervised objectives.
Autoregressive long video generation often adopts bounded-memory streaming for efficiency, typically combining local windows for short-term continuity with static early-frame sinks as long-range anchors. However, this fixed allocation keeps early frames cached even when the current visual state has substantially diverged from them, while discarding potentially more relevant intermediate history. As a result, the retained long-range context may become less adaptive and bias generation toward outdated cues; in severe cases, RoPE-induced phase re-alignment can homogenize inter-head attention and cause sink collapse, where content regresses toward sink frames. We propose DySink, a retrieval-based framework that maintains a compact memory bank and selects visually relevant historical frames as dynamic frame sinks. DySink couples adaptive retrieval with a sink anomaly gate, which detects excessive inter-head consensus over retrieved context and suppresses collapse-prone context. Experiments on minute-long videos show that DySink consistently improves dynamic degree over strong baselines while also achieving higher temporal quality. The code and model weights will be released at https://github.com/yebo0216best/DySink.